The Signal — November 16, 2026
Washington and Silicon Valley spent the same day pulling on opposite ends of Nvidia. Treasury Secretary Scott Bessent went on Fox News and said the board needs a “dramatic shakeup,” should hire a chief executive from outside the company, and owes answers inside ten days — adding that if Nvidia were a bank and had not complied, “a lot of the executives would be removed.” The Financial Times then reported the counter-campaign: Jensen Huang talking with confidants about a return, an ally assembling a “war room” of Silicon Valley executives to lobby the board and the White House, and a petition past 150,000 signatures carrying Elon Musk, Tim Sweeney, Travis Kalanick and Steve Chen. Underneath the personality fight, the thing that actually compounds moved: ten governments signed a declaration committing them to draw up a shared list of AI technologies and supply chains to be protected from China, and Stanford's AI Index editor published the two numbers that define this build-out — more than $1 trillion of global corporate AI investment so far in 2026, 46% above the whole of 2025, against 60% of American adults who say they would be uncomfortable with a new data center in their area. No lab shipped anything. The capital is arriving faster than the consent, and closing that gap is the most underpriced job in the stack.
The Tide
One confirmation, no shift. All four tides hold. governance-as-market-structure took the day's only structural evidence, and it arrived as a signed declaration rather than a rule — which is what market structure looks like when it is being built rather than imposed.
The chip regime acquired a membership list
The State Department published on Sunday that ten governments signed the San Francisco Declaration on the Secure Foundations of Enduring Prosperity — the United States, Australia, India, Israel, Japan, the Republic of Korea, Qatar, Singapore, the United Arab Emirates and the United Kingdom. The declaration's own text resolves to establish working groups, five of them proposed, and defines “sensitive technologies” broadly as the goods, software, materials, knowledge, components and inputs used to produce them. The sharper detail is the Financial Times': Demetri Sevastopulo reported that the signatories have agreed to draw up a joint list of AI-important technologies and supply chains needing extra protection to counter China, spanning chips, critical minerals and energy. That list is the FT's reporting rather than the published document, and it has not yet been written. Germany, Sweden, the Netherlands and the European Union are expected to sign shortly, though that too is an expectation relayed to the FT rather than a confirmed signature — and all four are already parties to the Pax Silica Declaration, signed in Washington in December 2025 by seven countries and since grown to roughly 25, so this is a new instrument inside an existing coalition rather than a first accession, which several aggregators are already conflating. Read it against the statutory track: the Senate's version of the defense authorisation carries three provisions pulling export discretion back from the executive — the AI OVERWATCH Act, the Chip Security Act and the MATCH Act. Anthropic has publicly urged their passage and OpenAI has now backed them too, both publicly and in direct contact with congressional offices, though reports differ on how long it held that position back. None of the three is law yet; the conference outcome is unresolved.
So what: A list is the most useful thing a regulator can give a builder, because you can read it. When it publishes, work out which side of it each of your critical inputs sits on — chips, minerals, power equipment, and the vendors behind them. Inputs on the list get allied protection and, with it, a procurement moat worth having; inputs off it are commodities you will be bidding for against everyone. Ten signatories with four more expected is also the first version of this regime a company could plan a decade against, which is a better operating environment than the deal-by-deal discretion it replaces.
Sources: US, allies agree to protect AI supply chains from China, FT says · MTS — Week In Review 11/15: The Fallen King · The San Francisco Declaration on the Secure Foundations of Enduring Prosperity
The Waves
The fight over who runs Nvidia became a public contest between the administration and the Valley
Treasury Secretary Scott Bessent appeared on Fox News' Sunday Morning Futures and went considerably further than an official usually goes about a private board. “They needed a dramatic shakeup,” he said; “this makes me think that as the board conducts a search for a new leader that they need to go outside the company”; he pressed repeatedly on what the culture at Nvidia had been and on whether the company had pushed illicit sales to goose revenue; and “I think we need answers in the next 10 days from the board and from the outside counsel.” Asked why Megaspeed had not been blacklisted, he redirected to whether Nvidia had a culture of compliance around US export restrictions, and said that if Nvidia were a bank that had not complied, the Office of the Comptroller of the Currency would be involved and “a lot of the executives would be removed.” On Friday, on CNBC, he had said Treasury, War, USTR and State had favoured putting Megaspeed on the entity list for months — “But Commerce has the pen.” The counter-campaign surfaced the same day: the Financial Times reported Huang has discussed a return with confidants, and that one ally is assembling a “war room” of Silicon Valley executives to lobby the board and seek White House support, on the theory that “everyone that matters in tech outside the frontier labs loves Jensen.” The standwithjensen.com campaign had passed 150,000 signatures by Sunday evening, including Musk, Epic's Tim Sweeney, former Uber chief Travis Kalanick and YouTube co-founder Steve Chen; tens of thousands of signers self-report as shareholders or current and former employees, which the site notes is unverified. The board, meanwhile, approached Arm chief executive Rene Haas about the job, per an FT report on Friday. Nvidia closed Friday at $162.85, down about 13% for the week (12.75%) and off roughly 32% from its record close of $239.24 on 6 October, and HSBC's Frank Lee moved the stock to hold on Friday, cutting his target to $170 on the view that it will struggle until a permanent chief executive is named and on export-rule risk to roughly 11% of revenue.
Roadmap implication: Tuesday's third-quarter call is the read, and it is a governance call disguised as an earnings call: interim chief executive Colette Kress makes her first public appearance in charge, and the question that matters is not the revenue line but which commitments made under Huang were institutional and which were his phone calls — the neocloud investments, the supply allocations, the sovereign deals. Put that question to Nvidia directly, and to anyone whose financing leans on it. The constructive read is that a board that now has a first-ever chair in Harvey C. Jones, outside counsel and a public deadline is a better counterparty across a ten-year build-out than one that had none of the three, and a supplier whose governance can be examined is one you can actually underwrite.
Sources: Treasury Secretary Scott Bessent says if Nvidia was a bank, 'a lot of the executives would be removed' — while Huang reportedly eyes return · Jensen Huang weighs Nvidia CEO comeback as ally plans 'war room': report
A trillion dollars of AI capital, and a country that has not agreed to host it
Sha Sajadieh, who edits Stanford's AI Index, published a Sunday essay previewing the next edition and it puts two numbers side by side that belong there. Global corporate AI investment has surpassed $1 trillion so far in 2026, exceeding the entire 2025 total by 46%. In the United States, 60% of adults say they would feel uncomfortable with a new data center in their area — and that includes majorities in both political parties, which is rare enough to be worth pausing on. On labour, the essay reports that in every country surveyed more people expect AI to eliminate jobs than create them, even though a majority globally still believe the benefits outweigh the drawbacks; and that employment declines among younger workers in AI-exposed occupations have continued into 2026, worst in customer service, and have now deepened even in occupations where AI supports rather than replaces human work, naming marketing and IT managers. The piece is careful to say these patterns do not establish causality. The full AI Index lands in April 2027.
Roadmap implication: Treat siting as a political input to your capacity plan rather than a real-estate line item, because bipartisan 60% discomfort means the binding constraint on the next gigawatt is consent, not capital — and capital is the thing you already have. That is also the opening, and it is a large one: whoever makes a data center something a town actively wants, through power it brings, revenue sharing, water discipline and jobs that are legible locally, gets permits that competitors cannot buy. On the labour finding, the honest planning move is to stop assuming assistive deployment is politically free; it is showing up in the same employment data as replacement, and saying so out loud is cheaper than being found out.
Sources: The Jagged Edge of the AI Boom
China is behind on the chip and ahead on the plug
The Information's Claudia Chong and Qianer Liu reported on Sunday from Ulanqab, in Inner Mongolia, where the renewable-energy firm Envision has started using an 800-volt direct-current system to power live computing workloads at its Galaxy campus — a deployment the company announced earlier this month and says is the first running in a functioning data center. Solid-state transformers convert 10,000-volt alternating current from the grid straight to an 800-volt DC bus, lifting conversion efficiency to 98.5% from a typical 94% to 95% — Envision senior vice president Kevin Huang's figures, given in an interview, so read them as the vendor's. Nvidia and its partners are still preparing the same architecture for next year's hardware. The reason this matters is rack power. The Information's figures: Blackwell racks require around 140 kilowatts, Vera Rubin NVL72 up to 227, and Nvidia's Kyber racks, due from 2027, about 600 — note that public spec sheets put Blackwell nearer 120 to 130, so treat 140 as the reported rather than the rated number, and Kyber as roadmap. Chinese operators are already at 400 to 500 kilowatts a rack because their chips do less individually, so more get linked into a single machine — ByteDance showed a rack needing up to 500 kilowatts in July. The first commercial deployment is a modest 5-megawatt pilot, per The Information, the system costs 10% to 15% more up front, and the company expects component costs to fall 20% to 25% as it scales over 18 months. The strategic detail is Huang's expectation of two power architectures, one for China and one following Nvidia's international design — the split in chips reproduced in the power layer — and Envision is already discussing 800-volt pilots with customers in Europe, the Middle East and Asia-Pacific.
Roadmap implication: If you are underwriting capacity for 2028, the 800-volt transition is a hard fork in the facility spec and not a refit you can defer: a hall built for today's conversion chain will not take a 600-kilowatt rack without rebuilding the power path. Ask your colocation provider now which architecture their next phase is designed to, and get it in writing. The export angle is the one most people are missing and it is an opportunity as much as a warning — power systems are not under the controls that chips are, so Chinese vendors can sell this into Europe and the Gulf while their chips cannot go the other way. And the 94%-to-98.5% delta is the shape of the whole opportunity: the gap between watts bought and work delivered is where the next layer of margin lives.
Sources: How a Chinese Energy Company Is Outpacing the U.S. on AI Power Technology · Envision Deploys SST-Based 800V DC Power at Ulanqab AI Data Center
The Ripples
Anthropic prices Wednesday and trades Thursday
Yahoo Finance's Sunday week-ahead, by Jake Conley and Ethan Wolff-Mann, lays out the calendar: Anthropic prices its offering on Wednesday 18 November and begins trading on the Nasdaq under the ticker ANT on Thursday 19 November. The company launched its roadshow on 10 November offering 825 million Class A shares at $70 to $78, for an implied valuation of about $1.4 trillion at the top of the range on a fully diluted basis — $1.404 trillion as reported at the time — down from earlier projections that ran as high as $2 trillion before the recent AI-sector turmoil. The Information's Martin Peers wrote on Sunday evening that demand is intense enough to price above the range, guessing $84 to $85 for roughly $1.5 trillion, and that the stock opens “well above $100 a share, possibly above $150”; that is his estimate, not a book. The filings are the part worth reading: $33.2 billion of revenue for the nine months to 30 September, and compute commitments of $895.4 billion. Google and Nvidia have each committed about $10 billion to a concurrent $26 billion private placement of non-voting Class N shares, with AMD at roughly $2.5 billion, Broadcom, Micron and Microsoft at about $1 billion each and TSMC around $500 million — and Bloomberg reported on 12 November that Nvidia's pledge came under fresh scrutiny after Huang's exit.
Do this now: Before Thursday, write down what you think this company is worth and what the first print would have to be for you to be a buyer, then do not move the number once the tape starts — opening prints on a listing this heavily anticipated are the worst possible place to form a view. If Anthropic is in your stack, the line to track is the $895.4 billion of compute commitments rather than the valuation, because that is the number that decides whether your rate card holds in 2028.
Sources: Nvidia's first post-Jensen earnings and Anthropic's IPO: What to watch this week · Anthropic launches roadshow for proposed initial public offering
Pony.ai is training on Chinese silicon because it can actually get it
In TechCrunch Mobility on Sunday, Kirsten Korosec published an interview with Pony.ai chief financial officer Leo Wang conducted during the week. Wang said demand for AI compute far exceeds supply and is a genuine pain point for the robotaxi company, that Pony.ai runs training on midtier rather than high-end hardware wherever it can, and that it is turning to domestically produced Chinese chips for AI training. On how widespread that is: “everyone, I would say, is also practicing kind of like a Plan B for a few years.” And on the trade-off: “although cost-effectiveness-wise, probably it's not the best choice yet.”
Do this now: If any part of your China roadmap assumes continued access to Nvidia parts, re-plan it this quarter on domestic silicon and price the performance gap honestly rather than hoping a carve-out arrives. The sentence that actually matters is “not the best choice yet” — substitution is happening on availability rather than price, which means it does not reverse when price catches up, it accelerates. For anyone selling tooling, inference software or evaluation into China, that is a market being created by constraint.
Sources: TechCrunch Mobility: The Nvidia backup plan for China's self-driving industry
Britain started paying for AI; its agents still are not buying
Azeem Azhar's Sunday edition of Exponential View carries two numbers that belong next to each other. One in twenty UK adults now pays for AI, spending more than $1.3 billion a year — up from 1% two years ago and 2.5% a year ago, concentrated in London and skewed to tech, finance, education and PR, with Claude's share quadrupling over the year. Against that adoption curve, Uber says four rides out of 320 million booked during Halloween week came from agents. Azhar's own framing is the useful part: his agent secured him a hard-to-get phone — the kind of job, he notes, that means refreshing a page every few minutes from 7am — and he would still open the app to book a car.
Do this now: Stop pricing agentic commerce off demos and price it off four-in-320-million, then go find the tasks where an agent is obviously better than a human with an app — patient monitoring of scarce inventory, finicky multi-constraint product searches, anything that rewards checking a page a hundred times. That is a short list today and it is the correct beachhead; the generic “agent books your Uber” demo is competing with a button that already works. Meanwhile a payment curve that went 1% to 2.5% to 5% in two years is the demand side of the cost-collapse tide showing up in consumer wallets, not just enterprise contracts.
Sources: 🔮 The $22 billion question about AI self-improvement #610
Read this edition and the full archive at excelsiorgroup.ai/insights/signal
The Signal — The Excelsior Group