Chip Stocks Sell Off on Amodei's Pacing Call While 2026 Capex Forecasts Keep Rising
The stated cause of today’s move is clean enough. Dario Amodei argued for a slower pace of frontier development, Sam Altman backed the idea of pacing, and AI-linked semiconductors sold off through Asia, Europe and the US premarket: Nvidia, AMD, Micron, Intel, ASML, the rest of the complex. What the tape is pricing is harder to defend than the headline suggests.
Pacing the frontier and slowing AI spending are two different propositions, and Altman said plainly that pacing does not mean stopping. The question worth asking is who would have to change behaviour for the second thing to follow from the first. Not the labs, whose compute budgets are already committed years out. The hyperscalers. And the hyperscalers have announced nothing of the kind.
The forecasts point the other way. TrendForce recently lifted its 2026 AI-server shipment growth estimate to nearly 31% year over year, and puts combined 2026 capital spending across nine major cloud providers north of $886 billion. Goldman models roughly $765 billion of annual AI capex in 2026, on a path toward $1.6 trillion a year by 2031 in the base case. PwC’s September outlook has global data-centre capex around $800 billion next year and climbing for decades after. These are estimates, and estimates move. But they move on power contracts, land, transformer lead times and memory allocations, not on whether one lab chief thinks model releases should be spaced out.
There is a category error underneath the selloff. AI has already stopped being a model-development story and become a buildout, with all the concrete that implies. Every generation of model consumes compute, networking, memory, storage, electricity and cooling in volumes that get contracted long before the model exists. And the models that already exist are being pushed into search, coding, enterprise software, security, advertising, support, robotics and autonomous systems, all of which run on inference rather than training. A slower improvement curve doesn’t erase that demand. It stretches the useful economic life of installed capacity, which is the opposite of the depreciation problem the bears usually raise.
So why did the stocks drop? Because semiconductor multiples don’t discount today’s demand; they discount future capacity decisions, which are exactly the thing a credible warning from inside the industry calls into question. Positioning was one-sided for uninterrupted acceleration. A name can fall 8% on a repriced expectation while its order book is untouched, and most of this complex has spent the year priced for the order book to keep compounding. The move is a multiple event. Nothing in it tells you the business changed.
There’s an irony in the safety framing that the tape ignored entirely. If regulators and labs get more serious about control, the spending response is not fewer GPUs. It is more inference capacity, more evaluation infrastructure, more monitoring, more logging, more security tooling, more redundant deployment. Caution is expensive, and it gets bought with the same silicon. Some of that money shifts away from raw frontier scaling, which matters for the mix inside the complex, though it says very little about the total.
The useful discipline here is to name what would actually change the thesis, because the pacing debate isn’t it. A hyperscaler cutting or even flattening its capex guide at the next print. Order pushouts showing up in equipment lead times rather than in commentary. DRAM and NAND contract prices rolling over on renewal, which is the fastest-reacting series in the chain and the one that would show demand softening before any executive admitted it. None of the three has happened. Until one does, a selloff on a podcast argument is a sentiment event wearing fundamental clothes.
That does not make the valuation concern illegitimate. The market may well have priced this too aggressively for the next few quarters, and a reset from those levels can be violent without meaning anything about 2029. It only means the larger claim being traded today, that AI investment itself is about to decelerate, currently has no support in the capital-spending data.
The cycle is entering a more awkward phase, from build as fast as possible to build as fast as is economically and politically sustainable. The second one is slower, messier, and still enormous. For Nvidia, AMD, Broadcom, the memory suppliers and everyone attached to the buildout, that distinction is most of the investment case.
Watch the next round of hyperscaler capex guidance. That’s the number that would have to break.