Latest issueTuesday 3 November 2026 · Issue 1 · 7 min read
Anthropic's $895B compute commitments, OpenAI's trimmed round, and the week pricing moved to outcomes
The week's biggest rounds
Bedrock Robotics — Autonomous systems for heavy construction equipment; Series C; $600 million; led by CapitalG and Emergence Capital; $4.1 billion post-money, reported by Crunchbase News.
Serval — Coding agents for enterprise automation; Series C; $150 million; led by Sequoia Capital; $3 billion valuation, reported by Crunchbase News.
General Compute — AI cloud running inference chips from Cerebras, SambaNova and Positron; Series A; $100 million; led by Blackbird and Abstract; reported by Crunchbase News.
Foundry Robotics — Builds and runs autonomous factories; Series A; $100 million; led by Marlinspike; reported by Crunchbase News. What the ranking hides is its own ceiling: the week's two biggest rounds overall were not AI — Re:Build drew $2.1 billion for manufacturing and Kalshi $1 billion for a predictions marketplace, reported by Crunchbase News — while the AI cheques clustered in robotics and agent tooling.
42AI — Behavioural world model that predicts individual decisions, live at brokerages; seed; $62.5 million; from the founders and Infinity Group; reported by TechCrunch.
Trajectory — Post-training tools for non-technical experts; Series A; $45 million; led by Sequoia with Nvidia and Bessemer; $300 million valuation, reported by TechCrunch.
Tutor Intelligence — Robots run by human teleoperators until the models earn autonomy; Series B, amount not disclosed; led by a16z.
Hyperexponential — Agentic underwriting software for nearly 600 insurers; Guidewire is buying it for $1.7 billion cash, around $2 billion with earnout; reported by TechCrunch.
The week's theme, explained
Anthropic S-1/A — Anthropic filed amended S-1 papers on November 2 with its third-quarter numbers: roughly $17 billion of quarterly revenue, twelve times a year earlier and more than the first half combined, against a $23.29 billion net loss that comes almost entirely from revaluing convertible notes — largely Amazon's — rather than from operations. The filing also lists purchase and lease commitments of about $895.4 billion, nearly all for compute and infrastructure. Banks are quoting investors a $1.4–1.5 trillion valuation for an IPO targeted before Thanksgiving, reported by Bloomberg, via Trending Topics.
This is the first complete public account of a frontier lab's economics, and it anchors every AI mark in a fund's book: twelvefold annual growth, a gross margin reported by Trending Topics around 85 percent, and a capital bill so large it dictates the IPO. For portfolio companies buying models, it confirms inference budgets keep flowing to a handful of hyperscalers. The risk factors now say Anthropic may imminently meet its own Responsible Scaling Policy threshold for "dramatically accelerated progress attributable to AI" — no current model is designated over it — and that it is weighing policy updates to keep humans in control of frontier development; TechCrunch notes skeptics read such lines as marketing as much as disclosure. SEC EDGAR
OpenAI cuts its raise target as big investors hold back. OpenAI has trimmed the target for its pre-IPO round from $30 billion to $20 billion, with a first close expected lower still, reported by the Financial Times via TechCrunch on October 31; the Wall Street Journal says $11 billion is secured and that some large investors responded tepidly. Investors from the UAE lead the round, the FT said.
Two weeks after Anthropic's prospectus, the two largest private AI raises are repricing together — a signal for late-stage marks, not a trading call. If the strongest asset in the category has to cut its ask, growth-stage AI valuations anchored to lab comparables deserve a harder look in the next markup memo. TechCrunch ties the coolness to Wall Street's new wariness of heavy AI infrastructure spending after a tech sell-off. TechCrunch
The pattern behind the numbers
Fewer, bigger cheques — and two-thirds of Series B money is AI. Crunchbase counts about $48 billion into U.S. Series B financings this year, a multiyear high, yet round counts are heading for a multiyear low: at least 110 deals of $100 million or more make up roughly a quarter of the total, and just 26 rounds came in under $10 million. Two-thirds of that money went to AI-focused startups, up from about half in 2025 — concentration, not breadth.
Judge a Series B pipeline on traction, not narrative: sub-$30 million rounds are disappearing, so companies without jumbo-A momentum face a thinner market at this stage than the headline dollar totals suggest. Crunchbase News
Board-meeting questions
Anthropic's GSA discount — Forty percent off Claude across every product for federal, state and local government, no minimum spend, usage-based billing, announced October 27, with roughly 500 government organizations already using Claude and discounts estimated to save $1 billion. Board question: does each portfolio company's public-sector price survive a frontier lab undercutting by 40% on pay-for-use terms? Good: the moat is workflow, integrations and compliance on cheap models. Worrying: the premium is for access to AI itself. Anthropic
Mistral Large 4 — Mistral released the weights of a one-trillion-parameter model on October 29 and prices the API at $0.70 per million input tokens and $2.10 per million output, under a license restricting model-as-a-service resale. Board question: which of our theses assume frontier capability stays closed and expensive? Good: the answer names moats that survive open weights — proprietary context, workflow depth, distribution. Worrying: it leans on exclusive model access or a fine-tune. Mistral AI
Squarespace's restructuring — Squarespace is cutting about 19% of its workforce, roughly 300 people, as prompt-to-site builders upend website creation — Lovable raised $400 million in August at a $13.3 billion valuation — while it readies an AI-native platform, Forma. Board question: for each portfolio company, which incumbent's budget do we take, and what is that incumbent's Forma? Good: names the incumbent and a speed advantage that survives their roadmap. Worrying: "we don't really have incumbents." TechCrunch
Are they keeping up?
Pricing: consumption billing is an operating model now. ICONIQ interviewed operators at Anthropic, ElevenLabs, Glean, Legora and Snowflake on what changes after a company moves to consumption pricing. The numbers: AI-native companies under $100 million ARR run 56% gross margins against 70% for non-AI peers, hybrid pricing is standard at roughly half of those surveyed, and half still set sales quotas on upfront-commit ACV while 30% now quota off forecasted consumption.
Behind: pure seats, finance touching price only at renewal, usage forecast by hand. On par: hybrid pricing, machine-led forecasts, some comp tied to consumption. Ahead: usage telemetry run as a product capability, with a CFO close to pricing — Glean's now spends much of his week on it — and RevOps spending close to 30% of its time on data, against 21% at seat-based peers. Snowflake once had a routine performance improvement erase $100 million from the forecast overnight. The sign to check: who owns the usage forecast, and could an engineer's optimization silently wipe out committed revenue? ICONIQ
Pricing: outcome billing is pulling budget away from seats. ServiceNow grew subscriptions 21% to $3,988 million in Q3 and still saw its shares drop 26.9%, with gross margin down to 70.5% from 77.3%; the next morning Sierra said it added over $150 million of ARR in the quarter and passed $400 million, charging for outcomes — a claim paid, a sale closed — rather than seats.
Same B2B budget, re-split: Gartner's April forecast has 2026 software spending up 15.1% while agentic AI spending grows 141% to $201.9 billion. The private-market read: seat-priced products need an answer for why every efficiency they sell doesn't eventually shrink their own seat count. Lemkin's bar for agent companies: land deals at least twice the size pre-AI competitors win, or you are splitting the budget rather than taking it. SaaStr
Pricing: three tests before the next price increase. Erten's a16z essay takes the bind cheaper inference creates — application margins improve, but so does the customer's best in-house alternative — and turns it into three tests: beat that alternative across the whole workflow, hold context that costs less to rent than to rebuild, and measure the full cost of a finished unit of work before setting price.
A ready agenda for pricing reviews. The budget is there: 94% of the 100 enterprise LLM decision-makers a16z surveyed in August 2026 expect spending to keep climbing, and 37% see it doubling or more inside a year — but only demonstrated work captures it. At Repsol, a blind eight-week pilot across 50 lawyers in 12 jurisdictions had Harvey beat both Microsoft Copilot and Repsol's own prototype; lawyers now save four to six hours a week, at 96% adoption. a16z
Gemini 4 Carbon — Google DeepMind announced Gemini 4 Carbon on October 28 — its next frontier model, aimed first at cybersecurity work — at an introductory $2 per million input and $10 per million output tokens, rising to $4 and $20 once the intro period ends, with early access limited to cyber defenders through the Fairwind Program.
Two frontier releases in two weeks at falling prices is the input-cost side of your portfolio's P&L: any company whose COGS model locks in teaser token prices is overstating next year's gross margin. Security-AI startups should also note a frontier lab aiming squarely at their category. Cached input is priced 95% off the input rate — a real lever for agent workloads that repeatedly re-read context. Google DeepMind