There’s a theory I’ve heard floated from some pretty well-respected people in finance that this article backs up. I haven’t heard anyone endorse this theory, but they do talk about it. As I...
Exemplary
There’s a theory I’ve heard floated from some pretty well-respected people in finance that this article backs up. I haven’t heard anyone endorse this theory, but they do talk about it.
As I understand it, the idea requires these assumptions:
A) Competitive AI systems have opex on the same or greater order of magnitude as capex.
B) Tech leaders believe AI is a “sport of kings”: the AI system the captures the majority of users will take all the profits (ala search engines and social media).
C) Most users always choose whichever AI is free regardless of relative quality.
So long as these three assumptions hold, there will always exist a tech leader willing to subsidize the opex of their AI system to capture the user base, and you can conclude that no AI company will ever win the sport, and every single company is going bankrupt.
In other words, if most of your debt is used to subsidize your opex, and a competitor will be willing to subsidize their opex to capture your user base, then you cannot convert to profitability without losing your users, and you must subsidize opex in perpetuity.
I think most investors believe either A or B fails. If A fails, the good AI systems will be so expensive to build that it precludes new competitors from entering the space (if this holds, then you only need to subsidize for longer than your competitors). If B fails, companies will find a way to be profitable without capturing a huge number of users. (This is why the belief is the important axiom, it informs the willingness of tech CEOs to subsidize their opex in the short-term.)
I do not know what OpenAI believes; they currently have the vast majority of users but they cannot remove their free plan without losing users. And they cannot subsidize longer than Google or Meta who can self-subsidize with their money printers.
I don’t know what I believe. But when you lose sight of unit economics, interesting games are bound to appear.
I'll leave it to the mods to decide if this post is more suitable for ~finance. However, the jaw-dropping scale of computing and other resources at issue here makes OpenAI a systemic risk on so...
I'll leave it to the mods to decide if this post is more suitable for ~finance. However, the jaw-dropping scale of computing and other resources at issue here makes OpenAI a systemic risk on so many fronts - financial markets, energy, environment, international relations, military... As usual, it's a "read the whole thing".
The tl;dr is that OpenAI's business model is wholly unsustainable. At a generous estimate, it's burning twice as much money as it's receiving in revenue. That's assuming the revenue statements are trustworthy - for instance, OpenAI is booking money from SoftBank that it's never yet received. The size of expenditures threatens to run SoftBank out of money.
OpenAI has massive growth in user numbers, but can't convert the vast majority of free users to paid at a sufficient rate for sustenance. Model efficiency is getting worse with each generation, not better.
Microsoft and Coreweave, OpenAI's primary datacenter partners, are pulling back expenditures (Microsoft) or unable to deliver sufficient resources (Coreweave). New computing facilities can't be built fast enough to support existing users as more demanding features are rolled out (e.g. reasoning models, image generation).
OpenAI is turning construction of its new Stargate datacenter over to partners (former crypto companies) who've never designed and run AI datacenters. It's unlikely they'll be able to meet tight timelines required for OpenAI to receive its next funding tranches. OpenAI has another $10 billion funding problem - it must convert the company from the mixed non-profit model to for-profit by the end of 2025 or lose that capital investment.
Multiple tech companies and funders are at risk of severe losses or complete business failure if OpenAI doesn't succeed - SoftBank, Microsoft, Oracle, Coreweave, Nvidia, to name the biggest.
Zitron's summary:
I want to be extremely blunt with the following points, as I feel like both members of the media and tech analysts have failed to express how ridiculous things have become. I will be repeating myself, but it's necessary, as I need you to understand how untenable things are.
SoftBank is putting itself in dire straits simply to fund OpenAI once. This deal threatens its credit rating, with SoftBank having to take on what will be multiple loans to fund OpenAI's $40 billion round. OpenAI will need at least another $40 billion in the next year.
This is before you consider the other $19 billion that SoftBank has agreed to contribute to the Stargate data center project, money that it does not currently have available.
OpenAI has promised $19 billion to the Stargate data center project, money it does not have and cannot get without SoftBank's funds.
Again, neither SoftBank nor OpenAI has the money for Stargate right now.
OpenAI needs Stargate to get built to grow much further.
I see no way in which OpenAI can continue to raise money at this rate, even if OpenAI somehow actually receives the $40 billion, which will require it becoming a for-profit entity. While it could theoretically stretch that $40 billion to last multiple years, projections say it’ll burn $320 billion in the next five years.
Or, more likely, I can’t see a realistic way in which OpenAI gets the resources it needs to survive. It’ll need a streak of unlikely good fortune, the kind of which you only ever hear about in Greek epic poems:
SoftBank somehow gets the resources (and loses the constraints) required to bankroll it indefinitely.
The world’s wealthiest entities — those sovereign wealth funds mentioned earlier, the Saudis and so on — pick up the slack each year until OpenAI reaches productivity (assuming it does).
It has enough of those mega-wealthy benefactors to provide the $320bn it needs before it reaches profitability.
Crusoe and CoreScientific turn out to be really good at building AI infrastructure — something they’ve never done before.
Microsoft walks-back its walk-back on building new AI infrastructure and recommits to the tens of billions of dollars of capex spending it previously floated.
Stargate construction happens faster than expected, and there are no supply chain issues (in terms of labor, building materials, GPUs, and so on).
If those things happen, I’ll obviously find myself eating crow. But I’m not worried.
In the present conditions, OpenAI is on course to run out of money or compute capacity, and it's unclear which will happen first.
I can promise you that Microsoft, Oracle, and NVIDIA will be just fine. NVIDIA is probably the most at-risk, as a lot of their valuation is tied up in 'Exponential AI GPU sales forever" hype...but...
I can promise you that Microsoft, Oracle, and NVIDIA will be just fine.
NVIDIA is probably the most at-risk, as a lot of their valuation is tied up in 'Exponential AI GPU sales forever" hype...but one would hope that they would be able to coast back to their pre-crypto immense profitability.
Oracle will also be fine for the forseeable future, given their immense ties to huge support contracts.
Microsoft will be propped up by their Windows/Office monopoly probably until the end of time unless huge swaths of the world actively ban using it.
I don't think any of those companies will go out of business, but any large losses will trouble an already vulnerable marketplace. AI investment has rapidly sucked the oxygen out of the atmosphere...
I don't think any of those companies will go out of business, but any large losses will trouble an already vulnerable marketplace. AI investment has rapidly sucked the oxygen out of the atmosphere for startups and capacity in other areas - biotech, clean energy, advanced manufacturing, infrastructure, housing...
Imagine what $320 billion over five years could accomplish otherwise, then imagine it vanishing in a puff of vapor if there is an OpenAI business failure. I'm not sure all that datacenter GPU capacity has signifcant value if the AI bubble pops.
Nvidia only makes 6% of its revenue from OpenAI now, but I don't know how much is at risk if there's a prolonged downturn in overall GPU demand.
As Zitron mentions, Microsoft owns OpenAI's intellectual property in models, but what's that worth if there's little general market revenue to be realized, or the costs of using them are unsustainable? [I don't have detailed information on why Microsoft pulled back from AI datacenter investment, but they may have seen this writing on the wall.]
There’s a theory I’ve heard floated from some pretty well-respected people in finance that this article backs up. I haven’t heard anyone endorse this theory, but they do talk about it.
As I understand it, the idea requires these assumptions:
A) Competitive AI systems have opex on the same or greater order of magnitude as capex.
B) Tech leaders believe AI is a “sport of kings”: the AI system the captures the majority of users will take all the profits (ala search engines and social media).
C) Most users always choose whichever AI is free regardless of relative quality.
So long as these three assumptions hold, there will always exist a tech leader willing to subsidize the opex of their AI system to capture the user base, and you can conclude that no AI company will ever win the sport, and every single company is going bankrupt.
In other words, if most of your debt is used to subsidize your opex, and a competitor will be willing to subsidize their opex to capture your user base, then you cannot convert to profitability without losing your users, and you must subsidize opex in perpetuity.
I think most investors believe either A or B fails. If A fails, the good AI systems will be so expensive to build that it precludes new competitors from entering the space (if this holds, then you only need to subsidize for longer than your competitors). If B fails, companies will find a way to be profitable without capturing a huge number of users. (This is why the belief is the important axiom, it informs the willingness of tech CEOs to subsidize their opex in the short-term.)
I do not know what OpenAI believes; they currently have the vast majority of users but they cannot remove their free plan without losing users. And they cannot subsidize longer than Google or Meta who can self-subsidize with their money printers.
I don’t know what I believe. But when you lose sight of unit economics, interesting games are bound to appear.
I'll leave it to the mods to decide if this post is more suitable for ~finance. However, the jaw-dropping scale of computing and other resources at issue here makes OpenAI a systemic risk on so many fronts - financial markets, energy, environment, international relations, military... As usual, it's a "read the whole thing".
The tl;dr is that OpenAI's business model is wholly unsustainable. At a generous estimate, it's burning twice as much money as it's receiving in revenue. That's assuming the revenue statements are trustworthy - for instance, OpenAI is booking money from SoftBank that it's never yet received. The size of expenditures threatens to run SoftBank out of money.
OpenAI has massive growth in user numbers, but can't convert the vast majority of free users to paid at a sufficient rate for sustenance. Model efficiency is getting worse with each generation, not better.
Microsoft and Coreweave, OpenAI's primary datacenter partners, are pulling back expenditures (Microsoft) or unable to deliver sufficient resources (Coreweave). New computing facilities can't be built fast enough to support existing users as more demanding features are rolled out (e.g. reasoning models, image generation).
OpenAI is turning construction of its new Stargate datacenter over to partners (former crypto companies) who've never designed and run AI datacenters. It's unlikely they'll be able to meet tight timelines required for OpenAI to receive its next funding tranches. OpenAI has another $10 billion funding problem - it must convert the company from the mixed non-profit model to for-profit by the end of 2025 or lose that capital investment.
Multiple tech companies and funders are at risk of severe losses or complete business failure if OpenAI doesn't succeed - SoftBank, Microsoft, Oracle, Coreweave, Nvidia, to name the biggest.
Zitron's summary:
I can promise you that Microsoft, Oracle, and NVIDIA will be just fine.
NVIDIA is probably the most at-risk, as a lot of their valuation is tied up in 'Exponential AI GPU sales forever" hype...but one would hope that they would be able to coast back to their pre-crypto immense profitability.
Oracle will also be fine for the forseeable future, given their immense ties to huge support contracts.
Microsoft will be propped up by their Windows/Office monopoly probably until the end of time unless huge swaths of the world actively ban using it.
I don't think any of those companies will go out of business, but any large losses will trouble an already vulnerable marketplace. AI investment has rapidly sucked the oxygen out of the atmosphere for startups and capacity in other areas - biotech, clean energy, advanced manufacturing, infrastructure, housing...
Imagine what $320 billion over five years could accomplish otherwise, then imagine it vanishing in a puff of vapor if there is an OpenAI business failure. I'm not sure all that datacenter GPU capacity has signifcant value if the AI bubble pops.
Nvidia only makes 6% of its revenue from OpenAI now, but I don't know how much is at risk if there's a prolonged downturn in overall GPU demand.
As Zitron mentions, Microsoft owns OpenAI's intellectual property in models, but what's that worth if there's little general market revenue to be realized, or the costs of using them are unsustainable? [I don't have detailed information on why Microsoft pulled back from AI datacenter investment, but they may have seen this writing on the wall.]
As for Oracle, f*ck them anyway.