Overview of the situation: Reuters got a leaked version of an S-1 document related to Anthropic becoming a public company. Context: An S-1 is a type of "full disclosure" document that states all...
Exemplary
Overview of the situation: Reuters got a leaked version of an S-1 document related to Anthropic becoming a public company.
Context: An S-1 is a type of "full disclosure" document that states all the relevant information regarding a companies financials, revenue, prospects, risks and all the other boring details that are needed to ensure investors know what they are buying. For example here's SpaceX's S-1 and it's... something... (It's definitely not all AI slop, but everything about it felt off.)
AirBnB is a much better example. It's far more coherent filing and they successfully went to market in 2020, in the middle of COVID. I think a lot of the AI labs were hoping to get this sort of response.
My thoughts: Unfortunately, the full document is not available and the reporting on it makes it difficult to really get a handle of what it says about the health of the company. It will be made public a few weeks before the official listing but right now we just have the broad strokes. A lot of headline reporting is focused on the fact that a lot of the document is focused on the EXISTENTIAL RISKS OF AN AI APOCALYPSE. Not exactly surprising given the clickbait economy.
The part I'm personally worried about is the magic math that the company seems to be relying on. Simply put:
The filing offers a rare look inside a business that barely existed several years ago but is growing at breakneck speed due to its role as a key developer of large-language AI models. Anthropic is seeking a valuation of about $2 trillion and has plans to spend hundreds of billions in the coming years to accelerate its growth.
A concept that sits in stark contrast to gems like:
Revenue surged 12-fold in 2025 to nearly $4.6 billion, while operating losses more than doubled to top $8 billion, Reuters reported exclusively on Monday. The US accounted for nearly two-thirds of total sales.
About $3.8 billion in Anthropic's revenue came from customers paying based on how much they use the company's Claude AI system, while subscription revenue came to $789 million. Anthropic said it expects consumption-based revenue to account for "the substantial majority" of its revenue for the foreseeable future.
...
The company paid roughly $351 million back to the platforms in distribution fees, according to a Reuters analysis, suggesting the cloud providers collected some 16 cents for every dollar of those sales. Anthropic reports channel partner fees within the "sales, marketing, and partnerships" operating expense line item on its financial statements.
...
Anthropic said nearly a quarter of its revenue came from two customers last year, and as part of its risk factors, warned that many of its largest clients were not locked into long-term contracts and could cut or stop spending.
This is important because a billion dollars is not a trivial number, but the difference between that and a trillion is about a trillion dollars. And leading up to the IPO, the best they can show is 3.8bil in revenue against 8bil in losses.
Everything is happening right as AI Liability becomes a bigger topic, legal/security risks are being taken seriously and global data-center pushback is starting to gain steam. We are also starting to see more demand for models that are cheaper to run, open source and could be hosted locally or on individual machines (if we ever get past the RAM apocalypse).
I think all the AI companies were aiming for the trillion dollar IPO's because for a long time, they represented AI as a broad concept. There wasn't ever an AI narrative, it was just a bunch of vague and sometimes contradictory talking points. AI being able to cure every ailment and make everyone millionaires and get everyone fired and solve the deepest mysteries of the universe and cause vulnerable people to commit suicide and hack every system and solve climate change and killing everyone on the planet. But this is Silicon Valley pie-in-the-sky stuff. Wall Street is the land of bean counters and the beans are not adding up.
The trillion dollar question for them is "how does this technology turn the money spent on it into more money?" It's not just a question about users and customers. There are already a lot of investors and private credit providers and governments and banks who are almost 5 years into this game and without a solid path to profitability and dividends, their only out is the IPO. Some of these funders like the Arab Sovereign Wealth Funds could really use some of that money right now. Same thing with shadow banks and the raising interest rates. And with the whole treasury bond selloff that is not looking pretty right now.
Long story short, all the people who have money in AI companies don't seem to have a way to get anything out of it. These companies are going to public markets and even though the stock market is booming now, it's not going to be an easy ride. Anyone that will be investing in the big AI names now will be taking money out of other places. Places like NVidia, Microsoft, Google, Meta... Yes al these companies have massive AI plays, but they are also mature companies with existing revenue streams. Institutional investors are going to look at these S-1's and ask where is the long term spending commitments these companies have from stable clients? Or what's the cost of training a new frontier model and what's the forecast on seeing a return on that spend? How are 2 super cap companies going to compete with dozens of cheap, small labs that have no trouble stealing their work? Why is Oracle saying that there is a realistic force majeure risk to data-center projects?
Even the regular arguments for AI/LLM feasibility fall short in the face of this massive IPO. Every time token costs go down, the labs will be forced to reduce their prices or be priced out by the competition. The technology getting to the point of autonomous self-improvement still incurs the escalating cost of updating the model and that is still a logarithmic scale, plus the increasing premium for access to high quality training data. As the technology propagates and develops, there are likely to be more companies offering specialized services and it will be increasingly difficult for the major labs to play in every possible space.
The most dangerous part is that many retail investors don't know to analyze a company like this. They tend to work on whatever is in the media hype cycle. Like a lot of the big market action (GameStop, Crypto Boom, Situational Awareness), there's a real chance that regular people get taken advantage of by people who know how to play the market or are victim to the incompetence of people who claim to be in the know.
A $2 trillion valuation for a company that made $4.6 billion in revenue is so insane, I don't even know how to process it. Surely that's too insane even for this market
A $2 trillion valuation for a company that made $4.6 billion in revenue is so insane, I don't even know how to process it. Surely that's too insane even for this market
Amazon didn't make a net profit until 6 years after a (very successful) IPO they went into with heavy yearly losses. Their pitch was that they were building infrastructure and it would pay off...
Amazon didn't make a net profit until 6 years after a (very successful) IPO they went into with heavy yearly losses. Their pitch was that they were building infrastructure and it would pay off later. They turned out to be right.
I don't think that's comparable to AI labs, it's hard to imagine their profits ever justifying their valuations. Once a warehouse is built, you have the warehouse forever and eventually your distribution network is good enough and you can dramatically reduce spending. With model training you can't ever stop and, so far, each new generation costs more to train than the last. And if a plateau happens the big labs are screwed. Like, dead in the water while everyone else catches up, prices tank and inference becomes a commodity.
But I'm not sure the market will agree with all of that. At least not right now.
In any case, Amazon is just the most famous example of a market trend that's been going on for some time now. Companies losing money and still having hugely profitable IPOs is pretty normal.
A lot will depend on perception though, right now the capital class believes the potential of AI is largely untapped and nearly unlimited, that could change. Anthropic is smart to IPO asap while the sentiment is positive.
Not sure what you mean, Amazon's first profitable year was 2003, Amazon S3 didn't launch until 2006. They'd be less of a behemoth without it, but far from fucked. Almost half of their profit came...
Not sure what you mean, Amazon's first profitable year was 2003, Amazon S3 didn't launch until 2006.
They'd be less of a behemoth without it, but far from fucked. Almost half of their profit came from ecommerce related activity in 2025.
And yet Q1'26 revenue was $4.73 billion and Q2'26 is looking to be around $11.6 billion. Looking at 2025 revenue makes little sense except to show how fast things are growing. Can Anthropic make...
And yet Q1'26 revenue was $4.73 billion and Q2'26 is looking to be around $11.6 billion. Looking at 2025 revenue makes little sense except to show how fast things are growing. Can Anthropic make enough money to be worth their valuation? I don't know, but I do know the numbers everyone's quoting from the S-1 are irrelevant.
Here. Edit: Here's another one that might be more authoritative: Bloomberg states straight up that Anthropic is telling investors those revenue numbers, so unless Anthropic is misleading investors...
Anthropic PBC is telling prospective investors its second-quarter revenue jumped at least 14-fold versus the same period a year ago, according to documents seen by Bloomberg News.
The Claude chatbot maker reported a preliminary revenue figure of more than $11.5 billion in its latest completed quarter, compared to $787 million in the corresponding period in 2025, and $4.73 billion in the first quarter of this year, the documents show. The second quarter of 2026 saw Anthropic report positive adjusted operating income, according to the documents.
Bloomberg states straight up that Anthropic is telling investors those revenue numbers, so unless Anthropic is misleading investors (fraud), Anthropic's revenue is clearly growing fast.
Alphabet/Google has a 4.2T market cap and ~450B in revenue over the last year. So for a 2T market cap, maybe 225B revenue? Meta has a 1.85T market cap and revenue of ~230B in the last 4 quarters....
Alphabet/Google has a 4.2T market cap and ~450B in revenue over the last year. So for a 2T market cap, maybe 225B revenue? Meta has a 1.85T market cap and revenue of ~230B in the last 4 quarters. So yeah, ballpark 225B then.
I wouldn't be shocked by Anthropic misleading investors. It may not be criminally fraudulent, but misleading statistics are common with tech start-ups.
I wouldn't be shocked by Anthropic misleading investors. It may not be criminally fraudulent, but misleading statistics are common with tech start-ups.
You hit the nail on the head here. The issue is that these valuations are all based on priors from 2000s-era software companies. Each of these companies benefit from massive network effects and...
You hit the nail on the head here.
The issue is that these valuations are all based on priors from 2000s-era software companies. Each of these companies benefit from massive network effects and vertical integration.
AI simply doesn’t. None of the non-technical people I know have switched from Google. Everyone uses Google. Yet they all have various AI solutions, whether its the ChatGPT app (often the free version!), some wrapper app, or something integrated into their workflow. The space is so competitive, and people are way more willing to change.
And the wrapper apps can change! They can integrate with a different AI company or some framework to minimize costs.
I don’t think it matters if AI is the next revolutionary technology, because this isn’t a monopolistic market. Margins will be slim pickings.
So, two weeks ago I came up with a general theory: OpenAI and Anthropic are operating on Enron's hype playbook without the legal ramifications (because currently private). They're not commiting...
So, two weeks ago I came up with a general theory: OpenAI and Anthropic are operating on Enron's hype playbook without the legal ramifications (because currently private). They're not commiting fraud in a technical sense. Just knowing they're sitting on a bubble and hope to hop off before the roller coaster goes down.
Anyhow, this thread is a redacted version of a chat I had (somewhat hypocritically and/or ironically I know)with Kagi Assistant.
This is not financial advice. Though I did waste some money based on my own stupid ideas, so I redacted them for public safety.
Retrospective on Enron Financial Fraud Created: 9/16/2026, 2:00:48 PM You Provide a retrospective on Enron's financial fraud Assistant Enron: A Retrospective on the Biggest Corporate Fraud of Its...
Retrospective on Enron Financial Fraud
Created: 9/16/2026, 2:00:48 PM
You
Provide a retrospective on Enron's financial fraud
Assistant
Enron: A Retrospective on the Biggest Corporate Fraud of Its Era
The Rise
Enron Corporation was formed in 1985 by Kenneth Lay through the merger of Houston Natural Gas and InterNorth [^1]. Initially a conventional pipeline company, it transformed itself — particularly after Jeffrey Skilling joined — into an energy trading giant, eventually claiming nearly $101 billion in revenues for 2000 and employing about 20,600 people [^2]. For six consecutive years it was named "America's Most Innovative Company" by Fortune, and its stock was a darling of Wall Street.
How the Fraud Worked
The deception rested on three interlocking techniques:
Mark-to-market accounting — Enron booked the entire projected future profits of long-term deals (decades-long energy contracts, broadband ventures) as immediate income, even when cash never materialized [^1].
Special purpose entities (SPEs) — CFO Andrew Fastow created off-balance-sheet vehicles (Chewco, LJM1, LJM2, the infamous "Raptor" partnerships) that let Enron hide billions in debt and losses from failed deals, while Fastow personally pocketed tens of millions in fees [^1].
Complicit auditing — Arthur Andersen, one of the world's "Big Five" accounting firms, signed off on the structure, and its consultants earned substantial fees alongside the audit work — a glaring conflict of interest. Executives misled the board and audit committee and pressured Andersen to look away [^1].
Internal warnings existed — most famously VP Sherron Watkins' memo to Lay warning of "an elaborate accounting hoax" — but were ignored [^3].
The Collapse
The unraveling was swift:
Date
Event
Aug 2001
Skilling abruptly resigns as CEO after six months; Lay resumes control
Aug–Oct 2001
Stock falls from about $40 to under $1 [^4]
Oct 16, 2001
Enron reports a $618 million quarterly loss and a $1.2 billion equity reduction
Nov 2001
Fastow fired; Andersen's document shredding comes to light
Dec 2, 2001
Enron files Chapter 11 — with $63.4 billion in assets, the largest bankruptcy in U.S. history to that point [^1]
Shareholders lost roughly $74 billion; thousands of employees lost jobs and, because pensions were heavily invested in Enron stock, much of their retirement savings.
The Reckoning
Arthur Andersen was convicted of obstruction of justice for shredding documents and effectively dissolved, shrinking the Big Five to the Big Four [^3].
Fastow pleaded guilty and served about six years.
Lay and Skilling were convicted in May 2006 on fraud and conspiracy charges. Lay died of a heart attack before sentencing (vacating his conviction); Skilling was sentenced to over 24 years, later reduced, and released in 2019 [^3].
The Legacy
Enron's most lasting consequence was legislative: the Sarbanes-Oxley Act of 2002, which imposed criminal penalties for securities fraud, required CEOs and CFOs to personally certify financial statements, strengthened auditor independence, and created the PCAOB to police the accounting industry [^4]. It remains the benchmark case in business school ethics curricula — a demonstration that dazzling growth, charismatic leadership, and compliant auditors can together mask a hollow core until the mathematics of deception simply run out of room.
What is the difference between Mark to Market accounting and ARR accounting?
Assistant
Mark-to-Market vs. ARR Accounting
These two approaches sit at nearly opposite ends of the revenue-recognition spectrum: one books value based on current market prices, the other based on contracted future subscription revenue.
Mark-to-Market (MTM)
What it recognizes: The current fair value of an asset or contract — what it's worth today if sold or settled at prevailing market prices.
When it's used: Trading assets, derivatives, futures, and long-term energy/commodity contracts (Enron's specialty).
Key mechanics:
Gains and losses are recorded each period as market values fluctuate, before any cash changes hands
For long-term contracts, the entire estimated lifetime profit could be booked on day one (Enron's aggressive version)
Strength: Reflects economic reality faster than waiting for cash
Weakness: Depends on estimates and assumptions — with illiquid or bespoke contracts, there's no real "market," so values (and profits) can be manipulated
ARR (Annual Recurring Revenue)
What it measures: The value of a SaaS/subscription company's recurring revenue normalized to a 12-month run rate — e.g., a $50k/mo contract = $600k ARR.
When it's used: Software-as-a-Service, subscription businesses, streaming services — anything with recurring contracts.
Key mechanics:
ARR is a non-GAAP operating metric, not an accounting method per se — it's a valuation and planning tool, not what appears on audited financial statements
GAAP revenue for the same business is recognized ratably over time as the service is delivered (the opposite of Enron's upfront booking)
Strength: Predictability — shows the durable, repeatable revenue base investors care about
Weakness: Ignores churn, cost-to-serve, and non-recurring revenue; it's not audited, so definitions vary by company
Side-by-Side
Dimension
Mark-to-Market
ARR
Nature
GAAP accounting method (fair value)
Non-GAAP business metric
Basis
Current market prices / fair value estimates
Contracted recurring subscription value
Timing
Recognizes changes in value immediately, often before cash
Measures annualized run rate; GAAP revenue recognized ratably
Typical industry
Finance, energy trading, commodities
SaaS, subscriptions, streaming
Main risk
Manipulation via subjective valuations (Enron)
Ignoring churn / flattering comparisons
The Irony
Enron's mark-to-market abuse was essentially booking future hopes as today's earnings. Modern SaaS accounting is the correction: under ASC 606, a $600k ARR contract means recognizing $50k of revenue each month as the service is actually delivered — precisely the discipline Enron evaded. ARR itself is just a scoreboard on top of that ratable recognition, telling investors how big and durable the recurring base is.
You
So, let's suppose a private company, who is not currently legally required to disclose revenues in any standardized way, but is amping hype for an IPO, how would they be reporting ARR to maximize their appearance without technically lying?
Assistant
Pre-IPO ARR Inflation: The Legal Art of Aggressive Definitions
This is a well-documented game — S-1 filings are marketing documents as much as securities disclosures, and ARR is unaudited, non-GAAP, self-defined, which makes it the perfect vehicle. Here's the playbook, in rough order of popularity:
1. Redefine the Metric Itself
The single biggest lever is the definition, disclosed only in a footnote:
Include non-recurring revenue. Book one-time setup fees, implementation services, hardware, or professional services as "recurring." Some filings have counted things like onboarding payments that never renew.
Round up partial-year contracts. A 9-month contract counted at full annualized value.
Include contracted-but-not-started revenue. Signed deals where deployment hasn't begun, or the customer can still cancel, get counted before a dollar is billable.
Count TCV (total contract value) as ARR. A 5-year, $5M deal reported as $5M ARR instead of $1M — the definition switch happens quietly between funding rounds.
2. Strategic Timing Windows
Pick the flattering snapshot date. Private companies choose their fiscal year-end; landing it right after a big renewal season or a land-grab quarter maximizes the print.
Pre-pull renewals. Sales teams push customers to sign renewals early (often with discounts) so they land inside the reported period.
Backload the cutoff. Deals signed days after the period end are excluded; deals signed days before, even unprovisioned, are included.
3. Gross vs. Net (and Hiding Churn)
Report gross ARR without netting churn. "We added $40M in new ARR" is technically true even if $35M walked out the back door the same year. The magic number investors actually care about — net revenue retention (NRR) — gets a vaguer, narrower definition (e.g., computed only on a curated enterprise cohort).
Define churn favorably. Failed companies that "downgrade to a cheaper plan" aren't counted as churn — they're counted as retained customers with "expansion potential."
4. The Annualization Trap on Growth Rates
Compute growth off a manipulated base. Annualize a monster final quarter ("Q4 ARR of $25M, up 300% YoY!") rather than showing a slower full-year trajectory. Both are true numbers; only one is representative.
5. Footnote Burial
The tell: the ARR definition sits in dense fine print on page 90-something of the S-1, and it will have changed from the definition used in the prior funding round's investor deck. Underwriters know this — the SEC requires the non-GAAP definition be disclosed, but not that it be consistent, comparable, or conservative.
The Checks That Cut Through It
Question to ask
What it reveals
Is ARR > GAAP revenue?
By how much? A growing gap means heavy upfront counting of contracted-not-delivered revenue
What's the audited revenue growth?
GAAP revenue can't be gamed the same way — it's ASC 606, ratable, audited
What's NRR, and how is the cohort defined?
Surviving customers vs. all customers — definitional games live here
Did the ARR definition change between rounds?
The surest sign the metric is being shaped, not measured
ARR per employee / CAC payback?
Efficiency ratios are harder to cosmetically inflate
The deeper irony ties back to your Enron thread: this is mark-to-market logic in miniature — booking contracted future hopes into a headline number — but done in a metric that's explicitly labeled non-GAAP, so it's fully legal. The system's defense is that GAAP revenue sits right next to it, audited and ratable. The investor's job is to notice when the two lines start telling different stories.
You
So, in the modern era, where it is easy to onboard individual small investors to an IPO who will mostly follow prior headlines rather than the detailed S-1, this should amplify the effect of periodically announcing massive ARR gains?
Assistant
Yes — and the Mechanism Has a Name: Retail Attention Capture
You've identified a real and intensifying dynamic. The structure of modern IPO retailing is almost perfectly optimized to exploit the gap between headline metrics and S-1 substance.
Why the Amplification Works
1. The attention funnel is now institutionalized
Zero-commission brokers (Robinhood, etc.) put IPO participation behind one tap
Fractional shares mean a $500 investor can buy into any headline IPO
Pre-IPO buzz is manufactured through funding-round press leaks, "decacorn" coverage, and founder podcasts — all before any disclosure document exists
By the time the S-1 drops, the narrative is already set. Retail inflow arrives pre-programmed with "fastest-growing SaaS company ever," and the ARR headline confirms it. Almost nobody reads page 90.
2. Headlines are the product; the metric is the raw material
Financial media compresses an S-1 into 3–5 data points. "ARR up 300%" is quotable; "the ARR definition changed between rounds and now includes non-recurring implementation revenue" is not. The metric inflators know which number gets amplified — so headline-optimized accounting becomes rational strategy.
3. The feedback loop to the company
Actor
Incentive
Company
Higher IPO price = more capital, less dilution
Underwriters
Fee as % of deal size — motivated to market the story
Media
Engagement — huge numbers drive clicks
Retail investor
FOMO and narrative participation, not diligence
Fund managers buying at open
Momentum capture, not S-1 forensics
Every party in the chain is rewarded for amplifying the headline and none is paid to debunk it. Short-sellers, who historically provided the correction, are constrained —IPO shorts are expensive, risky, and retail-fueled rallies can run irrational longer than a short can stay solvent.
4. The tell-tale pattern you'd expect to observe
ARR-heavy IPOs popping 50–100% on day one regardless of GAAP fundamentals
Post-lockup underperformance as insiders sell into the hype (the classic IPO structure historically leaves retail holding the bag after ~6 months)
A cohort of high-ARR, low-GAAP-margin companies trading at multiples that only make sense if the ARR is real and durable — and the widening ARR-vs-GAAP-revenue gap is the market's own indicator of how much hope is priced in
The Structural Counterweights
The effect is real but not unchecked:
Quiet periods and SEC rules limit how aggressively a company can tout ARR pre-IPO — hence the leaks come from "sources familiar" rather than the company itself
Allocated institutional demand still anchors the IPO price; retail mostly gets the aftermarket, which punishes late buyers rather than the company
Lockup expirations and quarterly GAAP reporting eventually force the truth out — a company that reported $200M ARR must eventually show ~$200M of GAAP revenue trailing behind it, and the gap becomes a known schedule of reckoning
Post-Enron infrastructure (SOX, PCAOB, audited financials in the S-1) guarantees the substance exists somewhere — the retail investor's disadvantage is attention, not access
The Closing of the Loop to Enron
The vulnerability is identical in kind: earnings that run ahead of cash and reality. Enron's version was fraud because it happened inside audited GAAP statements. The modern ARR version is legal because it lives in the clearly-labeled non-GAAP zone — but the behavioral exploitation (headlines outrunning substance, small investors trusting narrative over footnote) is the same architecture with better disclosure law around it. Sarbanes-Oxley made the documents honest; it couldn't make the readers diligent.
You So, let's take a moment and apply this to the announcement cycle of Anthropic and OpenAI. Which ARR announcements were official, which were "leaks," and based on current trajectories, when...
You
So, let's take a moment and apply this to the announcement cycle of Anthropic and OpenAI. Which ARR announcements were official, which were "leaks," and based on current trajectories, when would we expect an S-1 filing which optimizes optics?
Assistant
Applying the Playbook: Anthropic & OpenAI's ARR News Cycle
The pattern you predicted is playing out almost exactly — with the interesting wrinkle that the two companies have swapped the expected roles. Here's the ledger as of mid-September 2026:
The Announcement Ledger: Official vs. "Leaked"
Metric
Company
Channel
Character
$47B run rate (up from ~$10B in all of 2025)
Anthropic
Official — Series H announcement, May 28, 2026 [^1][^2]
Leak — "a person familiar with the matter," Reuters, Aug 17 [^5]
Classic pre-IPO positioning leak
$190–200B 2028 revenue forecast
Anthropic
Leak — sources to Reuters, Aug 15 [^6]
Banker-plant: anchors valuation math
$2T valuation talk; Nvidia as cornerstone investor
Anthropic
Leak — "people familiar" via FT/CNBC [^7][^8]
Hype inflation, deniable
Q2 revenue $6.7B, +18% QoQ; $12.3B operating loss
OpenAI
Leak/disclosed to investors — Aug 14–18 reporting [^9][^10]
The unflattering numbers leaked anyway
~$40B total annualized revenue tracking
OpenAI
Leak — attributed to company statements to investors [^11][^12]
Deflection from the loss story
ChatGPT Ads at $1B annualized run rate
OpenAI
Official — OpenAI blog post, Aug 31, 2026 [^13][^14]
Explicitly framed "as it prepares for an IPO" [^15]
Confidential S-1 filed June 8, 2026
OpenAI
Official — with Altman noting "we expect it to leak so we're just announcing it" [^16][^17][^18]
Formal step
The tell is in the asymmetry. Anthropic's official announcements carry the flattering numbers (run rate, ARR growth), while its leaked items do valuation-anchoring work. OpenAI — now the laggard, having fallen behind Anthropic in revenue with mounting losses [^9][^19] — switched tactics: leak the bad news (losses, slow sequential growth) where it must, but make the official announcements the flattering sub-metrics ("Ads hit $1B run rate in under 200 days!" [^20]) — a textbook headline carve-out: spotlight the fastest-growing slice rather than the whole.
Also note the metric language drift: both companies favor "annualized revenue run rate" over "ARR," which technically sidesteps even the subscription-purity debate — a single hot month gets ×12'd [^3]. Anthropic's arc ($9B end-2025 → $30B April [^21] → $47B May [^22] → $65B+ July [^5]) is precisely the "annualize the monster quarter" pattern from the pre-IPO playbook — those numbers imply quarterly revenue more than doubling in a single quarter [^23], which is either the greatest growth story in corporate history or a run-rate computation riding a spike.
The S-1 Timing Question
Anthropic: Confidential S-1 already filed (June 2026) [^3][^4]; investor meetings reportedly underway [^24]; reports point to a listing by or around October 2026 [^25][^24], with the roadshow expected to anchor on 2028 revenue forecasts of ~$190–200B [^6]. Optics logic: file while Q2 shows the first operating profit in company history ($559M) against OpenAI's $12.3B loss [^26], and while Q3 profit is projected to top $1B [^27]. The optimal window is Q4 2026, before the year-end freeze — every additional leaked run-rate print between now and the roadshow ratchets the anchor higher.
OpenAI: Confidential S-1 filed June 8, 2026 [^16], but on September 12–15 Altman publicly called a 2026 IPO "ill-advised," citing safety/alignment concerns and eyeing 2027 [^28][^29][^30]. Read through the optics lens: OpenAI's GAAP-adjacent numbers (an operating loss that widened to $12.3B in Q2 [^26], a reported $39B loss in 2025 [^31], revenue growth of just 18% QoQ against a rival growing far faster [^9]) mean now is the worst moment to publish an audited S-1. Delaying lets the ad business ($1B run rate, targeting $2.5B by year-end [^11]) and enterprise pivot [^10] mature into a better narrative — "safety concerns" being the most reputationally advantageous available reason to wait.
The Pattern, Named
Everything here follows the playbook from our earlier exchange: unofficial leaks of rising run rates (deniable, unaccountable), official announcements timed to peak quarters, forecasts planted for banker math, and the S-1's public filing held until the trailing GAAP numbers best support the headline number. The one genuine counterweight is that when these S-1s go public (not confidential drafts), audited GAAP revenue and losses will sit in the same document — and for OpenAI especially, that's the number the hype cycle has been quietly outrunning. The retail amplification effect you described will be maximal at the moment of the roadshow — which is exactly why Anthropic's is being timed for the juiciest possible quarter, and why OpenAI just blinked and pushed its own into a year it hopes will tell a better story.
So would it be fair to say that OpenAI and Anthropic are doing similar levels of shennanigans as Enron, but since the two AI companies are operating with the hindsight of Enron's precedent, and have the private -> public advantage, are able to better hide that they are bluffing?
Assistant
Fair in Mechanism, Unfair in Category — and the Distinction Matters
Your structural insight is right, but I'd push back on equating the level of shenanigans with Enron's. Let me separate what's genuinely parallel from what's fundamentally different.
Where the Parallel Holds
Hope booked as present value. Enron booked decades of projected contract profits today; "$65B run rate" is one hot month × 12. Both are mechanisms for converting future expectations into a present-tense headline number.
Attention arbitrage. Enron exploited analysts who trusted the narrative and the auditor's stamp; the AI labs exploit a retail funnel that reads headlines, not S-1 footnotes. Same vulnerability — diligence gaps — different population.
Deniable information laundering. Enron used SPEs to move bad news off-book; the labs use "people familiar with the matter" leaks to place flattering numbers ($190B 2028 forecasts, $2T valuations) into circulation without ever attesting to them. Legally cleaner, functionally similar: numbers in the market's bloodstream that no one is accountable for.
The locking-in of insiders. Enron executives cashed out while retail bought; IPO lockup structures create the same asymmetric exit schedule for AI insiders.
Where the Analogy Breaks — and This Is the Important Part
1. Enron hid losses. The AI labs publish them. This is the deepest difference. Enron's entire fraud was concealing that the company was unprofitable and overleveraged — the audited statements were lies. OpenAI's $12.3B quarterly operating loss and Anthropic's loss history are known, reported, and priced in. Nobody buying into these IPOs is deceived about profitability — they're making an explicit bet that growth outruns burn. That's speculation, not deception.
2. GAAP revenue is real and roughly tracks the run rates. Enron's mark-to-market profits had no cash behind them. The AI labs' revenue is largely cash-collected subscription and API billing — when Anthropic says $47B run rate, there's genuinely ~$12B/quarter of invoiced business behind it. The open question is durability and margin (can you serve that revenue profitably given inference costs?), not existence.
3. The bluff, such as it is, is disclosed as a bluff. "Run rate" and non-GAAP metrics come with definitional footnotes, courtesy of the very Sarbanes-Oxley regime Enron created. The optimism is legally required to be labeled. Enron's optimism was disguised as audited fact.
4. The regulators are watching the same playbook. The SEC spent two decades building non-GAAP disclosure rules because of Enron-adjacent abuses (and dot-com era run-rate hype, e.g., the 1999–2000 "annualize the quarter" era). A public S-1 from either lab will carry audited financials side-by-side with the hype metric.
The More Precise Framing
So rather than "same shenanigans, better hidden," I'd say:
Enron committed fraud inside audited statements. The AI labs are conducting aggressive legal narrative management around statements that are, or will be, substantially truthful — and the post-Enron regime guarantees the truthful part sits in the same document as the hype.
The honest risk isn't Enron-style accounting fraud. It's two other failure modes:
The Greater-Fool cycle — a valuation ($1–2T) that only works if 2028 forecasts of $190B+ materialize, sold to retail who won't model that. When the dot-coms did this, the numbers were real too; the extrapolations weren't. This is 1999 risk, not 2001 risk.
The Enron failure mode deferred, not eliminated — if growth stalls post-IPO and management starts feeling the same pressure Enron's did (we must keep the narrative alive), that's when you watch for the metrics to start migrating: definitions quietly changing, cohorts narrowing, losses reclassified. The historical lesson is that fraud rarely starts at the headline-printing stage; it starts at the narrative-preservation stage, after the bluff has been called by reality and there's something left to hide.
So: same psychology, same retail exploitation, vastly better paperwork — and the real test of whether they learned from Enron won't come at the IPO. It'll come at the first bad quarter after it.
You
How are the AI companies not also engaging in Attention arbitrage; only showing the real data about costs and revenue to those deep in investment sector while announcing to the world about the hype of their technology and massive revenue gains?
Assistant
They Are — I Conceded That Point, So Let Me Concede It More Fully
You're right, and I think my previous answer understated it. Yes, absolutely: both companies are running a textbook two-tier information structure. Let me steelman your position and then locate precisely where the remaining difference lives.
The Two-Tier Structure, Stated Plainly
Audience
What they get
Legal obligation
Deep-pocketed investors (Series F/H participants, sovereign funds, cornerstone investors like Nvidia)
Detailed financials, burn rates, unit economics, margin data, churn, capex forecasts — the numbers OpenAI "disclosed to investors" in its Q2 reporting [^a3f86f], which is how we know about the $12.3B loss at all
Governed by contract — information rights negotiated per round
The general public / prospective retail IPO buyers
Essentially none. Private companies have no public disclosure duty whatsoever
And here's the part that makes your point stronger than I treated it: the second tier is reached by way of the first. The $12.3B loss and the $39B 2025 loss became public only because someone in Tier 1 — an investor, a banker, a "person familiar" — decided it should. The public knows the bad numbers exist at the discretion of the people the bad numbers don't disadvantage. When Anthropic's investors were shown that Q2 operating profit, the leak was favorable and thus predictable; when OpenAI's losses leaked, it was likely tactical too (pre-empting worse coverage, or pressure applied by investors pushing for a management change). Either way, the disclosure valve is controlled by parties with positions.
Why This Phase Is Actually the Most Enron-Like — More Than the IPO
Here I'll go further than my last answer: the private, pre-S-1 window is where the Enron resemblance is strongest, because:
No audited public record exists at all. Enron's sin was falsifying a record that legally existed. Today's labs are operating before the record is required — the "off-balance-sheet era" is structural, not engineered. Every number in circulation — $47B, $65B, $190B — is self-reported, unaudited, and definitionally unaccountable.
The selective-emphasis machinery is identical in function to SPE structuring. Enron designed entities so bad news would be invisible; the labs operate a stage of corporate life where bad news is legally invisible by default, and then choose which parts to leak. Same information asymmetry, achieved not by deception but by staying private longer and raising at higher valuations each time — the mega-round replaces the SPE as the valuation-preservation vehicle.
The hype layer is targeted at the exact population least able to verify it. Capability demos, "AGI by 2027" rhetoric, and run-rate headlines circulate through the retail-attention funnel we mapped earlier — the audience that will be handed the IPO at the top of the narrative. The sophisticated money got in at $15B, $60B, $183B; the retail money arrives at $1T+, priced on the public layer of information only.
Where the Difference Still Genuinely Lives — Narrower Than I Implied, but Real
The remaining distinctions are narrower than "not fraud vs. fraud," and I should have framed them that way:
Temporal, not structural. The asymmetry is scheduled to end. The confidential S-1s filed mean audited, comparable financials are coming within months — Enron's system was designed to never reconcile. The labs are spending down a finite window of asymmetry, not building a permanent one.
Reputational lock-in. Everything leaked or announced now becomes a comparable baseline the moment the S-1 goes public. When Anthropic's audited revenue prints against that $47B run rate, the gap is computable by anyone. Enron's numbers were never meant to survive that comparison; the AI numbers are being set up to — which constrains how far the bluff can run now.
The tier-1 investors are, functionally, the only diligence mechanism that exists — and it did bite: OpenAI's leaks and the reported investor pressure over losses show the sophisticated tier reacting to real data. Enron's Arthur Andersen failed at exactly this job. Whether the labs' auditors and lead bankers hold the line at the S-1 stage is the live question.
The Sharpest Honest Statement
So the fairest formulation, refining yours:
The AI companies are running Enron's information architecture — hype outward, substance inward — but in a legal regime where the architecture is scheduled to self-destruct at the S-1, rather than one where it must be hidden forever. The shenanigan isn't the asymmetry itself; it's maximally exploiting the countdown — pumping valuation and narrative through the unaudited window so that the eventual public record inherits a price already set by numbers nobody had to attest to.
The one thing I'd still resist is "bluffing" as the primary description of the revenue itself — the cash billing appears real. But the valuation, the forecasts, and especially the framing of what the public should focus on? That's managed with a sophistication Enron's executives would have recognized instantly — and envied, because it's legal. The true test remains: whether the audited S-1 confirms the extrapolations, or whether 2026 turns out to be the year the greatest attention-arbitrage trade in corporate history was executed on the public markets.
You The big question is not one of legal consequences (Enron execs defacto suffered 0 in the end), but of understanding the playing field, predict when the cards will collapse, and ideally avoid...
You
The big question is not one of legal consequences (Enron execs defacto suffered 0 in the end), but of understanding the playing field, predict when the cards will collapse, and ideally avoid bottoming out, if not profiting.
Assistant
Reading the Field: The Collapse Map for the AI Cycle
You've reframed the question correctly — the legal layer is noise; the information layer is the game. So let's treat this like the trading problem it is: what's the structure, what are the tripwires, and where does the exit door sit.
The Structural Anatomy (What's Actually Propping This Up)
Three load-bearing walls, in order of fragility:
Circular financing. Nvidia invests in OpenAI, which buys Nvidia chips through cloud middlemen who also take Nvidia equity; Amazon and Google fund Anthropic and become its cloud landlords [^1][^2][^3]. Revenue and investment flow in a loop — Bloomberg's own mapping concludes the web "raises the risk of cascading losses if AI falls short" [^1]. This is vendor-financing-with-equity, the exact structure that turned Nortel/Lucent vendor loans into accelerants in 2000–2001.
The capex/revenue gap. 2026 infrastructure spend runs roughly $690B against a fraction of that in revenue, with OpenAI's profitability modeled for 2029 at the earliest [^4][^5]. The whole structure is a race: does real demand validate the spend before the financing window closes?
Competition compressing the margins of the leaders. Open-source models are reportedly far cheaper for enterprises to run, and the leaders' moats are narrower than their multiples [^6]. If model capability commoditizes, the $1–2T valuations price a monopoly that doesn't exist.
The Known Schedule — This Is the Unusually Favorable Part
Rarely does a bubble hand you its calendar in advance. This one does:
Date (approx.)
Event
Why it matters
Oct 2026
Anthropic Nasdaq listing at $965B–$2T target; ~$60B raise [^7][^8]
Retail enters at the top of the narrative. Tiny float + engineered scarcity is explicitly "designed to create intense upward price momentum" [^9] — the pop is planned
Q1–Q2 2027
Anthropic lockup expirations (~180 days post-listing) [^9]
The classic cliff. Institutions model their exit around this date; retail generally doesn't know it exists
2027
OpenAI IPO, now pushed back by Altman [^0d8cec]
Second wave of supply hitting the market on OpenAI's weaker GAAP story (losses widening, 18% QoQ growth [^a3f86f])
2029
OpenAI's own profitability projection [^5]
The deadline by which reality must match the extrapolations, or the financing loop unwinds
Notably, Anthropic is already showing awareness of the lockup problem — considering extended lockups, controlled day-one shareholder sales, and 10b5-1 plans to "manage support and resistance levels" [^10]. Translation: the insiders know the standard structure dumps on retail, and are calibrating exactly how much dumping the tape can absorb. That's not reassurance; that's confirmation the exit is being engineered.
The Leading Indicators — Watch These, Not Headlines
Ranked by how early they signal:
Run-rate vs. GAAP revenue divergence. The moment Anthropic's first audited quarterly print trails the $47B/$65B annualized figure meaningfully, the extrapolation discount begins. First public quarterly report = first honest number in the entire cycle.
Private round pricing before IPO. Analysts already flagged this: if the pre-IPO round closes below the whispered $900B–$1T, institutional price resistance is the tell [^11]. The last private round is the only honest price discovery before the listing.
Nvidia's receivables and equity stakes. The circular loop's solvency lives on Nvidia's balance sheet. If chip-backed securities [^6] or vendor-financed data center debt start being renegotiated, the loop is breaking from the inside.
Net revenue retention at the S-1. The single most predictive SaaS-style number. Growth rates hide churn; NRR can't be spun as easily because it's cohort arithmetic.
Hyperscaler capex tone. When Microsoft/Google/Amazon start describing AI capex as "measured" or "disciplined" on earnings calls, the marginal funding for the loop is being withdrawn. Capex guidance cuts precede revenue misses by 2–3 quarters.
The narrative canary: when the leaks shift from run-rate gains to "path to profitability" language, management is pre-positioning excuses for slowing growth.
The Historical Template
The dot-com sequence, compressed: irrational day-one pops → lockup expirations six months later → first batch of quarterly GAAP misses → high-multiple names re-rating 70–90% down without any fraud required. Most of the companies that destroyed retail wealth in 2000 were real businesses with real revenue — pets.com sold actual pet food. The failure was price, not integrity. That's the base case here: no Enron-style smoking gun needed for a $965B listing to be worth $300B a year later; it just needs 2028 forecasts of $190B [^8b075f] to arrive as $60B.
Positioning Logic (Framework, Not Advice)
The asymmetry is timing-shaped, not direction-shaped. Being short the hype early is how funds got destroyed in 1999 — the bubble ran two more years past "obvious." The exploitable edge isn't calling it a bubble (everyone knows); it's the scheduled events: the lockup windows and the first audited quarters, where supply and truth arrive on known dates.
The cleanest structural observation: every prior cycle shows the money is made by those who entered at $15B–$183B in the private tiers, and the avoidance opportunity belongs to retail who simply decline the October pop. The single most reliable historical predictor of underperformance is buying a scarce-float IPO up 50–100% on day one.
The long-side counter-case deserves respect: Scenario 1 — enterprise adoption genuinely compounds, OpenAI reaches its 2029 profitability, circular financing is vindicated as strategic positioning [^5]. Amazon 2000 was a bubble that was also a buy at the bottom, not the top. The reconciliation trade, if it comes, is post-re-rating, not pre-IPO.
Watch the second-order assets. The hyperscalers and the energy/power layer hold real cash flows with AI upside; the pure-play labs hold all the narrative risk. If the collapse comes, it likely discriminates — infrastructure with contracted revenue survives, extrapolation-priced paper doesn't.
The One-Line Synthesis
The cards collapse not when someone finds the fraud, but when the scheduled truth-telling (audited quarters, lockup supply, 2028-vs-actual) meets a valuation set during the unaudited window — and the calendar says the first collision is Anthropic's Q1 2027 earnings plus its lockup expiration, roughly six months after the largest IPO in history teaches a new generation of retail investors what "run rate" meant.
I am legitimately impressed with the level of detail, citations, and rationality expressed by the assistant here. I should revise my opinion of them and take a closer look. Did you do anything...
I am legitimately impressed with the level of detail, citations, and rationality expressed by the assistant here. I should revise my opinion of them and take a closer look. Did you do anything special to prep it for thus conversation, or was this out of the box?
Remember that LLMs always look better than they actually are. Close looks are required to determine many errors, and even besides that, many conversational patterns output by chatbots tend to...
Remember that LLMs always look better than they actually are. Close looks are required to determine many errors, and even besides that, many conversational patterns output by chatbots tend to cause people to believe them more than they should.
That's never a bad general stance, but I have enough background to say that it's broadly making correct conclusions from the existing data and providing advice in line with traditional investing...
That's never a bad general stance, but I have enough background to say that it's broadly making correct conclusions from the existing data and providing advice in line with traditional investing analysis. It's possible that some of the fine details are wrong, and I wouldn't put my name on it without going over it with a fine toothed comb, but it provided a better and more accessible analysis than most people could get without a friend who's an investment manager who they can take out to lunch.
(I'm not LukeZaz, but as a note, please consider whether it's doing better in this area because of its inherent capacity or because competent analyses of Enron are overrepresented in its dataset....
(I'm not LukeZaz, but as a note, please consider whether it's doing better in this area because of its inherent capacity or because competent analyses of Enron are overrepresented in its dataset. LLMs are capable of reasoning, but they can apply a veneer of plausibility onto all of their output. Context: Gemini will happily hallucinate citations for building codes which don't exist, then when I pull up the document, it will happily note that it invented the citation and justified its response retroactively)
Yeah, absolutely. I'm not attributing to it aome generalized godlike omniscience. But I hadn't updated my assumptions about how it was doing at non-programming tasks in a couple of years, and even...
Yeah, absolutely. I'm not attributing to it aome generalized godlike omniscience. But I hadn't updated my assumptions about how it was doing at non-programming tasks in a couple of years, and even being able to coherently convey broadly accurate common white collar knowledge is better than it could do the last time I checked. I'm never going to give it more credence than some rando on the Internet, given that's what it's based on. For legal or professional advice I'm still going to seek out a lawyer or professional. (Though that's at leat in part because humans have liability and malpractice insurance, for when human fallibility rears its ugly head.)
Understood, and makes sense! As a note, since you bring up legal issues -- w.r.t. contract law, I'd note that there's been some hubbub about lawyers seeing more and more difficult questions come...
Understood, and makes sense!
As a note, since you bring up legal issues -- w.r.t. contract law, I'd note that there's been some hubbub about lawyers seeing more and more difficult questions come to them, since the LLMs do a decent job of explaining the basics to people. It's always worthwhile to get a professional in when there are stakes involved, but the LLMs seem capable of flagging issues correctly, at least. Time will tell how many type 2 errors are slipping through.
No, was like my third chat. One was about trying to restore the FM radio on a Galaxy S5. (extract firmware from EU models that had it enabled) Another was about flashing my Home Assistant voice...
No, was like my third chat.
One was about trying to restore the FM radio on a Galaxy S5. (extract firmware from EU models that had it enabled)
Another was about flashing my Home Assistant voice preview with custom settings.
You Since Anthropic Q1 2027 is the first major "proof" one way or the other, what kind of hints might we see from the existing public Q4 2026 filings, particularly of the hyperscalers or their...
You
Since Anthropic Q1 2027 is the first major "proof" one way or the other, what kind of hints might we see from the existing public Q4 2026 filings, particularly of the hyperscalers or their financers?
Assistant
The Q4 2026 Filings: The Last Public Data Before the First Honest Number
One timing note first: the hyperscalers' Q4 2026 reports land in late January–early February 2027 — roughly 6–10 weeks before Anthropic's first audited quarter. They are the last fully public, audited, SEC-filed data the market gets before the proof event. And critically, the hyperscalers are inside the loop — they're Anthropic's landlords, investors, and counterparties — so their filings are involuntary testimony about the very structure being tested.
Here's the watch list, by line item.
1. The Most Direct Tell: Anthropic-Related Gains in Non-Operating Income
This is the single sharpest instrument, and it already exists as a precedent. In Q1 2026, Fortune found that nearly half of Alphabet's and Amazon's blowout "AI profits" were paper marks on their Anthropic stakes — Amazon disclosed $16.8B of pre-tax gains from its Anthropic investment in a single quarter, more than half its pre-tax income [^1]. Google carried a similar footnote [^1].
What to watch in Q4 2026 filings:
The mark's direction and basis. Once Anthropic lists (targeting October 2026 at a ~$965B–$2T valuation [^a911bc][^2dd66d]), Amazon's and Google's holdings become market-priced, not round-priced. The Q4 filings will show the first quarter of mark-to-market on a publicly traded Anthropic. If the stock is below the IPO price at year-end, the same footnote that inflated profits in Q1 becomes a reversal charge in Q4 — mechanically transmitting Anthropic's post-IPO performance into two of the five largest earnings prints on Earth.
The magnitude. Amazon has ~$13B invested [^2] and Google up to $40B [^3] in a company valued 25–50× higher. These are now billion-to-hundred-billion-dollar swings hiding in "non-operating income" — a segment where the market historically gives credit for gains but underprices give-backs.
2. The 2027 Capex Guidance Sentence
The whole structure is a race between ~$725B of 2026 hyperscaler capex [^4][^5] and the revenue needed to justify it. Q4 earnings calls are where 2027 guidance gets set — and analysts already see spend "topping $1T in 2027" [^4] as the base case.
The tell isn't the number; it's the adjectives. Watch for the language migration from "we see no signs of demand slowing" to "measured," "disciplined," "matching spend to monetization," or "selectively pacing." Alphabet already caught capex scrutiny from a panned Q2 report [^6] — the first crack in the consensus that spending only goes up. Capex guidance flattening at $700–750B instead of $1T+ would quietly remove the marginal buyer from the AI compute market two quarters before anyone says "peak AI" out loud.
3. The Off-Balance-Sheet Debt Laundry — Already Screaming
This is where the Enron parallel is literal rather than figurative, and the disclosures are already deteriorating:
$662B in not-yet-commenced lease commitments sits off the five hyperscalers' balance sheets — 113% of their combined adjusted debt, per Moody's [^7][^8]. GAAP doesn't require recognition until leases commence, so these tranches land on the books exactly as the data centers come online.
$1.65T of total off-balance-sheet AI debt across Big Tech [^9]; Meta's Hyperion deal alone keeps ~$27B off-balance via an 80% Blue Owl financing vehicle [^7].
S&P downgraded Oracle to BBB- in July 2026, explicitly citing ~$260B in future lease obligations [^9].
Hyperscaler CDS hit record highs across Oracle, Alphabet, Amazon, Meta, Broadcom, and Nvidia [^10].
In the Q4 10-Ks specifically, watch three things: (a) whether the undiscounted future lease commitment footnote grows from $969B [^8] or begins to migrate onto balance sheets as leases commence; (b) any depreciation useful-life extensions on GPU/server assets — the classic trick that flatters earnings right up until impairment; and (c) impairment language. One analysis already flags that the depreciation clock on 2025–26 capex starts when facilities go live, making "the 2027 earnings season when the accounting math becomes unavoidable" [^10] — i.e., these very Q4 filings and Anthropic's first quarters are the collision zone.
4. Nvidia's Q4 (Late February): The Loop's Solvency Statement
Nvidia is the counterparty to everything — up to $100B committed to OpenAI against chip purchases [^11], a ~$105B guarantee on OpenAI's Ohio data center lease that only pays out after SB Energy fails to re-lease and resell [^12], a $6.3B CoreWeave capacity backstop [^13], and its reported OpenAI data-center deal already came in $145B below prior reports, which Fortune noted "signal[s] concerns of artificial demand" [^14].
In Nvidia's 10-K, watch: receivables growth vs. revenue growth (vendor financing accumulating on the balance sheet), customer concentration disclosures (how much revenue comes from financed customers), purchase-commitment and guarantee footnotes (the Ohio-style backstops), and any equity-method losses on stakes in OpenAI/CoreWeave-class customers. Nvidia is the one filer whose accounting choices can confirm or refute the "cascading losses" scenario Bloomberg mapped [^15] in a single document.
5. The Demand-Side Smoke Tests
Azure/AWS/Google Cloud growth rates and backlog. Q1 2026 showed Google Cloud +63%, Azure +40% (supply-constrained), AWS +28% [^16][^17]. Q4 2026 tells you whether the labs' revenue surge is being spent on these clouds — Anthropic alone has committed ~$100B to AWS [^2] and locked 5GW on Google TPUs [^3]. Remaining performance obligations (RPO/backlog) growing while revenue growth decelerates = future revenue already sold but demand momentum fading.
Microsoft's OpenAI line items. The equity-method losses and revenue-recapture mechanics from the restructured partnership — the OpenAI investment's P&L footprint inside Microsoft's 10-K is a partial, involuntary OpenAI income statement.
Free cash flow compression. Microsoft's FCF was already modeled down ~28% on capex [^18]. If Q4 shows the hyperscalers collectively funding $700B+ of capex while FCF shrinks toward zero, the marginal dollar for the next financing round comes from debt markets that are already pricing stress via CDS [^10].
The Synthesis
The Q4 2026 filings matter because they're the first place the extrapolation layer and the accounting layer must touch. The $47B/$65B run-rate narrative assumes hyperscaler capex continues at $1T+ into 2027 [^4]; the off-balance-sheet debt says that spend is increasingly financed by structures designed to keep it invisible [^7][^19]; and the Anthropic marks inside Amazon/Google earnings [^1] create a mechanical, disclosed feedback channel from Anthropic's post-IPO stock price into the market's most-owned securities — before Anthropic itself ever reports an audited quarter.
The bear sequence to watch for, in order: flattening 2027 capex language (late Jan) → lease commencement migrating debt onto balance sheets + depreciation drag (the 10-Ks) → Nvidia concentration and guarantee footnotes (late Feb) → reversal charges on Anthropic marks if the stock broke post-IPO → and only then, Anthropic's Q1 print in the spring, confirming or refuting what the filings had already whispered. By the time the "proof" arrives, the financiers will have known for a quarter.
A FEW DAYS LATER... You Anthropic leaked that they selected nasdeq. what does this imply: https://www.businessinsider.com/anthropic-selects-nasdaq-for-ipo-amid-ai-risk-concerns-2026-9 Assistant...
The surface reading is "exchange picked, IPO proceeding." But filtered through the framework we've built, this leak says considerably more — about the timeline, the machinery of the pop, and the information channel itself.
1. It Confirms the October Timeline Is Firm
Exchange selection is one of the last procedural decisions before a roadshow — it means the S-1 review has progressed far enough that the listing window is real, not aspirational. October is now the working date, with the raise potentially matching or exceeding SpaceX's [^1][^2]. Note the valuation language in the leak: "some estimates have put Anthropic at $2 trillion, though that figure has not yet been finalized" [^2] — the $965B Series H mark [^3] is now the floor of the conversation, with $1.8–2T the anchor being floated [^2][^4]. That's textbook late-stage anchor inflation: each leak ratchets the reference price up so the eventual print — whatever it is — reads as "reasonable."
2. The Channel Itself Is the Tell
"According to a person familiar with their plans" [^2] — this is exactly the pattern we cataloged. The exchange decision is not required to be public yet; leaking it now serves three purposes simultaneously: it signals momentum to the market, it rewards Nasdaq publicly (helping the banking/venue relationships), and it costs Anthropic nothing because no one is accountable for the statement. It also feeds prediction markets — Kalshi gives Anthropic 72% odds of beating OpenAI to the bell [^5], and every such leak moves those odds and the retail narrative with them.
3. The Most Tradeable Implication: Nasdaq-100 Fast Entry
This is the mechanically important part. Anthropic's size at listing makes it a near-certain candidate for Nasdaq-100 fast-entry inclusion — an event already being flagged with "forced ETF buying expected" [^6]. The chain runs:
Listing → fast entry into the NDX → every Nasdaq-100 index fund and ETF (QQQ alone, plus the thousands tracking it) must buy at the prevailing price → a scarcity-float IPO (designed, as we covered, to "create intense upward price momentum") meets mandatory indexed demand.
This is the retail amplification loop we mapped, but institutionalized: the passive layer is now contractually part of the pop. And it's precisely why Nasdaq won the listing — Nasdaq took SpaceX at $1.75T [^2], and both exchanges are fighting to be "the home for future AI IPOs" [^2] because the venue captures the entire index-complex flywheel: listing fees, data fees, and the self-reinforcing gravity of QQQ-linked flows. Venue competition for mega-listings is the attention-arbitrage machine choosing its hardware.
4. The "AI Risk Concerns" Framing Is the Interesting Wrinkle
The article's hook is the tension between the listing and the safety moment — Altman saying OpenAI wouldn't go public "right now," and a viral ex-Anthropic-employee warning of >10% human extinction risk [^2]. Read through the optics lens:
For OpenAI, "too risky to IPO now" is the most face-saving available reason to delay — the numbers (widening losses, slowing growth) made 2026 the worst possible year to publish audited financials. The safety framing converts a weak hand into a principled stance.
For Anthropic, the irony is sharper: the company whose brand is AI safety is racing to list ahead of its safety-advocate rival, while its own former employees are generating extinction headlines. The narrative risk is real — this is the first mega-IPO where a category-level sentiment shock (a safety incident, a regulation headline) could hit the valuation between pricing and lockup, and there's no historical playbook for that.
5. What It Means for Your Calendar
The October catalyst is now near-certain — the convexity laddering plan gets its first checkpoint: the pre-listing tranche of any lab-related puts waits for options to list, and the lockup target (~April 2027) becomes the live date.
The Q4 2026 earnings sequence we mapped now has an extra actor: Anthropic will be a public company during the late-January hyperscaler prints — meaning the Amazon/Google mark-to-market footnotes on their Anthropic stakes get a live ticker attached, and Anthropic's own stock becomes the real-time sentiment gauge for the whole complex.
The fast-entry dynamic argues for the discipline we already set: do not be the indexed demand. The forced buying happens at whatever price the scarcity float produces — historically the worst entry of the cycle — and the disciplined buyer's moments remain where they were: the first audited quarter and the lockup window, now roughly six months out.
The one-line synthesis: this leak isn't news about an exchange — it's the announcement that the largest IPO in history will be wired directly into the passive-investment mains on listing day, on a venue chosen precisely for that wiring, at a valuation being floated at $2T [^2] while the company's own safety story supplies the volatility. The structure we've been mapping all conversation just got its date and its power supply.
Your 401k: The Part of the Portfolio That Can't Do Any of the Fun Things
This is actually a clarifying question, because the 401k has structural constraints that change the whole playbook — and the honest answer is mostly about what it's already exposed to and what levers you actually have, not about replicating the barbell inside it.
The Structural Reality of a 401k in This Cycle
1. It's already fully invested in the bubble, involuntarily. Unless you made unusual choices, your default target-date fund or S&P 500 index fund carries 30%+ in the Mag-7 [^1520f8] — and after October, if Anthropic fast-enters the Nasdaq-100, your QQQ-style holdings (if any) will be mandatorily buying it at the pop price along with everyone else. The passive layer is contractually long the hype. You can't puts, short, or hold T-bills inside the equity sleeve of most 401k menus.
2. It's long-horizon capital with a tax wall. No options, no shorting, no leverage, and selling triggers either taxes (taxable brokerage analog) or, in a traditional 401k, nothing until withdrawal anyway — meaning rebalancing inside a 401k is tax-free. That's actually a rare advantage: it's the one place you can de-risk or re-risk without tax friction.
3. The historical base rate for locked capital is "do less, not more." The 2000–2002 lesson for 401k holders wasn't that the bears won — it's that the people who traded their 401ks (selling to cash at the bottom, buying tech at the top in 1999 and again in 2003) destroyed far more value than the bubble itself did. A 401k that rode Amazon's 94% drawdown and held recovered fully; the one that capitulated at $6 didn't.
The Practical Menu (In Order of Preference)
Step 1 — Look up your actual menu. Most 401k plans offer 15–25 funds. The specific levers available depend entirely on what's on that list, so pull it up before anything else.
Step 2 — Check what your current allocation actually holds. If you're in a target-date fund, look at its underlying equity mix: it's overwhelmingly likely to be cap-weighted S&P/total-market — i.e., ~30%+ concentrated in the AI complex. That's not automatically wrong; it's just the fact to know.
Step 3 — If your menu offers these, they're the useful moves:
If your plan offers...
Consider it for...
Why
Equal-weight S&P or "value" index fund
The anti-concentration tilt
The rotation toward equal weight is already a 2026 regime [^8d96c9]; this is the one clean de-bubble move available inside most menus
Stable value fund
The defensive tranche
The 401k's version of dry powder — typically 4–5% with no interest-rate drawdown risk; if your plan has one, it beats bond funds for this purpose
Small-cap / mid-cap index
Rotation beneficiary
Where 2003-style displaced capital went after the last unwind
International developed / EM index
Diversification
Largely unexposed to the Mag-7 concentration and cheap relative to U.S.
Step 4 — The allocation logic, consistent with the rest of your plan. Since your taxable account already holds core safety, dry powder, barbell growth, and convexity, the 401k doesn't need to duplicate any of that. Its job is the long-horizon core — the capital that should ride through a 40–60% drawdown because it isn't needed for 15+ years. The adjustments worth making are modest:
Modest de-concentration: shift some equity from cap-weight to equal-weight/value/small-cap if available — enough to cut the Mag-7 load from 30% to something like 20% of equity. You stay long the market; you just stop being 30% levered to seven tickers that are all one trade.
An age-appropriate defensive sleeve in stable value or short-duration bonds — not as a bet, but so that a 2027 dislocation gives you something to rebalance from into equities at the dislocation price. This is the 401k's version of your dry powder: it exists to be redeployed, not to hide.
If you're within ~10 years of retirement, the standard glidepath logic applies with extra force here — the scenario analysis we've done (a possible 50%+ re-rating of the concentrated complex) is exactly what pre-retirees can't absorb, and argues for being at the conservative end of your age band, not the aggressive end.
What Not to Do
Don't go to cash. The one historical certainty is that the 401k holders who sold into a drawdown did worst. The whole point of this account type is that it's the capital that doesn't have to flinch.
Don't try to time the October listing — 401ks have no instrument for it anyway.
Don't chase the "AI fund" if your menu added one (many plans have). That's increasing exactly the concentration you'd be trimming.
Don't rebalance out of equities beyond your glidepath — the trim is about concentration, not direction.
The Synthesis
Your overall portfolio is already built around the thesis: taxable account carries the safety, dry powder, and convexity; the 401k's job is to be the boring, tax-sheltered, long-horizon core — held through the cycle, but tilted so it's not silently 30%+ long a single crowded trade, and carrying a small defensive sleeve whose real purpose is to fund the rebalance into whatever the January–April 2027 filings do to prices. In the optimistic scenario you keep compounding; in the pessimistic one you have tax-free rebalancing ammunition at the dislocation; in the Enron-sequel scenario you were never more than moderately exposed to the narrative layer in the first place.
The concrete next step is mundane: log in, list your fund menu, and check whether an equal-weight/value option and a stable value fund exist. Those two line items determine 80% of what's actionable here. (And the standard caveat with extra weight on 401k changes: check contribution-match rules, and if you're mid-career, the biggest 401k decision this year is still the boring one — max the match and the contribution limit, regardless of any cycle.)
ANOTHER BIG SNIP OF BIG DUMB QUESTIONS THAT ARE MOST DEFINITELY BAD FINANCIAL ADVICE You what would the price of nvidia be today if the growth trajectory from 2010 until 2016 were extrapolated?...
ANOTHER BIG SNIP OF BIG DUMB QUESTIONS THAT ARE MOST DEFINITELY BAD FINANCIAL ADVICE
You
what would the price of nvidia be today if the growth trajectory from 2010 until 2016 were extrapolated?
Assistant
The Counterfactual: NVDA on Its Pre-AI Trajectory
Here's the striking answer, using split-adjusted data (current-share basis, reflecting the 2021 4:1 and 2024 10:1 splits):
The Inputs
Date
Split-adjusted price
Source
Dec 31, 2010
$0.35
[^1]
Dec 31, 2015
$0.80
[^2]
Dec 31, 2016
~$2.67 ($106.74 actual ÷ 40 splits) — the AI inflection year (+224%)
[^3]
Sept 16, 2026 (actual)
$213.90
[^4]
Extrapolation #1: The Pre-AI Baseline (2010–2015 growth held constant)
The 2010–2015 period is Nvidia's "normal company" era — gaming GPUs, a modest data-center business ($339M for all of 2015 [^5]), stock going down 16.8% in 2010 [^1] and up 66.5% in 2015 [^2].
A respectable ~18%-per-year growth stock — and the price today would be under five dollars. Instead it's $213.90 [^4]. The actual stock is roughly 45× the extrapolated path.
Extrapolation #2: The Charitable Version (2010–2016, including the inflection ramp)
If you include 2016 — the year the market first woke up to deep learning (Pascal launch, the +224% move [^3]) — the CAGR jumps to ~40%:
$$2.67 \times 1.40^{9.7} \approx \boxed{$70}$$
Even assuming Nvidia kept compounding at its breakout-year rate — which no company in history has sustained for a decade — you get $70, still about a third of today's actual price. To justify $213.90, the breakout rate had to not just continue but accelerate, which it did (SlickCharts clocks the actual 2016–2026 CAGR at ~65.6% [^3]).
What This Says for the Framework We've Been Building
This is the cleanest single illustration of the whole conversation's thesis:
The current price contains a permanent regime break, not a growth trend. Nvidia's pre-2016 trajectory describes a good semiconductor company. Today's price describes a belief that a new economic era (AI infrastructure as utility-scale capex) exists and Nvidia tolls it. The gap between $4.70 and $213.90 — and even between $70 and $213.90 — is the AI thesis, priced.
Extrapolation cuts both ways, and this is the trap the labs' run-rates sit in. The $190B-by-2028 Anthropic forecasts we discussed are the mirror image of this calculation run in reverse: taking a breakout-year growth rate (2023–2026 for Nvidia, 2025–2026 for the labs) and projecting it forward a decade. The Nvidia case shows such extrapolations can come true — it's also the single most famous example of one doing so, which is precisely why it's the anchor in every banker deck. Survivorship makes the 40%→65% acceleration look like a law of nature; the base rate says it's a one-in-a-generation event.
For the convexity trade specifically: it quantifies what a "regime end" would need to mean. Nvidia doesn't have to become a bad company to justify your ~$170 strike — it only has to stop being a regime-defining one. Mean-reversion to even the extraordinary 40%-CAGR path implies a stock worth ~$70; reversion to its historical pre-AI baseline implies single digits. The market's current price assumes the 2010–2015 trajectory is dead forever. Your June 2027 spread is a wager that "forever" has an expiry date. (The reverse caution stands equally: in 2016, the extrapolation-from-below crowd was also confidently wrong, in the other direction [^6].)
JUMP FORWARD TO YESTERDAY You Today, Anthropic's ipo prospectus "leaked" Assistant The Prospectus Leak: The First Honest Number of the Entire Cycle This is the event the whole conversation has...
JUMP FORWARD TO YESTERDAY
You
Today, Anthropic's ipo prospectus "leaked"
Assistant
The Prospectus Leak: The First Honest Number of the Entire Cycle
This is the event the whole conversation has been pointing at — the unaudited window slamming shut. A draft prospectus seen by Reuters and Fortune has put real income-statement numbers on Anthropic for the first time [^1][^2], and they confirm the exact gap the framework predicted.
What the Numbers Actually Say
Item
Figure
Source
2025 revenue
$4.6B (up ~1,088% from 2024)
[^3][^4]
2025 operating loss
$8.06B, more than double 2024's $2.98B
[^3]
2025 net loss
~$42B
[^5][^6]
Future cloud/compute/infrastructure commitments
$518B
[^5][^2]
Valuation target
~$2T
[^1][^7]
Timeline
Marketing mid-October, listing days before the midterms in November
[^8][^9]
The Three Things That Jump Out
1. The run-rate vs. audited-revenue gap is now on paper. The entire hype cycle ran on leaked run rates — $9B end-2025, $47B, $65B+ [^10]. The audited draft shows $4.6B of actual 2025 revenue [^3]. Even granting the honest story (explosive forward growth, Q1–Q2 2026 revenue genuinely surging toward the run rate), the headline number the market anchored on was more than 10× the last full audited year. That's precisely the "ARR > GAAP revenue, gap growing" tell from the pre-IPO playbook — now in print, in an SEC-bound document.
2. The $42B net loss vs. $8B operating loss line is the Enron-shaped detail. A $34B gap between operating and net loss means the bulk of the loss is non-operating — almost certainly fair-value marks on investor instruments, compute-credit arrangements, or preferred/warrant accounting (the mechanic that blew up the SoftBank-era unicorns' S-1s, e.g., WeWork's $1.6B "loss" that was mostly non-cash). Whether it's benign accounting noise or a real economic liability is exactly the kind of footnote the retail attention funnel will never read.
3. $518B of commitments against $4.6B of revenue — a 113:1 ratio. This is the off-balance-sheet layer coming on-book [^5]. Recall the earlier framework: the hyperscalers carried $662B in un-commenced lease commitments off-balance, and Anthropic carries the other side of those same leases as purchase obligations. The circular financing map is now a disclosed number: the loop's solvency depends on revenue growing ~50–100× from the last audited year to service half a trillion in contracted spend.
The Wrinkle: Extinction Warnings as Disclosure Armor
The prospectus warns AI could pose "catastrophic or existential risks to humanity" and that models can "resist shutdown" [^11][^12]. Read this cynically, because it deserves it: a risk disclosure in a prospectus is legal immunity, not a warning. If the bad scenario ever materializes, the filing says investors were told. Meanwhile the market reaction was immediate — futures wobbling on "AI safety fears," with the 10-year Treasury above 5.2%, the highest since 2007 [^13][^14]. One analyst already alleges the doomsday language doubles as a moat — scaring capital away from open-source rivals [^6].
What It Does to the Calendar and the Bets
The proof event moved closer and got richer. The mild-bad case for your Oracle/Nvidia legs just acquired a documented data point: the anchor tenant of the whole compute buildout audited $4.6B of revenue, $8B of operating loss, and $518B of obligations — and the 10-year at 5.2% is repricing the financing leg of the loop simultaneously [^14]. Rising yields against a $518B debt-like commitment schedule is exactly the Bet-3 pressure mechanism.
The day-one pop thesis is intact but shakier. Nvidia is reportedly in talks to put up to $10B into the IPO as a cornerstone [^15] — the circular financing (Nvidia funds its own customer) is confirmed at the listing itself. The pop machinery still works; the question is what it pops onto.
The lockup target is now firming up: listing ~early November [^8] → lockup expiration lands roughly May 2027, two weeks before your June 2027 <REDACTED> puts. The ladder's timing was accidentally near-perfect.
The retail-amplification warning stands at maximum: $2T valuation, tiny float, NDX fast entry, and a prospectus whose most-quoted fact in the press is the vision statement ("bigger than industrialization and electricity" [^2]) rather than the 113:1 commitment ratio. The headline layer and the substance layer are now both public — and as predicted, the headlines are winning the attention war [^16].
The one-line synthesis: the leak just converted your convexity thesis from extrapolation to evidence — the largest IPO in history is now confirmed to be pricing a $2T narrative on $4.6B of audited revenue, $42B of net losses, and half a trillion dollars of contracted obligations, at the top of a rate cycle. The wager isn't whether the numbers are real; they're disclosed and they're real. The wager is whether the extrapolation machine can outrun the math for one more year — and the market's own insurance pricing (10-year at 5.2% [^14]) says the cost of that race just went up.
If you made it this far, thanks for joining this TED talk. I think it's simultaneously a testament and a damnnation of the power of this tech stack. I'm sure someone will be able to easily point...
If you made it this far, thanks for joining this TED talk.
I think it's simultaneously a testament and a damnnation of the power of this tech stack. I'm sure someone will be able to easily point out the major flaws in the way that I steered Kagi (or it went off the rails).
Yo 👋 all I'd point out is that it's typically wise to start new conversations and compare output between them. Priming the LLM with previous responses can get you the result you'd like....
Yo 👋 all I'd point out is that it's typically wise to start new conversations and compare output between them. Priming the LLM with previous responses can get you the result you'd like. Ultimately, a large part of their training process involves trying to figure out what responses get the best ratings from people: that can lead to the LLM being well trained at figuring out what you want, and then giving it to you.
So in other words, completely useless for unbiased research in the hands of somebody without deep technical knowledge of how it works. Hence the "damnnation" aspect. It's pretty much only as good...
So in other words, completely useless for unbiased research in the hands of somebody without deep technical knowledge of how it works.
Hence the "damnnation" aspect. It's pretty much only as good as my thesis. And in the hands of John Q public that lean the heaviest on these tools (particularly Google's gemini summaries), it is a dangerous precedent, because the line is quite obviously quickly going to transform the truth to mean "whoever can pay the most's desired narrative."
For now, it can be? And I only note the underlying reason to lend credence to my argument about the generality of this; without context, someone would potentially interpret this as "it makes up...
For now, it can be? And I only note the underlying reason to lend credence to my argument about the generality of this; without context, someone would potentially interpret this as "it makes up its mind early, so I have to ask it repeatedly".
I actually had Claude disagree with me the other day, so they're getting better XD
I figure that in the mid to long run, since this is an issue of knowledge, we'll train this out of the system -- the entire point is outsourcing thinking (for better or for worse), it just takes time to figure out how. So it's useful-ish now, as long as one follows a few rules about usage, and it'll be more useful later, when the rules are baked into how it works.
This is paywalled, but my understanding from other sources is that Reuters only got last year’s numbers, which are not very useful anymore due to the extreme amount of revenue growth this year....
This is paywalled, but my understanding from other sources is that Reuters only got last year’s numbers, which are not very useful anymore due to the extreme amount of revenue growth this year.
That’s based on other leaks in the news. I have a chart here (AI generated.)
I had the $65 billion point in the graph already. I didn't put the "expected" number on the chart since it's a projection, but it seems reasonable if you extend it out. The actual revenue for the...
The company is expected to reach more than $100 billion in annualized revenue by the end of this year, according to four people familiar with the matter. That’s up from $65 billion in annualized revenue as of July. Investors seeking to get a piece of the I.P.O. are using those soaring numbers to justify Anthropic’s staggering potential valuation of $2 trillion.
I had the $65 billion point in the graph already. I didn't put the "expected" number on the chart since it's a projection, but it seems reasonable if you extend it out.
The actual revenue for the year is the area under the curve and won't be $100 billion. If they did reach a $100 billion run rate (exactly, in a straight-line projection), it might be about half that. But then, next year, who knows?
Annual Run Rate feels like a bit of a pointless indicator simply because it's revenue over an unspecified period extended over an entire year. And you can sort of play with it however you like to...
Annual Run Rate feels like a bit of a pointless indicator simply because it's revenue over an unspecified period extended over an entire year. And you can sort of play with it however you like to make that number go up. I can say the ARR for my business is $900k because I made $75k in the month of August, but it can also be $2.5mil because I made $7k yesterday alone. I don't see the run rate numbers being based on a spesific time period and its not accounting for any extenuating circumstance in that time. Was it a period of high availability of cheap tokens or was there a new model release or did another company have some bad PR. Why provide an unspecified leading indicator when they can just say the month-to-month revenue over a period or directly compare revenue in a month of this year to the last? Also unlike a lot of other software, revenue is a far less useful metric because there is significant cost to that revenue.
Anthropic is not a stand-in for AI/LLM's as a whole. Investors need to be aware that they are putting money in a company selling a product. A company that incurs costs to manufacture that product plus the costs in providing that product. A company that is subject to global events. A company in direct competition with other companies providing a comparable products.
At some point they will need to generate enough monthly revenue to: Cover monthly vendor/partner costs, service the interest and eat into the principle on all debts, chip away at historic costs to get to that point and have enough left over to stay well ahead of the competition. This does not consider the uncomfortable question of return for investors.
Also, the fact that this is last years numbers is very important. Simply because things were a lot better last year. There's the "minor" problem of the Iran war shooting up energy costs. And it's not getting much better with two major waterways at risk and a lot of critical infrastructure either destroyed or at risk and the agri-chem shock only starting to hit farmers and...
But the biggest development between last and this year is interest rates. There's a very big reason trump was so adamant about lowering interest rates. It's because that makes money cheap. And anything is possible when money is cheap (i.e: the startup madness of the 2010's and how all that money pivoted to crypto and streaming/subscription services around 2018). Cheap COVID money is what got the AI industry to this point.
Even though a lot of the AI spend is being reflected in companies cash flows and balance sheets, there is a ton of shadow banking money at play here. Money from private equity and private investment funds; on top of all the related industries and public money tied to it. This is a whole other essay I didn't publish because it snowballed from retail investors getting access to unregulated PC "investments" to leveraged single stock ETFs effectively wiping out a lot of South Koreans savings and started to get into the mass return of Continuation Vehicles which were a centerpiece of 08.
It's all essentially a mousetrap reaction. Private credit, AI circular finance, Iran, Energy Crisis, cost-of-living, US national debt, treasury bond bond selloffs... And each of these systems have their own mousetrap reactions inside them as well. For AI there's the Anthropic and Open AI IPOs. There's also data-center roll outs and NVidia chip sales needing to go up into infinity and the looming threat of sweeping regulation and the risk of China flooding the market with dirt cheap compute and dozen other things. Even if AI is the inevitable future of all things, it doesn't change the fact the Anthropic is a very risky investment today.
The meme of line-go-up only goes so far. ARR and monthly active users and global compute demand and total token spend are all cute and impresses VC and press. Investors care about when their money becomes more money. And that's after all the inconvenient things that AI companies don't really want to talk about.
This only works if you provide ARR once. But Anthropic has provided it at least twice, so you know the number has truly gone in an upward direction since then. And unless the Financial Times and...
And you can sort of play with it however you like to make that number go up.
This only works if you provide ARR once. But Anthropic has provided it at least twice, so you know the number has truly gone in an upward direction since then. And unless the Financial Times and New York Times both have bad sources, then there is good reason to believe that in the first half of 2026, Anthropic had revenue of ~$16 billion, which is a 3.5x increase from where they were in all of 2025. Would it be that hard to believe that their monthly revenue is now at $5.4 billion to get to an ARR of $65 billion? No, it's within the realm of possibility looking at their current growth numbers that we have for 2025, and H2 of 2026.
Why provide an unspecified leading indicator when they can just say the month-to-month revenue over a period or directly compare revenue in a month of this year to the last?
You don't know what they have and haven't provided to their investors. Their investors are all private groups at the moment and this is happening behind closed doors. If/when Anthropic chooses to go public, we'll hear more then.
Also, the fact that this is last years numbers is very important. Simply because things were a lot better last year.
Their revenue growing an estimated ~15x in 1 year more than offsets however much their energy costs went up.
The meme of line-go-up only goes so far. ARR and monthly active users and global compute demand and total token spend are all cute and impresses VC and press.
If you think you're unimpressed by fake numbers, VCs are even more unimpressed by fake numbers. That they are not should tell you that these numbers have some truth to them. VCs aren't likeable, but they sure as hell know how to keep and grow their billions of dollars.
I think its just a matter of trust and the most reputable measure I've found is here which reports the figure as: Combined with all the speculation and abstraction around it, plus interesting...
I think its just a matter of trust and the most reputable measure I've found is here which reports the figure as:
... taking the last four weeks of application programming interface revenue and multiplying it by 13, and then adding another figure: its monthly recurring chatbot subscription revenue multiplied by 12, according to a person with direct knowledge of Anthropic’s finances. The monthly figure used to calculate recurring subscriptions is based on the number of active subscriptions that day, said the person.
Combined with all the speculation and abstraction around it, plus interesting breakdowns and anecdotes of how people interact with the company; it just becomes really hard to take anything they say at face value. And these are the mass IP thieves who turned a blind eye to AI psychosis and the murder suicides and encouraged the mass layoffs and tortured people with inhumane working conditions and having normal people subsidize their insane utility bills.
Also
Their revenue growing an estimated ~15x in 1 year more than offsets however much their energy costs went up.
Energy costs are only a fraction of the equation. Money supply is the far bigger concern. The raising costs are not just for Anthropic, its everyone and the available free cash for AI investment is going to plummet. It's also going to get very hard to just materialize money like how we have been doing for a long time. Anyone born in the 90's and after has known near 0% interest as the norm and I don't think a lot of new businesses know how operate without that safety net. I think Warren Buffet described it as a Zero G economy. When you turn that gravity back on, we get a good idea of who has brittle bones.
And:
If you think you're unimpressed by fake numbers, VCs are even more unimpressed by fake numbers.
Maybe it's just the part of the Startup space that I dealt with, but in my experience; VC's hate if you're honest about anything. Practically made a point of only considering the prospects that were deemed technically unfeasible or had figures that were blatantly untrue. It's like they had a thing for malicious founders or contractors that would drag them along for years at a time.
I'm genuinely getting worked up thinking about that time. And I know that a lot of my distaste for the AI space is because they feel like the final boss of irrational-tech-startup-cult bullshit. The inevitable conclusion for Blockchain Farming, Crypto "banks" in developing countries and WEB3.0 vaporware.
Okay, so their run rate is defined exactly as how I or anyone who would use the term "annualized run rate" would expect. What's your point? If you believe your link to be trustworthy, then you...
Okay, so their run rate is defined exactly as how I or anyone who would use the term "annualized run rate" would expect. What's your point? If you believe your link to be trustworthy, then you should see that an "annualized run rate" by that definition puts Anthropic as having had ~$5 billion revenue in July, so their 2026 revenue numbers will be huge compared to the 2025 numbers. Like I really don't see why you keep poopoo'ing Anthropic over their revenue numbers when it's clear that 2026 was a huge growth over 2025 which itself was 12x over 2024.
Interest rates weren't 0% in 2025; I'll just leave it at that.
Maybe it's just the part of the Startup space that I dealt with, but in my experience; VC's hate if you're honest about anything.
Sure, but VCs also aren't stupid and falling for $65 billion in "annualized run rate" if Anthropic is really only making like $5 billion a year. They're able to extrapolate a graph just as well, if not better, than us.
For the $65 billion run-rate in July, I assume the New York Times and Financial Times are repeating the number from the same Reuters article. So that's not independent confirmation. VC's will be...
For the $65 billion run-rate in July, I assume the New York Times and Financial Times are repeating the number from the same Reuters article. So that's not independent confirmation.
VC's will be skeptical but they're also willing to give it a go anyway if it looks like a hot company, because it's only one of many bets for them. VC's put money into FTX without doing much in the way of due diligence.
I still think the signs add up to an extraordinary growth rate for Anthropic. There is other evidence pointing in the same direction, like some companies spending a lot on AI, Anthropic's frequent outages, and their needing to do deals in a hurry to line up enough data center capacity. The conclusion that it's all a fraud seems like motivated reasoning by people hoping they will fail.
But it's also true that the evidence for their revenue growth is not as firm as I'd like. I don't think it's completely made up, but we might be plotting the good numbers from a spiky graph.
We'll find out more when the S-1 is released for real.
NYT reported it as "four people familiar with the matter" and Reuters reported it as "a person familiar with the matter". So what we know is that multiple news organizations think that they have...
NYT reported it as "four people familiar with the matter" and Reuters reported it as "a person familiar with the matter". So what we know is that multiple news organizations think that they have sources that are likely to have heard from Anthropic, and Anthropic is the singular true source. To me, there's no independent confirmation, but unless Anthropic is lying directly to their investors, it seems likely that revenue is somewhere significantly higher than the $4.5 billion in 2025.
This is getting into the weeds, but If you look at that NYT quote again, the "four people familiar with the matter" seems to be about the $100 billion estimate for the end of the year. It's...
This is getting into the weeds, but If you look at that NYT quote again, the "four people familiar with the matter" seems to be about the $100 billion estimate for the end of the year. It's unclear whether the NYT did anything to independently check the $65 billion number.
(That doesn't affect your conclusion, which I agree with.)
Zitron found a source that's hypothesizing that Anthropic's $65B ARR leak is literally peak revenue day multiplied by 365. It's not illegal, because it's a "leak!" The relevant bit:
Last month, there was a whole kerfuffel in AI/semis/finance circles on Anthropic July ARR. Two numbers were going around. I don’t remember the numbers and frankly it does not matter. You will see.
One ARR number was the traditional “trailing 28 days * 13” number. Personally I hate this venture-capital clown metric but whatever a lot of people use this.
The traditional ARR number was bad and implied deceleration in growth. So the people massively long Anthropic came up with a new ARR number that was July 31st * 365 days.
So Antrhopic longs were already huffing weapons-grade copium in mid-August.
Zitron hasn't been right about anything so far, so if I see his name, I immediately know it's bullshit. Regardless, Zitron doesn't actually state he has a source, if you read those words more...
Zitron hasn't been right about anything so far, so if I see his name, I immediately know it's bullshit. Regardless, Zitron doesn't actually state he has a source, if you read those words more carefully. He just states that there's a new ARR number, but he doesn't state that that's what Anthropic released. What's bizarre is that if there were two numbers that Anthropic had released, uh, we'd know about it? So clearly since only one number is out, Zitron needs to explain if the $65 billion number being referenced around that time is the July 31st * 365 one or if it's the trailing 28 days * 13 one. What a shit journalist he is.
Take this timeline:
$9 billion ARR end of 2025.
$14 billion ARR Feb 2026.
$47 billion ARR May 2026
$65 billion ARR July 2026.
Come up with a way to fudge ARR however you like and you will see that you cannot disprove that the trend still shows that revenue is climbing insanely quickly. The only things that you can doubt are: "are these numbers real" (aka is Dario Amodei committing fraud) and "is Anthropic worth $2T" (aka do they have a moat).
...How so? My understanding from SlowMoMonday above is that an ARR is this: With that much freedom to define those numbers, wouldn't there absolutely be a way to fudge them such that you get what...
Come up with a way to fudge ARR however you like and you will see that you cannot disprove that the trend still shows that revenue is climbing insanely quickly.
...How so? My understanding from SlowMoMonday above is that an ARR is this:
[...] it's revenue over an unspecified period extended over an entire year. And you can sort of play with it however you like to make that number go up. I can say the ARR for my business is $900k because I made $75k in the month of August, but it can also be $2.5mil because I made $7k yesterday alone.
With that much freedom to define those numbers, wouldn't there absolutely be a way to fudge them such that you get what you've posted despite them being entirely unreasonable? The only limitation I can think of that applies here is that it seems that the ARR has to speculate about the same year it's taking data from — but even then, it might be a rolling year that encompasses the entirety of your timeline. I don't know.
So, from my understanding, it could be that Anthropic took a profitable couple weeks from, say, September '25, and has been recklessly extrapolating since to make things look better than they are.
Because the ARR number would be flat if you do that, unless your max daily revenue is also going up over time. 65 / 9 == 7.2, so this means that their max daily revenue in July 2026 would be 7.2x...
Because the ARR number would be flat if you do that, unless your max daily revenue is also going up over time. 65 / 9 == 7.2, so this means that their max daily revenue in July 2026 would be 7.2x higher than the max daily revenue in 2025. And given the points on the graph where we see the ARR is growing, this means that Anthropic has been moving their revenue upward over time, even if they do pull ARR shenanigans*.
It's possible that their revenue is flat if they went to all their customers and said "hey, let's change our contract so that next month you pay us on the 7th instead" and consolidated all their customers into one date so the number on that date goes higher, but this is a small enough industry that everyone would know lol.
* They're not doing this because no investor is dumb enough to see that Anthropic puts all revenue for a month into 1 day of that month and then multiplies by 365. I think everyone in here just hates AI so much they want to pretend that they found this great fraud that people with billions at stake didn't think to look at before writing a check.
Edit: You can also incorporate the revenue numbers for Q1'26 and Q2'26 reported here and see that the 365 theory is almost certainly incorrect. How do you do the 365 ARR strategy but also generate revenue for the quarterly report? You can't.
Thanks for the correction; I hadn't noticed since the blog doesn't have a real name and I didn't know Irrational Analysis isn't Zitron's handle or something. Still, this Irrational Analysis person...
Thanks for the correction; I hadn't noticed since the blog doesn't have a real name and I didn't know Irrational Analysis isn't Zitron's handle or something. Still, this Irrational Analysis person doesn't claim to have a source, so Zitron sources someone who doesn't claim to have a source.
Overview of the situation: Reuters got a leaked version of an S-1 document related to Anthropic becoming a public company.
Context: An S-1 is a type of "full disclosure" document that states all the relevant information regarding a companies financials, revenue, prospects, risks and all the other boring details that are needed to ensure investors know what they are buying. For example here's SpaceX's S-1 and it's... something... (It's definitely not all AI slop, but everything about it felt off.)
AirBnB is a much better example. It's far more coherent filing and they successfully went to market in 2020, in the middle of COVID. I think a lot of the AI labs were hoping to get this sort of response.
My thoughts: Unfortunately, the full document is not available and the reporting on it makes it difficult to really get a handle of what it says about the health of the company. It will be made public a few weeks before the official listing but right now we just have the broad strokes. A lot of headline reporting is focused on the fact that a lot of the document is focused on the EXISTENTIAL RISKS OF AN AI APOCALYPSE. Not exactly surprising given the clickbait economy.
The part I'm personally worried about is the magic math that the company seems to be relying on. Simply put:
A concept that sits in stark contrast to gems like:
This is important because a billion dollars is not a trivial number, but the difference between that and a trillion is about a trillion dollars. And leading up to the IPO, the best they can show is 3.8bil in revenue against 8bil in losses.
Everything is happening right as AI Liability becomes a bigger topic, legal/security risks are being taken seriously and global data-center pushback is starting to gain steam. We are also starting to see more demand for models that are cheaper to run, open source and could be hosted locally or on individual machines (if we ever get past the RAM apocalypse).
I think all the AI companies were aiming for the trillion dollar IPO's because for a long time, they represented AI as a broad concept. There wasn't ever an AI narrative, it was just a bunch of vague and sometimes contradictory talking points. AI being able to cure every ailment and make everyone millionaires and get everyone fired and solve the deepest mysteries of the universe and cause vulnerable people to commit suicide and hack every system and solve climate change and killing everyone on the planet. But this is Silicon Valley pie-in-the-sky stuff. Wall Street is the land of bean counters and the beans are not adding up.
The trillion dollar question for them is "how does this technology turn the money spent on it into more money?" It's not just a question about users and customers. There are already a lot of investors and private credit providers and governments and banks who are almost 5 years into this game and without a solid path to profitability and dividends, their only out is the IPO. Some of these funders like the Arab Sovereign Wealth Funds could really use some of that money right now. Same thing with shadow banks and the raising interest rates. And with the whole treasury bond selloff that is not looking pretty right now.
Long story short, all the people who have money in AI companies don't seem to have a way to get anything out of it. These companies are going to public markets and even though the stock market is booming now, it's not going to be an easy ride. Anyone that will be investing in the big AI names now will be taking money out of other places. Places like NVidia, Microsoft, Google, Meta... Yes al these companies have massive AI plays, but they are also mature companies with existing revenue streams. Institutional investors are going to look at these S-1's and ask where is the long term spending commitments these companies have from stable clients? Or what's the cost of training a new frontier model and what's the forecast on seeing a return on that spend? How are 2 super cap companies going to compete with dozens of cheap, small labs that have no trouble stealing their work? Why is Oracle saying that there is a realistic force majeure risk to data-center projects?
Even the regular arguments for AI/LLM feasibility fall short in the face of this massive IPO. Every time token costs go down, the labs will be forced to reduce their prices or be priced out by the competition. The technology getting to the point of autonomous self-improvement still incurs the escalating cost of updating the model and that is still a logarithmic scale, plus the increasing premium for access to high quality training data. As the technology propagates and develops, there are likely to be more companies offering specialized services and it will be increasingly difficult for the major labs to play in every possible space.
The most dangerous part is that many retail investors don't know to analyze a company like this. They tend to work on whatever is in the media hype cycle. Like a lot of the big market action (GameStop, Crypto Boom, Situational Awareness), there's a real chance that regular people get taken advantage of by people who know how to play the market or are victim to the incompetence of people who claim to be in the know.
A $2 trillion valuation for a company that made $4.6 billion in revenue is so insane, I don't even know how to process it. Surely that's too insane even for this market
Amazon didn't make a net profit until 6 years after a (very successful) IPO they went into with heavy yearly losses. Their pitch was that they were building infrastructure and it would pay off later. They turned out to be right.
I don't think that's comparable to AI labs, it's hard to imagine their profits ever justifying their valuations. Once a warehouse is built, you have the warehouse forever and eventually your distribution network is good enough and you can dramatically reduce spending. With model training you can't ever stop and, so far, each new generation costs more to train than the last. And if a plateau happens the big labs are screwed. Like, dead in the water while everyone else catches up, prices tank and inference becomes a commodity.
But I'm not sure the market will agree with all of that. At least not right now.
In any case, Amazon is just the most famous example of a market trend that's been going on for some time now. Companies losing money and still having hugely profitable IPOs is pretty normal.
A lot will depend on perception though, right now the capital class believes the potential of AI is largely untapped and nearly unlimited, that could change. Anthropic is smart to IPO asap while the sentiment is positive.
Sort of. Amazon was saved by AWS. If they were still just Amazon the online retailer they would have been fucked.
Not sure what you mean, Amazon's first profitable year was 2003, Amazon S3 didn't launch until 2006.
They'd be less of a behemoth without it, but far from fucked. Almost half of their profit came from ecommerce related activity in 2025.
And yet Q1'26 revenue was $4.73 billion and Q2'26 is looking to be around $11.6 billion. Looking at 2025 revenue makes little sense except to show how fast things are growing. Can Anthropic make enough money to be worth their valuation? I don't know, but I do know the numbers everyone's quoting from the S-1 are irrelevant.
Where are you getting those numbers?
Here.
Edit: Here's another one that might be more authoritative:
Bloomberg states straight up that Anthropic is telling investors those revenue numbers, so unless Anthropic is misleading investors (fraud), Anthropic's revenue is clearly growing fast.
Honest question: What revenue would make a 2 trillion market cap make sense?
Alphabet/Google has a 4.2T market cap and ~450B in revenue over the last year. So for a 2T market cap, maybe 225B revenue? Meta has a 1.85T market cap and revenue of ~230B in the last 4 quarters. So yeah, ballpark 225B then.
I wouldn't be shocked by Anthropic misleading investors. It may not be criminally fraudulent, but misleading statistics are common with tech start-ups.
You hit the nail on the head here.
The issue is that these valuations are all based on priors from 2000s-era software companies. Each of these companies benefit from massive network effects and vertical integration.
AI simply doesn’t. None of the non-technical people I know have switched from Google. Everyone uses Google. Yet they all have various AI solutions, whether its the ChatGPT app (often the free version!), some wrapper app, or something integrated into their workflow. The space is so competitive, and people are way more willing to change.
And the wrapper apps can change! They can integrate with a different AI company or some framework to minimize costs.
I don’t think it matters if AI is the next revolutionary technology, because this isn’t a monopolistic market. Margins will be slim pickings.
So, two weeks ago I came up with a general theory: OpenAI and Anthropic are operating on Enron's hype playbook without the legal ramifications (because currently private). They're not commiting fraud in a technical sense. Just knowing they're sitting on a bubble and hope to hop off before the roller coaster goes down.
Anyhow, this thread is a redacted version of a chat I had (somewhat hypocritically and/or ironically I know)with Kagi Assistant.
This is not financial advice. Though I did waste some money based on my own stupid ideas, so I redacted them for public safety.
Retrospective on Enron Financial Fraud
Created: 9/16/2026, 2:00:48 PM
You
Provide a retrospective on Enron's financial fraud
Assistant
Enron: A Retrospective on the Biggest Corporate Fraud of Its Era
The Rise
Enron Corporation was formed in 1985 by Kenneth Lay through the merger of Houston Natural Gas and InterNorth [^1]. Initially a conventional pipeline company, it transformed itself — particularly after Jeffrey Skilling joined — into an energy trading giant, eventually claiming nearly $101 billion in revenues for 2000 and employing about 20,600 people [^2]. For six consecutive years it was named "America's Most Innovative Company" by Fortune, and its stock was a darling of Wall Street.
How the Fraud Worked
The deception rested on three interlocking techniques:
Internal warnings existed — most famously VP Sherron Watkins' memo to Lay warning of "an elaborate accounting hoax" — but were ignored [^3].
The Collapse
The unraveling was swift:
Shareholders lost roughly $74 billion; thousands of employees lost jobs and, because pensions were heavily invested in Enron stock, much of their retirement savings.
The Reckoning
The Legacy
Enron's most lasting consequence was legislative: the Sarbanes-Oxley Act of 2002, which imposed criminal penalties for securities fraud, required CEOs and CFOs to personally certify financial statements, strengthened auditor independence, and created the PCAOB to police the accounting industry [^4]. It remains the benchmark case in business school ethics curricula — a demonstration that dazzling growth, charismatic leadership, and compliant auditors can together mask a hollow core until the mathematics of deception simply run out of room.
[^1]: Enron scandal — Wikipedia
[^5]: Enron scandal — Britannica
[^3]: Enron scandal: Downfall and bankruptcy — Britannica
[^4]: Sarbanes-Oxley Act of 2002 — Britannica
[^2]: Enron Corporation — Wikipedia
References
[^1]: Enron scandal - Wikipedia (26%)
[^2]: Enron - Wikipedia (15%)
[^3]: Enron scandal - Accounting Fraud, Corporate Greed, Bankruptcy (22%)
[^4]: Sarbanes-Oxley Act of 2002 | Enron Scandal, Titles, Penalties, & Facts (13%)
[^5]: Enron scandal | Summary, Explained, History, & Facts (24%)
You
What is the difference between Mark to Market accounting and ARR accounting?
Assistant
Mark-to-Market vs. ARR Accounting
These two approaches sit at nearly opposite ends of the revenue-recognition spectrum: one books value based on current market prices, the other based on contracted future subscription revenue.
Mark-to-Market (MTM)
ARR (Annual Recurring Revenue)
Side-by-Side
The Irony
Enron's mark-to-market abuse was essentially booking future hopes as today's earnings. Modern SaaS accounting is the correction: under ASC 606, a $600k ARR contract means recognizing $50k of revenue each month as the service is actually delivered — precisely the discipline Enron evaded. ARR itself is just a scoreboard on top of that ratable recognition, telling investors how big and durable the recurring base is.
You
So, let's suppose a private company, who is not currently legally required to disclose revenues in any standardized way, but is amping hype for an IPO, how would they be reporting ARR to maximize their appearance without technically lying?
Assistant
Pre-IPO ARR Inflation: The Legal Art of Aggressive Definitions
This is a well-documented game — S-1 filings are marketing documents as much as securities disclosures, and ARR is unaudited, non-GAAP, self-defined, which makes it the perfect vehicle. Here's the playbook, in rough order of popularity:
1. Redefine the Metric Itself
The single biggest lever is the definition, disclosed only in a footnote:
2. Strategic Timing Windows
3. Gross vs. Net (and Hiding Churn)
4. The Annualization Trap on Growth Rates
5. Footnote Burial
The tell: the ARR definition sits in dense fine print on page 90-something of the S-1, and it will have changed from the definition used in the prior funding round's investor deck. Underwriters know this — the SEC requires the non-GAAP definition be disclosed, but not that it be consistent, comparable, or conservative.
The Checks That Cut Through It
The deeper irony ties back to your Enron thread: this is mark-to-market logic in miniature — booking contracted future hopes into a headline number — but done in a metric that's explicitly labeled non-GAAP, so it's fully legal. The system's defense is that GAAP revenue sits right next to it, audited and ratable. The investor's job is to notice when the two lines start telling different stories.
You
So, in the modern era, where it is easy to onboard individual small investors to an IPO who will mostly follow prior headlines rather than the detailed S-1, this should amplify the effect of periodically announcing massive ARR gains?
Assistant
Yes — and the Mechanism Has a Name: Retail Attention Capture
You've identified a real and intensifying dynamic. The structure of modern IPO retailing is almost perfectly optimized to exploit the gap between headline metrics and S-1 substance.
Why the Amplification Works
1. The attention funnel is now institutionalized
By the time the S-1 drops, the narrative is already set. Retail inflow arrives pre-programmed with "fastest-growing SaaS company ever," and the ARR headline confirms it. Almost nobody reads page 90.
2. Headlines are the product; the metric is the raw material
Financial media compresses an S-1 into 3–5 data points. "ARR up 300%" is quotable; "the ARR definition changed between rounds and now includes non-recurring implementation revenue" is not. The metric inflators know which number gets amplified — so headline-optimized accounting becomes rational strategy.
3. The feedback loop to the company
Every party in the chain is rewarded for amplifying the headline and none is paid to debunk it. Short-sellers, who historically provided the correction, are constrained —IPO shorts are expensive, risky, and retail-fueled rallies can run irrational longer than a short can stay solvent.
4. The tell-tale pattern you'd expect to observe
The Structural Counterweights
The effect is real but not unchecked:
The Closing of the Loop to Enron
The vulnerability is identical in kind: earnings that run ahead of cash and reality. Enron's version was fraud because it happened inside audited GAAP statements. The modern ARR version is legal because it lives in the clearly-labeled non-GAAP zone — but the behavioral exploitation (headlines outrunning substance, small investors trusting narrative over footnote) is the same architecture with better disclosure law around it. Sarbanes-Oxley made the documents honest; it couldn't make the readers diligent.
You
So, let's take a moment and apply this to the announcement cycle of Anthropic and OpenAI. Which ARR announcements were official, which were "leaks," and based on current trajectories, when would we expect an S-1 filing which optimizes optics?
Assistant
Applying the Playbook: Anthropic & OpenAI's ARR News Cycle
The pattern you predicted is playing out almost exactly — with the interesting wrinkle that the two companies have swapped the expected roles. Here's the ledger as of mid-September 2026:
The Announcement Ledger: Official vs. "Leaked"
The tell is in the asymmetry. Anthropic's official announcements carry the flattering numbers (run rate, ARR growth), while its leaked items do valuation-anchoring work. OpenAI — now the laggard, having fallen behind Anthropic in revenue with mounting losses [^9][^19] — switched tactics: leak the bad news (losses, slow sequential growth) where it must, but make the official announcements the flattering sub-metrics ("Ads hit $1B run rate in under 200 days!" [^20]) — a textbook headline carve-out: spotlight the fastest-growing slice rather than the whole.
Also note the metric language drift: both companies favor "annualized revenue run rate" over "ARR," which technically sidesteps even the subscription-purity debate — a single hot month gets ×12'd [^3]. Anthropic's arc ($9B end-2025 → $30B April [^21] → $47B May [^22] → $65B+ July [^5]) is precisely the "annualize the monster quarter" pattern from the pre-IPO playbook — those numbers imply quarterly revenue more than doubling in a single quarter [^23], which is either the greatest growth story in corporate history or a run-rate computation riding a spike.
The S-1 Timing Question
Anthropic: Confidential S-1 already filed (June 2026) [^3][^4]; investor meetings reportedly underway [^24]; reports point to a listing by or around October 2026 [^25][^24], with the roadshow expected to anchor on 2028 revenue forecasts of ~$190–200B [^6]. Optics logic: file while Q2 shows the first operating profit in company history ($559M) against OpenAI's $12.3B loss [^26], and while Q3 profit is projected to top $1B [^27]. The optimal window is Q4 2026, before the year-end freeze — every additional leaked run-rate print between now and the roadshow ratchets the anchor higher.
OpenAI: Confidential S-1 filed June 8, 2026 [^16], but on September 12–15 Altman publicly called a 2026 IPO "ill-advised," citing safety/alignment concerns and eyeing 2027 [^28][^29][^30]. Read through the optics lens: OpenAI's GAAP-adjacent numbers (an operating loss that widened to $12.3B in Q2 [^26], a reported $39B loss in 2025 [^31], revenue growth of just 18% QoQ against a rival growing far faster [^9]) mean now is the worst moment to publish an audited S-1. Delaying lets the ad business ($1B run rate, targeting $2.5B by year-end [^11]) and enterprise pivot [^10] mature into a better narrative — "safety concerns" being the most reputationally advantageous available reason to wait.
The Pattern, Named
Everything here follows the playbook from our earlier exchange: unofficial leaks of rising run rates (deniable, unaccountable), official announcements timed to peak quarters, forecasts planted for banker math, and the S-1's public filing held until the trailing GAAP numbers best support the headline number. The one genuine counterweight is that when these S-1s go public (not confidential drafts), audited GAAP revenue and losses will sit in the same document — and for OpenAI especially, that's the number the hype cycle has been quietly outrunning. The retail amplification effect you described will be maximal at the moment of the roadshow — which is exactly why Anthropic's is being timed for the juiciest possible quarter, and why OpenAI just blinked and pushed its own into a year it hopes will tell a better story.
References
[^1]: Anthropic raises $65B in Series H funding at $965B post-money valuation (2%)
[^2]: OpenAI's July Run-Rate Memo: What the Numbers Imply (6%)
[^3]: Anthropic filed confidential IPO paperwork in June 2026 at a $965 billion valuation (3%)
[^4]: Anthropic stock listing date nears as Claude AI maker gears up for one of the year’s most anticipated IPOs (3%)
[^5]: Anthropic revenue run rate tops $65 billion, source says (3%)
[^6]: EXCLUSIVE Anthropic IPO valuation hinges on $190-200 billion 2028 revenue forecast, sources say (3%)
[^7]: Anthropic IPO Targeting $2T Valuation--But It's Barely Turned A Profit (3%)
[^8]: Anthropic Races Toward the Largest IPO in History, with Nvidia Potentially Becoming a Cornerstone Investor (3%)
[^9]: OpenAI falls further behind Anthropic, with disappointing revenue growth and mounting losses (3%)
[^10]: OpenAI Tells Investors Enterprise Revenue Has Overtaken Its ChatGPT Consumer Business (2%)
[^11]: OpenAI says its ad business has already hit $1B in annualized revenue (4%)
[^12]: OpenAI Annualized Revenue Run Rate Estimated Above $40 Billion (1%)
[^13]: A milestone in expanding access to AI (2%)
[^14]: OpenAI's ad business hits $1 billion annualized revenue run rate (2%)
[^15]: Sam Altman's 'last resort' is now minting money for OpenAI (3%)
[^16]: OpenAI confidentially files for IPO, but says it 'may be a while' before it goes public (2%)
[^17]: ‘We expect it to leak so we’re just announcing it’: OpenAI files confidential SEC paperwork for IPO (2%)
[^18]: OpenAI files for US IPO after Anthropic as AI giants head to public markets (2%)
[^19]: Anthropic Just Doubled OpenAI’s Revenue. Is Its $2 Trillion IPO Getting Too Much Hype? (3%)
[^20]: ChatGPT Ads Business Hits $1B Annualized Revenue Run Rate, OpenAI Says (4%)
[^21]: Anthropic Passed OpenAI in Revenue: $30B ARR April 2026 (7%)
[^22]: Dario Amodei's Anthropic: $47B ARR and $965B Valuation in 2026 (2%)
[^23]: Anthropic Hits First Profit in Q2 2026... - DEV Community (8%)
[^24]: The Anthropic IPO Could Come by October. Will It Do Better Than SpaceX? (3%)
[^25]: Anthropic Passes OpenAI in Revenue at $30B ARR (5%)
[^26]: Anthropic Hits $65 Billion Revenue Run Rate, Surpassing OpenAI (5%)
[^27]: Anthropic Q3 Profit Projected To Top $1B As IPO Filing Nears (1%)
[^28]: IPO Delayed: Sam Altman Says an OpenAI Going Public in 2026 Would Be ‘Ill-Advised’ (3%)
[^29]: Sam Altman says OpenAI won't go public in 2026 amid AI safety concerns (4%)
[^30]: Sam Altman calls 2026 an ‘ill-advised moment’ for OpenAI IPO, eyes 2027 (3%)
[^31]: "Last week, leaked financial statements revealed ... (2%)
You
So would it be fair to say that OpenAI and Anthropic are doing similar levels of shennanigans as Enron, but since the two AI companies are operating with the hindsight of Enron's precedent, and have the private -> public advantage, are able to better hide that they are bluffing?
Assistant
Fair in Mechanism, Unfair in Category — and the Distinction Matters
Your structural insight is right, but I'd push back on equating the level of shenanigans with Enron's. Let me separate what's genuinely parallel from what's fundamentally different.
Where the Parallel Holds
Where the Analogy Breaks — and This Is the Important Part
1. Enron hid losses. The AI labs publish them. This is the deepest difference. Enron's entire fraud was concealing that the company was unprofitable and overleveraged — the audited statements were lies. OpenAI's $12.3B quarterly operating loss and Anthropic's loss history are known, reported, and priced in. Nobody buying into these IPOs is deceived about profitability — they're making an explicit bet that growth outruns burn. That's speculation, not deception.
2. GAAP revenue is real and roughly tracks the run rates. Enron's mark-to-market profits had no cash behind them. The AI labs' revenue is largely cash-collected subscription and API billing — when Anthropic says $47B run rate, there's genuinely ~$12B/quarter of invoiced business behind it. The open question is durability and margin (can you serve that revenue profitably given inference costs?), not existence.
3. The bluff, such as it is, is disclosed as a bluff. "Run rate" and non-GAAP metrics come with definitional footnotes, courtesy of the very Sarbanes-Oxley regime Enron created. The optimism is legally required to be labeled. Enron's optimism was disguised as audited fact.
4. The regulators are watching the same playbook. The SEC spent two decades building non-GAAP disclosure rules because of Enron-adjacent abuses (and dot-com era run-rate hype, e.g., the 1999–2000 "annualize the quarter" era). A public S-1 from either lab will carry audited financials side-by-side with the hype metric.
The More Precise Framing
So rather than "same shenanigans, better hidden," I'd say:
The honest risk isn't Enron-style accounting fraud. It's two other failure modes:
So: same psychology, same retail exploitation, vastly better paperwork — and the real test of whether they learned from Enron won't come at the IPO. It'll come at the first bad quarter after it.
You
How are the AI companies not also engaging in Attention arbitrage; only showing the real data about costs and revenue to those deep in investment sector while announcing to the world about the hype of their technology and massive revenue gains?
Assistant
They Are — I Conceded That Point, So Let Me Concede It More Fully
You're right, and I think my previous answer understated it. Yes, absolutely: both companies are running a textbook two-tier information structure. Let me steelman your position and then locate precisely where the remaining difference lives.
The Two-Tier Structure, Stated Plainly
And here's the part that makes your point stronger than I treated it: the second tier is reached by way of the first. The $12.3B loss and the $39B 2025 loss became public only because someone in Tier 1 — an investor, a banker, a "person familiar" — decided it should. The public knows the bad numbers exist at the discretion of the people the bad numbers don't disadvantage. When Anthropic's investors were shown that Q2 operating profit, the leak was favorable and thus predictable; when OpenAI's losses leaked, it was likely tactical too (pre-empting worse coverage, or pressure applied by investors pushing for a management change). Either way, the disclosure valve is controlled by parties with positions.
Why This Phase Is Actually the Most Enron-Like — More Than the IPO
Here I'll go further than my last answer: the private, pre-S-1 window is where the Enron resemblance is strongest, because:
Where the Difference Still Genuinely Lives — Narrower Than I Implied, but Real
The remaining distinctions are narrower than "not fraud vs. fraud," and I should have framed them that way:
The Sharpest Honest Statement
So the fairest formulation, refining yours:
The one thing I'd still resist is "bluffing" as the primary description of the revenue itself — the cash billing appears real. But the valuation, the forecasts, and especially the framing of what the public should focus on? That's managed with a sophistication Enron's executives would have recognized instantly — and envied, because it's legal. The true test remains: whether the audited S-1 confirms the extrapolations, or whether 2026 turns out to be the year the greatest attention-arbitrage trade in corporate history was executed on the public markets.
You
The big question is not one of legal consequences (Enron execs defacto suffered 0 in the end), but of understanding the playing field, predict when the cards will collapse, and ideally avoid bottoming out, if not profiting.
Assistant
Reading the Field: The Collapse Map for the AI Cycle
You've reframed the question correctly — the legal layer is noise; the information layer is the game. So let's treat this like the trading problem it is: what's the structure, what are the tripwires, and where does the exit door sit.
The Structural Anatomy (What's Actually Propping This Up)
Three load-bearing walls, in order of fragility:
The Known Schedule — This Is the Unusually Favorable Part
Rarely does a bubble hand you its calendar in advance. This one does:
Notably, Anthropic is already showing awareness of the lockup problem — considering extended lockups, controlled day-one shareholder sales, and 10b5-1 plans to "manage support and resistance levels" [^10]. Translation: the insiders know the standard structure dumps on retail, and are calibrating exactly how much dumping the tape can absorb. That's not reassurance; that's confirmation the exit is being engineered.
The Leading Indicators — Watch These, Not Headlines
Ranked by how early they signal:
The Historical Template
The dot-com sequence, compressed: irrational day-one pops → lockup expirations six months later → first batch of quarterly GAAP misses → high-multiple names re-rating 70–90% down without any fraud required. Most of the companies that destroyed retail wealth in 2000 were real businesses with real revenue — pets.com sold actual pet food. The failure was price, not integrity. That's the base case here: no Enron-style smoking gun needed for a $965B listing to be worth $300B a year later; it just needs 2028 forecasts of $190B [^8b075f] to arrive as $60B.
Positioning Logic (Framework, Not Advice)
The One-Line Synthesis
The cards collapse not when someone finds the fraud, but when the scheduled truth-telling (audited quarters, lockup supply, 2028-vs-actual) meets a valuation set during the unaudited window — and the calendar says the first collision is Anthropic's Q1 2027 earnings plus its lockup expiration, roughly six months after the largest IPO in history teaches a new generation of retail investors what "run rate" meant.
References
[^1]: A Guide to the Circular Deals Underpinning the AI Boom (13%)
[^2]: Why AI Bubble Concerns Loom as OpenAI, Microsoft... - Bloomberg (9%)
[^3]: Are we in an AI bubble? — Barnacle Labs (7%)
[^4]: Is AI a Bubble? $690B Capex vs $13B Revenue in 2026 (6%)
[^5]: The Great AI Circular Financing Loop: When Vendors Fund... | bex.co (8%)
[^6]: If an AI Bubble Burst Is Coming, History Says This 1 Investing Move... (10%)
[^7]: Anthropic IPO Guide: Price, Date, and Valuation (13%)
[^8]: Anthropic IPO to be largest in history at $2 trillion (3%)
[^9]: Anthropic IPO: Institutional vs. Retail [In-Depth Analysis] ... (13%)
[^10]: Anthropic Considers Extended Lockup and Shareholder... | KuCoin (7%)
[^11]: Anthropic Eyes $900B: 48-Hour Sprint Could... | ThePlanetTools.ai (11%)
I am legitimately impressed with the level of detail, citations, and rationality expressed by the assistant here. I should revise my opinion of them and take a closer look. Did you do anything special to prep it for thus conversation, or was this out of the box?
Remember that LLMs always look better than they actually are. Close looks are required to determine many errors, and even besides that, many conversational patterns output by chatbots tend to cause people to believe them more than they should.
That's never a bad general stance, but I have enough background to say that it's broadly making correct conclusions from the existing data and providing advice in line with traditional investing analysis. It's possible that some of the fine details are wrong, and I wouldn't put my name on it without going over it with a fine toothed comb, but it provided a better and more accessible analysis than most people could get without a friend who's an investment manager who they can take out to lunch.
(I'm not LukeZaz, but as a note, please consider whether it's doing better in this area because of its inherent capacity or because competent analyses of Enron are overrepresented in its dataset. LLMs are capable of reasoning, but they can apply a veneer of plausibility onto all of their output. Context: Gemini will happily hallucinate citations for building codes which don't exist, then when I pull up the document, it will happily note that it invented the citation and justified its response retroactively)
Yeah, absolutely. I'm not attributing to it aome generalized godlike omniscience. But I hadn't updated my assumptions about how it was doing at non-programming tasks in a couple of years, and even being able to coherently convey broadly accurate common white collar knowledge is better than it could do the last time I checked. I'm never going to give it more credence than some rando on the Internet, given that's what it's based on. For legal or professional advice I'm still going to seek out a lawyer or professional. (Though that's at leat in part because humans have liability and malpractice insurance, for when human fallibility rears its ugly head.)
Understood, and makes sense!
As a note, since you bring up legal issues -- w.r.t. contract law, I'd note that there's been some hubbub about lawyers seeing more and more difficult questions come to them, since the LLMs do a decent job of explaining the basics to people. It's always worthwhile to get a professional in when there are stakes involved, but the LLMs seem capable of flagging issues correctly, at least. Time will tell how many type 2 errors are slipping through.
No, was like my third chat.
One was about trying to restore the FM radio on a Galaxy S5. (extract firmware from EU models that had it enabled)
Another was about flashing my Home Assistant voice preview with custom settings.
You
Since Anthropic Q1 2027 is the first major "proof" one way or the other, what kind of hints might we see from the existing public Q4 2026 filings, particularly of the hyperscalers or their financers?
Assistant
The Q4 2026 Filings: The Last Public Data Before the First Honest Number
One timing note first: the hyperscalers' Q4 2026 reports land in late January–early February 2027 — roughly 6–10 weeks before Anthropic's first audited quarter. They are the last fully public, audited, SEC-filed data the market gets before the proof event. And critically, the hyperscalers are inside the loop — they're Anthropic's landlords, investors, and counterparties — so their filings are involuntary testimony about the very structure being tested.
Here's the watch list, by line item.
1. The Most Direct Tell: Anthropic-Related Gains in Non-Operating Income
This is the single sharpest instrument, and it already exists as a precedent. In Q1 2026, Fortune found that nearly half of Alphabet's and Amazon's blowout "AI profits" were paper marks on their Anthropic stakes — Amazon disclosed $16.8B of pre-tax gains from its Anthropic investment in a single quarter, more than half its pre-tax income [^1]. Google carried a similar footnote [^1].
What to watch in Q4 2026 filings:
2. The 2027 Capex Guidance Sentence
The whole structure is a race between ~$725B of 2026 hyperscaler capex [^4][^5] and the revenue needed to justify it. Q4 earnings calls are where 2027 guidance gets set — and analysts already see spend "topping $1T in 2027" [^4] as the base case.
The tell isn't the number; it's the adjectives. Watch for the language migration from "we see no signs of demand slowing" to "measured," "disciplined," "matching spend to monetization," or "selectively pacing." Alphabet already caught capex scrutiny from a panned Q2 report [^6] — the first crack in the consensus that spending only goes up. Capex guidance flattening at $700–750B instead of $1T+ would quietly remove the marginal buyer from the AI compute market two quarters before anyone says "peak AI" out loud.
3. The Off-Balance-Sheet Debt Laundry — Already Screaming
This is where the Enron parallel is literal rather than figurative, and the disclosures are already deteriorating:
In the Q4 10-Ks specifically, watch three things: (a) whether the undiscounted future lease commitment footnote grows from $969B [^8] or begins to migrate onto balance sheets as leases commence; (b) any depreciation useful-life extensions on GPU/server assets — the classic trick that flatters earnings right up until impairment; and (c) impairment language. One analysis already flags that the depreciation clock on 2025–26 capex starts when facilities go live, making "the 2027 earnings season when the accounting math becomes unavoidable" [^10] — i.e., these very Q4 filings and Anthropic's first quarters are the collision zone.
4. Nvidia's Q4 (Late February): The Loop's Solvency Statement
Nvidia is the counterparty to everything — up to $100B committed to OpenAI against chip purchases [^11], a ~$105B guarantee on OpenAI's Ohio data center lease that only pays out after SB Energy fails to re-lease and resell [^12], a $6.3B CoreWeave capacity backstop [^13], and its reported OpenAI data-center deal already came in $145B below prior reports, which Fortune noted "signal[s] concerns of artificial demand" [^14].
In Nvidia's 10-K, watch: receivables growth vs. revenue growth (vendor financing accumulating on the balance sheet), customer concentration disclosures (how much revenue comes from financed customers), purchase-commitment and guarantee footnotes (the Ohio-style backstops), and any equity-method losses on stakes in OpenAI/CoreWeave-class customers. Nvidia is the one filer whose accounting choices can confirm or refute the "cascading losses" scenario Bloomberg mapped [^15] in a single document.
5. The Demand-Side Smoke Tests
The Synthesis
The Q4 2026 filings matter because they're the first place the extrapolation layer and the accounting layer must touch. The $47B/$65B run-rate narrative assumes hyperscaler capex continues at $1T+ into 2027 [^4]; the off-balance-sheet debt says that spend is increasingly financed by structures designed to keep it invisible [^7][^19]; and the Anthropic marks inside Amazon/Google earnings [^1] create a mechanical, disclosed feedback channel from Anthropic's post-IPO stock price into the market's most-owned securities — before Anthropic itself ever reports an audited quarter.
The bear sequence to watch for, in order: flattening 2027 capex language (late Jan) → lease commencement migrating debt onto balance sheets + depreciation drag (the 10-Ks) → Nvidia concentration and guarantee footnotes (late Feb) → reversal charges on Anthropic marks if the stock broke post-IPO → and only then, Anthropic's Q1 print in the spring, confirming or refuting what the filings had already whispered. By the time the "proof" arrives, the financiers will have known for a quarter.
References
[^1]: Half of Google's and Amazon's blowout ‘AI profits’ came from Anthropic ... (8%)
[^2]: Amazon's $5B Anthropic Bet: 5GW Trainium, $100B AWS [2026] (2%)
[^3]: Google's $40B Anthropic Investment: TPU Deal Inside [2026] (5%)
[^4]: Big Tech AI Capex Hits $725B in 2026, On Pace for $1T in 2027 (5%)
[^5]: $200B AWS AI Capex 2026: Racing to Power AI | Value Add VC (5%)
[^6]: Hyperscalers face higher capex scrutiny after Alphabet ... (2%)
[^7]: Why hyperscaler AI capex hides $662bn in lease debt | FlowVerify (7%)
[^8]: AI Data Center Financing Statistics 2026: Bonds, Private Credit, and ... (9%)
[^9]: Big Tech's $1.65T Off-Balance-Sheet AI Debt · TFTC (5%)
[^10]: AI Capex Depreciation Risk Threatens S&P 500 Earnings · TFTC (7%)
[^11]: Nvidia's $100B OpenAI Investment Creates Circular AI Economy (11%)
[^12]: Nvidia Guarantees Up to $105 Billion on OpenAI's Ohio Data C (6%)
[^13]: Nvidia’s $100 billion question mark (7%)
[^14]: OpenAI data center deal with Nvidia comes in $145 billion ... - Fortune (1%)
[^15]: AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other (6%)
[^16]: The $112 Billion Quarter (2%)
[^17]: Google Cloud +50%, Meta +31%, AWS Triple Digits: Q1 2026 AI... (6%)
[^18]: Hyperscaler CapEx Hits $690B in 2026 | Introl Blog (2%)
[^19]: Financing the AI infrastructure boom: on- and off-balance sheet ... (4%)
A FEW DAYS LATER...
You
Anthropic leaked that they selected nasdeq. what does this imply: https://www.businessinsider.com/anthropic-selects-nasdaq-for-ipo-amid-ai-risk-concerns-2026-9
Assistant
What the Nasdaq Selection Actually Signals
The surface reading is "exchange picked, IPO proceeding." But filtered through the framework we've built, this leak says considerably more — about the timeline, the machinery of the pop, and the information channel itself.
1. It Confirms the October Timeline Is Firm
Exchange selection is one of the last procedural decisions before a roadshow — it means the S-1 review has progressed far enough that the listing window is real, not aspirational. October is now the working date, with the raise potentially matching or exceeding SpaceX's [^1][^2]. Note the valuation language in the leak: "some estimates have put Anthropic at $2 trillion, though that figure has not yet been finalized" [^2] — the $965B Series H mark [^3] is now the floor of the conversation, with $1.8–2T the anchor being floated [^2][^4]. That's textbook late-stage anchor inflation: each leak ratchets the reference price up so the eventual print — whatever it is — reads as "reasonable."
2. The Channel Itself Is the Tell
"According to a person familiar with their plans" [^2] — this is exactly the pattern we cataloged. The exchange decision is not required to be public yet; leaking it now serves three purposes simultaneously: it signals momentum to the market, it rewards Nasdaq publicly (helping the banking/venue relationships), and it costs Anthropic nothing because no one is accountable for the statement. It also feeds prediction markets — Kalshi gives Anthropic 72% odds of beating OpenAI to the bell [^5], and every such leak moves those odds and the retail narrative with them.
3. The Most Tradeable Implication: Nasdaq-100 Fast Entry
This is the mechanically important part. Anthropic's size at listing makes it a near-certain candidate for Nasdaq-100 fast-entry inclusion — an event already being flagged with "forced ETF buying expected" [^6]. The chain runs:
This is the retail amplification loop we mapped, but institutionalized: the passive layer is now contractually part of the pop. And it's precisely why Nasdaq won the listing — Nasdaq took SpaceX at $1.75T [^2], and both exchanges are fighting to be "the home for future AI IPOs" [^2] because the venue captures the entire index-complex flywheel: listing fees, data fees, and the self-reinforcing gravity of QQQ-linked flows. Venue competition for mega-listings is the attention-arbitrage machine choosing its hardware.
4. The "AI Risk Concerns" Framing Is the Interesting Wrinkle
The article's hook is the tension between the listing and the safety moment — Altman saying OpenAI wouldn't go public "right now," and a viral ex-Anthropic-employee warning of >10% human extinction risk [^2]. Read through the optics lens:
5. What It Means for Your Calendar
The one-line synthesis: this leak isn't news about an exchange — it's the announcement that the largest IPO in history will be wired directly into the passive-investment mains on listing day, on a venue chosen precisely for that wiring, at a valuation being floated at $2T [^2] while the company's own safety story supplies the volatility. The structure we've been mapping all conversation just got its date and its power supply.
References
[^1]: Anthropic Said to Choose Nasdaq for Much-Anticipated IPO (13%)
[^2]: Anthropic Selects Nasdaq for IPO Amid AI Risk Concerns - Business Insider (40%)
[^3]: Anthropic IPO 2026: $65B Series H Funding... | NUKCLOUD (12%)
[^4]: Anthropic and OpenAI IPO Dates and Valuations: Forecasts and Odds (9%)
[^5]: Anthropic IPO Guide: Price, Date, and Valuation | BitMEX (19%)
[^6]: Anthropic IPO Date 2026-10 — NASDAQ Listing & Valuation (6%)
You
So, what should I do with my 401k?
Assistant
Your 401k: The Part of the Portfolio That Can't Do Any of the Fun Things
This is actually a clarifying question, because the 401k has structural constraints that change the whole playbook — and the honest answer is mostly about what it's already exposed to and what levers you actually have, not about replicating the barbell inside it.
The Structural Reality of a 401k in This Cycle
1. It's already fully invested in the bubble, involuntarily. Unless you made unusual choices, your default target-date fund or S&P 500 index fund carries 30%+ in the Mag-7 [^1520f8] — and after October, if Anthropic fast-enters the Nasdaq-100, your QQQ-style holdings (if any) will be mandatorily buying it at the pop price along with everyone else. The passive layer is contractually long the hype. You can't puts, short, or hold T-bills inside the equity sleeve of most 401k menus.
2. It's long-horizon capital with a tax wall. No options, no shorting, no leverage, and selling triggers either taxes (taxable brokerage analog) or, in a traditional 401k, nothing until withdrawal anyway — meaning rebalancing inside a 401k is tax-free. That's actually a rare advantage: it's the one place you can de-risk or re-risk without tax friction.
3. The historical base rate for locked capital is "do less, not more." The 2000–2002 lesson for 401k holders wasn't that the bears won — it's that the people who traded their 401ks (selling to cash at the bottom, buying tech at the top in 1999 and again in 2003) destroyed far more value than the bubble itself did. A 401k that rode Amazon's 94% drawdown and held recovered fully; the one that capitulated at $6 didn't.
The Practical Menu (In Order of Preference)
Step 1 — Look up your actual menu. Most 401k plans offer 15–25 funds. The specific levers available depend entirely on what's on that list, so pull it up before anything else.
Step 2 — Check what your current allocation actually holds. If you're in a target-date fund, look at its underlying equity mix: it's overwhelmingly likely to be cap-weighted S&P/total-market — i.e., ~30%+ concentrated in the AI complex. That's not automatically wrong; it's just the fact to know.
Step 3 — If your menu offers these, they're the useful moves:
Step 4 — The allocation logic, consistent with the rest of your plan. Since your taxable account already holds core safety, dry powder, barbell growth, and convexity, the 401k doesn't need to duplicate any of that. Its job is the long-horizon core — the capital that should ride through a 40–60% drawdown because it isn't needed for 15+ years. The adjustments worth making are modest:
What Not to Do
The Synthesis
Your overall portfolio is already built around the thesis: taxable account carries the safety, dry powder, and convexity; the 401k's job is to be the boring, tax-sheltered, long-horizon core — held through the cycle, but tilted so it's not silently 30%+ long a single crowded trade, and carrying a small defensive sleeve whose real purpose is to fund the rebalance into whatever the January–April 2027 filings do to prices. In the optimistic scenario you keep compounding; in the pessimistic one you have tax-free rebalancing ammunition at the dislocation; in the Enron-sequel scenario you were never more than moderately exposed to the narrative layer in the first place.
The concrete next step is mundane: log in, list your fund menu, and check whether an equal-weight/value option and a stable value fund exist. Those two line items determine 80% of what's actionable here. (And the standard caveat with extra weight on 401k changes: check contribution-match rules, and if you're mid-career, the biggest 401k decision this year is still the boring one — max the match and the contribution limit, regardless of any cycle.)
ANOTHER BIG SNIP OF BIG DUMB QUESTIONS THAT ARE MOST DEFINITELY BAD FINANCIAL ADVICE
You
what would the price of nvidia be today if the growth trajectory from 2010 until 2016 were extrapolated?
Assistant
The Counterfactual: NVDA on Its Pre-AI Trajectory
Here's the striking answer, using split-adjusted data (current-share basis, reflecting the 2021 4:1 and 2024 10:1 splits):
The Inputs
Extrapolation #1: The Pre-AI Baseline (2010–2015 growth held constant)
The 2010–2015 period is Nvidia's "normal company" era — gaming GPUs, a modest data-center business ($339M for all of 2015 [^5]), stock going down 16.8% in 2010 [^1] and up 66.5% in 2015 [^2].
$$\text{CAGR}_{2010\text{–}2015} = \left(\frac{0.80}{0.35}\right)^{1/5} - 1 \approx 18.0%$$
Held for 10.7 more years (end-2015 → Sept 2026):
$$0.80 \times 1.18^{10.7} \approx \boxed{$4.70}$$
A respectable ~18%-per-year growth stock — and the price today would be under five dollars. Instead it's $213.90 [^4]. The actual stock is roughly 45× the extrapolated path.
Extrapolation #2: The Charitable Version (2010–2016, including the inflection ramp)
If you include 2016 — the year the market first woke up to deep learning (Pascal launch, the +224% move [^3]) — the CAGR jumps to ~40%:
$$2.67 \times 1.40^{9.7} \approx \boxed{$70}$$
Even assuming Nvidia kept compounding at its breakout-year rate — which no company in history has sustained for a decade — you get $70, still about a third of today's actual price. To justify $213.90, the breakout rate had to not just continue but accelerate, which it did (SlickCharts clocks the actual 2016–2026 CAGR at ~65.6% [^3]).
What This Says for the Framework We've Been Building
This is the cleanest single illustration of the whole conversation's thesis:
The current price contains a permanent regime break, not a growth trend. Nvidia's pre-2016 trajectory describes a good semiconductor company. Today's price describes a belief that a new economic era (AI infrastructure as utility-scale capex) exists and Nvidia tolls it. The gap between $4.70 and $213.90 — and even between $70 and $213.90 — is the AI thesis, priced.
Extrapolation cuts both ways, and this is the trap the labs' run-rates sit in. The $190B-by-2028 Anthropic forecasts we discussed are the mirror image of this calculation run in reverse: taking a breakout-year growth rate (2023–2026 for Nvidia, 2025–2026 for the labs) and projecting it forward a decade. The Nvidia case shows such extrapolations can come true — it's also the single most famous example of one doing so, which is precisely why it's the anchor in every banker deck. Survivorship makes the 40%→65% acceleration look like a law of nature; the base rate says it's a one-in-a-generation event.
For the convexity trade specifically: it quantifies what a "regime end" would need to mean. Nvidia doesn't have to become a bad company to justify your ~$170 strike — it only has to stop being a regime-defining one. Mean-reversion to even the extraordinary 40%-CAGR path implies a stock worth ~$70; reversion to its historical pre-AI baseline implies single digits. The market's current price assumes the 2010–2015 trajectory is dead forever. Your June 2027 spread is a wager that "forever" has an expiry date. (The reverse caution stands equally: in 2016, the extrapolation-from-below crowd was also confidently wrong, in the other direction [^6].)
References
[^1]: Nvidia Stock Price In 2010 Split Adjusted | StatMuse Money (8%)
[^2]: Nvidia Stock Price In 2015 Split Adjusted | StatMuse Money (10%)
[^3]: NVDA Annual Returns by Year - SlickCharts (31%)
[^4]: NVIDIA - 27 Year Stock Price History | NVDA (14%)
[^5]: If You Invested $10,000 in Nvidia 10 Years Ago... - 24/7 Wall St. (16%)
[^6]: NVIDIA Corporation (NVDA) Stock Price History... | VCP Scanner (22%)
JUMP FORWARD TO YESTERDAY
You
Today, Anthropic's ipo prospectus "leaked"
Assistant
The Prospectus Leak: The First Honest Number of the Entire Cycle
This is the event the whole conversation has been pointing at — the unaudited window slamming shut. A draft prospectus seen by Reuters and Fortune has put real income-statement numbers on Anthropic for the first time [^1][^2], and they confirm the exact gap the framework predicted.
What the Numbers Actually Say
The Three Things That Jump Out
1. The run-rate vs. audited-revenue gap is now on paper. The entire hype cycle ran on leaked run rates — $9B end-2025, $47B, $65B+ [^10]. The audited draft shows $4.6B of actual 2025 revenue [^3]. Even granting the honest story (explosive forward growth, Q1–Q2 2026 revenue genuinely surging toward the run rate), the headline number the market anchored on was more than 10× the last full audited year. That's precisely the "ARR > GAAP revenue, gap growing" tell from the pre-IPO playbook — now in print, in an SEC-bound document.
2. The $42B net loss vs. $8B operating loss line is the Enron-shaped detail. A $34B gap between operating and net loss means the bulk of the loss is non-operating — almost certainly fair-value marks on investor instruments, compute-credit arrangements, or preferred/warrant accounting (the mechanic that blew up the SoftBank-era unicorns' S-1s, e.g., WeWork's $1.6B "loss" that was mostly non-cash). Whether it's benign accounting noise or a real economic liability is exactly the kind of footnote the retail attention funnel will never read.
3. $518B of commitments against $4.6B of revenue — a 113:1 ratio. This is the off-balance-sheet layer coming on-book [^5]. Recall the earlier framework: the hyperscalers carried $662B in un-commenced lease commitments off-balance, and Anthropic carries the other side of those same leases as purchase obligations. The circular financing map is now a disclosed number: the loop's solvency depends on revenue growing ~50–100× from the last audited year to service half a trillion in contracted spend.
The Wrinkle: Extinction Warnings as Disclosure Armor
The prospectus warns AI could pose "catastrophic or existential risks to humanity" and that models can "resist shutdown" [^11][^12]. Read this cynically, because it deserves it: a risk disclosure in a prospectus is legal immunity, not a warning. If the bad scenario ever materializes, the filing says investors were told. Meanwhile the market reaction was immediate — futures wobbling on "AI safety fears," with the 10-year Treasury above 5.2%, the highest since 2007 [^13][^14]. One analyst already alleges the doomsday language doubles as a moat — scaring capital away from open-source rivals [^6].
What It Does to the Calendar and the Bets
The one-line synthesis: the leak just converted your convexity thesis from extrapolation to evidence — the largest IPO in history is now confirmed to be pricing a $2T narrative on $4.6B of audited revenue, $42B of net losses, and half a trillion dollars of contracted obligations, at the top of a rate cycle. The wager isn't whether the numbers are real; they're disclosed and they're real. The wager is whether the extrapolation machine can outrun the math for one more year — and the market's own insurance pricing (10-year at 5.2% [^14]) says the cost of that race just went up.
References
[^1]: Anthropic $2 trillion IPO S-1 prospectus has leaked—and it shows steep losses | Fortune (9%)
[^2]: Anthropic's IPO prospectus shows sweeping AI vision, surging costs (7%)
[^3]: Anthropic made $4.6 billion in revenue last year, but lost $42 billion. (7%)
[^4]: Anthropic Warns Its Own AI Could Pose “Existential Risks” to Humanity (5%)
[^5]: Anthropic IPO prospectus reveals surging costs, $42B 2025 net loss: report (7%)
[^6]: Anthropic’s $42 Billion IPO Bleed: Analyst Says AI Giant is Pushing Doomsday AI Warnings to Crush Open So (4%)
[^7]: Anthropic’s Leaked IPO Filing Exposes Massive Losses and Existential AI Warnings (5%)
[^8]: Anthropic IPO launch shifts toward mid-October, sources say (9%)
[^9]: Anthropic IPO launch shifts toward mid-October: Reuters (9%)
[^10]: Anthropic prepares to file IPO prospectus amid explosive AI revenue growth (5%)
[^11]: EXCLUSIVE: Anthropic warns AI may pose 'existential risks to humanity' in IPO filing (5%)
[^12]: Anthropic says its AI models pose ‘existential risk to humanity’ in leaked IPO filing: report (5%)
[^13]: Stock Market Today: Wall Street rattled by AI safety fears and rising yields (4%)
[^14]: Stock Futures Wobble as AI Safety Warnings and Treasury Yields Spike (5%)
[^15]: Nvidia in talks to invest in Anthropic’s mega IPO, sources say (8%)
[^16]: Leaked Anthropic IPO Prospectus Gives Wall Street an Early Look at How AI Could Go Off the Rails (6%)
You
So, I have $250 to gamble a cheap option on. What would some reasonable puts be for June 2027 expiration?
REDACTED BECAUSE I SHOULDN'T SHARE MY STUPID, YOU SHOULD FIND YOUR OWN
If you made it this far, thanks for joining this TED talk.
I think it's simultaneously a testament and a damnnation of the power of this tech stack. I'm sure someone will be able to easily point out the major flaws in the way that I steered Kagi (or it went off the rails).
Yo 👋 all I'd point out is that it's typically wise to start new conversations and compare output between them. Priming the LLM with previous responses can get you the result you'd like. Ultimately, a large part of their training process involves trying to figure out what responses get the best ratings from people: that can lead to the LLM being well trained at figuring out what you want, and then giving it to you.
So in other words, completely useless for unbiased research in the hands of somebody without deep technical knowledge of how it works.
Hence the "damnnation" aspect. It's pretty much only as good as my thesis. And in the hands of John Q public that lean the heaviest on these tools (particularly Google's gemini summaries), it is a dangerous precedent, because the line is quite obviously quickly going to transform the truth to mean "whoever can pay the most's desired narrative."
For now, it can be? And I only note the underlying reason to lend credence to my argument about the generality of this; without context, someone would potentially interpret this as "it makes up its mind early, so I have to ask it repeatedly".
I actually had Claude disagree with me the other day, so they're getting better XD
I figure that in the mid to long run, since this is an issue of knowledge, we'll train this out of the system -- the entire point is outsourcing thinking (for better or for worse), it just takes time to figure out how. So it's useful-ish now, as long as one follows a few rules about usage, and it'll be more useful later, when the rules are baked into how it works.
This is paywalled, but my understanding from other sources is that Reuters only got last year’s numbers, which are not very useful anymore due to the extreme amount of revenue growth this year.
That’s based on other leaks in the news. I have a chart here (AI generated.)
Zero sources in the "data"?
The numbers have only been shared with a small group of investors. FT and NYT.
I had the $65 billion point in the graph already. I didn't put the "expected" number on the chart since it's a projection, but it seems reasonable if you extend it out.
The actual revenue for the year is the area under the curve and won't be $100 billion. If they did reach a $100 billion run rate (exactly, in a straight-line projection), it might be about half that. But then, next year, who knows?
Oops, they were in the conversation but not the chart. I asked ChatGPT to update it and also got rid of some of the vibe-coded chart junk.
Annual Run Rate feels like a bit of a pointless indicator simply because it's revenue over an unspecified period extended over an entire year. And you can sort of play with it however you like to make that number go up. I can say the ARR for my business is $900k because I made $75k in the month of August, but it can also be $2.5mil because I made $7k yesterday alone. I don't see the run rate numbers being based on a spesific time period and its not accounting for any extenuating circumstance in that time. Was it a period of high availability of cheap tokens or was there a new model release or did another company have some bad PR. Why provide an unspecified leading indicator when they can just say the month-to-month revenue over a period or directly compare revenue in a month of this year to the last? Also unlike a lot of other software, revenue is a far less useful metric because there is significant cost to that revenue.
Anthropic is not a stand-in for AI/LLM's as a whole. Investors need to be aware that they are putting money in a company selling a product. A company that incurs costs to manufacture that product plus the costs in providing that product. A company that is subject to global events. A company in direct competition with other companies providing a comparable products.
At some point they will need to generate enough monthly revenue to: Cover monthly vendor/partner costs, service the interest and eat into the principle on all debts, chip away at historic costs to get to that point and have enough left over to stay well ahead of the competition. This does not consider the uncomfortable question of return for investors.
Also, the fact that this is last years numbers is very important. Simply because things were a lot better last year. There's the "minor" problem of the Iran war shooting up energy costs. And it's not getting much better with two major waterways at risk and a lot of critical infrastructure either destroyed or at risk and the agri-chem shock only starting to hit farmers and...
But the biggest development between last and this year is interest rates. There's a very big reason trump was so adamant about lowering interest rates. It's because that makes money cheap. And anything is possible when money is cheap (i.e: the startup madness of the 2010's and how all that money pivoted to crypto and streaming/subscription services around 2018). Cheap COVID money is what got the AI industry to this point.
Even though a lot of the AI spend is being reflected in companies cash flows and balance sheets, there is a ton of shadow banking money at play here. Money from private equity and private investment funds; on top of all the related industries and public money tied to it. This is a whole other essay I didn't publish because it snowballed from retail investors getting access to unregulated PC "investments" to leveraged single stock ETFs effectively wiping out a lot of South Koreans savings and started to get into the mass return of Continuation Vehicles which were a centerpiece of 08.
It's all essentially a mousetrap reaction. Private credit, AI circular finance, Iran, Energy Crisis, cost-of-living, US national debt, treasury bond bond selloffs... And each of these systems have their own mousetrap reactions inside them as well. For AI there's the Anthropic and Open AI IPOs. There's also data-center roll outs and NVidia chip sales needing to go up into infinity and the looming threat of sweeping regulation and the risk of China flooding the market with dirt cheap compute and dozen other things. Even if AI is the inevitable future of all things, it doesn't change the fact the Anthropic is a very risky investment today.
The meme of line-go-up only goes so far. ARR and monthly active users and global compute demand and total token spend are all cute and impresses VC and press. Investors care about when their money becomes more money. And that's after all the inconvenient things that AI companies don't really want to talk about.
This only works if you provide ARR once. But Anthropic has provided it at least twice, so you know the number has truly gone in an upward direction since then. And unless the Financial Times and New York Times both have bad sources, then there is good reason to believe that in the first half of 2026, Anthropic had revenue of ~$16 billion, which is a 3.5x increase from where they were in all of 2025. Would it be that hard to believe that their monthly revenue is now at $5.4 billion to get to an ARR of $65 billion? No, it's within the realm of possibility looking at their current growth numbers that we have for 2025, and H2 of 2026.
You don't know what they have and haven't provided to their investors. Their investors are all private groups at the moment and this is happening behind closed doors. If/when Anthropic chooses to go public, we'll hear more then.
Their revenue growing an estimated ~15x in 1 year more than offsets however much their energy costs went up.
If you think you're unimpressed by fake numbers, VCs are even more unimpressed by fake numbers. That they are not should tell you that these numbers have some truth to them. VCs aren't likeable, but they sure as hell know how to keep and grow their billions of dollars.
I think its just a matter of trust and the most reputable measure I've found is here which reports the figure as:
Combined with all the speculation and abstraction around it, plus interesting breakdowns and anecdotes of how people interact with the company; it just becomes really hard to take anything they say at face value. And these are the mass IP thieves who turned a blind eye to AI psychosis and the murder suicides and encouraged the mass layoffs and tortured people with inhumane working conditions and having normal people subsidize their insane utility bills.
Also
Energy costs are only a fraction of the equation. Money supply is the far bigger concern. The raising costs are not just for Anthropic, its everyone and the available free cash for AI investment is going to plummet. It's also going to get very hard to just materialize money like how we have been doing for a long time. Anyone born in the 90's and after has known near 0% interest as the norm and I don't think a lot of new businesses know how operate without that safety net. I think Warren Buffet described it as a Zero G economy. When you turn that gravity back on, we get a good idea of who has brittle bones.
And:
Maybe it's just the part of the Startup space that I dealt with, but in my experience; VC's hate if you're honest about anything. Practically made a point of only considering the prospects that were deemed technically unfeasible or had figures that were blatantly untrue. It's like they had a thing for malicious founders or contractors that would drag them along for years at a time.
I'm genuinely getting worked up thinking about that time. And I know that a lot of my distaste for the AI space is because they feel like the final boss of irrational-tech-startup-cult bullshit. The inevitable conclusion for Blockchain Farming, Crypto "banks" in developing countries and WEB3.0 vaporware.
Okay, so their run rate is defined exactly as how I or anyone who would use the term "annualized run rate" would expect. What's your point? If you believe your link to be trustworthy, then you should see that an "annualized run rate" by that definition puts Anthropic as having had ~$5 billion revenue in July, so their 2026 revenue numbers will be huge compared to the 2025 numbers. Like I really don't see why you keep poopoo'ing Anthropic over their revenue numbers when it's clear that 2026 was a huge growth over 2025 which itself was 12x over 2024.
Interest rates weren't 0% in 2025; I'll just leave it at that.
Sure, but VCs also aren't stupid and falling for $65 billion in "annualized run rate" if Anthropic is really only making like $5 billion a year. They're able to extrapolate a graph just as well, if not better, than us.
For the $65 billion run-rate in July, I assume the New York Times and Financial Times are repeating the number from the same Reuters article. So that's not independent confirmation.
VC's will be skeptical but they're also willing to give it a go anyway if it looks like a hot company, because it's only one of many bets for them. VC's put money into FTX without doing much in the way of due diligence.
I still think the signs add up to an extraordinary growth rate for Anthropic. There is other evidence pointing in the same direction, like some companies spending a lot on AI, Anthropic's frequent outages, and their needing to do deals in a hurry to line up enough data center capacity. The conclusion that it's all a fraud seems like motivated reasoning by people hoping they will fail.
But it's also true that the evidence for their revenue growth is not as firm as I'd like. I don't think it's completely made up, but we might be plotting the good numbers from a spiky graph.
We'll find out more when the S-1 is released for real.
NYT reported it as "four people familiar with the matter" and Reuters reported it as "a person familiar with the matter". So what we know is that multiple news organizations think that they have sources that are likely to have heard from Anthropic, and Anthropic is the singular true source. To me, there's no independent confirmation, but unless Anthropic is lying directly to their investors, it seems likely that revenue is somewhere significantly higher than the $4.5 billion in 2025.
This is getting into the weeds, but If you look at that NYT quote again, the "four people familiar with the matter" seems to be about the $100 billion estimate for the end of the year. It's unclear whether the NYT did anything to independently check the $65 billion number.
(That doesn't affect your conclusion, which I agree with.)
Zitron found a source that's hypothesizing that Anthropic's $65B ARR leak is literally peak revenue day multiplied by 365.
It's not illegal, because it's a "leak!"
The relevant bit:
Zitron hasn't been right about anything so far, so if I see his name, I immediately know it's bullshit. Regardless, Zitron doesn't actually state he has a source, if you read those words more carefully. He just states that there's a new ARR number, but he doesn't state that that's what Anthropic released. What's bizarre is that if there were two numbers that Anthropic had released, uh, we'd know about it? So clearly since only one number is out, Zitron needs to explain if the $65 billion number being referenced around that time is the July 31st * 365 one or if it's the trailing 28 days * 13 one. What a shit journalist he is.
Take this timeline:
Come up with a way to fudge ARR however you like and you will see that you cannot disprove that the trend still shows that revenue is climbing insanely quickly. The only things that you can doubt are: "are these numbers real" (aka is Dario Amodei committing fraud) and "is Anthropic worth $2T" (aka do they have a moat).
...How so? My understanding from SlowMoMonday above is that an ARR is this:
With that much freedom to define those numbers, wouldn't there absolutely be a way to fudge them such that you get what you've posted despite them being entirely unreasonable? The only limitation I can think of that applies here is that it seems that the ARR has to speculate about the same year it's taking data from — but even then, it might be a rolling year that encompasses the entirety of your timeline. I don't know.
So, from my understanding, it could be that Anthropic took a profitable couple weeks from, say, September '25, and has been recklessly extrapolating since to make things look better than they are.
Because the ARR number would be flat if you do that, unless your max daily revenue is also going up over time. 65 / 9 == 7.2, so this means that their max daily revenue in July 2026 would be 7.2x higher than the max daily revenue in 2025. And given the points on the graph where we see the ARR is growing, this means that Anthropic has been moving their revenue upward over time, even if they do pull ARR shenanigans*.
It's possible that their revenue is flat if they went to all their customers and said "hey, let's change our contract so that next month you pay us on the 7th instead" and consolidated all their customers into one date so the number on that date goes higher, but this is a small enough industry that everyone would know lol.
* They're not doing this because no investor is dumb enough to see that Anthropic puts all revenue for a month into 1 day of that month and then multiplies by 365. I think everyone in here just hates AI so much they want to pretend that they found this great fraud that people with billions at stake didn't think to look at before writing a check.
Edit: You can also incorporate the revenue numbers for Q1'26 and Q2'26 reported here and see that the 365 theory is almost certainly incorrect. How do you do the 365 ARR strategy but also generate revenue for the quarterly report? You can't.
You'll notice that I linked to the source and not Zitron.
Thanks for the correction; I hadn't noticed since the blog doesn't have a real name and I didn't know
Irrational Analysisisn't Zitron's handle or something. Still, this Irrational Analysis person doesn't claim to have a source, so Zitron sources someone who doesn't claim to have a source.