68% of the Selloff Back in Three Sessions (18 September 2026): Nvidia Told the AI Safety Summit It Will Sell Twice as Many Chips

On 14 September the market marked down the suppliers of AI compute by about 5% on an essay arguing the industry should slow the improvement of model capabilities. This note argued at the time that the ordering of the damage looked like a positioning unwind rather than a demand reassessment, and set a test: a demand reassessment persists and shows up in supplier guidance, while a positioning unwind retraces. Three sessions later the Philadelphia Semiconductor Index had recovered 68% of the fall — and the catalyst came from inside the safety summit itself, where Nvidia's chief executive said he expects to sell twice as many chips next year as this year.

What actually happened

The original note drew a distinction it could not then resolve. One reading of 14 September was a demand reassessment that happened to be badly targeted. The other was a positioning unwind, in which the hardest falls land where the crowding is greatest rather than where the fundamental exposure is greatest. The two are distinguishable, and not by argument: a demand reassessment persists and eventually appears in supplier guidance; a positioning unwind retraces and leaves no trace in any company's order book.

Here is the tape.

Two things in that column are worth separating. The first is the total: 468.21 points of the 692.72 lost, about 68%, recovered inside three sessions. The second is the distribution. Tuesday and Wednesday returned 114.83 points between them — under a quarter of the recovery — while the index sat essentially still through a Federal Reserve decision. Thursday alone returned 353.38. Sectors that have genuinely repriced their demand outlook do not behave like that, and neither do sectors that simply bounce on reflex. This was two days of nothing followed by a catalyst.

At the single-name level the recovery overshot where the fall had been deepest.

Every one of them is now above where it stood before the essay was published. The industry index is not, because it carries names that did not participate in Thursday's move — but the securities the demand story was supposedly about have already been marked back through their starting point.

The catalyst said it at the safety summit

On Thursday 17 September, King Charles III hosted senior representatives from Nvidia, OpenAI, Anthropic and Google DeepMind at Dumfries House in Cumnock, Scotland, for a meeting on AI safety organised by The King's Trust, The King's Foundation and the Sustainable Markets Initiative. Al Jazeera reported the King telling attendees there "seems urgency in adequately considering the existential dangers of such technologies falling into the wrong hands, and being used in potentially catastrophic ways." The summit was not expected to produce binding agreements, and did not.

Speaking to reporters before it, Huang said: "I expect Nvidia to sell twice as many chips as this next year as we do this year." The reason he gave was demand geography — "you can see that in almost every single country that we're in, people want to invest in AI" — per CNBC. At the same event, on the summit's own subject, he said: "When a product is not safe, we should hold it back and keep engineering it."

Twice as many chips is not twice the money

The forecast needs deflating before it is useful, and the deflation is the interesting part.

Nvidia does not disclose total unit shipments. Its best-known products are data-centre GPUs — Blackwell, and the Vera Rubin platform now in full production — but it also sells CPUs, switch and optical networking silicon, laptop parts, Jetson modules for robots and vehicles, and the processor inside a games console. The last time a number was attached to any of this was last October, when Huang said Nvidia had shipped 6 million Blackwell GPUs over four quarters. "Chips" is therefore a count across products whose prices differ by orders of magnitude, and a doubling of that count says nothing on its own about revenue.

Set it against the money. Nvidia's own second-quarter release reported revenue of $96.221bn, data-centre revenue of $89.0bn, and guided the October quarter to $108.0bn plus or minus 2%, with no China data-centre compute assumed — a guide whose index arithmetic we examined at the time. CNBC reported the company had indicated roughly 70% growth for the fiscal year ending January 2028, about $673bn. Working backwards, $673bn at +70% implies a fiscal 2027 base near $396bn; against $81.6bn, $96.2bn and $108.0bn for the first three quarters, that implies a fourth quarter near $110bn.

So: units up 2.0x, revenue up 1.7x. Average revenue per chip falls about 15%. That is what a broadening mix looks like — proportionally more of the cheaper silicon, not less of the expensive kind — and it is close to the opposite of what a reader would take from the headline. The caveat is real and belongs in the open: a unit remark to reporters and a revenue indication are not measured on the same basis, were not published together, and neither is a formal guide. The arithmetic is a sanity check on the headline, not a forecast derived from it.

What the essay actually committed anyone to

The document that started this was a personal essay, not a filing. Amodei's central claim was stated plainly: "We must slow the pace at which we improve the capabilities of AI models." The plan had three steps, descending in feasibility — embedded external evaluators with permanent employee-level access, which Anthropic committed to unilaterally; coordination among frontier labs in democratic countries, which would need government cover to be lawful; and an attempt by those governments to coordinate with authoritarian ones, whose verification problem the essay conceded.

The endorsements were narrower than the headlines suggested. OpenAI's Sam Altman wrote: "I agree with Dario that we need to pace the frontier." Elon Musk wrote: "Dario is right." Google DeepMind's Demis Hassabis said the direction was correct while the details needed working through. None of the four announced a delayed model, a reduced training run, or a cancelled order. One week on, none has.

The pledge touches one workload; a datacentre runs several

The contracts have kept arriving regardless. On 14 September, while suppliers were being sold, The Information reported that the customer behind a $13.7bn six-year compute contract RUM Group had disclosed in an August securities filing was Anthropic itself — the essay author's own company. That sits alongside a $45bn arrangement with Nscale, a $35bn deal with Lambda, a $9bn deal with Riot, and a July partnership with AMD covering up to 2 gigawatts of Instinct MI450 Series GPUs, per AMD's announcement. Behind them sits Microsoft's $329.1bn of signed-but-not-commenced leases, a figure we unpacked in the footnote that predicted the leak, and the backlog arithmetic of the neocloud tier covered in the $103bn note. A lease is a contract with a counterparty. A blog post is not a termination notice.

The supply side spoke, and said the opposite

Thursday's move had a second input, and it ran through supply rather than demand. Intel chief executive Lip-Bu Tan told an AI infrastructure summit in Santa Clara that the company can currently meet only about half of customer CPU demand, attributing the shortfall to inference workloads, and warned that the global memory shortage could worsen next year. Separately, reports of exploratory talks between SK hynix and Intel over US memory manufacturing added a second thread, which we examined in the Korean memory note.

The follow-through reached Seoul the next morning. On 18 September Samsung Electronics closed at 261,000 won, up 8,500 won or 3.37%, and SK hynix at 1,857,000 won, up 112,000 won or 6.42%, per Seoul Economic Daily — the latter now 9.4% above the 1,697,000 won it closed at on 14 September. The reported reasoning there was generational rather than volumetric: Vera Rubin uses HBM4, its successor is expected to use HBM4E, and high-bandwidth memory prices step up with each generation, so a doubling of Nvidia units is read in Seoul as a mix question first.

None of this is audited guidance. A chief executive at a conference is not a company filing, and that distinction is the whole reason the test set out in the original note has not yet been settled.

What still has not been asked

Four observable places would show a genuine slowdown, and one week on, none of them has reported.

Hyperscaler capital expenditure guidance is the first — a real downshift appears as a cut to a range, on a call, with a reason attached. Second is the lease disclosure: signed-but-not-commenced leases are the longest-dated commitment in the chain, and a pause shows up as that number growing more slowly, or as contractual conditions being invoked. Third is supplier guidance, issued under securities law rather than on a personal website — and that one now has a date. Micron reports fiscal fourth-quarter results on Wednesday 30 September, the first full supplier outlook since the essay, from the company most directly levered to the memory tightness Intel described. Fourth is the release cadence itself, the only direct test of the commitment, which will take months to observe and will always be deniable.

Does any of this reach currencies?

Still mostly no, and the week proved it. The route from a technology-sector drawdown to the eight majors runs through risk sentiment — one of the five factors scored on the live meter — and it is fast, shallow and quick to reverse. A sector that fell 5.86% and recovered 68% of it inside three sessions, while the S&P 500 finished the stretch within a quarter of a percent of where it started, is not a currency event.

What was a currency event was the rate calendar. The Federal Reserve raised its target range 25 basis points to 3.75%-4.00% on 16 September in a 12-0 vote, its first increase since 2023, with the statement noting that inflation remains elevated. The Bank of England held Bank Rate at 3.75% on a 6-3 split, with three members voting for an immediate rise to 4%. The Bank of Japan followed on Friday. We mapped all three in the decisions note; for the dollar's own scorecard the USD factor page is the place to look, and the methodology page explains what the meter is and is not doing.

The question to hold

What has to be true for an AI pacing commitment to matter to the instruments you can actually trade? Not that the labs are sincere — they may well be. Not that the risks are real — that is a separate argument with its own merits. What has to be true is that a slower capability frontier translates into fewer chips bought, less capacity leased, or less power contracted, within a horizon the market discounts.

One week of evidence points the other way, and it now includes the sector's own price. The most accelerator-levered listed company fell least and is back above where it started. The two names that fell hardest are both more than 5% above their pre-essay prices. The largest supplier stood at a safety summit and forecast twice the units. A second supplier said it can fill about half its orders. And the selloff that was supposed to mark the turn in AI spending has two-thirds of its damage repaired without a single company amending a contract.

That does not make the selling wrong, and the sector is still 1.90% below where it was — a residual discount, not a resolved question. It makes the selling a statement about positioning and relative winners rather than about the level of AI capital spending, and those two things resolve very differently. The tape has now answered twice. On 30 September, for the first time, an order book answers under securities law.

Educational macro context only — not investment advice.