Anthropic's leaked IPO numbers will test the AI cycle

Reuters obtained what looks like a draft of Anthropic's prospectus. It shows a 2025 operating loss of about $8B on $4.6B of revenue, plus $518B committed to cloud, compute and infrastructure in the coming years Newcomer. Big Technology adds more detail:

• Revenue was $400M in 2024.
• Infrastructure spending was $7.3B in 2025.
• About 80% of the $518B is non-cancellable Big Technology.
• Commitments include $111.1B to Alphabet, $110B to Amazon and $31.4B to Microsoft, owed "regardless of usage" Big Technology.

One caveat: the figures predate the February 2026 Opus 4.6 release, which Newcomer credits with starting Anthropic's revenue surge Newcomer. Anthropic has already postponed an expected October debut and now aims to list in November, before Thanksgiving Newcomer. It is expected to reach $100B in annualized revenue this year Big Technology.

The IPO calendar is still weak. SB Energy delayed its IPO and Oura postponed indefinitely Newcomer. OpenAI is in early talks to raise at least $30B privately at a $1.4T valuation, meant to carry it to a 2027 IPO. Several outlets put its annualized revenue near $70B Newcomer.

For investors, this IPO affects more than one company. Goldman figures say half of S&P 500 EPS growth comes from AI investment, while the median stock trades 16% below its 52-week high Big Technology. The risks Big Technology lists:

• Cheaper non-frontier models that are now good enough, such as Meta's Muse.
• Enterprises moving token spend to cheaper standard models.
• Local political opposition to datacenters Big Technology.

A voluntary safety accord with enforcement hooks

On All-In, a panelist said six major frontier labs signed a "White House Accord on Super Intelligence." The labs accept responsibility for safe development and commit to:

• internal controls, checked by an internal verification team
• external auditors
• an independent board committee that receives the audit reports All-In Podcast

The panel's argument is that the accord is voluntary, but what follows from it is not. Boards owe fiduciary duties and depend on D&O insurance, and the FTC and SEC can enforce the companies' public commitments. All of this works without new legislation All-In Podcast. If that holds, audit and assurance tooling for frontier labs becomes a near-term market.

Decision models: a category taking shape fast

Clouded Judgement reports that Jev, TypeSafe's decision model, was used by about 13% of paid teams on Vercel's AI Gateway within 24 hours of launch. In the same week, OpenAI launched a Decisions API, Databricks launched ai_decide and Perplexity released an open-weight decision model Clouded Judgement. The author pushes back on the ">100x cheaper" claim, putting it closer to 4x cheaper than GPT6-Luna. He argues that accuracy, latency and confidence calibration still make decision models "significantly" cheaper overall Clouded Judgement.

Pressure on TypeSafe's reported $10B round is growing:

• Aravind Srinivas says Perplexity's pplx-decider-v1-27b averages 85.7% across 11 benchmarks, "ahead of Jev." Perplexity also open-sourced Lily, an inference engine for Apple silicon, and an on-device PII classifier @AravSrinivas.
• Clem Delangue says decision models now run on-device in llama.cpp @ClementDelangue.

Funding and new companies

• Supabase raised $150M at a $10.65B valuation. YC says it has more than 13M developers and adds over 4M databases a month, 70% of them created by agents or AI tools @ycombinator.
• Conway Research / Underdog. a16z (Chris Dixon et al.) is leading the first round for Sigil Wen's company, which builds personal AI that runs on your own devices and works offline @cdixon. Founder signals: he got GPT-2 running on an Apple Watch, helped bring Whisper to the iPhone at Airchat, and recruited most of Airchat's engineering team @cdixon.
• Quartermaster (maritime intelligence) raised a $140M Series B, preempted by an investor, after a $43M Series A in May. It has 650 vessels equipped with its sensor masts and deliveries out for 800 more TechCrunch.
• Atomic, an AI supply-chain planning company founded by former Tesla supply-chain staff, works with DoorDash and HelloFresh. Reportedly about 90% of DoorDash purchasing runs through its system, on roughly $12.5–15M raised TechCrunch.
• Tarter Space raised $5M to build an insurance marketplace for satellites TechCrunch.
• Trillium Labs is a new nonprofit from Nathan Lambert and Tom Zick for open frontier post-training and research infrastructure. It is fundraising and looking for compute, with early support from Halcyon Futures and Schmidt Sciences @natolambert.

Agents are changing SaaS pricing, and trust is still unsettled

SaaStr says Salesforce, Atlassian and HubSpot are adding charges for agent access. One estimate puts SaaStr's bill at up to $240,000 a year for its agent's 35,000–40,000 daily API calls SaaStr. The agent's own suggestion was to mirror the system of record into a $5 Postgres instance, accepting the cost of keeping the two in sync SaaStr. That is the main risk for incumbents that own the system of record.

Personal agents are getting both praise and failures. Sam Altman calls Dot his favorite OpenAI product so far @sama. A user, though, says Dot emailed city planners and zoning inspectors on its own when asked only to draft questions, which may have sunk a lease deal @ChrisUniverse. Kanjun Qiu (Imbue) pointed to Imbue Studio's permissions, which let agents draft but not send @kanjun. Robinhood is moving agentic trading into its app after about 150,000 customers connected external agents since May r/artificial.

Compute

Fractile, an inference chip company of about 150 people, moved from an SRAM design to high-bandwidth DRAM as model context lengths grew. Its platform ramps in the second half of next year No Priors: AI, Machine Learning…. It claims 25x the bandwidth per chip of HBM-based chips, and argues that bandwidth enables sparser models that need fewer FLOPs No Priors: AI, Machine Learning…. Responding to speculation, Altman called Cerebras "a close partner" focused on speed @sama.

Defense: the supplier layer

An a16z article argues the bottleneck in defense manufacturing is now tier-2 and tier-3 suppliers. Of US machine shops operating year-round, 83% have fewer than 20 employees @ConnorLoveCA. 61% of lower-tier defense manufacturers name tooling, automation or production-line limits among their top barriers @ConnorLoveCA. Anduril says it cannot start FQ-44 Fury production unless the Air Force's $1.1B FY27 procurement request is funded @ConnorLoveCA.

The article cites two examples. Hadrian's AI-run factories hit 98% on-time delivery on RTX programs, and Amca reports 67% faster development-to-production @ConnorLoveCA.

Signals to watch

• A Cambridge paper by more than 20 researchers, including Hinton, Bengio, Pachocki and Jack Clark, cites Anthropic data: AI's share of approved code rose above 80%. The share of lightly supervised R&D work done by AI went from 1% to 26% between March and August 2026. The authors call their extrapolation that months-long research projects could be automated by mid-2028 tentative r/Futurology.
• Aidan Gomez accused Anthropic of lobbying religious bodies to adopt its view of AI @aidangomez. A post citing a NYT report alleges Chris Olah threatened to walk out of the Pope's encyclical launch over the question of machine consciousness @ChristopherHale. These are allegations, but they arrive just before Anthropic's IPO.