Tonight, in the middle of a state banquet in Washington, the people whose companies are moving $30.7 trillion of market value will be seated in the same room as the two leaders whose handshake everyone is watching. President Xi Jinping is in town — the first US state visit in eleven years — and President Trump went to Joint Base Andrews to greet him personally, a courtesy no sitting president has paid a foreign leader (save the pope) in roughly six decades.

“Tonight, in the middle of a state banquet in Washington, the people whose companies are moving $30.7 trillion of market value will be seated in the same room as the two leaders whose handshake everyone is watching.”

The official agenda is the "three Ts": trade, technology, Taiwan. Treasury Secretary Scott Bessent already delivered the pre-summit deliverable — the tariff truce that was due to expire November 10 gets a two-month extension to January 10. On the AI file, the most concrete proposal on the table is a notification "hotline" for national-security-level AI incidents, plus talks on threat-intelligence sharing between labs. No one expects a treaty.

But markets don't wait for communiqués. While diplomats negotiated sentence wording, the tape was doing its own diplomacy — pricing exactly which parts of the AI economy benefit from detente, which got hurt, and what valuations say about how much optimism is already spent. I pulled the full 36-name AI basket (ChartRow, close of Tuesday September 23) and ran the numbers. Here is what the market signed before the summit even opened.

A Basket That Doesn't Move as One

Headline first: the median name in the AI basket is +50.3% year-to-date, against an S&P 500 total return of roughly +12% through mid-September. The AI trade is still the trade. But the dispersion inside the basket is almost absurd — from Sandisk at +665% to Oracle at -25.8% in the same calendar year. That is a 691-point spread between the best and worst member of the same "theme."

The category medians tell you where the money actually went:

Read that table from the bottom. The two groups everyone talks about most — the hyperscalers writing the $100 billion capex cheques, and the power companies everyone assumed would get rich supplying the data centres — are the two worst-performing groups in the basket. Meta's Muse moment (1.5 million iOS downloads in its first two weeks, 642k daily actives, per Sensor Tower) has lifted the stock 35% in a month, yet Microsoft is up just 3.5% for the year and Oracle is down 26%. Meanwhile the quiet companies making the atoms — memory dies, hard-drive platters, server boxes — printed four consecutive triple-digit years-of-favor.

Two Sessions Moved the Whole Tape

Yesterday's close was the last one before the summit, and it added a second data point to yesterday's CPU story: the categories that ran hardest in the prior week are still running. Memory's median YTD crept from +242% to +255% in two trading days. ARM's one-week gain expanded to +36%; Intel, +21%. The two-day delta is small in absolute terms, but the direction is unanimous — every category that was green on September 21 is greener on the 23rd. The market's pre-summit position is not defensive; it leans into the trade completing, not into it breaking.

The one thing that did not move: the energy names. Constellation Energy at -25% YTD is now the second-worst stock in the entire basket, and Vistra is down 14.5%. The thesis that "AI needs power, buy power" has quietly broken on the tape, for reasons the PJM capacity-auction math suggests run the other way for incumbents — costs rising faster than rate-base relief, moratoriums in Maine, ratcheting efficiency rules in Virginia. The market bought the picks and shovels of compute. It sold the utilities until the contracts are signed.

The Valuation Question Nobody Asks the Banquet Guests

If you want a single number for how much optimism the market has already paid itself, look at where each summit-sensitive name trades relative to its own ten-year history. Price-to-sales percentile is a crude tool, but it is honest:

Three of the names in the play are at or above the 99th percentile of their own decade-long price-to-sales range: AMD (100), TSMC (100), Micron (99), Intel (99). Dell sits at 100. Sandisk is off the scale entirely — no percentile even computes because its 1-year move of +1,607% has blown through every reference point in its trading history.

Contrast Nvidia: +21% YTD, trading at the 8th percentile of its own P/S range. The company at the center of the summit's chip agenda — the H200 licensing regime, the $5.6 billion Inspur subsidiary loophole, the distillation accusations — is the cheapest major name in the basket relative to its own history. The market has decided Nvidia's growth is already in the price of everything else, and that the chip king's own multiple has already had its party.

That asymmetry is the actual summit trade. If the hotline gets announced and chips stay rationed, the names at the 99th percentile need perfection to hold their prices — any disappointment in export-control relief is a valuation air pocket. If the summit collapses, those same names have the most to lose. Nvidia, at the 8th percentile, is the only large position in the basket where being wrong about the summit is cheap.

What the Hotline Is Worth

Let's be clear-eyed about the deliverables. A US-China AI notification hotline is not arms control — during the Cold War, the Washington-Moscow link went up after near-catastrophes, not before them. What it is, is a floor. Jonathan Czin of Brookings put it plainly: expect "some extension of the gentleman's agreement" and nothing more. China's Foreign Ministry spent September 14 publicly rejecting Dario Amodei's pacing-plan framing as "fearmongering," and Beijing's actual ask at the table is the thing Washington least wants to give: relief on advanced chip sales.

Note who showed up and who didn't: no Chinese CEOs in Xi's delegation, while the leaders of the major US AI labs are banquet guests. The symbolism writes itself — Washington is treating AI as a domestic industrial champion conversation with a foreign-policy wrapper, and Beijing is refusing to legitimise the framing by sending its own champions.

For a mid-market business buying AI services, the practical read from the data is narrower than the headlines:

  • The truce extension to January 10 buys one quarter, not a year. Put "what happens if export-control enforcement tightens on January 11" into your vendor questionnaire — specifically: does your AI vendor run inference on Alibaba Cloud, ByteDance-adjacent, or Inspur-adjacent hardware that could be caught by a subsidiary-loopback crackdown.
  • Memory and hardware concentration is real. The 99th-percentile names are also the physically irreplaceable ones. If TSMC or Micron pricing power survives a summit disappointment, that cost flows through to your GPU-hour bill in 2027.
  • Cheapest doesn't mean best-positioned. Nvidia's 8th-percentile valuation is a market verdict on known risk (China revenue capped, loopholes exploited, custom silicon competition). You can profit from that without pretending the risk isn't in the number.

The Bottom Line

The summit will produce a hotline, a truce extension, and photographs. The market produced its verdict before the appetizers: it bought the atoms of AI — memory, chips, servers — up 190% to 260% median, sold the utilities that were supposed to feed them, and left Nvidia, the actual subject of the negotiations, cheaper than it has been against its own history in a decade. Diplomacy moves in communiqués. Capital moves first, and tells you what it believes while the diplomats are still agreeing on the wording.

Data: ChartRow AI-stocks basket (36 names), closes through 2026-09-23; category medians computed from the same snapshot. S&P 500 total return ~+12.3% YTD through Sep 18, 2026 (ChartRow SPY series). Summit facts: AFP/Japan Times/RTE reporting Sep 23-24, 2026; Bessent truce extension to Jan 10; AI hotline proposal per Bessent-He preliminary talks (Sep 20-23). Not investment advice.