OpenAI’s missed goals, the Elon Musk lawsuit, and the latest hyperscaler spending spree all point to the same constraint: power, not demand, is now shaping the AI market.
OpenAI missed its user and revenue targets, but the argument around that miss quickly turned into something larger: whether AI demand is outrunning the physical grid that has to support it. David Sacks and Chamath Palihapitiya said the real constraint is power, not interest, while David Friedberg argued that hyperscalers are winning because they control the infrastructure. The same theme kept returning in a different form, from coding tools to cyber models to the latest cloud earnings: the companies that can secure compute, energy, and capital are the ones that can keep scaling.
OpenAI’s missed targets landed less like a verdict than a test: does a shortfall in users and revenue mean the company is losing momentum, or only that its ambitions are now running into the limits of physical infrastructure? David Sacks argued that the answer is the second one, and that the market is reading the wrong signal. The real story, in his telling, is not weak demand but scarce compute and power.
I have a little bit of a contrarian take on this. OpenAI had a really bad week in the press, but if you look at what’s happening at the product level, it’s been a pretty good couple of weeks for them.
4:55
Sam may end up being right, but for the wrong reason. He missed on consumer, but enterprise is going gangbusters and is giving him the ability now to catch up on code.
6:34
Sacks’s case rests on a simple trade. He acknowledged the Wall Street Journal’s report that OpenAI missed a 1 billion weekly active user target and fell short of its 2025 revenue goal, then shifted to product quality, saying chat GPT 5.5 is getting stronger reviews while Anthropic’s Opus 4.7 is, in his words, being rolled back by some users. He also pointed to OpenAI’s larger compute base, arguing that the company’s earlier spending commitments now look like an asset in the coding market, where demand is concentrated and performance differences are easier to feel.
Chamath Palihapitiya pushed the argument further, saying the missing numbers do not point to falling demand at all. In his view, the bottleneck is power, from the grid to the hardware that feeds model inference, and the symptom is not weaker appetite but queues, caps, and slower rollout. He said the same pattern is already visible at Anthropic, which he described as forced into capacity deals and routing workarounds to get enough supply.
Everything in this market is power constrained. The reason these folks may miss a number or a forecast has nothing to do with demand.
8:41
It is entirely due to the supply of the power necessary to generate the output token. The problem is only getting worse.
OpenAI’s missed targets quickly became a proxy fight over a deeper limit: how much AI growth is being throttled by power, not demand. Chamath Palihapitiya argued that the market is misreading the shortfall, because the bottleneck now sits in electricity, grid hardware, and supply chains, while the customer appetite for models and enterprise tools keeps rising.
I think they're going to be fine. I think this is a multi-trillion dollar company. I think Anthropic is a multi-trillion dollar company.
8:09
There is one very specific choke point that is constraining everything, which is access to the power that's necessary to drive these tokens.
8:41
His claim was not that demand is weak. It was that the industry has outrun the physical system needed to serve it, so missed forecasts can reflect scarce capacity rather than fading interest. He pointed to Anthropic’s deals with Amazon and Google as evidence that the scarce asset is no longer attention or software talent, but the ability to secure enough compute behind the scenes.
Everything in this market is power constrained. The reason that these folks may miss a number or a forecast have nothing to do with demand.
8:51
The limiting resource is power. Power, which then powers compute, which then provides tokens.
10:58
David Friedberg pushed the same line into a more strategic frame. In his view, the companies that win are not the ones with the flashiest models, but the ones that control the industrial inputs, and he said that means the hyperscalers are positioned to benefit most. He named Oracle, Amazon, Meta, Microsoft, and Google as the firms that can sit on the scarce infrastructure while smaller model makers negotiate for access.
OpenAI’s new cyber model lands in the conversation as both proof and warning. David Friedberg argues that GPT 5.5 cyber shows the frontier models are no longer just writing code or answering prompts, they are now stepping directly into offense and defense, where the same tool can harden a system or break it open.
The frontier models have reached the point where they’re capable of automating cyber activities just like they’re capable of automating coding.
21:55
A model could power up a cyber attacker or cyber defender the same way it can power up a coder.
21:55
The speakers kept returning to the same bottleneck: the most capable systems are only as useful as the infrastructure behind them. Friedberg said Anthropic’s Mythos made a splash as a cyber proof of concept, but OpenAI’s model looks closer to something enterprises can actually deploy because it appears less compute-constrained.
It’s not going to be just Mythos. Within six months or so, all the frontier models are going to have Mythos-level cyber capability.
22:14
I think 5.5 might be the first cyber model that cyber defenders actually get to use.
22:14
The argument around Big Tech’s latest earnings was not really about whether the businesses are healthy. It was about what kind of businesses they are becoming. As the four giant cloud platforms prepared to spend more than $700 billion on capex, the speakers treated that figure as evidence that AI is pulling the industry out of its old software-light, cash-rich shape and into something far more industrial.
I think we’re seeing a very important structural shift in the capital markets. The last 20 or 30 years has been that the Mag 7 just kind of ran away with it, because they had these very asset-light business models.
43:33
Now all of a sudden the pendulum is swinging violently in the other direction. These companies will now get levered, they’re going to have more debt, and they’re going to look like this big bulky industrial business in five years.
43:44
Friedberg’s case was that the old model is being replaced by one in which capital intensity, borrowing, and long-lived infrastructure matter as much as code. He pointed to Microsoft’s deal to restart part of Three Mile Island as a sign that cloud companies are now bargaining directly for power, not just renting it indirectly through data centers. If that is right, the big platforms stop looking like pure software monopolies and start looking like utilities with better margins.
The issue we had in 2000 was dark fiber. You had all this infrastructure being built out and it wasn’t being used. There’s no dark GPUs today.
46:05
What’s driving the capex now is the voracious demand for compute for tokens, and the demand is now pulling forward this additional investment in infrastructure.
46:54
That was the line of defense against the dot-com comparison. Friedberg argued that this spending is not speculative supply waiting for an audience, as in the fiber buildout two decades ago, but a response to current demand that is already straining capacity. He tied the cloud earnings to the claim that AI is not only consuming more infrastructure, it is forcing the platforms to reinvest faster just to keep up.
By the time the conversation reached the Supreme Court, the mood had shifted from market frenzy to institutional awe. Friedberg described walking into the marble chamber as if entering a place that still commands ritual respect, even as the speakers kept circling back to how fragile that respect may be. The same anxiety that hangs over AI regulation also hangs over the court itself: how much authority do Americans still grant to systems that are under visible strain?
I think it was one of the most amazing experiences I’ve ever had. That building, you walk in, it’s like sacred.
6:38
Enjoy it while you can. I think the Supreme Court is one of the last highly functional institutions in the United States.
7:34
Friedberg’s account of the Monsanto case turned that abstract reverence into a concrete fight over who gets to define risk. The question was whether the EPA’s label should preempt state failure-to-warn claims, after Bayer, which now owns Monsanto, had already paid out $10 billion and still faced 90,000 cases. He framed the hearing as a contest between federal uniformity and state power, with the Trump administration’s solicitor general urging the court to take the case.
The EPA sets the label for pesticides. Does this cause cancer or not? What are the warnings?
7:05
If the states get to interpret federal law and ignore federal regulatory bodies, it opens up a whole new can of worms.
7:18
Why do they say power matters more than demand?
They argue that AI demand is already there, but electricity, grid equipment, and data-center buildouts are limiting how many tokens models can actually serve.
Why is OpenAI’s product performance still seen as strong?
David Sacks says recent releases and coding improvements may offset the missed consumer targets, especially if enterprise usage keeps rising.
What did they think about the Musk v. Altman case?
They treated it as hard to handicap, but said the diary excerpts and nonprofit-to-profit dispute could create settlement pressure or delay an IPO.
Why are the hyperscalers central to the story?
Because Google, Microsoft, Amazon, and Meta control the infrastructure that AI companies need, and they can finance the buildout better than most startups.
What was the Supreme Court discussion doing here?
It served as a contrast case, showing how institutions with clear procedures and legal constraints still shape high-stakes outcomes in a way AI companies increasingly do too.
AI-assisted summary of All-In Podcast's podcast, verified against the original transcript.