The AI Trade’s First Threat From Inside: What a…

The AI Trade’s First Threat From Inside: What a…

Every previous scare in the AI trade came from outside the industry. A cheaper Chinese model would raise the question of whether US companies were overspending on chips; a jump in bond yields would pressure the valuations. This time the threat came from the people building the technology.

Over one weekend, the chief executives of Anthropic, OpenAI and xAI publicly agreed that AI development should be deliberately slowed, and on Monday the stocks that depend on it fell hard, Nvidia (NASDAQ: NVDA) down about 2.9% before the US open, with Broadcom off 3.2%, AMD 5.7%, Intel 5% and Marvell 7%.

The reason the call lands as a market event rather than a philosophical one is the number on the other side of it. The largest US hyperscalers plan to spend roughly $700 billion on AI infrastructure in 2026, up about 77% from around $410 billion in 2025, as compiled from company guidance by Fortune and Value Add VC. That spending is a bet that capability keeps accelerating fast enough to justify it. If the builders themselves slow the acceleration, the payoff moves further out while the bills keep arriving.

The largest US AI infrastructure builders plan to spend roughly $700 billion in 2026, up 77% from 2025. Data: company guidance, compiled by Fortune and Value Add VC · Chart: FinanceFeeds.

The $700 Billion Bet That Assumes Acceleration

The scale of the spending is what makes the timing of a slowdown so awkward. Amazon is guiding to about $200 billion of 2026 capex, Microsoft toward $190 billion, Alphabet $175 billion to $205 billion, Meta $125 billion to $145 billion, and Oracle around $50 billion, the overwhelming majority of it going to Nvidia GPUs, custom silicon, data centers and the power to run them. The logic behind building years ahead of demand is that each new model generation unlocks enough new revenue to justify the last round of spending.

That logic was already under strain before this weekend. Barclays has modeled negative free cash flow for the hyperscalers in 2027 and 2028, calling the prospect “somewhat shocking” but “likely what we eventually see for all companies in the AI infrastructure arms race,” CNBC reported. Meta shares fell more than 9% earlier this year when it raised its capex guidance, the first real investor rebellion against the spending curve. The market was already asking whether the returns would arrive before the patience ran out, and a deliberate slowdown is a direct answer in the wrong direction.

Why a Slowdown From Inside Is Different

The distinction from past selloffs is the source. When a cheaper model out of China rattled the trade, the question was whether US firms were overbuilding; the demand thesis itself stayed intact. This time the demand thesis is what is in question, because the companies generating the demand are the ones proposing to temper it.

Anthropic’s Dario Amodei published an essay, “We Must Pace the Frontier,” arguing for a deliberate slowing of frontier capability gains. OpenAI’s Sam Altman posted that he agreed and would “pace the frontier,” and xAI’s Elon Musk replied “Dario is right,” an alignment among three fierce competitors within 48 hours that FinanceFeeds detailed alongside OpenAI ruling out a 2026 listing on safety grounds.

A slower capability cycle does not cancel the $700 billion; the contracts are signed and the chips are shipping. What it changes is the timeline for the payoff. If the interval between major capability jumps lengthens deliberately, the revenue that is supposed to catch up with the spending arrives later, while the spending, already committed, does not pause with it.

AI-linked shares and index futures fell worldwide as the slowdown calls spread, with SoftBank, OpenAI’s largest backer, the hardest hit. Data: CNBC, CNN, company reporting · Chart: FinanceFeeds.

The Market Repriced the Bet, Not the Business

The selloff reached across the AI supply chain, and it is worth reading as a repricing rather than a verdict. SoftBank, OpenAI’s largest backer, fell as much as 13% in Tokyo, South Korea’s KOSPI dropped 3.3% with SK Hynix down more than 6%, and in Europe ASML fell around 6%. The move also arrived alongside a separate oil spike, with WTI up about 3% above $103 after a Saudi pipeline shutdown, and a Federal Reserve decision due this week, so no single cause explains it.

Analysts framed it as the market reassessing expectations, not the underlying business. The pullback in Nvidia “looks more like a risk repricing than a broken business,” Investing.com wrote, with the catalyst “serious but still mostly narrative-driven.” Saxo’s Neil Wilson said analysts would be “scrabbling around to assess likely impact on earnings and valuations” if labs coordinate a material slowdown, and investor Michael Burry, characteristically, called the slowdown talk “self-serving.”

Investor Takeaway

The threat is to the timeline, not the demand: a deliberate slowdown lengthens the payback on $700 billion of committed capex without stopping the spending, which is why infrastructure and memory names fell hardest.

The Earnings Engine on the Other Side of the Ledger

The awkwardness is that the same AI build-out is delivering record profits at the same moment it is being asked to slow. S&P 500 full-year earnings are now projected to grow about 32% in 2026, up from roughly 24% before second-quarter reporting began, with 86% of companies beating estimates, the highest rate in years, and Bloomberg Intelligence naming the AI infrastructure buildout as the clearest driver, per Bloomberg. Communication Services’ 2026 growth forecast alone jumped to 51% from 26%.

Much of the surge was concentrated and partly one-off: Alphabet accounted for 92% of one week’s earnings increase on a large gain, and Amazon booked $53.4 billion largely tied to its Anthropic stake, so excluding the two, growth was closer to 32% than 50%, FactSet noted. The tension is the story: AI is simultaneously the profit engine lifting the whole index and the spending commitment straining cash flow, and the people running it just introduced doubt into both sides of that ledger at once.

Coordination Is Unlikely, Which Cuts Both Ways

A genuine, sustained slowdown would require agreement that looks improbable. China’s state-backed Global Times dismissed Amodei’s essay as a “Cold War playbook” aimed at Beijing, noting it also urged tighter US chip export controls, as Reuters reported, and President Donald Trump brushed the calls aside, saying “whoever wins AI wins.”

At the same time, the pressure is escalating at the level of heads of state: King Charles will convene the leaders of Nvidia, Google DeepMind, OpenAI and Anthropic this week at Dumfries House in Scotland to discuss shared safety principles, Buckingham Palace confirmed, and the US and China are due to discuss frontier-AI safety in mid-September, possibly at a Trump-Xi summit on September 24.

For the trade, that unresolved split cuts both ways. If coordination fails, as China’s and Trump’s positions suggest it might, the competitive pressure that drove the $700 billion does not ease, and the capex bet may hold. But the debate itself, now reaching from the labs to the palace, keeps a risk premium on the most crowded trade in the market, which is exactly what Monday’s repricing was.

Investor Takeaway

A coordinated slowdown looks unlikely: with China rejecting the call and Trump dismissing it, the arms-race logic behind the $700 billion stays intact, which supports the spenders even as the safety debate pressures sentiment.