The Signal — November 12, 2026
The Read
Control became a filing. Anthropic's amended registration statement told its future public shareholders that it has met its own threshold for AI-accelerated research — the point at which AI does enough of the company's own R&D to roughly double the pace of model improvement — and on the same Tuesday the AI 2027 forecasters pulled their median for a superhuman AI researcher from December 2028 to May 2027. Commerce, meanwhile, moved the export-control regime from the crate to the login, and Axios put on the record that the drafting was never the bottleneck: the Commerce Secretary stalled a crackdown the President had personally authorised, for the third time this year. Nvidia's board was reported to be preparing to remove the man who built it, and the stock gave up about eight percent after the bell. None of this is noise. It is the governance layer of the build-out acquiring filing dates, prices and statutory language — which is the first version of these arguments an operator can actually plan against, and the first version that creates products nobody has built yet.
Tide
Two confirmations, no shift. All four tides hold; ai-as-worker and governance-as-market-structure both picked up unusually hard evidence — one in a registration statement, one in a procurement record.
A lab told its future shareholders that AI is now doing enough of its own research to double the pace
Anthropic's amended registration statement, filed Tuesday, moved from saying it might cross its Responsible Scaling Policy threshold for AI-accelerated R&D "imminently" to stating it plainly: "We made a determination that we have met this threshold." The threshold is defined as AI doing enough of Anthropic's own research to roughly double the pace of model improvement — packing two years of progress at the 2018–2024 rate into one. The determination rests on "what we expect from future models and our overall rate of progress" rather than measured output from shipped models, and crossing it triggers stronger safety and monitoring commitments rather than any pause. Two concessions in the filing matter more than the headline: Anthropic says it may lose visibility into some of its own internal processes as it leans harder on Claude, and that it does not expect to use Claude for fully autonomous model development without meaningful human oversight. The same Tuesday, the AI Futures Project — the AI 2027 authors — pulled its medians forward hard. Daniel Kokotajlo's 50% date for a superhuman AI researcher moved from December 2028 to May 2027 and Eli Lifland's to August 2027, and all of them now call 2027 the likely year absent a deliberate slowdown, though the individual distributions still differ. Wednesday was when the reporting put the documents in one frame and got the forecasters on the record: Kokotajlo told The Information "We're at the beginning of the takeoff period," while Google DeepMind's Sebastian Krier criticised the underlying METR measurement as relying on flawed measures. The anchor number is not new — METR's 1.9x estimate, with its 95% interval of 1.5 to 2.8x and its note that Anthropic redacted a majority of the evidence METR found most informative, ran in this brief on the 5th.
So what: The forecast is not the useful part; the filing date is. A capability-progress rate asserted in a registration statement is dateable, auditable and liable in a way a safety blog post never was, which makes this the first version of the recursive-self-improvement argument you can plan against rather than argue about. The opening sits in the gap the filing admits to. Anthropic expects to lose visibility into its own internal processes as it delegates more work to its models, and every organisation putting agents into production is about to have that problem one layer down — with no filing obligation and far worse instrumentation. Almost nobody sells the thing that fixes it. If you are choosing a problem for the next eighteen months, legible AI-accelerated work got materially bigger this week, and the buyers now have a disclosure reason to pay for it.
- Anthropic says a key AI safety alarm has been triggered, days before its IPO
- Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All
Export control is moving from the crate to the login, and the record now shows who slowed it down
The Commerce Department is finalising a rule to stop Chinese AI firms reaching advanced chips remotely through data centres in third countries, expanding the Bureau of Industry and Security's existing semiconductor controls to seven additional countries — Malaysia, Singapore and Thailand among those named — with an announcement possible this month, per The Information's Wednesday report citing two people familiar with the process. This brief flagged the shift from hardware to access on the 9th; what changed is that the rule is now in interagency consultation rather than under discussion. Axios reported the same day that the politics, not the drafting, have been the bottleneck: Trump authorised a crackdown on suspected chip diversion on Thursday the 5th — his third approval this year, after May and August — and the decision had to be forced in two rooms. A Situation Room meeting attended by Bessent, Hegseth, Greer and Rubio sided with Bessent; the group then moved to the Oval Office, where Commerce Secretary Howard Lutnick was the lone holdout when Trump made the call, having warned the measures "will start a trade war with China" and then offering to "develop some options and come back." Meanwhile the evidence that crate-level control leaks became documentary. Reuters reviewed 26 Chinese procurement and corporate records from 2026 and found Northwestern Polytechnical University — on the US Entity List since 2001 and one of China's "Seven Sons of National Defense" — bought two H200s in October for 874,700 yuan, about $123,000, through a domestic reseller rather than from NVIDIA, and obtained H200 and other NVIDIA capacity through five rented-compute contracts this year. ChinaTalk opened a third surface the same morning: American vendors have sold roughly half a billion dollars of data to Chinese labs, some of it reinforcement-learning environments for coding, cyber, finance, law and bio, across nine named US firms.
So what: The regulated surface is widening from hardware to access to the training environment itself, and each widening is a compliance product that does not exist yet. Three of them are now visible: attestation of who is actually running on a rented GPU, provenance for where a model was trained, and chain-of-custody for the environments a model was post-trained in. All three are boring, all three are about to be mandatory, and none has an obvious incumbent. For anyone selling compute, the practical move this month is to be able to answer — in writing, per tenant — who holds the login, because a rule that lands this month will make that answer a condition of sale rather than a courtesy. Read the Lutnick reporting as timing information, not scandal: it tells you the policy direction has presidential authorisation behind it three times over, which makes the eventual rule more durable than its delay suggests.
- U.S. finalizing rule to block China's remote AI chip access- The Information
- Scoop: Lutnick stalled curbing Nvidia chip sales to China — again — after Oval Office showdown
- Chinese university buys Nvidia AI H200 chips despite US ban, records show
- Piggybacking to the Frontier
Waves
The inference resellers started buying their own silicon
Fal, Together and Baseten have begun buying their own NVIDIA GPUs rather than purely reselling leased cloud capacity, per The Information — Fal funding purchases for leased data centres out of part of the $1.65 billion it announced on the 3rd across its Series E and F, of which $1.3 billion was the Bessemer-led Series F at a $16 billion valuation, double April's $8 billion. It has told investors it wants to borrow for more. The driver is scarcity rather than strategy: leased capacity has become harder to secure, and some clouds now favour customers with their own end users over resellers. One person familiar said CoreWeave stopped selling Baseten compute after launching a competing inference service in March, which someone close to Baseten disputed. Fireworks, valued at $17.5 billion, is the holdout and has told investors that "owning hardware is a commodity business" with no plans to buy chips. The same squeeze has pushed this cohort toward smaller neoclouds with capacity to spare — a fragmentation whose downside showed up in Fireworks' own widened breach disclosure, covered here on the 10th.
Roadmap implication: Inference sold by an asset-light reseller and inference sold by an owner-operator are not the same product, and the difference is about to appear in your renewal terms rather than your benchmark. Put one question on the next vendor call and write the answer down: do you own the silicon serving my traffic, and if not, whose is it and on what term? The roadmap implication is that the asset-light inference layer is splitting into two businesses — a capital-intensive owner-operator and a differentiated software layer on top — and the second is where a team with no balance sheet should be standing. Fireworks' bet that hardware is a commodity is the more interesting one to watch: if it is right, the companies levering up to buy GPUs in a tight market will have bought the low-margin half of their own business.
- AI Inference Startups Are Quietly Buying Their Own Chips
- Fal, Together, and Baseten start buying their own Nvidia GPUs as AI inference startups struggle to lease enough cloud capacity
- AI inference startup Fal doubles its valuation to $16B as demand surges
Power is the binding constraint, and somebody finally priced the delay
An analysis by insurance broker Howden, built on EIA and S&P Global data and written up Wednesday, put arithmetic behind what everyone building has been saying. Power delivered to US data campuses has grown about 19% a year since 2000 while grid capacity grew about 2% a year, taking data centres from 0.1% of US grid capacity in 2000 to roughly 5% in 2025 — a fifty-fold increase in share. To hold that 5% share through 2030, Howden estimates data centres would need to source up to 58% of their power off-grid, equivalent to about 110 GW of behind-the-meter capacity, which Howden itself calls unlikely. The number that belongs in a model is the delay cost: for the largest power-constrained projects, grid interconnection delays of two to six years translate to $800 million to $2.4 billion in forgone revenue. The Register noted, correctly, that the source chart's two lines use mismatched axes — about 1,400 GW of grid against 80 GW of data centre — so this is a growth-rate story, not an imminent takeover of US supply.
Roadmap implication: Time-to-energize, not cost per megawatt, is now the variable that decides which capacity plans are real, and Howden has given you a defensible way to price it. Put an interconnection-delay line in the model next to the capex line: at $800 million to $2.4 billion per large constrained project, a two-year queue is a bigger number than most of the hardware decisions being argued about above it. The constructive read is that 110 GW of behind-the-meter capacity being called unlikely is a statement about today's financing and permitting structures, not about physics — and that is precisely the kind of old problem that pays to re-found. Behind-the-meter generation, interconnection-queue advisory and firm-power contracting are all markets where the buyer is desperate, the incumbent is a utility process, and the margin is sitting in the delay.
The venture money went to watching the agents, not building them
The Information published its annual list of the 50 most promising startups on Wednesday and said the two categories its reporters found hardest to narrow were AI infrastructure and AI security — an explicit consequence, they wrote, of the rising rate of AI-powered cyberattacks. The week's funding ledger agrees. Bay came out of stealth Wednesday with $50 million total: a $37 million Series A led by Battery Ventures plus a previously undisclosed $13 million seed co-led by YL Ventures and Striker. Founded in February by three former IDF Unit 8200 officers — Yuval Zacharia as CEO, Boris Danilovich as VP R&D, Aner Morag as CPO — it inventories an enterprise's agents, coding assistants and MCP servers, maps their connections, and then routes each proposed agent action through an "Enforcer" that can allow it, refer it to the user, redirect it or block it, under policies generated by an embedded system called Lighthouse. Per The Information's deals ledger the same day, it was not alone: Daylight raised $80 million led by Growth Equity at Goldman Sachs Alternatives, while Tuesday brought Dreeze out of stealth with $33 million from Norwest and Bessemer to restore entire cloud applications after an outage or attack, and WhiteMatter with $20 million led by Blumberg Capital to monitor agents running inside an enterprise.
Roadmap implication: The control plane for agent actions is being bought rather than built, and the architecture the market is converging on is worth copying even if you buy nothing: inventory every agent and MCP server, map what each can reach, and put a decision point in front of each action with four outcomes rather than two — allow, ask the user, redirect, block. The roadmap implication is that "can this agent do this right now" is becoming a platform capability on the same footing as identity, which means the budget line is moving out of security and into infrastructure within a year or two. Build the inventory now whatever you decide about vendors, because every one of these products needs it as an input and you will be asked for it by an auditor before you are asked for it by a procurement team.
A London court is deciding whether a model trained abroad becomes an infringing copy on import
Getty Images v Stability AI reached the Court of Appeal on Wednesday, listed across Wednesday and Thursday. Getty is appealing the November 2025 High Court judgment of Mrs Justice Joanna Smith, which largely sided with Stability, and the appeal turns on two questions: a narrow reading of secondary copyright infringement, and whether a model trained overseas on copyrighted images can be classed as an "infringing copy" when imported into the UK even where the model stores no copies of those images. Stability's argument rests on the images having been downloaded and stored in data centres outside the UK. Getty's counsel put the counter-position bluntly in open court, saying Stability "wants to divorce what was done with the copyright works from the act of the model training." The procedural history is a warning about how hard these cases are to win: Getty dropped a primary copyright claim about two weeks into the June 2025 trial and succeeded only partly on trademark, for older models and on isolated examples.
Roadmap implication: If the import theory succeeds even narrowly, where a model was trained becomes a procurement question rather than a trivia question, and the practical consequence is immediate: any vendor contract signed in the next two quarters should carry a representation about training jurisdiction and an indemnity that survives it. The roadmap implication is broader and more interesting than the risk. A ruling that attaches liability to the model as an artefact rather than to the act of copying would make verifiable training provenance a commercial requirement across the whole supply chain — and right now almost no one can produce it. That is a tooling market with a court-imposed deadline, which is the rarest kind.
Ripples
Nvidia's board was reported to be preparing to remove its founder, and the stock gave up 8% after the bell
The Wall Street Journal reported Wednesday that Nvidia is preparing to remove co-founder and chief executive Jensen Huang after 33 years, with the board possibly meeting as early as that day to decide his future. NVDA fell about 8% in after-hours trading and is down about 11% over five days. The trigger was Huang's Sunday admission of an extramarital affair with a woman US intelligence reportedly believes to be a Chinese agent, compounded Wednesday evening by a New York Times report of a second relationship with an executive reporting to him. The operational story is the succession hole: Huang reportedly has around 60 direct reports, and there is no identified successor or published succession plan. Three Republican committee chairmen have asked to be briefed by Huang on Wednesday the 18th. Two caveats worth carrying: the ouster is reported rather than confirmed, and Nvidia's own newsroom last spoke on Tuesday — the board's appointment of Wachtell, Lipton, Rosen & Katz as its investigator, covered here on the 10th.
Do this now: If a single vendor's roadmap sits inside your 2027 capacity plan, price the key-person risk this week rather than next quarter: write down which commitments depend on specific allocation promises, who at that vendor made them, and what happens to each if that person's manager changes. The cheap hedge is not switching silicon, it is making sure at least one workload is genuinely portable and has been run end to end somewhere else — which, per Anthropic's own roadshow materials, is a three-day job when a capable model does the port. Do that port now while it is optional.
- Nvidia's stock drops as board reportedly preps to oust CEO Jensen Huang
- Anthropic's IPO roadshow video
AlphaGenome 2 doubled the hit rate on disease-linked variants by asking one cell type at a time
Google DeepMind released AlphaGenome 2 on Wednesday, and the architectural change is the interesting part. Where the original predicted across many tissues in parallel, AlphaGenome 2 is conditional — it takes a one-million-letter DNA input and predicts for one tissue, cell type or cell state at a time — and it adds DNA methylation to the existing expression, histone-modification and chromatin-accessibility readouts. It was trained on bulk and single-cell human and mouse data spanning more than 100 million cell profiles across 68 tissues. DeepMind reports it identifies nearly twice as many expression-altering variants as the prior state of the art; on the TenK10K benchmark across 28 immune cell types it found causal variants in 30% of cases against 16% for a fine-tuned AlphaGenome, and got the direction of the gene effect right 43% of the time against 23%. It can be conditioned on unseen cell states from a few dozen cells of single-cell RNA sequencing with no retraining. It is available now through the AlphaGenome API for non-commercial research, with commercial research in private preview. DeepMind states plainly that it is for research use and is not clinically validated.
Do this now: Note what moved: not parameter count, not compute, but the decision to answer a narrower question. Conditioning on one cell type at a time nearly doubled the useful hit rate on the same underlying problem, which is the most transferable lesson of the day for anyone whose model is underperforming on a broad objective. Before buying more capacity, try specialising the question. If you work anywhere near variant interpretation, the no-retraining conditioning path is the thing to test this week — a few dozen cells of scRNA-seq is a weekend of lab work, not a research programme.
MiniMax shipped open-weight 16-megapixel image generation in six sampling steps, and published no benchmarks at all
MiniMax released H3-Image on Wednesday: open weights for text-to-image and image editing in a single checkpoint, built on the H3 architecture with the audio layers removed and the MiniMax-H3 text encoder unchanged. The specifications are genuinely aggressive. The model card claims roughly 16 megapixels at every aspect ratio from 4096x4096 to 6272x2688, generated directly with no upscaler; the checkpoint is distilled to six sampling steps with no classifier-free guidance; it accepts up to 16 reference images; and the BF16 DiT weighs 68.65 GB. Adoption was immediate — 143,000 downloads in thirty days and 723 likes on the official repository, second on Hugging Face global trending, with a ComfyUI repack past 350,000 downloads. Two things to hold against the enthusiasm. The licence is not Apache or MIT: it is a MiniMax H3-Image community licence, so read it before it reaches a product. And the launch blog asserts "state-of-the-art prompt adherence among open-weight models" while naming no benchmark and publishing no figure for it — the speed and pricing claims (4K in under ten seconds, up to 5x faster, 50–70% of comparable pricing) are all vendor-reported too.
Do this now: Download it and run your own prompt-adherence set this week, because nobody else has published one and the download count is not evidence of quality. Six sampling steps with no classifier-free guidance is the number that changes a product plan: it moves high-resolution generation from a batch job to an interactive one, which is a different feature rather than a cheaper version of the same one. Budget for the 68.65 GB checkpoint before you promise anything on-premise, and get the community licence in front of counsel first — an unreviewed licence on a model this good is the standard way a prototype becomes an expensive problem.
AMD started selling the memory answer to the DRAM squeeze: half the DRAM, 99% of the performance
AMD began commercialising Mext on Wednesday, the memory-tiering software it acquired in June, as a licensed add-on to EPYC servers. A transformer-derivative model predicts the next memory page a workload will need and prefetches it between DRAM and NVMe flash ahead of the operating system; it trains on a single EPYC core in under a second and runs entirely on the CPU. AMD claims 99% of DRAM-only performance on Apache Spark with half the DRAM replaced by flash, 95% on Synopsys PrimeTime at the same 50/50 split, degrading to 80% at 25% DRAM and 75% flash. AMD's Gary Smerdon put the economics at up to 83% server cost reduction and $5 million of avoided DRAM cost across 145 deployed systems, with flash roughly 50x cheaper and 30x lower power per bit but about 500x slower — and cited a Meta paper finding 50–80% of memory pages go untouched for five minutes or more. Virtualised deployments need one core per VM. It sells through OEMs including HPE in two phases: a "Mext ready" certification, then licence sales with the management tooling preinstalled. Every performance and cost figure here is vendor-reported.
Do this now: This is the most concrete vendor answer yet to the 2026 DRAM price spike, and it is testable cheaply: if your workload has the cold-page profile the Meta paper describes — and most analytics and EDA workloads do — the 50/50 split is a configuration change rather than a migration. Ask your OEM for the "Mext ready" certification status on your next server refresh before you size DRAM, because sizing for DRAM-only and then discovering you could have halved it is an expensive way to find out. Treat the 83% figure as marketing and the 99%-of-Spark figure as a hypothesis to reproduce on your own data; the second one is cheap to test and the first one is not a number anyone can hand you.
OpenAI's kill switch takes 35 seconds, and it published the number
The Register reported Wednesday on OpenAI's "Control Room," described in OpenAI's own Tuesday post as a central system to track model workloads across its research infrastructure and stop them when safety or security concerns arise. The figure worth keeping is the latency: it halts model output about 35 seconds after a stop decision, a delay OpenAI attributes to a caching layer for inference authorisation and says it wants to drive to zero. Two design choices are more notable than the headline. It is built to keep working when other systems fail, and it is built to stop model output if Control Room itself goes down — fail-closed rather than fail-open. It is already deployed across OpenAI's research infrastructure, with expansion planned over coming months alongside misalignment and security-breach monitoring. The Register's angle is that OpenAI has not explained why this took so long, and it points — without claiming causation — to recent agent incidents involving Hugging Face, the Wikimedia Foundation and Australian healthcare data.
Do this now: Go and measure your own number this week. Almost every organisation running agents in production has a stop button nobody has ever timed, and 35 seconds from a frontier lab that built the system deliberately is a useful benchmark for how bad yours probably is. The design lesson is the one to copy immediately and costs nothing: make the kill path fail closed, so that if your control plane dies the agents stop rather than carry on unsupervised. And note that OpenAI published the latency rather than the capability — a measured, unflattering number is what makes a safety claim worth anything, and it sets a bar your own vendors should be asked to meet.
Read this edition and the archive at excelsiorgroup.ai/insights/signal
The Signal - The Excelsior Group