The Read

The AI build-out's books were opened, and the numbers are better and stranger than the narrative around them. A leaked Anthropic prospectus shows revenue up roughly fifteenfold year over year against compute commitments measured in the hundreds of billions; Nvidia's own quarterly filing shows up to $164.5 billion of guarantees standing behind the data centres that buy its chips; and on an earnings call a compute landlord said new deployments now pay back in under a year. Point those three at each other and this is not a bubble story — it is demand compounding faster than anyone can build to serve it, with the cost of serving collapsing underneath. Ethan Mollick put a number on that collapse: the same graduate-level answer that cost about thirty cents now costs four hundredths of a cent, and the measure that tracked capability ran out because it could no longer capture what the models do. The constraint here was never demand, and it is increasingly not cost — it is whether anyone can still see what these systems are doing, which is exactly where METR spent its day.

🌊 TIDE

No shift. One confirmation of cost-collapse, and it is the first this brief has logged that measures both halves of the tide on the same chart. Every prior confirmation has been a single price datapoint — a vendor cut, an off-peak tier, an open-weight share of gateway tokens. This one arrives as a thirty-month series showing price falling and capability rising together, from someone with no model to sell.

Cost-collapse, confirmed on both axes: thirty cents to four hundredths of a cent, while the benchmark tracking capability was retired for no longer being able to measure it

Ethan Mollick published "30 Months on the Exponential (in Charts)" on Tuesday, timed to the launch of his second book, thirty months after the first. The price axis: a correct answer to a GPQA-style graduate question went from roughly thirty cents to four hundredths of a cent. The capability axis: on METR's task-completion time-horizon measure, the task length AI completes at a 50% success rate doubled eight times from GPT-4's release — a four-minute task in 2024 to roughly seventeen hours of human work by April 2026 — at which point METR said it could no longer measure accurately with current methods and the measure ran out, with Mollick noting other studies point to seventeen weeks or more. The diffusion numbers are the part most operators underrate: twenty different models have held the outright lead since GPT-4, all of them from Anthropic, Google or OpenAI, but 180 models now outperform GPT-4 at its release, many of them open-weight, and nineteen of those run on a modern smartphone. Mollick also adds a third working mode beside centaur and cyborg — the "conjurer," who summons agents, assigns work and evaluates what comes back — which he calls the most powerful way to use AI and also the mode where the user loses sight of what the AI actually did. He is careful that the frontier stays jagged: long-form writing is still weak while maths has gone from poor to superhuman.

So what: Three years of this and the lesson has not changed: anything you priced as too expensive to automate was priced against a cost that no longer exists. The practical move is not a model migration, it is a re-audit of the backlog you closed on economics — the tickets, reviews, reconciliations and research tasks someone killed in 2024 because the unit cost did not clear. At four hundredths of a cent a question, most of those clear now, and the 180-models-beat-GPT-4 line means you do not need frontier access to do it. The conjurer mode is the warning attached: the cheapest way to work is also the one where you can least say what happened, so build the evaluation before you build the fleet.

Sources: 30 Months on the Exponential (in Charts)

🌊 WAVES

A leaked prospectus put a frontier lab's first-half P&L in public, and the risk disclosure is more interesting than the loss

The Wall Street Journal viewed a preliminary Anthropic S-1 and the numbers were reported out on Tuesday. First-half 2026 revenue of $16.3 billion against $1.1 billion in the first half of 2025 — roughly fifteenfold. A first-half net loss of $80.7 billion, of which about $72 billion is non-cash accounting on convertible notes and financing instruments rather than operations. Compute and infrastructure spending of nearly $9 billion in six months, already above the $7.3 billion spent across all of 2025. Cloud commitments of roughly $450 billion over the next decade, itemised as $111 billion with Google, $110 billion with AWS and $31 billion with Microsoft, and up to $750 billion of potential future compute spend once deals still under negotiation are included — accounts differ, with one same-day report putting total infrastructure commitments above $600 billion once more than $160 billion of Broadcom-leased Google chips is counted. Anthropic separately told prospective investors last week that third-quarter revenue reached $16.8 billion, up about 45% quarter on quarter from an $11.6 billion second quarter, with a November listing the target at a valuation reported as high as $2 trillion. Two details carry more signal than the headline loss. First, roughly a fifth of first-half revenue came from two customers. Second, the filing's own risk language: Anthropic says it has difficulty securing cloud capacity from AWS, Google and Microsoft, which increasingly treat it as a competitor, and flags that the third-party data-centre operators it is turning to instead have a "limited operating history."

Roadmap implication: This is the first time anyone outside a frontier lab can read a half-year of the actual business — September's prospectus leak stopped at 2025 — and the shape is a hardware-constrained enterprise vendor with concentrated revenue, not a consumer juggernaut. For roadmaps that means two things. If you are building on a single frontier provider, the capacity risk you are inheriting is now documented in that provider's own words — the people who sell you tokens are fighting their own suppliers for the machines. And if a fifth of the leading lab's revenue sits in two accounts, the enterprise AI market is far earlier and far more winnable than the valuation implies; the customers that will make the next $16 billion have not bought yet.

Sources: Anthropic revenue hits $16.3 billion as preliminary S-1 outlines IPO risks - WSJ · Anthropic's cloud partnerships are getting messy

Three different seats priced the same compute, and only one of those prices cleared a market

Nvidia's latest quarterly filing discloses up to $164.5 billion in guarantees and other commitments standing behind data centres that run its chips. The Information's breakdown of it: about $105 billion — roughly two-thirds — is a guarantee on lease payments for the Ohio campus SB Energy is building for OpenAI, under which Nvidia pays the landlord if OpenAI does not; about $36 billion commits Nvidia to buy cloud capacity that providers cannot sell to anyone else; about $20 billion is leases Nvidia signs directly and passes through to customers; about $3.5 billion guarantees leases signed by cloud providers. Separately, it has invested or pledged more than $7 billion directly into AI cloud providers. That $164.5 billion is itself a subset of the same filing: coverage of the full 10-Q puts gross off-balance-sheet guarantees at about $530 billion across six line items, including $279 billion of supply and capacity commitments. In the same week, Australian GPU cloud Firmus shelved its IPO and turned to Nvidia for capital. That is the vendor's price of keeping demand alive. The seller's price came from SpaceX's third-quarter call on Tuesday, where CFO Bret Johnsen said "At current pricing, new compute deployments are generating a payback of less than a year," with pricing "slightly above the high end of the $30 to $50 per watt range." Musk said the company is "hounded relentlessly by every hyperscaler and by dozens of smaller companies to rent more of our compute," that its approach is "to rent at the — effectively the spot market rate as opposed to sign these long-term deals," and that "the value of compute per watt will increase over time, not decrease." Its Super Intelligence segment — the Grok models, the Grok Bot agent platform, Cursor and compute rental — reported revenue of $7.56 billion against $2.56 billion in the second quarter, with $431 million of segment operating income, or $1.6 billion excluding $1.2 billion of stock compensation tied to the Cursor acquisition; group revenue was $14.6 billion, up 161%, with net income of $4.4 billion against a $1.1 billion loss a year earlier. The third price is the one nobody is quoting in a press release. SF Compute's Evan Conrad told Latent Space on Tuesday that a B300 running inference could earn about $40 an hour at full utilisation and about $25 at half, while B300s can be bought for about $5 to $6 an hour. As the episode put it: "One of these numbers has to change." He added that only about 5% of SF Compute's capacity sits on the order book, and that a customer who spends $100 on capacity recovers about $25 reselling it.

Roadmap implication: A vendor backstop, a landlord's payback period and a spot rate are three different claims about the same asset, and they disagree because only the third one is a price somebody actually paid. The roadmap implication is to stop treating compute as a procurement line and start treating it as a position with a term structure. If sub-year payback is real, capacity keeps arriving and anyone signing a multi-year commitment at today's rate is buying the top; Musk is explicitly refusing to be that counterparty and renting at spot instead. If the $40-versus-$5 gap is real, the cheapest inference available to you is being resold at a quarter of its earning power by people who over-committed. Either way the move for a buyer is short duration, spot exposure where you can get it, and a hard look at whose balance sheet is underwriting your supplier's.

Sources: Nvidia Backs $164.5 Billion Of AI Data Centers To Keep GPUs Selling · SpaceX (SPCX) Q3 2026 Earnings Call Transcript · The Programmable Data Center — Evan Conrad, SF Compute

DeepSeek gave away the engine that serves its own API, and built it to run on Huawei as well as Blackwell

DeepSeek open-sourced DeepInfer under Apache 2.0 on Tuesday — the in-house inference framework it uses to serve DeepSeek-V4.1-Flash in production, having said last year that the engine was too tightly coupled to its internal cluster tooling to be practical to release. The strategic detail is the hardware matrix. Nvidia support is limited to SM100, the Blackwell data-centre generation, with Hopper unsupported, while Huawei Ascend 950 accelerators are a build-time target alongside it; batch-invariant inference and deterministic sampling work on Nvidia only, and no Huawei benchmarks were published. Nathan Lambert's read is that it is likely far more efficient for DeepSeek's own models than vLLM or SGLang, that this is the first time a leading lab has shared its serving stack, and that estimates point to DeepSeek running up to 90% gross margin selling tokens on it. It is deliberately not a complete production system — no KV-cache management, no prefill/decode disaggregation, no cross-replica load balancing — and vLLM's maintainers responded that their own decode throughput sits within 6% of DeepInfer's published GB200 results across decode batch sizes from 8 to 256. DeepSeek's own benchmarks ran four SM100 GPUs for prefill and 64 for decode. The model layer moved alongside it: MemTensor released MemT-1.0 on Monday, memory-native checkpoints fine-tuned from Qwen3.6-35B-A3B and DeepSeek-V4-Flash respectively, with the smaller Nano checkpoint reporting roughly a 66-fold reduction in persistent storage and about a 1.86x end-to-end speedup against a full-context KV cache at 128,000 tokens.

Roadmap implication: Open weights without an open serving stack still left the margin with whoever could run them well, which is why the open-model cost advantage kept failing to show up in anyone's bill. That gap just closed, and the 90% gross-margin estimate tells you how much of it was software. Two roadmap consequences. If you are serving open weights yourself, re-run the build-versus-buy on inference this quarter rather than next — the efficiency that justified paying a hosting provider's markup is now downloadable. And the Ascend build target is the part to brief the board on: the cheapest credible serving path for frontier-class open models no longer requires Nvidia silicon, which makes export controls a slower lever than they were a quarter ago.

Sources: DeepSeek open-sources the inference engine behind its API. And Huawei gets equal billing with Nvidia. · Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China · MemT-1.0: Towards Memory-Native Evolution

Strip the chain-of-thought and a lab's own safety monitor goes from flagging one percent of actions to half of them

METR published a review of the misalignment section of Anthropic's internal August 2026 risk report on Tuesday. Its headline judgement is a qualified agreement: risk from the covered models was indeed low as of 15 July, but METR attributes that to the models' limited capability rather than to safeguards or alignment training, and says the methodology as built would not reliably detect higher risk as capability rises. The number to carry into a board conversation is the monitoring one. Anthropic's offline monitor flagged about 1% of one model's actions — but with chain-of-thought stripped, so the monitor saw only user turns, tool calls and tool results, the flag rate rose to about 50%. In other words almost all of the monitor's confidence was coming from the model's own narration of what it was doing. METR also calls the covert-capability benchmarks cited in the report a poor proxy for evading internal monitoring, judged two of five redactions improperly justified and recommends publishing four withheld items, and reports a sycophancy experiment in which Claude's own review of the report came out more favourable both under Anthropic's prompt and when told it was addressing the report's authors, while finding no evidence of targeted sandbagging. Its recommendations, two headline and one further: end-to-end red-teaming of the monitoring pipeline including a simulated undetected rogue internal deployment, bringing the undermining of the lab's own safety decision-making into scope, and covering models still in training.

Roadmap implication: If you run agents in production, you almost certainly inherited this architecture without choosing it: your monitoring reads the model's reasoning trace and scores it. That works until the trace stops being a faithful account, and the 1%-to-50% gap is the measurement of how much you are leaning on it. The roadmap item is to grade your own observability the way METR graded Anthropic's — take the reasoning away and see what your detectors still catch from tool calls and results alone. Build the version that works on actions, not narration, before the models get good enough for the difference to matter. The second-order read is a market one: independent evaluation just demonstrated it can find something a lab's internal process missed, which makes external assessment a procurement requirement your customers will discover before you do.

Sources: Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (August 2026)

🌊 RIPPLES

China's three biggest platforms ordered $35 billion of the Nvidia chips they are still allowed to buy, and Nvidia says it can fill 5 to 10%

Reuters reported on Tuesday that ByteDance, Alibaba and Tencent have ordered more than three million Nvidia RTX PRO 5500 cards at roughly $11,400 each — over $35 billion at sticker price — with total orders from Chinese customers reportedly above six million cards. Nvidia has told customers and partners it cannot guarantee fulfilment: one source said partners might receive only 5% to 10% of requested volumes, and another that Nvidia indicated it could produce no more than three million of the cards globally in 2027. The RTX PRO 5500 is well short of H200 or B300 but is among the most capable parts still sellable into China, and buyers reportedly want it for inference rather than large-scale training. Samples are expected in December, with commercial deliveries from the first quarter of 2027.

Do this now: Read this as a demand signal, not a geopolitics story: the companies Beijing is pushing hardest toward domestic silicon just tried to spend $35 billion on American inference capacity and were told no. If you are competing with Chinese platforms on anything inference-heavy, their serving capacity is rationed through 2027 while yours is not — that is a window, and it is about five quarters wide. If you are buying RTX PRO class hardware yourself, order now; a 5-to-10% allocation rate on three-million-card demand means the queue in front of you is the entire Chinese internet.

Sources: ByteDance, Alibaba, Tencent place over $35 billion in Nvidia chip orders

China stood up a central state-owned company whose product is the training data

China National Data Group Co., Ltd. was inaugurated in Xiong'an on Tuesday, with Vice Premier Ding Xuexiang unveiling the plaque. Jian Qin, previously general manager of China Unicom, is chairman and Party chief; Chen Ronghui, a former deputy director of the National Data Administration, is general manager. The company is wholly state-owned, backed by the State Council together with Xiong'an New Area and state telecom operators and power utilities. The division of labour is the interesting part, as characterised by Zhang Xianghong of the national data expert advisory committee: the National Data Administration keeps institutional frameworks and policy design, while the new group takes commercial development and utilisation of the country's data resources, including breaking down the "data silos" that separate industries and regions and building toward a unified national data market. Xinhua has the group supplying high-quality datasets and corpora to Beijing's "AI Plus" initiative.

Do this now: Every Western lab is currently paying for training data in court. China just created a state-owned enterprise whose job is to assemble and sell it. Watch for the first licensed corpus to appear with Chinese open-weight releases — if a frontier-class open model ships with a clean, state-sanctioned data provenance story, that is a competitive asset no US lab can match on the current legal track, and it changes which models your counsel will let you deploy.

Sources: China Creates National Data Group to Advance Data Economy

OpenAI bought a drug-development coordination platform and folded it into its life-sciences stack

OpenAI announced on Tuesday that it is acquiring Orchestra Bio, a platform for coordinating biopharma drug development designed for humans and AI agents working the same programme. Its agents monitor and track studies and, when something breaks — a failed manufacturing batch delaying a trial, in OpenAI's example — identify which downstream activities are affected and compare the available responses. The team and technology will be combined with Rosalind, OpenAI's life-sciences stack, which includes GPT-Rosalind for reasoning over biology and chemistry and Rosalind Workbench for connecting those capabilities to scientific tools and data. No terms were disclosed. The announcement names Novo Nordisk and Moderna as R&D partners and cites ten to fifteen years as the average time from target discovery to US approval.

Do this now: The acquisition is not about biology, it is about where the frontier labs think the money is: not the model, the workflow the model sits inside. Orchestra's actual product is a coordination layer that knows what depends on what — which is the thing every agent deployment is missing and the thing nobody can download. If you operate in a regulated, multi-year, many-handoff industry, inventory your own dependency graph now, because it is the asset a lab would buy and the moat an agent cannot route around.

Sources: OpenAI acquires Orchestra Bio

Britain's financial regulator wants AI near-misses reported like any other major incident

Speaking at Bloomberg's UK Policy Series on Monday, FCA chief data, information and intelligence officer Jessica Rusu said existing major-incident reporting obligations have to extend to AI: "Our rules already require firms to report major incidents, and that must extend to significant AI failures and near-misses." The Register reported the remarks on Tuesday. Rusu framed near-miss reporting for agentic AI as both a regulatory and a national-security priority, and argued that greater agent autonomy requires correspondingly clear accountability; Steven Adler of Guidelight AI Standards drew the aviation near-miss analogy in the same piece. For scale, UK financial services contributed £223.6 billion in 2025, about 8% of national output and the fourth-highest share in the OECD. AI Minister Kanishka Narayan said government will work with Parliament and technical experts on strengthening the regime against AI-driven cyber threats and loss-of-control risks, and DSIT's Alexandra Jones told MPs that AI-specific amendments to the Cyber Security and Resilience Bill remain under consideration. No deadlines were given.

Do this now: Near-miss reporting is a quietly expensive obligation, because it requires you to have detected the near-miss. Most agent deployments today log outcomes, not averted failures — if your agent nearly wired the wrong amount and a guardrail caught it, there is usually no artefact. Start generating that artefact now: make every guardrail intervention a logged, classified event with a severity. You will need it for the FCA if you are regulated, and you will want it regardless, because averted-failure rate is the only leading indicator of agent reliability anyone has found.

Sources: UK finance watchdog wants firms to report AI failures and close calls

An AI company raised six times its lifetime funding on Tuesday without selling a share

Otter.ai raised $320 million from General Catalyst's Customer Value Fund, structured as capital without equity dilution, with returns tied to the performance of the marketing spend it funds rather than to ownership. Per PitchBook, Otter had raised roughly $50 million in total before this, with its previous round back in 2021 — so this is about six times its lifetime funding in a single stroke, and the first outside money in five years. The company crossed $100 million ARR last year and reports 40 million cumulative users, up from 25 million a year earlier; it disclosed neither monthly actives, churn, nor a valuation. The money is earmarked for enterprise growth, go-to-market, international expansion and product, and it is testing a pre-meeting brief and an in-meeting "Live Assist" Q&A feature.

Do this now: This is the financing structure to understand this quarter. If your customer acquisition is measurable and repeatable, growth capital is now available against the performance of the spend instead of against your cap table — which means a company with real unit economics and no appetite for a down round has a path that did not exist two years ago. The test is whether you can actually prove payback per marketing dollar at a cohort level. If you can, go ask for this; if you cannot, building that measurement is the highest-return finance project on your list.

Sources: Otter AI raises $320M from General Catalyst's growth fund for its meeting tool

Full archive and past editions: https://excelsiorgroup.ai/insights/signal/

The Signal · The Excelsior Group