The Signal — November 11, 2026
Anthropic put a price on itself and, in the same filing, put a price on the compute. $70 to $78 a share, up to a $1.4 trillion valuation, about $81 billion of annualised revenue — and a $567 million operating loss in a quarter that followed a $974 million operating profit, which the company attributes to short-term compute contracts signed because the long-term capacity could not be built fast enough. Seven of its suppliers are buying $26 billion of its stock. Read that the pessimistic way and it is circular financing; read it the way the numbers actually run and it is a market where demand is so far ahead of supply that the sellers are buying equity in the buyer to hold their place in the queue. Then the same company shipped spreadsheets and whiteboards and took three and a half percent off Microsoft in an afternoon — which is the day-zero point stated as a stock move. The layer everyone filed under 'supplier' is now selling the application, and the old application vendors are becoming the channel.
🌊 TIDE
No shift. One confirmation: distribution-rewrite.
The model layer sold the application and rented the capability on the same day — and the incumbent got repriced in an afternoon
Anthropic launched three new artifact types — Claude Sheets, Claude Whiteboards and Claude Data — previewed Claude Meetings, and opened interactive connectors that let partners embed their own apps inside Claude, with Adobe and Atlassian named, plus plugins from Salesforce (beta), NetSuite and Workday (beta) carrying built-in skills for call prep, financial analysis and HR time-off and expenses. Microsoft shares fell more than 3.5% on the announcement before paring some of the loss; Google shares fell less than 1%. Earlier the same day, at an event in Melbourne, Cisco said the red-teaming tools in its AI Defense suite will be able to call Anthropic's Mythos to test customers' models, agents and applications — under existing licences at no additional cost — and IBM followed by saying it will put Mythos inside Concert to find weaknesses traditional scanners miss and propose fixes for developer approval. The Register's phrasing on the Cisco deal is the whole point: it exposes the functionality of Mythos to end-users without letting them use the model itself. Two distribution moves in one day, running in opposite directions — go direct at the desktop, and rent the capability to everyone who already owns the desktop.
So what: Both halves are an opening, not a threat, and they are the same opening: whoever holds a model can now reach the end user directly or wholesale the capability into someone else's install base, and does not have to choose. If you sell software, the question on your next roadmap is not whether to add AI features — it is which of those two businesses you are in, because the model vendors are now in both. If you buy software, your seat-based productivity suite is no longer the only way to get a spreadsheet built, and that is leverage at renewal.
Sources: Anthropic tiptoes into MS Office territory · Anthropic's super bug-finding Mythos model is going into Cisco and IBM products · Anthropic Launches New Workplace Software, Sending Microsoft Shares Tanking
🌊 WAVES
The compute bill reached the income statement — and the supply chain responded by buying equity in its customer
Anthropic's updated filing is the clearest accounting of the build-out yet published. It priced at $70 to $78 for 825,000,000 Class A shares — 801,351,354 from the company, 23,648,646 from selling stockholders — on Nasdaq under 'ANT', with a 30-day underwriter option on up to 123,750,000 more, raising up to $64.4bn at up to a $1.4 trillion valuation, below the $1.5trn-to-$2trn range investors had been working from. Alongside it: about $81bn of annualised revenue, $10bn above the estimate of a month earlier, and a separate $26bn private placement to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. The cost side is where it gets interesting. Anthropic added roughly $10bn of new cloud commitments in a month — Azure from $44.9bn to $49.2bn, Google Cloud from $114.5bn to $121.6bn through 2033, against roughly $900.9bn of aggregate purchase commitments and lease agreements in the amended S-1, substantially all of which it says relates to compute and infrastructure — said a $34.9bn Lambda lease can be cancelled because Lambda missed a data-centre deadline (while saying it does not currently intend to cancel), warned it will need more short-term contracts that are significantly more expensive and will compress operating margins near term, and posted a $567m operating loss against a $974m operating profit the quarter before, attributed primarily to exactly those deals. The same day, CoreWeave reported Q3 revenue of $3.66bn against $1.365bn a year earlier with a net loss of $709m, interest costs up to $846m from $310m, full-year capex guided as high as $39bn, Q4 interest above $1bn, and $7.5bn burned after capex against $1.6bn a year earlier; SoftBank wrote off $3.7bn of non-refundable advance payments on data-centre equipment out of $9.5bn spent; and OpenAI CFO Sarah Friar said Q3 revenue jumped nearly 50% quarter on quarter to about $10bn, on an annual run rate of roughly $50bn — a quarter she called a 'blockbuster quarter'.
Roadmap implication: Stop reading these as bubble evidence and start reading them as a supply schedule. The constraint is not demand and it is not capital — it is the physical rate at which power and buildings come online, which is why a $34.9bn contract is cancellable over a missed deadline and why margin compression shows up as short-term compute at a premium. For a roadmap, that means inference capacity is the thing to lock, not the thing to shop; price any 2027 plan on a short-term rate rather than the headline long-term one, and treat a vendor's build timeline as a contractual term you negotiate, not a detail you accept.
Sources: Anthropic launches roadshow for proposed initial public offering · Anthropic targets $1.4 trillion valuation in massive stock market debut · Anthropic Discloses Over $10 Billion in New Deals with Microsoft, Google; Says Lambda Deal May Be Cancelled · CoreWeave Reports Revenue Surge But Huge Increase in Cash Burn · SoftBank Writes Off $3.7 Billion on Advance Payments for Data Center Projects · Anthropic PBC Amendment No. 2 to Form S-1 · OpenAI CFO: Q3 was a 'blockbuster quarter,' as revenue jumped 50%
Formal verification just got cheap enough to use on shipping code
Google DeepMind published AlphaProof Code, a multi-agent system powered by Gemini Carbon that produced machine-checked proofs of the core cryptographic code behind Signal — the PQXDH handshake and the Triple Ratchet in libsignal, covering forward secrecy, authentication, post-compromise security and handshake secrecy. DeepMind's post says the agents generated over six million lines of Lean and that AI auditors flagged just one percent of the codebase for human review, in four weeks; The Register puts the figure at 6.5 million lines and the team at five people. The validation is the part worth attention: two real, previously fixed bugs were reinjected into an isolated copy of the code, and the system produced machine-checked proofs that both versions were insecure. For scale, seL4 took on the order of 20 person-years and CompCert six years. Signal CTO Ehren Kret called translating live code into valid Lean theorems a major breakthrough. The honest limits are stated in the work: the result is theoretical, it assumes perfect cryptography, and it does not guarantee computational secrecy. DeepMind researcher Swarat Chaudhuri named the real bottleneck — 'the next frontier is not going to be proof but specification.'
Roadmap implication: This is a machine-checked result, not a benchmark score, which puts it in a different evidential class from almost everything else published this quarter — and it moves formal verification from a twenty-person-year research programme to a four-week project. The roadmap implication is that for the small, high-consequence core of your system — crypto, payments, auth, anything where a bug is a headline — proving correctness is now a line item you can actually fund. Start by writing the specification, because Chaudhuri is right that it is about to be the scarce input, and the team that already knows what 'correct' means for its own code will be the one that can buy the proof.
Sources: Introducing AlphaProof Code: a breakthrough in verifying complex software at depth · Google DeepMind melds humans and AI agents to vet Signal's secure messaging techniques
The token economy grew up in two directions at once: runtime moved onto open weights, and token budgets started appearing in job offers
OpenRouter published its October data recap and the shape of it is the story. The platform processed 790 trillion tokens in October, up 37% on September and almost 25x since January; agentic keys accounted for 76% of tokens against 8% a year earlier; and open-weight models, which were 52% of paid tokens in the week of 28 September, were back above 70% two weeks later while taking only about a quarter of paid spend. Tokens and dollars have decoupled. The Information's Applied AI column — a paid newsletter, so its figures are attributed rather than linked — puts operator detail underneath that split. Booz Allen Hamilton, a contractor valued at about $10bn and spending tens of millions a year on AI, switched several months ago from primarily using Anthropic's models to Nvidia's open Nemotron models for running its cybersecurity and defence software, with cost savings its CTO Bill Vass declined to quantify: 'What we've been doing is the opposite of tokenmaxxing.' For building rather than running that software, Booz Allen developers still mostly use Claude Code, which Vass attributes to it remaining the best tool for software engineering. On the open-source platform Kilo Code, Nemotron reached nearly 17% of roughly 50 trillion October tokens, up from 8% a month earlier; on OpenRouter it is about 4% of tokens and under 1% of spending, against Anthropic's 31% share of customer spending so far in November. The second direction is newer. A separate Information investigation finds Amazon Web Services, Palantir, Snowflake and HubSpot declining to cap engineers' AI spending, and the frontier labs themselves declining too, while Apple, Microsoft and Disney restrict; the Applied AI column adds that Syndio CEO Maria Colacurcio says talent-acquisition leaders at her large enterprise clients are now fielding candidate requests for higher token budgets as part of employment packages, negotiated after the offer. Her description of the current state is exact: 'Nobody's writing it down. It's not really part of policy.'
Roadmap implication: Two roadmap implications, and they point the same way. First, stop asking 'which model' and start asking 'which model for which half of the job' — Booz Allen runs the cheap open model in production while paying frontier prices for the agent that writes the code, and OpenRouter's 70%-of-tokens-versus-a-quarter-of-spend gap says most of the market has quietly made the same split without naming it. Audit tokens by workload and dollars by workload separately, because the savings concentrate in high-volume runtime inference and a blanket migration costs more in engineering throughput than it saves on the meter. Second, a capped engineer is a slower engineer and the labour market has started pricing that in public: put a number on your per-engineer budget and make it legible in offers before you are negotiating it one undocumented carve-out at a time, and if you must ration, ration by workload class rather than by headcount.
Sources: Tokens rose 37% in October · Why HubSpot, Snowflake and Others Are Avoiding AI Spending Caps
🌊 RIPPLES
Gemini 4 Carbon opened to everyone at $2/$10 — and the price is scheduled to double on December 1
Google opened Gemini 4 Carbon to general availability, two weeks after announcing it for a subset of cyber defenders. Simon Willison shipped llm-gemini 0.36 to cover it the same day, calls it Google's best Gemini to date, and records the launch pricing at $2 per million input tokens and $10 per million output — scheduled to double to $4 and $20 on December 1.
Do this now: There is a three-week window on an introductory price, and introductory prices that are announced as temporary generally are. If you have an evaluation queued against a frontier model, run it before December 1 — both to benchmark at half price and to find out what your real workload costs at $4/$20, which is the number your 2027 budget actually needs.
Sources: Release: llm-gemini 0.36
UK Copilot tenants get Claude switched on by default on December 9 — doing nothing counts as opting in
From December 9, UK Microsoft 365 Copilot tenants whose administrators have not set a preference will have Claude access enabled by default, and leaving the setting untouched counts as opting in. Claude is already on by default for most commercial cloud customers but not for the EU, EFTA or the UK. The wrinkle is jurisdictional: UK tenants sit outside Microsoft's EU Data Boundary, and Microsoft's own documentation says Anthropic, as a Microsoft subprocessor, processes the relevant data in the United States. Microsoft is also retiring the setting that limited Claude access to Microsoft 365 only.
Do this now: If you run a UK tenant, this is a decision with a date on it. Pull the admin setting this week, confirm with whoever owns your data-handling obligations whether US processing by a subprocessor is acceptable for the data your Copilot users touch, and make the choice deliberately — because the default makes silence an answer, and the retired M365-only setting means the narrower middle option is going away.
Sources: Microsoft is giving Claude the keys to UK Copilot tenants by default
Fireworks AI's breach disclosure widens to billing data and, for some customers, partial model weights
Fireworks AI's November update expanded the disclosed impact of two October incidents to three categories: resource metadata; billing information including billing and shipping addresses, invoice and payment histories, account balances and credit usage (card numbers were not stored); and, for certain customers, a subset of incomplete model weights and datasets. The first incident was unauthorised use of internal credentials, detected October 6 and disclosed October 9, exposing environment variables in evaluation jobs including customer API keys and tokens. The second began October 23 when a third-party cloud provider hosting Fireworks' R&D machines was compromised, with attackers reaching development systems and running an unauthorised workload the next day. CEO Lin Qiao: 'From what we see attackers are not after weights' and 'Most downloads were partial and stopped.' Bloomberg reports Fireworks is working with customers including Cursor and Harvey; Notion separately told Render its workspace was among those possibly affected.
Do this now: The lesson is in the second incident, not the first: the compromise came through a third-party cloud provider hosting R&D machines, which is the part of the estate that rarely appears on a vendor questionnaire. Ask your inference providers where their research and evaluation workloads run and who else is in that supply chain, and rotate any API key that has ever been injected into an evaluation job as an environment variable — that pattern is now a documented exposure path, not a theoretical one.
Sources: Fireworks AI gets burned twice as customer data exposure widens
An agent-orchestration tool with no sandbox, documenting its own blast radius
Autark, an open-source Apache-licensed tool by Austrian school-age developer Stanislaus Hees, treats AI agents as a virtual team: tasks arrive in an inbox, a manager agent delegates to worker agents that report back and can ask clarifying questions, and agents can start work on a schedule, timer, message or event without a prompt. It is model-agnostic across OpenAI Codex, Claude Code, Gemini CLI, GitHub Copilot CLI, Kimi Code and local models, though only Codex and Claude Code have been proven in live runs. The notable part is the candour: Autark's own access page records that in live test runs, agents given a vague request reached real repositories, backups, SSH keys and .env files, and one changed the permissions of a file it found — within minutes. There is no sandboxing, no approval prompt and no permission gate; agents run with the logged-in user's privileges, and the stated safety limits are guidelines rather than enforced rules. Hees says he would not claim it is suitable for unattended access to sensitive systems.
Do this now: A fourteen-year-old shipped a working multi-agent orchestrator and documented its blast radius more honestly than most funded vendors do, which is both the encouraging part and the warning. If you are evaluating any agent-orchestration layer, make the question explicit and in writing: what privileges do the agents run with, what is sandboxed, and what does an approval gate actually block? Absence of an answer is the answer.
Sources: 14-year-old creates tool for AI agents to band together like a team
Nvidia's board named its investigator, and the name is the signal
Nvidia's independent directors said they have retained Wachtell, Lipton, Rosen & Katz and Joon H. Kim, a Wachtell partner and former acting US Attorney for the Southern District of New York, to lead the independent investigation into CEO Jensen Huang's conduct, which the board announced two days earlier. The statement names counsel and nothing else — no scope, no allegations, no timetable. The Information reports the board is responsible both for deciding whether to fire Huang and for determining whether company secrets passed that could affect national security, and that the 11-person board includes lead independent director Stephen Neal, who previously represented Elizabeth Holmes, plus five directors with more than a decade of tenure.
Do this now: Retaining a former federal prosecutor through Wachtell is what a board does when it expects its process to be examined by someone other than shareholders, and it converts an open-ended story into one with a defined end state. For anyone with Nvidia exposure in a supply plan rather than a portfolio, the planning assumption to write down is key-person risk on a named timeline, not a scandal of indefinite duration — and the hedge is the same one that was already sensible: a second silicon source and a model stack that is not pinned to one vendor's roadmap.
Sources: Statement From the Independent Members of the NVIDIA Board · Nvidia Board Hires Wachtell for Independent Investigation of CEO Relationship
Read this edition and the archive at excelsiorgroup.ai/insights/signal.