Ten Days That Changed the Course of AI: An Inference Rotation, Not a Crash

The ten days that began with the release of GPT-6 Astra and ended with a global rout of AI-chip stocks look like whiplash. On September 3, OpenAI shipped the most capable model it has ever broadly deployed, then spent the following week rolling it out across ChatGPT, Codex, and the API. Nine days later, the CEOs of Anthropic and OpenAI, joined by Elon Musk, publicly urged the industry to slow down. On Monday, September 14, that call sent AI-linked stocks tumbling — SoftBank slid as much as 13% in Tokyo, SK Hynix fell about 6%, ASML more than 4%.

Two bookends that appear to say opposite things. They do not. Read the way an operator reads it, the ten days delivered one message: the AI cycle has rotated from training to inference, and the market reacted to that single rotation twice — once with euphoria over the demand, once with fear over the rhetoric — while largely missing where the contest actually moved.

GPT-6 was an inference event

Start with the model, because its mechanics decide the economics. GPT-6 Astra reasons before it answers: it produces a long internal chain of thought on every request, and that chain is pure inference — tokens generated per response, at scale, constantly. OpenAI calls it the first of its frontier models to reach "Critical" on the company's cybersecurity preparedness scale, and admits the monitoring that capability requires comes at "significant compute cost."

That is the training-to-inference inflection made concrete. For years the argument over AI infrastructure was about who could build the biggest training run; the dollar followed the largest model into the cluster. A reasoning model in wide use flips the denominator. The expensive, recurring workload is now serving a token — and whoever serves a token cheaper owns a growing share of the industry's entire spend.

The slowdown call hit sentiment, not demand

Amodei's essay, titled "We Must Pace the Frontier," warned that within six to twelve months agent swarms could be capable of taking over the internet and called for external monitoring, industry-wide and global regulation. Altman said the industry needed to "pace the frontier"; Musk answered "Dario is right." The debate went public beside the resignation of an Anthropic researcher who said builders believe the technology "could kill us all by the end of the decade."

Markets read the headlines, not the operating record. Check what happened to the commitments across those same ten days. Nvidia's Vera Rubin — its inference-generation platform, built for roughly ten times the performance per watt of the prior generation — was ramping into full production, not pausing. Nvidia had committed up to roughly $105 billion to back OpenAI's data-center campus in Ohio, a stake in keeping OpenAI building with Nvidia systems even as OpenAI builds chips of its own. None of that moved. The price action quietly agreed: Nvidia is up over the past week. The rout was a sentiment shock on top of an intact demand signal.

That is the discipline that separates a ten-day panic from a changed thesis. A demand signal does not flip on an op-ed. What would flip it — a cancelled order, a pushed delivery date, a utilization drop — did not appear. What appeared was the opposite: more silicon, committed.

The contest that actually moved: the inference cost curve

Here is the change those ten days made visible, and it is worth separating from the ordinary "AI is growing" noise. The biggest buyers of compute are now engineering around Nvidia's price on the exact workload that just became dominant. In June, OpenAI and Broadcom unveiled Jalapeño, a custom inference chip cut at TSMC's 3nm node, reticle-sized, designed from scratch to serve large-language models. Broadcom's chief executive put the early result at roughly 50% lower cost per inference token than current Nvidia GPUs. OpenAI reported 1.5 to 1.9 times the peak token throughput per watt of commercial systems in its own tests and plans to deploy the chip by year-end alongside Nvidia and AMD accelerators. Meta, meanwhile, plans to begin manufacturing its own Iris inference chip this September, designed with Broadcom and fabbed by TSMC.

This is the dark-horse pattern the market keeps underestimating. Nvidia does not need a monopoly on AI accelerators to compound — its software and full-stack moat are formidable — but the inference rotation is precisely where that moat is thinnest. Custom silicon is not a catalogue of specs; it is a supply chain choosing to buy less from the incumbent on the workload that now sets the industry's economics. Nvidia's answer is Vera Rubin's efficiency, and it is a real answer. But the live question is no longer whether the AI buildout is real — that is settled and committed. It is whether the cost-per-token gap custom silicon is reaching holds against Nvidia's own inference generation, and whether all these commitments convert into delivered capacity on schedule.

To read those ten days as the end of the AI trade is to read the panic instead of the delivery schedule. The fabs, the order books, and the guarantees kept moving the same direction they had been — the buildout is accelerating, and the balance of that rising cycle is being settled over who can make a token cheapest. The useful divide is the one Amodei and Altman never mentioned: a slowdown that gets enforced through policy is a real long-tail risk to this cycle, but an op-ed, however sincere, is not delivery. I will try to trust the one of those that shows up in the financial statements.