On OpenAI's IPO delay: This is good and bad simultaneously.
Good: their current IPO timeline is far too compressed and they're racing to do so much in a short period of time, it buys them time to try and improve financials - valuation, profitability, etc - before finally releasing their S-1.
Bad: they need the money now because of the cash they're burning and IPOs are the final frontier for a company with OpenAI’s valuation to do that, their financials weren't good + attracts more public scrutiny from investors (they can hide a little longer), SpaceX's IPO may have sucked the oxygen (liquidity) out of the markets for now, hence Google frontrunning them with the $85bn equity raise. Also Anthropic have better numbers + a non-zero chance of reaching profitability, even if it’s small.
Sarah Friar was ultimately right with wanting to float OpenAI in 2027, but she can’t walk back that “looking for a government backstop” comment back in Nov ‘25.
As SpaceX lockup periods expire, SpaceX price could collapse. That will make an OpenAI IPO less attractive, because investors will see what happened to SpaceX stock post-IPO and wonder if the same could happen to OpenAI.
The Chinese models stop improving if they can’t distill the frontier labs, no? Very amusing - frontier labs stole knowledge - Chinese labs cribbed from what they stole - frontier labs can’t charge as much as it costs to maintain the theft algorithm, collapse? - Chinese models can’t distill anymore.
On device small models or some such survives. People in Memphis stop having to breathe nasty data center air.
I commented in part because I’d enjoy hearing from someone with more knowledge of this topic. That doesn’t seem to be you, given your assertion (that only Anthropic has made this claim, with no evidence) is clearly wrong. OpenAI made the claim first and both do have evidence. I don’t trust either of these companies, but distillation is a training technique they all use. And it’s part of the lack of moat.
I'm not saying it isn't used - and OpenAI hasn't been nearly as adamant about claiming it as Anthropic.
I'm saying there is no evidence that China's models are all based on it.
And frankly, even if that were true, I'd call it leveling the playing field.
I also reiterate that I haven't seen actual evidence - that is, a report detailing how and when it was done from actual logs. All I've seen are public claims. Provide a link to a technical report from either company which establishes their case.
I’m not criticizing the distillation? It’s very ironic that something based on copying is incredibly easy to copy and therefore impossible to own. That’s my point.
I'd agree with that - which is why I say it's leveling the playing field and that's fine by me - as long as it's not ALL the Chinese do. And I'm fairly sure it isn't. They do produce their own training methods after all.
I happen to be opposed to any form of "intellectual property". If you don't want it copied, don't put it on the Internet. That's what the Internet is for.
QUOTE:The company's VP of vehicle hardware engineering, Charles Poon said that leaders overlooked the deep experience of veterans who survived many product cycles. Poon admitted that simply replacing them with AI was a huge mistake, and that while AI is "a fantastic tool," it remains "only as good as the information you use to train it."
END QUOTE
QUOTE:
Ford's realization that AI cannot magically design and test quality vehicles without senior human oversight is just the tip of the iceberg. When Careerminds looked at companies that conducted AI-driven layoffs, researchers found out that 35.6% of those companies had to rehire more than half of the employees they previously fired. Another 32.7% had to rehire between 25% and 50% of them.
In 2024, Sebastian Siemiatkowski, CEO of Klarna, proudly announced that its new chatbot was doing the work of 700 full-time customer service agents. As a result, the fintech company froze hiring and cut hundreds of positions. But by mid 2025, and into 2026, Klarna was scrambling to recruit human agents again because customer satisfaction had plummeted. It turns out, while AI is very good at answering basic questions like how to check an account balance, when faced with complex customer issues that require nuance, the thing usually resorts to the unhelpful, robotic corporate jargon we all know and love.
Re: your comment to @noahsmith. Thing is, Gary, you can't know either. You write with such certainties in a market that is highly uncertain. And I'm not sure you can be that certain.
I have had a team (humans plus AI) tracking every type of leading signal for revenue or usage, and even though we revised up our forecasts for Anthropic, we were still caught off guard by how quickly revenue grew at the start of this year.
The chinese model data zero hedge shared from bloomberg is our data from openrouter -- but we make the point that it is from a non-representative sample. It's far more nuanced than this commentary put out.
Since you've referred specifically to two things we worked on, my view is easily found on Exponential View and the brief commentary around the AI revenue report early this week, let me comment:
I see very few signs -- from talking to close to roughly a 100 senior (C-suite, c minus 1 and c minus 2) execs in the US and Europe of them reducing spending or commitment to AI. Definitely ambitious to get bigger and better results. Also want to introduce some deliberate discipline, because it's silly to pay for licences for tools that some people aren't using. They may start shifting workloads to some open-source models. But even if they do that, it just shifts the margin to the hyperscalers, for whom serving those models can be a very good business.
I'd encourage you to build your own assumptions-based model to look at that.
There has been an overdue realisation in firms that they need an internal portfolio of models and software to get real results, and that getting real results is not as light a lift as replacing one SaaS vendor with another.
I'm also continually surprised by the types of LLM-based tools I've seen even stodgy government contractors build that are proven to work for them.
I am uncertain about what the real impact of cost control will be. I found one company of any size who believes their AI bill is approaching 1% of their revenues. But they also tell me that close to 40% of their multi-billion dollar real business is now related to AI. And no one who I've spoken to in even a wider group than the 100 or so I discussed above expects to spend less in the next 12 months.
Finally, we expect this years top line, full year, industry revenue to end up above $175bn (the current annualised run rate), which will be around three times higher than calendar year 2025.
Increasing efficiency of serving on existing hardware, let alone how much more efficient the new hardware is, I would expect the general margins for the hyperscalers to improve. Unless they feel as they might do reasonably that they're still in a market share game and can afford to keep margins lower for now. That's just what companies have done since time immemorial.
Your "you can't know either" framing doesn't engage Gary's specific claims including peak tokenmaxxing, pre-Chinese competition, the SpaceX subsidy. Symmetric uncertainty doesn't address the specific argument; it just skips it. Any rebuttals to his points?
As for the C-suite anecdotes: 100 execs talking to an analyst who covers AI aren't a random draw from the population of companies. Did you actively seek out organizations quietly scaling back with the same rigor? Was this from conversation notes or a structured methodological survey? What specifically did you ask and what were these "signs"? What does "may start shifting" even imply?
On the numbers. $175B is Annualized Run Rate? The most recent month extrapolated forward. Worse it may be cherry picked for best month. Your own report acknowledges on p.2 that "the demand side has been obscure" and "even public companies bury AI revenue inside segment totals." It also puts the actual trailing 12-month figure at $110bn?
Further, citing the extrapolated number without even offering a rough MoE undersells that caveat. Are you off by +/- 5%? 10%? At what confidence?
For example: OpenAI's CFO Friar listed $20B+ ARR as of year-end 2025, but the financials independently verified by the FT show actual 2025 revenue was $13.07B. That's a $7B difference on 13B actual.
This is not a tiny MoE; that gap isn't a forecasting error. It's what happens when you get to choose what to annualize and report it. This discrepancy is for ONE company. How does that propagate across all firms in your $175B analysis? What is your MoE?
But anyway, Gary's argument wasn't that revenue isn't growing, it's that revenue growth != profitability. The same framing Gary criticized Noah Smith for using.
Your own report's closing question is whether margin can service the CapEx buildout and it shows depreciation absorbing 81% of hyperscaler GenAI revenue before other costs. That's not Gary's worry grafted onto your data; that's your own data's open question.
That cost question is what's actually on the table. It's not a handwave. Not a "they'll figure it out". Your comment sidesteps it entirely.
Again, as your own report said "the open question is whether cheapening artificial intelligence can create enough volume and margin to service the buildout."
"I see very few signs -- from talking to close to roughly a 100 senior (C-suite, c minus 1 and c minus 2) execs in the US and Europe of them reducing spending or commitment to AI. Definitely ambitious to get bigger and better results."
Except they're not getting those results. And the issues with AI-generated code quality and cybersecurity are growing by the day.
This is being called "AI psychosis".
Companies are not looking before they're leaping into the AI playpen
93% of organizations report infrastructure incidents attributable to AI
"The findings are unambiguous: organizations are using AI to generate infrastructure code at a rate their governance frameworks were never designed to handle,” said Paweł Hytry, co-founder and CEO of Spacelift, in a statement.
The consequences of these incidents, respondents say, consist of reworking AI-generated changes (37 percent), security misconfigurations that reached production (36 percent), compliance violations (36 percent), infrastructure drift attributable to AI changes (35 percent), and incidents caused by agentic systems (33 percent).
END QUOTE
The assumption by many is that LLM foibles will be corrected "in the next model". Except they can't be by definition. They can be ameliorated to some degree by deterministic additional controls - which are expensive to install and which many, perhaps most companies using AI, don't know how to do.
And then again, when the Iran war oil shock hits and the recession arrives, consumers will be paying less for AI. How much less is uncertain, of course, but that it will be less is probable. So revenues, let alone profitability, will go down.
And as I've said before, you can't run massive data centers cheaply when oil costs over $100 a barrel - and oil experts expect oil to remain north of that price, despite any recent fall-off due to the supposed "negotiations" based on the Memorandum of Understanding.
Yeah- it is definitely a concern. I spoke with a bunch of CISOs about this about four weeks ago, a
reasonably broad group from US, Europe
and Asia. They do see it, as
you rightly point out, a priority to get the governance of this right, and figure out how to manage incipient risks, but again none planned to use LLMs less either in their work or across the firm. And all expected the types of changes they would
make would be people and
process not just faith in the tech. The one exception was the importance of dealing with Mythos (which was available to a few of them at the time).
The oil price at the moment is being manipulated by the US National Reserve draw down as well as the assumption that "the Strait is open" - which it is but at ten percent of pre-war volume which is pretty much where it's been all along as Iran has been letting ships go through all along except for limited periods and except for the limited number of ships the US blockade was turning back. Probably half the ships previously got past the US blockade by hugging the coastal waters of Pakistan where the US blockade could not go.
The US blockade is lifted now, but Iran controls the rate of passage. All ships must register with the Iran-Oman management, although at the moment no tolls are being collected. Still, it's going to be a slow affair getting the volume back up anywhere near the pre-war volume. It will likely take months.
Also, the oil price is the futures market price which is a bet by traders that the Strait will be open permanently in the next few months. That ain't gonna happen. And the price will fluctuate based on events. With events like the day or so - a ship tried to run past the Iran control and got hit, the US retaliated with air strikes and Iran retaliated with unspecified strikes - it's clear the Strait could be closed again at any time.
Oil experts expect the price to not stay at this level once the Reserve runs out which is due in some weeks. And of course, the war is likely to re-ignite at some level soon.
At the very least it seems apparent that the US and Israel are cooperating in using Lebanon as the means to restart the war. The deal between the US, Israel and the Lebanese government - which is under US control - is intended to maintain Israeli control of a portion of Lebanon. Hezbollah has rejected the deal and Israel is continually violating it. The MoU specifies that Israel must withdraw. So when Israel doesn't, Iran will be forced to declare the MoU dead and attack Israel. This is exactly what the US and Israel want - to restart the war and blame Iran for it. This is standard US and Israel "negotiation" tactics.
The other possibility based on Lebanon is the US is attempting to find some way to "disarm" Hezbollah - either by using the Lebanese army - which ain't gonna happen - or Syria's current Al Qaeda government - which also ain't gonna happen. I suspect at some point the US will have to put its own Marines and Army on the ground in Lebanon to assist Israel. I've been saying that since 2006, the last time Israel failed to dislodge Hezbollah. But if the US tries that, the Iran war will definitely restart.
And we haven't even covered the "negotiations" themselves which probably less than ten percent chance of succeeding in producing a lasting deal.
All of which pretty much guarantee high oil prices for all of this year and probably well into next year, if not further.
I'm sorry, but right now the prices for future delivery at different times are around $70 per barrel in 1000 barrel lots and bit including other charges. But a barrel can be obtained for $30 less than it could be when the lest price was $100 per barrel, Thus is why gas prices have dropped by about a third of their previous rise. When you say "you can't run massive data centers cheaply when oil costs over $100 a barrel - and oil experts expect oil to remain north of that price" -- they are not north of that price, they are far south of that price, so they can't remain there. You say "when the Iran war oil shock hits and the recession arrives..." but the oil shock already did arrive, prices were well above $100 per barrel, and we did not enter a recession. I'm not saying we won't have a recession soon, I don't know, and you've presented no evidence you do. If oil prices rise to 100 I will be surprised (unless Trump returns to carpet bombing) and give you credit.
Paul Samuelson observed "Wall Street indexes predicted nine out of the last five recessions.” You're claiming confidently that they are missing one! Though they may be predicting hard times for tech. A recession is always possible, but the past three months have seen more new jobs than all of 2025, consumer spending is so high the Fed may raise interest rates, and wages have risen much more than inflation, providing room for more spending. Spending may be more in the US if international travel declines. $70 per barrel is the rate for binding contracts for future delivery. It will be delivered, unless maybe there is an escape clause for "acts of God, Trump, or the Ayatollah." It is not a prediction market, it is contract pricing. It's why gas prices have dropped.
The point isn't that you are wrong, it's that you might sprinkle in more "could"s and "might"s and supporting data, and fewer "will"s and "zero percent"s that can diminish credibility. I agree with your general long-term outlook. Let Gary be an object lesson. He is not shy about reminding everyone that his long term predictiions have been borne out, but none of his 2024 predictions for 2025 were. They may be here by 2027. Your predictions for next year could arrive then, or by 2028, or not at all. Being a little humble provides time to hedge; mundane tools can be designated "symbolic" additions, as with chess and protein-folding and coding, despite modeling small fragments of cognition or the world. (The real challenge is social modeling, and some tech companies seem to know it.)
I came here to say half of that. I'd only add that I was in enterprise software my entire career and I've either slung or heard these kinds of arguments when things were going off the rails. AI assisted I'd say, as well
So many inconsistencies in this. For just one easy example, what does this mean?
They may start shifting workloads to some open-source models. But even if they do that, it just shifts the margin to the hyperscalers, for whom serving those models can be a very good business.
I mean I understand what you’re saying. Why hit the trivial stuff with the $$$ AI gold plated Claude code, but even at my level (rookie) this argument doesn’t make sense. They will farm out everything. Everything that can run on like Deepseek or Gemini to organize and optimize, but that is exactly what he’s saying. The hard stuff? If I’m a semi intelligent manager? Why would I pay for the AI vs a human at a known cost?
It all coming apart. The cost to implement I capable senior engineer is probably 10 to 30x cheaper and reliable than any slop coded product. Otherwise you’d be hearing about MASSIVE layoffs. Massive productivity.
My 18 year old kid? He won’t even acknowledge using AI.
I’ve followed Marcus since I saw him on Steve Eisman a few months ago and as a former operative engineer it took me about 10 minutes of research to realize he was right.
I think we are mixing up a couple of different things here. The pricing pressure openAi and anthropic face is independent of the margin the hyperscalers can make serving open source models.
And your point about cost is important - AI spend is getting more attributable not less attributable. See for example Brian Armstrong’s recent description of how CoinBase manages and monitors spend.
We touch on a bunch of these specific questions, including scenarios for faster or slower open source uptake in the main analysis. intelligence.exponentialview.co
Both OpenAI (per Ed Zitron) and Anthropic (per Ed Elson) officially spend billions per year on 'sales and marketing'. They say they spend roughly the same on sales and marketing as the Coca-Cola Company.
Coca-Cola Company world wide, that is...
Where? Is that really 'sales & marketing'? Or might, maybe might, those cost mostly be the unsustainable subsidies on inference they offer? I also noticed before that enormous amounts are spent on R&D, but we have no idea how much of that R&D is actually hidden cost of inference (i.e. in A/B testing). OpenAI has done a deal that seems to be meant to show increase of revenue for OpenAI (70% joint venture) while leaving any profits (of reselling tokens OpenAI delivers at a huge discount to the joint venture) at the joint venture partner. Then there are all these circular investment deals.
Both firms are desperate to show profitability ("1 month") and "rapid revenue growth" for their IPO. But none of it hides that the brute force of Generative AI is hideously (and increasingly as the scale goes up) inefficient and costly and the serious (non-slop) use is probably only economically viable for certain niches (like hunting for vulnerabilities in code where a single task may costs many tens of thousands of $ and that thus aren't really available for that small hacker in his bedroom, but are available for large enterprises or governments)
Generative AI will stay. But there is a large gap between what is serious use and what is economically viable.
That analysis of Ed Zitron in https://www.wheresyoured.at/exclusive-openai-financials/ about the actual numbers of OpenAI is an eye opener for two reasons: The catastrophic financial losses they incur and the desperate measures they fall back to to massage the numbers into something a *little bit* less terrible to look at.
If they proceed with their IPO and you want to buy: You have been warned.
I didn't think I'd find myself supporting anything this 2nd Trump administration did, but I have to admit I am loving this new turn toward treating AI companies as though their CEOs' cynical apocalyptic "OMG IT'S GETTING TOO POWERFUL EVERYONE BE CAREFUL (and also pls purchase a subscription)!!!!!" marketing schtick was serious. Altman and Amodei been pulling this shit the entire time and only just now has the government decided to show them what playing along looks like. Chef's kiss.
Hahahahaha. Azeem is a perma AI bull (just like he was/is wrt crypto).
The fact it took his TEAM "several months to construct" that data, in world replete with AI productivity sales pitches (of which he engages in), is somewhat ironic... It's not like he was constructing a regional climate model (extremely complex), but... basic revenue data? He is extremely well connected, so if it took so long because his connections were holding out on him... also does not bode well.
Some good points here and I don’t dismiss them all. But one is just silly: the specter of useless/out-of-date GPUs. 6-7 year old GPUs are still humming day and night; and there’s still a very active market for them. Such is the undersupply at the moment. Yes someday that may level off, but that point is not in sight. “Obsolete” vs the best, is one thing. ‘Obsolete’ for doing useful work is vey different, and thus far very resilient.
Seems like every day another Marcus GenAI prediction becomes undeniable fact. Shoulda listened to Gary, Sama, Musk and all the rest of you grifters who harassed, lied about, and dissed him.
It's almost like the SpaceX IPO capitalized on Anthropic's temporary rise, and left everyone else (OpenAI, Anthropic, and retail investors) holding the bag...
The bubble needs to burst . Too much BS is clouding focus on actual advances like protein folding and other expert systems
Go back to the drawing board and make stuff tbe world needs, not a parasitic bot that captures IP in the name of competition and guesstimates the next alphabet
Sharp, thanks for sharing this.
OpenAI ~ WeWork...awesome! Spot on!
On OpenAI's IPO delay: This is good and bad simultaneously.
Good: their current IPO timeline is far too compressed and they're racing to do so much in a short period of time, it buys them time to try and improve financials - valuation, profitability, etc - before finally releasing their S-1.
Bad: they need the money now because of the cash they're burning and IPOs are the final frontier for a company with OpenAI’s valuation to do that, their financials weren't good + attracts more public scrutiny from investors (they can hide a little longer), SpaceX's IPO may have sucked the oxygen (liquidity) out of the markets for now, hence Google frontrunning them with the $85bn equity raise. Also Anthropic have better numbers + a non-zero chance of reaching profitability, even if it’s small.
Sarah Friar was ultimately right with wanting to float OpenAI in 2027, but she can’t walk back that “looking for a government backstop” comment back in Nov ‘25.
As SpaceX lockup periods expire, SpaceX price could collapse. That will make an OpenAI IPO less attractive, because investors will see what happened to SpaceX stock post-IPO and wonder if the same could happen to OpenAI.
Even if OpenAI was a good bet financially, I personally am a little uncomfortable investing in human extinction: https://www.youtube.com/watch?v=_eYTkvZqbnQ
The Chinese models stop improving if they can’t distill the frontier labs, no? Very amusing - frontier labs stole knowledge - Chinese labs cribbed from what they stole - frontier labs can’t charge as much as it costs to maintain the theft algorithm, collapse? - Chinese models can’t distill anymore.
On device small models or some such survives. People in Memphis stop having to breathe nasty data center air.
I've yet to see any evidence of Chinese "stealing" from US models.
ALL of this "distilling" stuff comes from Anthropic - with zero public evidence - who have long since destroyed their credibility about anything.
I commented in part because I’d enjoy hearing from someone with more knowledge of this topic. That doesn’t seem to be you, given your assertion (that only Anthropic has made this claim, with no evidence) is clearly wrong. OpenAI made the claim first and both do have evidence. I don’t trust either of these companies, but distillation is a training technique they all use. And it’s part of the lack of moat.
I'm not saying it isn't used - and OpenAI hasn't been nearly as adamant about claiming it as Anthropic.
I'm saying there is no evidence that China's models are all based on it.
And frankly, even if that were true, I'd call it leveling the playing field.
I also reiterate that I haven't seen actual evidence - that is, a report detailing how and when it was done from actual logs. All I've seen are public claims. Provide a link to a technical report from either company which establishes their case.
I’m not criticizing the distillation? It’s very ironic that something based on copying is incredibly easy to copy and therefore impossible to own. That’s my point.
I'd agree with that - which is why I say it's leveling the playing field and that's fine by me - as long as it's not ALL the Chinese do. And I'm fairly sure it isn't. They do produce their own training methods after all.
I happen to be opposed to any form of "intellectual property". If you don't want it copied, don't put it on the Internet. That's what the Internet is for.
Distillation predates LLMs by quite a long time
And Steve Earl seems to know a lot about it
https://m.youtube.com/watch?v=xvaEJzoaYZk&pp=0gcJCUECo7VqN5tD&ra=m
LOL
Replying to myself, this is the kind of well informed explanation I was looking for: https://substack.com/@sayash/note/c-284224204?r=k9l1i&utm_medium=ios&utm_source=notes-share-action
Another example of where industry is not getting the benefits;
Ford admits AI couldn't replace experienced engineers, rehiring veterans to improve quality and cut recall costs.
https://www.neowin.net/news/ford-execs-say-they-made-a-mistake-when-they-replaced-human-engineers-with-ai/
QUOTE:The company's VP of vehicle hardware engineering, Charles Poon said that leaders overlooked the deep experience of veterans who survived many product cycles. Poon admitted that simply replacing them with AI was a huge mistake, and that while AI is "a fantastic tool," it remains "only as good as the information you use to train it."
END QUOTE
QUOTE:
Ford's realization that AI cannot magically design and test quality vehicles without senior human oversight is just the tip of the iceberg. When Careerminds looked at companies that conducted AI-driven layoffs, researchers found out that 35.6% of those companies had to rehire more than half of the employees they previously fired. Another 32.7% had to rehire between 25% and 50% of them.
In 2024, Sebastian Siemiatkowski, CEO of Klarna, proudly announced that its new chatbot was doing the work of 700 full-time customer service agents. As a result, the fintech company froze hiring and cut hundreds of positions. But by mid 2025, and into 2026, Klarna was scrambling to recruit human agents again because customer satisfaction had plummeted. It turns out, while AI is very good at answering basic questions like how to check an account balance, when faced with complex customer issues that require nuance, the thing usually resorts to the unhelpful, robotic corporate jargon we all know and love.
END QUOTE
He ain't never had Noahopinion worth listening to.
Re: your comment to @noahsmith. Thing is, Gary, you can't know either. You write with such certainties in a market that is highly uncertain. And I'm not sure you can be that certain.
I have had a team (humans plus AI) tracking every type of leading signal for revenue or usage, and even though we revised up our forecasts for Anthropic, we were still caught off guard by how quickly revenue grew at the start of this year.
The chinese model data zero hedge shared from bloomberg is our data from openrouter -- but we make the point that it is from a non-representative sample. It's far more nuanced than this commentary put out.
Since you've referred specifically to two things we worked on, my view is easily found on Exponential View and the brief commentary around the AI revenue report early this week, let me comment:
I see very few signs -- from talking to close to roughly a 100 senior (C-suite, c minus 1 and c minus 2) execs in the US and Europe of them reducing spending or commitment to AI. Definitely ambitious to get bigger and better results. Also want to introduce some deliberate discipline, because it's silly to pay for licences for tools that some people aren't using. They may start shifting workloads to some open-source models. But even if they do that, it just shifts the margin to the hyperscalers, for whom serving those models can be a very good business.
I'd encourage you to build your own assumptions-based model to look at that.
There has been an overdue realisation in firms that they need an internal portfolio of models and software to get real results, and that getting real results is not as light a lift as replacing one SaaS vendor with another.
I'm also continually surprised by the types of LLM-based tools I've seen even stodgy government contractors build that are proven to work for them.
I am uncertain about what the real impact of cost control will be. I found one company of any size who believes their AI bill is approaching 1% of their revenues. But they also tell me that close to 40% of their multi-billion dollar real business is now related to AI. And no one who I've spoken to in even a wider group than the 100 or so I discussed above expects to spend less in the next 12 months.
Finally, we expect this years top line, full year, industry revenue to end up above $175bn (the current annualised run rate), which will be around three times higher than calendar year 2025.
Increasing efficiency of serving on existing hardware, let alone how much more efficient the new hardware is, I would expect the general margins for the hyperscalers to improve. Unless they feel as they might do reasonably that they're still in a market share game and can afford to keep margins lower for now. That's just what companies have done since time immemorial.
Your "you can't know either" framing doesn't engage Gary's specific claims including peak tokenmaxxing, pre-Chinese competition, the SpaceX subsidy. Symmetric uncertainty doesn't address the specific argument; it just skips it. Any rebuttals to his points?
As for the C-suite anecdotes: 100 execs talking to an analyst who covers AI aren't a random draw from the population of companies. Did you actively seek out organizations quietly scaling back with the same rigor? Was this from conversation notes or a structured methodological survey? What specifically did you ask and what were these "signs"? What does "may start shifting" even imply?
On the numbers. $175B is Annualized Run Rate? The most recent month extrapolated forward. Worse it may be cherry picked for best month. Your own report acknowledges on p.2 that "the demand side has been obscure" and "even public companies bury AI revenue inside segment totals." It also puts the actual trailing 12-month figure at $110bn?
Further, citing the extrapolated number without even offering a rough MoE undersells that caveat. Are you off by +/- 5%? 10%? At what confidence?
For example: OpenAI's CFO Friar listed $20B+ ARR as of year-end 2025, but the financials independently verified by the FT show actual 2025 revenue was $13.07B. That's a $7B difference on 13B actual.
This is not a tiny MoE; that gap isn't a forecasting error. It's what happens when you get to choose what to annualize and report it. This discrepancy is for ONE company. How does that propagate across all firms in your $175B analysis? What is your MoE?
But anyway, Gary's argument wasn't that revenue isn't growing, it's that revenue growth != profitability. The same framing Gary criticized Noah Smith for using.
Your own report's closing question is whether margin can service the CapEx buildout and it shows depreciation absorbing 81% of hyperscaler GenAI revenue before other costs. That's not Gary's worry grafted onto your data; that's your own data's open question.
That cost question is what's actually on the table. It's not a handwave. Not a "they'll figure it out". Your comment sidesteps it entirely.
Again, as your own report said "the open question is whether cheapening artificial intelligence can create enough volume and margin to service the buildout."
It's not a minor niggle; it's the entire point.
"I see very few signs -- from talking to close to roughly a 100 senior (C-suite, c minus 1 and c minus 2) execs in the US and Europe of them reducing spending or commitment to AI. Definitely ambitious to get bigger and better results."
Except they're not getting those results. And the issues with AI-generated code quality and cybersecurity are growing by the day.
This is being called "AI psychosis".
Companies are not looking before they're leaping into the AI playpen
93% of organizations report infrastructure incidents attributable to AI
https://www.theregister.com/devops/2026/06/24/companies-are-not-looking-before-theyre-leaping-into-the-ai-playpen/5261819#
QUOTE:
"The findings are unambiguous: organizations are using AI to generate infrastructure code at a rate their governance frameworks were never designed to handle,” said Paweł Hytry, co-founder and CEO of Spacelift, in a statement.
The consequences of these incidents, respondents say, consist of reworking AI-generated changes (37 percent), security misconfigurations that reached production (36 percent), compliance violations (36 percent), infrastructure drift attributable to AI changes (35 percent), and incidents caused by agentic systems (33 percent).
END QUOTE
The assumption by many is that LLM foibles will be corrected "in the next model". Except they can't be by definition. They can be ameliorated to some degree by deterministic additional controls - which are expensive to install and which many, perhaps most companies using AI, don't know how to do.
And then again, when the Iran war oil shock hits and the recession arrives, consumers will be paying less for AI. How much less is uncertain, of course, but that it will be less is probable. So revenues, let alone profitability, will go down.
And as I've said before, you can't run massive data centers cheaply when oil costs over $100 a barrel - and oil experts expect oil to remain north of that price, despite any recent fall-off due to the supposed "negotiations" based on the Memorandum of Understanding.
Yeah- it is definitely a concern. I spoke with a bunch of CISOs about this about four weeks ago, a
reasonably broad group from US, Europe
and Asia. They do see it, as
you rightly point out, a priority to get the governance of this right, and figure out how to manage incipient risks, but again none planned to use LLMs less either in their work or across the firm. And all expected the types of changes they would
make would be people and
process not just faith in the tech. The one exception was the importance of dealing with Mythos (which was available to a few of them at the time).
Hasn't the price of oil already dropped to around $70 per barrel?
The oil price at the moment is being manipulated by the US National Reserve draw down as well as the assumption that "the Strait is open" - which it is but at ten percent of pre-war volume which is pretty much where it's been all along as Iran has been letting ships go through all along except for limited periods and except for the limited number of ships the US blockade was turning back. Probably half the ships previously got past the US blockade by hugging the coastal waters of Pakistan where the US blockade could not go.
The US blockade is lifted now, but Iran controls the rate of passage. All ships must register with the Iran-Oman management, although at the moment no tolls are being collected. Still, it's going to be a slow affair getting the volume back up anywhere near the pre-war volume. It will likely take months.
Also, the oil price is the futures market price which is a bet by traders that the Strait will be open permanently in the next few months. That ain't gonna happen. And the price will fluctuate based on events. With events like the day or so - a ship tried to run past the Iran control and got hit, the US retaliated with air strikes and Iran retaliated with unspecified strikes - it's clear the Strait could be closed again at any time.
Oil experts expect the price to not stay at this level once the Reserve runs out which is due in some weeks. And of course, the war is likely to re-ignite at some level soon.
At the very least it seems apparent that the US and Israel are cooperating in using Lebanon as the means to restart the war. The deal between the US, Israel and the Lebanese government - which is under US control - is intended to maintain Israeli control of a portion of Lebanon. Hezbollah has rejected the deal and Israel is continually violating it. The MoU specifies that Israel must withdraw. So when Israel doesn't, Iran will be forced to declare the MoU dead and attack Israel. This is exactly what the US and Israel want - to restart the war and blame Iran for it. This is standard US and Israel "negotiation" tactics.
The other possibility based on Lebanon is the US is attempting to find some way to "disarm" Hezbollah - either by using the Lebanese army - which ain't gonna happen - or Syria's current Al Qaeda government - which also ain't gonna happen. I suspect at some point the US will have to put its own Marines and Army on the ground in Lebanon to assist Israel. I've been saying that since 2006, the last time Israel failed to dislodge Hezbollah. But if the US tries that, the Iran war will definitely restart.
And we haven't even covered the "negotiations" themselves which probably less than ten percent chance of succeeding in producing a lasting deal.
All of which pretty much guarantee high oil prices for all of this year and probably well into next year, if not further.
I'm sorry, but right now the prices for future delivery at different times are around $70 per barrel in 1000 barrel lots and bit including other charges. But a barrel can be obtained for $30 less than it could be when the lest price was $100 per barrel, Thus is why gas prices have dropped by about a third of their previous rise. When you say "you can't run massive data centers cheaply when oil costs over $100 a barrel - and oil experts expect oil to remain north of that price" -- they are not north of that price, they are far south of that price, so they can't remain there. You say "when the Iran war oil shock hits and the recession arrives..." but the oil shock already did arrive, prices were well above $100 per barrel, and we did not enter a recession. I'm not saying we won't have a recession soon, I don't know, and you've presented no evidence you do. If oil prices rise to 100 I will be surprised (unless Trump returns to carpet bombing) and give you credit.
You need to listen to Art Berman, an oil expert with thirty years experience in the field, a recognized expert. Here are two interviews with him.
Oil Supply Crisis Won't End Quickly /Lt Col Daniel Davis & Art Berman
https://www.youtube.com/watch?v=IyVUgJ_5H5k&t=1s
What a New 'Energy War' within Iran War will do to Oil Prices /Lt Col Daniel Davis & Art Berman
https://www.youtube.com/watch?v=_h6izbAULTo
In addition, Moody's has estimated that there is a 49% chance of a recession this year - and their calculations were done BEFORE the Iran war started.
Paul Samuelson observed "Wall Street indexes predicted nine out of the last five recessions.” You're claiming confidently that they are missing one! Though they may be predicting hard times for tech. A recession is always possible, but the past three months have seen more new jobs than all of 2025, consumer spending is so high the Fed may raise interest rates, and wages have risen much more than inflation, providing room for more spending. Spending may be more in the US if international travel declines. $70 per barrel is the rate for binding contracts for future delivery. It will be delivered, unless maybe there is an escape clause for "acts of God, Trump, or the Ayatollah." It is not a prediction market, it is contract pricing. It's why gas prices have dropped.
The point isn't that you are wrong, it's that you might sprinkle in more "could"s and "might"s and supporting data, and fewer "will"s and "zero percent"s that can diminish credibility. I agree with your general long-term outlook. Let Gary be an object lesson. He is not shy about reminding everyone that his long term predictiions have been borne out, but none of his 2024 predictions for 2025 were. They may be here by 2027. Your predictions for next year could arrive then, or by 2028, or not at all. Being a little humble provides time to hedge; mundane tools can be designated "symbolic" additions, as with chess and protein-folding and coding, despite modeling small fragments of cognition or the world. (The real challenge is social modeling, and some tech companies seem to know it.)
I came here to say half of that. I'd only add that I was in enterprise software my entire career and I've either slung or heard these kinds of arguments when things were going off the rails. AI assisted I'd say, as well
So many inconsistencies in this. For just one easy example, what does this mean?
They may start shifting workloads to some open-source models. But even if they do that, it just shifts the margin to the hyperscalers, for whom serving those models can be a very good business.
I mean I understand what you’re saying. Why hit the trivial stuff with the $$$ AI gold plated Claude code, but even at my level (rookie) this argument doesn’t make sense. They will farm out everything. Everything that can run on like Deepseek or Gemini to organize and optimize, but that is exactly what he’s saying. The hard stuff? If I’m a semi intelligent manager? Why would I pay for the AI vs a human at a known cost?
It all coming apart. The cost to implement I capable senior engineer is probably 10 to 30x cheaper and reliable than any slop coded product. Otherwise you’d be hearing about MASSIVE layoffs. Massive productivity.
My 18 year old kid? He won’t even acknowledge using AI.
I’ve followed Marcus since I saw him on Steve Eisman a few months ago and as a former operative engineer it took me about 10 minutes of research to realize he was right.
I think we are mixing up a couple of different things here. The pricing pressure openAi and anthropic face is independent of the margin the hyperscalers can make serving open source models.
And your point about cost is important - AI spend is getting more attributable not less attributable. See for example Brian Armstrong’s recent description of how CoinBase manages and monitors spend.
We touch on a bunch of these specific questions, including scenarios for faster or slower open source uptake in the main analysis. intelligence.exponentialview.co
This is classic hubris—overconfidence in one's course of action leading to self-inflicted pain and agony.
(But maybe frog with a box cutter scratching the chips is the real problem.)
Both OpenAI (per Ed Zitron) and Anthropic (per Ed Elson) officially spend billions per year on 'sales and marketing'. They say they spend roughly the same on sales and marketing as the Coca-Cola Company.
Coca-Cola Company world wide, that is...
Where? Is that really 'sales & marketing'? Or might, maybe might, those cost mostly be the unsustainable subsidies on inference they offer? I also noticed before that enormous amounts are spent on R&D, but we have no idea how much of that R&D is actually hidden cost of inference (i.e. in A/B testing). OpenAI has done a deal that seems to be meant to show increase of revenue for OpenAI (70% joint venture) while leaving any profits (of reselling tokens OpenAI delivers at a huge discount to the joint venture) at the joint venture partner. Then there are all these circular investment deals.
Both firms are desperate to show profitability ("1 month") and "rapid revenue growth" for their IPO. But none of it hides that the brute force of Generative AI is hideously (and increasingly as the scale goes up) inefficient and costly and the serious (non-slop) use is probably only economically viable for certain niches (like hunting for vulnerabilities in code where a single task may costs many tens of thousands of $ and that thus aren't really available for that small hacker in his bedroom, but are available for large enterprises or governments)
Generative AI will stay. But there is a large gap between what is serious use and what is economically viable.
That analysis of Ed Zitron in https://www.wheresyoured.at/exclusive-openai-financials/ about the actual numbers of OpenAI is an eye opener for two reasons: The catastrophic financial losses they incur and the desperate measures they fall back to to massage the numbers into something a *little bit* less terrible to look at.
If they proceed with their IPO and you want to buy: You have been warned.
One can only hope that when they do IPO, federal regulators and potential investors will recognize the financial tricks for what they are: fraud.
I didn't think I'd find myself supporting anything this 2nd Trump administration did, but I have to admit I am loving this new turn toward treating AI companies as though their CEOs' cynical apocalyptic "OMG IT'S GETTING TOO POWERFUL EVERYONE BE CAREFUL (and also pls purchase a subscription)!!!!!" marketing schtick was serious. Altman and Amodei been pulling this shit the entire time and only just now has the government decided to show them what playing along looks like. Chef's kiss.
Hahahahaha. Azeem is a perma AI bull (just like he was/is wrt crypto).
The fact it took his TEAM "several months to construct" that data, in world replete with AI productivity sales pitches (of which he engages in), is somewhat ironic... It's not like he was constructing a regional climate model (extremely complex), but... basic revenue data? He is extremely well connected, so if it took so long because his connections were holding out on him... also does not bode well.
Some good points here and I don’t dismiss them all. But one is just silly: the specter of useless/out-of-date GPUs. 6-7 year old GPUs are still humming day and night; and there’s still a very active market for them. Such is the undersupply at the moment. Yes someday that may level off, but that point is not in sight. “Obsolete” vs the best, is one thing. ‘Obsolete’ for doing useful work is vey different, and thus far very resilient.
Sam Altman probably just figures he needs time to find another gig before OpenAI IPOs (and potential investors open a eye.)
Seems like every day another Marcus GenAI prediction becomes undeniable fact. Shoulda listened to Gary, Sama, Musk and all the rest of you grifters who harassed, lied about, and dissed him.
Can they delay until next year? Part of Amazon funding stipulated AGI or IPO by Q4 - any reasonable person knows AGI is not possible
Ant/ floats at the largess of Musk and his glut of hardware w/o other buyers for his compute
The frontier models may just die on the vine. I’ll shed a tear
It's almost like the SpaceX IPO capitalized on Anthropic's temporary rise, and left everyone else (OpenAI, Anthropic, and retail investors) holding the bag...
The bubble needs to burst . Too much BS is clouding focus on actual advances like protein folding and other expert systems
Go back to the drawing board and make stuff tbe world needs, not a parasitic bot that captures IP in the name of competition and guesstimates the next alphabet