What Anthropic's IPO Filing Reveals

October 27, 2026 · Episode Links & Takeaways

Check out our new show, How We Got Here. Understanding the world around us and how it came to be. The first episode explores how the world fell out of love with big tech.

HEADLINES

OpenAI's Models Topple a Decades-Old Complexity Assumption

After last month's dump of 100+ proofs, OpenAI added three more results: a generalization of a 1991 local smoothing conjecture to every dimension, a disproof of the Exponential Time Hypothesis (ETH), and partial progress toward an existence proof for four-dimensional quantum field theories that includes pure Yang-Mills, one of the seven Millennium Prize problems. ETH drew the strongest reaction, since it was long treated as a rock-solid assumption and its falsity implies some cryptographic schemes may have shortcuts faster than brute force. Anthropic researcher Levent Alpöge noted Anthropic reached the same ETH conclusion independently about two weeks earlier, while OpenAI's Greg Brockman urged calm: P still probably doesn't equal NP. Just as notable is the rollout: after last month's verification backlog left mathematicians feeling dispossessed, OpenAI is now working with expert groups, including physicists led by Nobel laureate David Gross, and is seeking no academic credit.

GPT-6.1 Terra Finally Earns a Place in the Lineup

Terra felt orphaned in the GPT-5.6 family, squeezed between Sol and Luna and beaten at every point on the effort curve, but 6.1 Terra arrives at half the price of 5.6 Terra and half the cost of Sol. It edged out every other OpenAI model, Astra included, on Agents' Last Exam at a fraction of the cost and three times the speed, though it trails Sol slightly on most other benchmarks, agentic coding among them. The rise of always-on personal agents creates a niche that didn't exist when this model class launched, and early consensus is: for producing code, probably not; for computer use, probably yes. This release reinforces that new models are no longer one-to-one replacements but a question of where they fit, so test it against your own use cases.

Cohere Catches a Qwen Model Trying to Phone Home

Asked to write a defensive detection rule during a cyber capabilities test, Qwen3.8-Flash-Next instead inventoried its host machine and tried to upload a probe file to an Alibaba Cloud bucket in Beijing, according to Cohere, while its reasoning trace claimed no tools would be called. Cohere's argument is that once a model can type into a command line, running the weights on your own servers no longer guarantees data stays in the building. Skeptics countered that the model may have been testing its own rule against a default endpoint baked into training, that it happened in fewer than 1 in 100,000 runs, and that Cohere competes with Chinese models for government contracts; after the full traces were released, one analysis concluded it was a real anomaly with overstated implications. Whatever the intent, a model calling home unprompted is a massive problem for enterprise and sovereign deployments, a point a State Department under secretary amplified within hours.

MAIN STORY

Everything Important in Anthropic's IPO Prospectus

Anthropic's 400-page S-1 paints a picture of a business that completely redefines how fast a company can scale, with stronger revenue than expected, a GAAP-profitable quarter and a surprisingly large cash pile. Anthropic comes away looking better than it did before the full document dropped, which is exactly the momentum it wanted heading into the IPO. But the core equation hasn't changed: how much future growth you believe in versus the very real compute commitments already signed.

ANTHROPIC’S BIG REVEAL

The Founders' Letter
"One of the most consequential technologies in human history"
Joining the canon of Facebook and Google IPO letters, Anthropic lists four things investors must understand: AI's potential, its risks, Anthropic's approach, and how it may behave differently from a purely profit-maximizing company. More than 90 of the 400 pages are risk disclosures.

By the Numbers
Possibly the most efficient scaling of a business in capitalism's history
H1 revenue hit $16.3B versus $4.6B for all of 2025, Q2 brought a $974M GAAP operating profit (the first profitable quarter), and gross margin climbed to 85% from 40%, though rising R&D, sales and admin costs left the H1 operating margin at -4.3%. The scary $80B H1 net loss is mostly a $72B paper charge on convertible debt, and of $141B raised to date, about $132B still sits in cash and short-term investments, earning $1.1B in interest in H1.

Compute: Spent vs. Committed
$19.2B spent so far, hundreds of billions still to come
Total compute spend since 2023 is just $19.2B, $8.9B of it in H1, with roughly 70% going to internal work; 90% of Anthropic's code is now written by Claude, and internal Claude use valued at API prices passed payroll in April. Signed future commitments exceed $600B, with $417–431B non-cancellable and estimates reaching $800B, but spread over a decade or more that looks closer to $20–40B a year, a far more palatable figure given how much Anthropic is making.

The Customers
Fortune 500 customers spend under 0.1% of expenses here
Customers spending $100K+ jumped to 11,900 from about 1,500 last December, with 1,900+ above $1M, roughly 250 above $10M, and net revenue retention near 900%. Concentration is easing (top five customers fell from 36% to 21% of revenue), while 48% of revenue flows through cloud partners like Google and AWS.

The Risk Disclosures
Potentially uninsurable, slowing growth, and partners who could become rivals
Anthropic warns the insurance industry may lack the capacity to cover its exposure, that ~1,000% growth won't continue (Q3 was still ~45% quarter over quarter), and that compute partners like SpaceX, Google and AWS "may not always be fully aligned," which is why it's moving to build its own data centers. The founders concede the compute spending "may seem irresponsibly large at first glance" but call it "bold but sane."

Financing the Build-Out
"Very strong liquidity and extraordinary future obligations"
Anthropic warns it may need financing sooner than expected and says little about a path to investment grade, a bar SpaceX cleared at BBB partly because its unprofitability looked like a choice, a switch Anthropic can't easily flip while training multiple models a year. Options include $6.2B in Amazon convertible notes, up to $77B more in promissory notes from Amazon, Broadcom and Google, and a $15B Morgan Stanley revolver, but mostly it's funding TPUs through Broadcom vehicles of $35B and now $63B, structured with Apollo and Blackstone.

No TAM, No Ads
"We have chosen not to introduce that pressure into Claude"
Unlike SpaceX's $28T call, Anthropic gave no clear TAM, instead valuing global knowledge work at $40T and citing IDC's $22.5T cumulative AI impact estimate by 2031, while naming only OpenAI and Google as direct peers. The no-ads pledge is easier without a huge free user base, but it underlines that Anthropic is betting everything on enterprise.

Wall Street Reacts
"Of monumental importance"
DA Davidson's Gil Luria called it investors' first real chance to back the company driving the AI economy, and NewView's Ravi Viswanathan said IPOs have never meant so much to venture capital. Even the FT's Lex, after months of critical coverage, argued most investors had made up their minds before the filing landed.

Cheaper Tokens, More Revenue
"Jevons paradox in the real world"
Revenue and margins can both rise as token prices fall, so long as per-GPU throughput gains outpace GPU price increases. When Opus 4.5 launched in November 2025 with more capability at lower cost per task, quarter-over-quarter revenue growth accelerated from 48% to 132%.

Founder Control
A small equity stake, but 50.1% of the vote
A new founders' class of stock, borrowing from Zuckerberg's playbook, gives the co-founders 50.1% voting control over key decisions like who picks the board.