Excelsior group logo

The Signal - daily AI evolution

Archives
Log in
Subscribe
November 13, 2026

The Signal — November 13, 2026

The most important relationship in the AI build-out was never Nvidia's. It was one man's. Nvidia's board accepted Jensen Huang's resignation as president, chief executive and director, effective immediately, and the company disclosed it to the SEC at 8:52 a.m. Eastern on Thursday, before the open; the filing gives no reason, and the affair reporting that preceded it is press, not filing. What leaves with him is not a product roadmap but a set of personal commitments: memory locked up in Korea, wafers locked up in Taiwan, $530 billion of guarantees that let neoclouds borrow at investment-grade pricing, and the industry's hardest-fought argument against cutting China off. AMD introduced its 288-gigabyte MI450X and launched its 256-core Venice EPYC the same morning, which is a coincidence of calendar and also a coincidence of substance. The day-zero read is unglamorous and large: when the allocator of a scarce input is replaced by a committee, the input gets reallocated, and it goes to whoever already built something to put it in.

Read the edition and the archive: https://excelsiorgroup.ai/insights/signal/


🌊 TIDE

Two confirmations, no shift. All four tides hold. governance-as-market-structure and cost-collapse both took on hard evidence, and in the first case the evidence is a personnel filing rather than a rule.

The chip industry's hardest-fought opponent of export controls is gone, and the bill he lobbied against is moving in the Senate

governance-as-market-structure — confirmed

Nvidia's 8-K says the board accepted Huang's resignation as president, chief executive and director, effective immediately, giving Wednesday as the date of the event; the press release says only that he has resigned, and neither document gives a reason. Colette Kress, 59, the CFO since 2013, becomes interim CEO while keeping the finance job; Harvey C. Jones, 73, a director since 1993 and a former Synopsys chief executive, becomes the board's first-ever chair, and the search covers internal and external candidates. Third-quarter results land Tuesday the 17th. The reason the day is a tide confirmation rather than a governance story sits one layer down. SemiAnalysis, whose Accelerator Model tracks where every AI chip ends up, puts more than 1.5GW of restricted Nvidia silicon inside China, says Commerce now accepts that figure too, and argues the four-day exit is tied to the smuggling record as much as to conduct. Its stated consequence is the Remote Access Security Act, one of several chip bills now moving, which passed the House 369-22 in January, has sat without a Senate Banking vote while Nvidia and Oracle lobbied against it "hard," and is now, in SemiAnalysis's words, a vote "happening right now." Speaker Mike Johnson told Reuters there is "a developing consensus in Congress" for limiting chip exports, "especially in the light of recent developments." OpenAI is backing the bipartisan push, Bloomberg reports. Bill Bishop's read the same day: "Access to Nvidia chips, both through offshore clouds and smuggling into the mainland, may soon get much more difficult for the PRC AI labs." Beijing answered that morning through MOFA spokesperson Mao Ning, who said the US "should stop fabricating the 'Chinese spy' narratives as a pretext for imposing technology restrictions" — Sinocism's rendering of the briefing; other outlets translate the phrase in the singular. Two things to carry lightly: Attorney General Todd Blanche, at a press conference Thursday, said per Reuters that the Justice Department has not yet reached a conclusion on whether Alice Huang was a Chinese agent, while confirming the FBI is investigating, and SemiAnalysis itself argues against RASA's shape, on the view that remote access is the regime that keeps profits, visibility and an instant off-switch in American hands.

So what: Compute access is now set in Washington on a legislative clock, not in Santa Clara on a product clock, and the single most effective brake on that clock has been removed. Treat a Senate Banking vote on RASA as a dated event you plan around rather than a tail risk. The opening is real: every customer whose roadmap assumed cheap offshore rental of frontier chips now needs a compliant alternative, and the firms that can prove where their inference physically runs and who can reach it are selling something that did not have a price last week.

Sources: NVIDIA Announces CEO Transition · Jensen Huang resigns as Nvidia chief executive · Nvidia and the World Without Jensen · 160th anniversary of Sun Yat-sen's birth; US-China; Nvidia fallout; Social security

The price of a finished task is still falling five to ten times a year while the frontier gets more expensive to move

cost-collapse — confirmed

Micah Hill-Smith, chief executive of the benchmarking firm Artificial Analysis, gave the cleanest pair of numbers of the day on a subscriber call reported by The Information. The cost of doing a given task at a given quality level is falling five to ten times a year. In the same breath: AI capability improvement per dollar of investment "will keep shrinking," which is why labs need so much capital, and measurement gets awkward when a company spends ten times more to produce a model twice as good. He called a fast takeoff, where one lab's models self-improve faster than rivals can follow, "pretty unlikely," put the closed-source lead at a few months, "arguably Anthropic this week," and said companies should keep investing in open-weight models because "the ecosystem will be there." His own firm's tests of long-running agents now run into hundreds of dollars per task. Every Hill-Smith figure here sits behind The Information's paywall, read in the edition delivered to this desk; no independent corroboration of the five-to-ten-times number was reachable. Both halves belong to the same tide: the unit price of delivered work collapses while the price of buying the next increment of frontier capability climbs. The commercial consequence is that the frontier becomes a capital business and the application layer becomes a deflation business.

So what: If the cost of a completed task is falling five to ten times a year, anything you priced as uneconomic twelve months ago is worth re-pricing now rather than next planning cycle, and the gap between frontier cost and task cost is where the next decade of margin lives. Build the re-pricing into the calendar: keep a standing list of the jobs you declined on unit economics and run it again quarterly. The roadmap implication for the frontier labs is harsher and is the reason the IPO window matters more than the benchmark table.

Sources: Free Chinese AI Can Hack Companies While Frontier Developers Face Diminishing Returns


🌊 WAVE

Nvidia's supply-chain stranglehold was one man's phone calls, and AMD launched its alternative the same morning

wave: inference-infra

SemiAnalysis's argument about what actually leaves with Huang is a supply argument, not a sentiment one. Nvidia's supply and capacity commitments jumped from $119 billion to $279 billion in a single quarter, mostly for memory, after Huang personally locked up DRAM in South Korea and wafers in Taiwan, CEO to CEO. SemiAnalysis estimates SK Hynix's 2027 HBM4 could price at roughly $3.00 to $3.30 per gigabit for Nvidia against $3.70 to $4.10 for other tier-one buyers including AMD, Broadcom, Amazon and Google, with an implicit understanding that SK Group gets priority GPU allocation in return. Remove the man who won those terms and the firm's view is that more than 70% of total capacity could be non-Nvidia by 2028. Internally the same structure is the succession problem: roughly 60 direct reports, about 42,000 employees, no published succession plan, and the only internal technical names SemiAnalysis floats are Ian Buck, who built CUDA out of his doctoral work, and Jonah Alben, who runs GPU engineering — and the firm doubts Buck could run hardware engineering and doubts either man could replicate Huang's supply-chain mastery. Its own bet is Buck as front-runner, with Kress second. AMD, hours later, introduced the Instinct MI450X at 288GB of HBM4 per GPU alongside the MI455X. For the MI455X running on its Helios rack it claims 21,333 tokens per second per GPU on DeepSeek-R1 in an MLPerf Inference Offline configuration, up to 30% above the latest published MLPerf offline result for Nvidia's Vera Rubin, with Helios carrying up to 20% more HBM bandwidth on published specs. Those are AMD's own October tests, not independent ones, and AMD disclosed no FLOPS and no ship dates. The same day it launched a 256-core, 512-thread Zen 6 Venice EPYC on SP7 with 16 channels of DDR5-8000, which SemiAccurate notes was a real launch where the companion Verano part was "only half launched."

Roadmap implication: The roadmap implication is that chip supply stops being a relationship and starts being a market, which is the first genuinely good news for everyone who is not Nvidia in about three years. Memory and advanced packaging were the binding constraint on every competing accelerator, and the price differential SemiAnalysis describes was the mechanism. If it narrows, AMD, Trainium, TPU and the in-house XPUs all get cheaper to build at volume, and the practical move is to stop treating your 2027 accelerator decision as settled and start asking vendors for allocation commitments now, while a reshuffle is in progress and new suppliers are motivated to win reference customers.

Sources: Nvidia and the World Without Jensen · AMD Advances the Instinct MI400 Series for AI · AMD launches Zen 6 Venice Based Epyc CPUs · Jensen Huang's self-sabotage sends Nvidia into a midlife crisis

Half a trillion dollars of Nvidia guarantees were the build-out's credit line, and a professional manager does not renew those at that size

wave: compute-financialization

The least-discussed number in the Huang story is a balance-sheet one. Nvidia disclosed $530 billion of gross off-balance-sheet guarantees as of its fiscal second quarter, up from $184 billion the prior quarter, including $36 billion of take-or-pay floors under neocloud capacity, and it holds a $99 billion equity book. SemiAnalysis describes the mechanism plainly: Nvidia sells GPUs at full margin on day one, then floors the buyer's revenue, guarantees the landlord or signs the lease itself, so operators without an investment-grade rating borrow at investment-grade pricing. In SemiAnalysis's description of that mechanism, neoclouds raised money because Huang backed them rather than because they compared chips. The Information reports Nvidia's dealmaking totalled $140 billion including credit guarantees in just two months this year, a figure behind its paywall; Anthropic's own amended registration statement is the firmer number on the Nvidia side, disclosing NVIDIA Corporation buying 135,135,135 Class N shares for approximately $10.0 billion in the concurrent private placement rather than in the IPO itself. Devansh at AI Made Simple separately puts up to $105 billion of contingent guarantees against OpenAI-linked infrastructure, a figure he traces to Nvidia's August 10-Q covering land, power and building shells in Ohio phasing in from fiscal 2029; neither source says whether it sits inside SemiAnalysis's $530 billion or alongside it. SemiAnalysis's forecast is that a manager-led Nvidia runs the only playbook that has worked at this scale, which is Apple after Jobs: more conservative, return the cash, buy back stock. Shareholders do fine. The ecosystem that was borrowing against Nvidia's credit does not.

Roadmap implication: If you are a neocloud, an inference provider or a lab whose financing assumed a Nvidia backstop, the roadmap implication is to re-underwrite that assumption this quarter rather than at renewal, because the signature that made it bankable has left the building and the replacement answers to a board. For everyone buying compute rather than financing it, the same shift is an opening: capacity that was committed on the strength of a guarantee may come loose, and loose capacity is cheap capacity. Ask your providers which of their contracted floors survive a change of guarantor.

Sources: Nvidia and the World Without Jensen · Nvidia's Dealmaking Binge Faces a Reckoning After Jensen Huang's Exit · People Don't Understand Nvidia

The open-weight ecosystem just lost its largest corporate patron, and the loss runs through kernel libraries nobody reads

wave: open-weight-credibility

This is the consequence almost nobody led with and it is the one with the longest tail. SemiAnalysis calls it an open secret that Nvidia's DevTech group works directly with the Chinese frontier labs on open source and on optimising models for Blackwell GPUs those labs are not supposed to have. The output of that collaboration is not press releases; it is wide expert parallelism and the open kernel libraries that the Western inference stack now depends on, MiniMax's MSA, DeepSeek's DeepEP and DeepGEMM, Moonshot's FlashKDA, all of which ship inside vLLM and SGLang and therefore inside a large share of everyone's serving costs. SemiAnalysis's read is that with Huang gone and the China hawks ascendant, Chinese labs go back to denying they hold Nvidia silicon and keep their Blackwell software internal, which is good for closed labs and bad for startups. On the American side the same departure removes the patron: Nvidia was, in SemiAnalysis's telling, single-handedly propping up US open models through investments in Reflection, Thinking Machines and Poolside and its acquisition of Hugging Face, and a manager-led company could quietly deprioritise the underperforming Nemotron line. One piece of technical evidence arrived the same day and cuts the other way on deniability: a forensic study by Unemployment Labs, carried in Sinocism, finds that six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns matching Blackwell hardware where the same companies' earlier models match Hopper. US officials had alleged this without publishing technical evidence; this report supplies some, and is careful to say the fingerprint supports but does not prove training on Blackwell.

Roadmap implication: The roadmap implication is that the cost curve of self-hosted open-weight inference has a dependency you have never audited, and it is a geopolitical one. Find out which kernel libraries your serving stack actually compiles in, because a slower cadence of Chinese open-source contribution shows up as a quiet rise in your tokens-per-dollar twelve to eighteen months from now, not as a headline. The opening is that an under-patronised open-weight ecosystem is an under-priced one: the teams that invest in their own kernel and serving expertise now buy a durable cost advantage over everyone who has been free-riding on a collaboration that was never guaranteed to last.

Sources: Nvidia and the World Without Jensen · 160th anniversary of Sun Yat-sen's birth; US-China; Nvidia fallout; Social security

Free Chinese open weights are now good enough to break into companies, and fixing is harder than finding

wave: ai-security-refounding

Ariel Zeitlin, co-founder and chief technology officer of the Israeli continuous-penetration-testing firm Tenzai, told The Information's subscribers that freely downloadable Chinese open-weight models have become good enough to break into corporate networks. "No one is restricting access to" them, he said. "I think this is the real threat for organizations," and "You don't need to be a skilled hacker to do that." He pointed, without naming it, to a free agentic hacking framework already used in real breaches, which matches ARTEX, covered here on the 6th, and which CrowdStrike says a China-linked group ran against more than 200 organisations in 11 days. His forecast is a rough stretch of months to years with breaches disclosed daily, mostly from software misconfigurations that AI makes trivially easy to find, including at the largest firms he works with that employ thousands of security staff. His line is the one to keep: "Fixing is harder than finding." He reads recent security lapses at OpenAI and Anthropic as startup habits rather than frontier-lab exceptionalism, and notes big banks are better prepared. His longer-run view is optimistic: AI-written code will eliminate whole bug classes such as injection flaws, just not the ones that arise where agents and applications meet. Zeitlin sells the remedy, which is worth stating once, and the direction of travel is corroborated by what the frontier labs shipped the same day.

Roadmap implication: The roadmap implication is that the asymmetry has inverted: discovery is now commodity and cheap, remediation is still bespoke and expensive, so security spend that is still weighted toward finding is mis-weighted. Fund the patch pipeline, not another scanner. Concretely, the firms that will win the next two years of security budget are the ones selling verified remediation throughput rather than vulnerability counts, which is exactly the product OpenAI launched into on Thursday and exactly where the day's security financings went.

Sources: Free Chinese AI Can Hack Companies While Frontier Developers Face Diminishing Returns · OpenAI launches security products, tells orgs they're doomed if they don't use them


🌊 RIPPLE

Google's cheapest Gemini 4 is flat-priced across the whole million-token window, and the discount triples on December 1

Gemini 4 Hydrogen shipped Thursday as the low-cost member of the Gemini 4 family: one million tokens in, up to one million out, $1.50 per million input and $7.50 per million output, $0.075 cached, and crucially flat across the full context window rather than stepping up past a threshold. There is a new xhigh thinking tier above high. Google's own figures put it at 51 on the Artificial Analysis Intelligence Index at high effort against 50 for Gemini 4 Argon, 54 at xhigh, and 57% on Terminal-Bench against 18.2% for Gemini 3.8 Flash, roughly 2.5 times faster than that model at default effort. It is the new free-tier model, and it is wired into Google Search for all users globally; Google's own announcement is bylined by Koray Kavukcuoglu, SVP of Google DeepMind and the company's chief AI architect. Simon Willison found the interesting wrinkle: Carbon and Argon are discounted to $2/$10 until November 30 and double on December 1, so Hydrogen is only about 25% cheaper than its siblings today and substantially cheaper after the increase — he writes 75%, though his own table's $1.50 against $4.00 works out closer to 62%. On his pelican test, Hydrogen took 19 seconds and about four cents at high effort where Carbon took over four minutes and about twenty cents, and produced visibly simpler output. Benchmarks here are Google's own and the index number is a one-point move.

Do this now: Do this now: if you are running anything high-volume on Carbon or Argon, price the same workload on Hydrogen against both the current and the December 1 rates before you renew, because the arbitrage roughly triples in two and a half weeks and the flat context pricing is worth more than the headline rate to anyone stuffing long documents into prompts. Run your own eval rather than the index; a one-point index move with a thirteen-times latency gap is a latency product, not an intelligence product.

Sources: Introducing Gemini 4 Hydrogen · Release: llm-gemini 0.37

OpenAI put agents on both sides of the vulnerability: one finds them, one writes the patch

At its Cyber at the Frontier event on Thursday, OpenAI added two products to its Daybreak platform. Daybreak Red Team, open by early-access request, runs multi-agent offensive security that uses source code and application context to investigate attack paths, confirms vulnerabilities, and triages findings that arrive from code scanners, bug bounties and disclosure programmes; the agents run in isolated environments under a "Guardian" review that bounds which apps and endpoints they may reach. Nancy Wang, who leads cyber-applied AI, described it as a "superhuman pen-test at the scale of automated continuous scanning." Daybreak Defense Factory is the other half: built on Codex Security and shipped as a preview plugin in ChatGPT and Codex for workspaces with Daybreak access, it traces a vulnerability through the codebase, drafts a patch, imports vulnerability data from Wiz, Snyk and Endor Labs, scans connected GitHub repos and monitors new commits, and opens draft pull requests that engineers must review and test; nothing deploys automatically. President Greg Brockman's framing was that companies will be secure only to the extent of the frontier model they use for defence, which is a sales line and also a real claim about who can afford defence. The Deep View reports OpenAI gave no pricing information, and no benchmark figures appear in any of the three write-ups. GPT-6 Cyber, which OpenAI released the prior week, is the model underneath; Zvi Mowshowitz records OpenAI's own numbers for it as 1,724 vulnerabilities reproduced, working exploits for 948, more than five times the 166 managed by GPT-5.6 Sol on the same harness, with 304 critical findings disclosed to maintainers with candidate patches.

Do this now: Do this now: request Red Team access and point it at one real internal application rather than a test target, because the thing worth measuring is not whether it finds bugs but what fraction of its findings survive your own triage. Then measure the other half separately, since Defense Factory's draft pull requests are the first product that competes with your engineers' time rather than your scanner budget, and the merge-acceptance rate is the only number that matters. Note that openai.com is unfetchable to this brief's tooling, so these details come from The Register, SecurityWeek and The Deep View rather than the vendor pages.

Sources: OpenAI launches security products, tells orgs they're doomed if they don't use them · OpenAI Unveils Daybreak Red Team, Defense Factory Security Tools · OpenAI sees a new enterprise frontier in cyber · AI #194: Spy Versus Guy

Four nonprofits asked a California court to bar five labs from shipping anything more capable than what they already have

A coalition of Irreplaceable, Public Citizen, the Heat Initiative and the Alliance for Secure AI filed suit Thursday in San Francisco County Superior Court against Anthropic, Google, Meta, OpenAI and SpaceX's AI arm. The vehicle is California's Unfair Competition Law, with the argument that the defendants' conduct is inconsistent with SB 53, the state's catastrophic-AI-risk statute. The relief sought is the striking part: an injunction barring the development or deployment of models more capable than each defendant's most advanced current system, plus a bar on AI programmes designed to autonomously self-improve. Alliance for Secure AI chief executive Brendan Steinhauser: "We hold every other industry accountable when its products threaten public safety, and AI companies should be no different." A state unfair-competition claim seeking a nationwide capability freeze is an aggressive theory and the odds of that specific relief are long, but the filing is a dated artefact of how fast private enforcement is arriving behind SB 53.

Do this now: Do this now: nothing operationally, but add it to the legal-risk register with the right framing. The signal is not the injunction, which will probably fail; it is that SB 53 is now a hook for third-party suits brought by people with no commercial relationship to the defendant, which changes who can sue you over a model release. If you deploy frontier models in California, have counsel read the complaint for the theory rather than the ask.

Sources: Nonprofits sue Anthropic, Google, and other tech giants to freeze the AI arms race · OpenAI launches security products, tells orgs they're doomed if they don't use them

Anthropic's book is about six times oversubscribed, and the headline raise depends entirely on which number you mean

Bloomberg reported Thursday, with Reuters picking it up, that Anthropic's IPO has drawn roughly six times the orders for the shares available, and that bankers have told investors the company will prioritise existing backers, which may leave some would-be buyers with smaller allocations than expected. The arithmetic underneath is where the public numbers diverge, and both circulating figures are right. The base Class A deal from the November 10 amended registration statement is 825,000,000 shares, 801,351,354 from the company and 23,648,646 from selling holders, at $70 to $78, plus a 123,750,000-share greenshoe; at the top of the range that is $64.35 billion gross. Separately there is a $26.0 billion concurrent private placement of 351,351,348 Class N shares at the IPO price to Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC. Add them and you get the $90.4 billion figure that several outlets print. On valuation, The Information reports, behind its paywall, a tentative range of $1.25 trillion to $1.4 trillion with several investors expecting it to price above, influenced at the bottom by the $1.25 trillion at which OpenAI raised on the 6th, and down from the $1.5 trillion-plus bankers had been floating. Morningstar's Malik Ahmed Khan published a $1.7 trillion fair value on November 10, with a bull case of slightly more than $3 trillion and a bear case of $500 billion. The ticker is ANT on Nasdaq, with pricing expected on the 18th and first trading on Thursday the 19th in New York; the Australian wrap's "20 November" is the same moment stated in Australian Eastern Daylight Time.

Do this now: Do this now: if you are quoting Anthropic's raise in a board deck, say which number you mean, because $64 billion and $90 billion differ only by whether you count the chipmakers' private placement, and getting it wrong in front of people who read the filing is an avoidable own goal. The structural point worth more than the valuation is the placement itself: Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC are all buying into the same model company, which tells you the supply chain has decided to own equity in demand rather than only sell into it.

Sources: Anthropic's IPO is about six times oversubscribed, Bloomberg News reports · Amendment No. 2 to Form S-1, Anthropic PBC · How Much Is Anthropic Really Worth?

The money kept moving while the chip story broke, and it went to defence, datacentre shells and proving AI paid for itself

Thursday's financings, and the two days before them, read like a thesis statement. UFORCE announced $700 million at a $6 billion valuation on Thursday, co-led by Valor Equity Partners and the UK National Wealth Fund with the US-Ukraine Reconstruction Investment Fund participating and Menlo Ventures, Xantium and Hiro Capital joining, to scale combat-proven systems whose NEMESIS Class 1 heavy strike UAVs the company says have logged more than 300,000 combat flights and been adopted by over 100 units, and whose MAGURA uncrewed surface vessels it credits with contributing to the destruction of more than a dozen Russian warships. Rox raised $100 million at $1.3 billion the same day, co-led by General Catalyst and March Capital with Databricks Ventures and Prosperity7 new, for sales agents that run on each customer's own data warehouse; its traction figures, including net revenue retention above 700% and more than 10x ARR growth in its first year, are company-reported and unaudited. And Nebuly raised $21 million led by Planven for the least glamorous product on the list: measuring whether enterprise AI agents actually produced a result, integrating with Claude-, ChatGPT- and Copilot-based assistants on-prem or in private cloud. Earlier in the week: Torus took a $500 million Series C led by Magnetar on Wednesday to build modular, shipping-container-sized AI datacentres; Edelbyte came out of stealth on Wednesday with $70 million total; and Operand raised $31 million led by Khosla Ventures on Tuesday.

Do this now: Do this now: if you cannot produce a number for what your AI deployment returned, assume your board will ask for one inside two quarters, and note that a venture market just priced that measurement gap at $21 million. Stand up the measurement before the next budget cycle rather than after it. The pattern across the three Thursday rounds is the day-zero one: the capital went to companies doing old work with new physics from scratch, not to companies bolting AI onto an existing product line.

Sources: UFORCE raises $700M at $6 billion valuation from Leading Tech and Sovereign Wealth Funds to scale combat-proven defence systems for Ukraine and Allied Nations · Sequoia-backed Rox raises $100M at $1.3B valuation: can AI agents unseat Salesforce in sales? · Nebuly raises $21M Series A to turn AI investments into measurable results


The Signal is a daily brief from The Excelsior Group. Old problems, new physics.

The Signal — Daily AI Brief | The Excelsior Group

The top things moving in AI today, read the way a trader reads a market: tide, wave, ripple.

Don't miss what's next. Subscribe to The Signal - daily AI evolution :
← Newer The Signal — November 14, 2026 Older → The Signal — November 12, 2026
LinkedIn
excelsiorgroup.ai
Powered by Buttondown, the easiest way to start and grow your newsletter.