Deep dive · AI infrastructure

AI is a Bubble. This Isn’t the Pin.

When three AI bosses said the race should slow down, most people read it as the end of the trade. Run it through the numbers and it does close to the opposite, buying the labs time they didn’t have, and it landed while the market was already leaning the wrong way.

Julius Stark · 14 September 2026 · Updated 27 September · 9 min read

A cracked bubble marked AI bursting above a trading floor of panicking traders, with a Nasdaq board showing a steep decline — The scene a lot of people spent that weekend bracing for. I think they had the timing wrong.

Update, 27 September. The Fed raised rates by a quarter point on 16 September to 3.75–4.00%, its first increase since 2023, on a unanimous 12–0 vote. The market took it better than the weekend mood suggested: from the 11 September close, semiconductors (SMH) are up 6.7% and the Nasdaq 100 4.1%, against 0.9% for the S&P. The rates side has kept getting worse, though. The ten-year is now 5.17% and the 10Y–2Y gap 0.36pp, both further along than when this was written. Anthropic’s annualised revenue was also reported above $100bn on 18 September, well past the roughly $65bn used below, which strengthens the argument. The argument below is as published, with a few figures corrected since.

The short version

Slowing the frontier cuts the largest cost these labs carry, which helps their numbers rather than hurting them.
OpenAI burns about 57% of its revenue and Anthropic about 33%, so profitability is the live problem, and this is one of the few levers that moves it.
The buildout is capped by memory and power supply, so a pause in model releases doesn’t stop the chips shipping.
Semiconductors were already 15.4% off their high and positioning was bearish into the Fed, which left more room to rally than to fall.

The setup

Most people I talk to went into the weekend of 12 September leaning bearish. Oil was above $100, the Fed was due to meet on the Wednesday, and the ten-year sat at 4.95%. Then on the Friday, Dario Amodei, Sam Altman and Elon Musk, who between them run most of the Western frontier labs, agreed that AI was moving too fast and the industry should slow down.

If you were looking for a reason to believe the AI trade was over, that looked like one, and the headlines ran with it. I read it the other way. When a market is already positioned for bad news and the bad news arrives, there often isn’t much selling left to do, and a lot of rallies have started from exactly that point.

The bearish read needs the buildout to stop, and nothing agreed on Friday stops it.

What they actually agreed

Amodei’s proposal has four parts. Outside evaluators get employee-level access to Anthropic’s safety work, other labs are invited to do the same, governments coordinate on standards, and legislation follows, covering chip export restrictions and pre-release testing with the power to block a model. Altman said he agreed the industry needs to “pace the frontier”, and Musk replied that Dario was right.

Read literally, it is a safety document, and nothing in it mentions money. What matters for markets is what pacing the frontier does to the accounts. Slowing model development slows the activity these companies spend most on, whatever the stated reason, and it does so at the point where that saving is most useful to them.

I’m not alone in reading it commercially. David Sacks, until recently the administration’s AI lead, backed a slowdown in substance but rejected the altruistic framing and argued the companies weren’t acting in good faith. Chamath Palihapitiya said the plan concentrates power with Anthropic. When people inside the industry describe the motive as commercial, I think that is worth taking at face value.

What it does to the numbers

The timing matters because these businesses have plenty of revenue and not enough profit.

57%of revenue — OpenAI’s 2026 burn

33%of revenue — Anthropic’s

~$25bnOpenAI’s reported 2026 cash burn

Anthropic is running at roughly $65bn of annualised revenue and OpenAI at about $40bn. Those are extraordinary figures for companies this young, and both still lose money, because training the next model costs more than the last one earns.

Anthropic has taken its gross margin from minus 94% in 2024 to somewhere between 44% and 60%, and has already reported its first adjusted operating profit, in the second quarter of 2026. SemiAnalysis puts the third-quarter GAAP operating profit at around $1bn, with sustained cash breakeven pencilled in for 2028. OpenAI, whose consumer users mostly pay nothing, isn’t expected to get there until 2029 or 2030.

Training costs sit in research and development, and that line is what stands between both companies and a profit. If the frontier slows, the line stops compounding, breakeven comes forward and the runway lengthens. For a company deciding whether it needs to raise again next year, that is a large change rather than a rounding one.

The saving won’t arrive overnight, though. Most lab compute is contracted years ahead: Anthropic pays SpaceX $1.25bn a month to 2029, and OpenAI has more than $600bn of commitments. That cost doesn’t fall because a model release slips. The benefit comes as capacity shifts from training to paid inference, which is the same cost earning more revenue, and as those contracts roll off and get renewed on the labs’ terms.

It also matters well beyond the labs. Forecasts for Nvidia, the memory makers and the data-centre operators all rest on how long their biggest customers can keep buying, and those forecasts currently show growth rolling over from 2028. If the labs last three years longer than the bears assume, every one of those estimates gets pushed out.

The build doesn’t stop

The obvious objection is that less spending means fewer chips and fewer data centres, and a worse year for everyone who sells into them. At the moment that doesn’t follow, because money isn’t the constraint.

High-bandwidth memory has been the bottleneck all year, and SK Hynix has warned the shortage could run past 2030. Advanced packaging is booked out, and grid connections rather than budgets decide when a new campus switches on. When an industry is already building as fast as it physically can, announcing that you’ll buy less doesn’t change what gets built this year.

So the labs get the accounting benefit without the physical build slowing: they look healthier on paper while the chips keep shipping. That combination is why I keep landing on the bullish side.

What was already priced

This is the part I think most people missed.

	Close, 11 Sep	52-week high	From high
S&P 500 (SPY)	764.29	779.37	−1.9%
Nasdaq 100 (QQQ)	714.88	748.65	−4.5%
Nvidia	218.29	236.54	−7.7%
Semiconductors (SMH)	568.53	671.83	−15.4%

Semiconductors had fallen eight times as far as the index. That is the corner of the market a real spending slowdown would hit first, and it had been marked down quietly for weeks while the S&P sat within 2% of its high. The sell-off people spent the weekend worrying about had largely happened already, in the place you would expect it, so the bearish case wasn’t a new idea arriving on Monday so much as a trade that was already well advanced.

One figure worth correcting. A lot of the weekend commentary said three or four rate hikes were priced in. Futures actually showed about a 65% chance of a single quarter-point hike on the Wednesday, and J.P. Morgan Wealth Management had moved to one hike from none. That cuts both ways: the bulls had less of a cushion than they claimed, and there was less fear in the price than the “it’s all priced in” crowd assumed.

How I’m wrong

There are three ways this could fall over, and I would rather set them out than have you find them.

A regulator could call it a cartel. Three dominant firms publicly agreeing to slow output is an antitrust problem waiting for someone to name it. Sacks has already come close, and a headline like that would arrive without warning and isn’t priced anywhere.

Revenue could slow with the models. My case assumes costs fall while sales keep growing, but sales come from shipping better models. If the frontier genuinely slows, growth may slow with it while the compute contracts, signed years ahead, don’t, which is a worse position than either problem on its own.

The Fed may not cooperate. With oil at $104, ISM manufacturing down to 54.6 in August from 55.6, and the prices index at 71.1, the 23rd straight month of rising prices, you have softening growth alongside sticky input costs. That is a poor backdrop for expensive stocks whatever three executives said on Friday.

I would also point to the yield curve, the input in my own framework that has deteriorated most. With the ten-year at 4.95% and the two-year at 4.56%, the gap has narrowed from about half a point in late August to under 0.40, with yields rising across the board. That is tightening into a slowdown, and it is the kind of move that takes a curve towards inversion.

What I’m watching

The Fed on Wednesday, and what the curve does afterwards. A second hike entering the futures strip would do more damage here than any announcement about model pacing.
Q3 capex guidance from the hyperscalers. If the pacing agreement were really about spending it would show up here first, and unchanged guidance would rule out the bearish reading.
Memory pricing. A genuine slowdown would ease the bottleneck, so if contract prices keep rising, nothing physical has changed.
Whether anyone breaks ranks. One lab shipping a frontier model on the old schedule would end the arrangement.
Semiconductors against the index. If SMH keeps lagging while the S&P holds up, the market is pricing a spending slowdown regardless of what the proposal says.

Where I land

I do think it is a bubble. The financing is circular, the customers are also the investors, and the valuations assume revenue will arrive faster than the supply chain can physically deliver it. I have written about parts of that before and none of it has changed.

But bubbles rarely end because the people running them agree to be careful; they end when someone can’t raise money. Friday’s announcement did the opposite. It cut the cost line that was pushing these companies towards another funding round, at the moment their profitability was being questioned most, and it did so while the market was positioned for the reverse and the most exposed part of it was already down 15%.

The headline was bearish but the mechanics look bullish to me, and gaps like that tend to be worth something. When the end does come, I expect it to look like a failed funding round or a neocloud that can’t refinance, rather than a statement about safety.

Not investment advice, and not an allegation of wrongdoing by any person or company named. Analysis of publicly reported figures and my own interpretation of them. Index and single-stock levels are closing prices for 11 September 2026. Treasury yields are constant-maturity values for 11 September 2026 from the Federal Reserve H.15 release. The 27 September update uses closing prices for 25–26 September and the FOMC statement of 16 September. Revenue, burn and breakeven figures for OpenAI and Anthropic are third-party estimates, not company disclosure, and are labelled as such in the sources below. The view that the slowdown is commercially motivated is my reading of the effect; the published proposal is framed on safety grounds and does not discuss spending. Verify everything against primary sources before acting on it.

Sources

← All reports