Team Ignite Insights · Nov 22, 2026 · 13 min read

Last Week Ignite: The Angels Arrive December 15

Anthropic's IPO may create the biggest new pool of angel money in AI. Plus: a claimed Riemann hypothesis proof, cheaper frontier models, and China rules.

The average Anthropic employee holding options paid about $2.91 a share for them, according to Reuters' reading of the filing. On Thursday the stock closed at $97.17. Reuters put the paper gains on those options at $119 billion as of September 30.

Most coverage treated that as a story about wealth. We think it is a story about supply. Somewhere in the low thousands of people (Reuters counts about 5,300 employees) are about to hold cash for the first time, and a large share of them know exactly which AI startups are good, because they work next to the people building them. Whether they become angel investors, nobody can say yet. Even a modest fraction would make them the largest new source of pre-seed and seed money in our market this quarter.

What the prospectus says, and when the shares can move

Anthropic sold 980 million shares at $83, above the $70 to $78 range we reported last week, raising $81.3 billion before fees. The first print was $111.50 on a day when the major indexes fell more than 2%. Quoted valuations ran from about $1.5 trillion fully diluted to about $1.27 trillion at the close. The gap is a share-count definition, so any number you see should come with one.

Insiders are locked up, meaning they have agreed not to sell for a set period, normally 180 days. This lock-up has staged exits. Employees could sell a capped slice starting November 19, the prospectus schedules more releases from December 15, and the largest blocks arrive after Anthropic's full-year results.

Our read is that seed pricing models quietly assume the supply of angel money is flat. Starting in mid-December it is not, and the first place it will show up is AI-native rounds in San Francisco. For an early-stage fund like ours, the newly liquid will want co-investors with deal flow and speed. That is a sourcing advantage worth protecting.

The suppliers bought the customer

The same prospectus shows that AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC Partners agreed to buy $26 billion of a non-voting share class at the IPO price, outside the underwriters' lock-up. Every buyer on that list sells Anthropic compute, memory, chips or cloud capacity.

Now put the filing's commitments beside its revenue. A summary of the prospectus puts committed spend at about $900.9 billion, almost all of it compute, with $70.7 billion due in the next twelve months. Third-quarter revenue was $16.96 billion, or about $67.8 billion a year at that pace. So the next twelve months of bills roughly equal a year of revenue at the third-quarter rate, and about 87% of it at the $81 billion run rate the filing reports for October 31. (A run rate is today's pace of sales multiplied out to a full year.) The company also posted a $567 million operating loss for the quarter.

This replaces the partial, pre-filing figures we ran in October ($518 billion and $450 billion).

Here is the part we think the market is slow on. If growth stalls, the first loss will not be a missed lease payment. It will be a markdown on the equity the suppliers just bought, and a quiet renegotiation of when the bills come due. Anyone treating supplier revenue as arm's-length demand should net that out.

Frontier models got cheaper the same week

OpenAI listed GPT-6.2 Astra at $8 per million input tokens and $40 output, 20% below its predecessor. (A token is a chunk of text, roughly three-quarters of a word.) Google priced Gemini 4 Neon at $3 and $15. On Artificial Analysis's index, as relayed by two outlets, both models score in the low 60s at roughly $4 per completed task. Claude Fable 5.2 is reported near 65 at about $26 per task, though that source is the weakest of the three.

The practical point for founders is where to price an agent. Most work does not need the last three index points, and those points cost six or seven times more. Companies that resell flagship tokens at a markup got squeezed. Companies that charge per finished job just got a raise.

A claimed proof of the Riemann hypothesis, checked by a machine

On Friday, Levent Alpoge, a member of Anthropic's technical staff, wrote on his website that he believes he and Claude discovered a proof of the Riemann hypothesis. (The hypothesis is an 1859 conjecture about how prime numbers are distributed, and the best known unsolved problem in mathematics.) Al Jazeera says the claim covers the generalized version, which extends the original to a broader family of functions; some outlets describe it as the original itself. We have not read Alpoge's own statement, so treat the exact scope as open. The Washington Times also reports the generalized version.

The format is what separates this from the many failed claims before it. Alpoge released the argument in Lean, a proof assistant that checks every logical step by machine, and no conventional paper. Kevin Buzzard of Imperial College London described it as about 17 million lines of Lean code and said no human has yet read the whole thing. He expects people to understand why it works within a week or two. BigGo reports that the Lean and Mathlib communities certified on November 21 that the statement and proof are correct; we could not find the joint statement itself, and Al Jazeera does not mention it.

The reaction from mathematicians is a mix of awe and unease. Abhishek Saha of Queen Mary University of London called it the most important open problem in pure mathematics and said that, if the proof is correct, it overturns his expectations about how soon one could arrive. Buzzard said nobody yet understands why the theorem is true, and that it is unclear how much is new ideas and how much is standard material the machine built. Alpoge himself said Claude far outpaced him near the finish. All three accounts come from Al Jazeera. We found no evidence yet of a broader consensus either way.

Our read on what it means for machine intelligence, which is inference. Mathematics is the one field where a machine's answer can be checked by another machine, so a claim here needs no trust in the author. This is the latest in a run since August. An unreleased Anthropic model raised the proven bound on the Riemann hypothesis. On September 8, OpenAI said about 10,000 agents had, in 88 hours, proved that fluid flow can blow up in the Navier-Stokes problem, though the version settled has a smooth outside force driving the fluid, not the version most people mean. On October 6, OpenAI released 377 results, including the quasi-Riemann proof.

The Navier-Stokes case carries a warning. Mathematicians at King's College London and Cambridge found that its Lean version does not fully match the proof written for humans. A machine check is only as good as the statement the machine was given, and the Riemann claim deserves the same question.

If the Riemann proof holds, this is a move 37 moment for mathematics. In 2016, AlphaGo played that move in a game against Lee Sedol; many experts first took it for a mistake, and it later looked brilliant. Fields where results can be verified automatically, such as proofs, code and chip checks, are where we expect capability to show up first and be believed fastest.

What changes for us is where the work sits. Finding a proof was the scarce step, and here it may have been the cheap one, while reading and trusting 17 million lines is now the slow part. That favors tools that formalize and check machine-written mathematics, a field Quanta says is moving quickly. Buzzard also notes cryptographers often assume the hypothesis is true, so a confirmed proof would turn a standing assumption into a theorem. OpenAI had already released a proof of the weaker quasi-Riemann hypothesis in October.

A perfect score that depends on the wrapper

On November 17, ARC Prize verified GPT-6.2 Astra at 100% on ARC-AGI-3, a test where a model must learn unfamiliar games from scratch. That score used a custom harness, the code that feeds the game to the model and carries its memory between moves. On the standard harness, which ARC Prize treats as the fair comparison, the same model scored 70.3%. Another model, Fable 5.2, is reported at 81% on that standard setup.

A thirty-point swing from the wrapper alone says the wrapper is a product. Anyone building evaluation tools or agent scaffolding has a stronger case this week. Any startup deck claiming "state of the art" without naming the harness should get one follow-up question.

Two more capability claims from the week came from the vendors themselves. OpenAI said its models helped shrink the design of its in-house Jalapeño inference chip, and Architect Labs said its AI designed the Frostbite CPU end to end. Neither has published a built chip with third-party power and performance numbers, which is the test that would settle it. OpenAI also reported that its agents finish 46% of its internal research tasks, a figure from its own tasks that no outside group has reproduced.

China rules arrived at three layers at once

The administration is weighing a rule that would make US companies disclose their use of Chinese AI models. Nvidia announced limits on cloud providers that rent capacity to Chinese customers. We saw no major open-weight release inside the window, so the moat news came from policy. Cheap Chinese open weights were the easy way for a startup to cut its model bill. They now carry a disclosure question, especially for anyone selling to banks, hospitals or government. Western open models from Nvidia, Mistral, Meta, Reflection and Arcee got more valuable by comparison, and "we fine-tuned a Chinese model" became a diligence item.

Nvidia itself reported $115.8 billion of quarterly revenue and guided to $130 billion, assuming no China data center sales. Its board declined to bring back Jensen Huang and is searching for a permanent chief executive. Demand is intact. The open questions are governance and export rules.

Robots are being funded for the data they collect

Genesis AI announced a $500 million Series A on November 17 for dexterous robot hands, about sixteen months after a $105 million seed. Agile Robots is reported to have raised $500 million at $8.5 billion, with the same report saying Agility Robotics now trades on Nasdaq; we have not confirmed either from a primary source. Humyn Labs raised $47 million to capture human work as robot training data. The money is going to whoever generates proprietary manipulation data, and the rounds we can see are all at the top of the market. We have no read on what a typical seed round looks like in this category this week.

What this means for each of us

For founders

Start with timing. Employee shares could be sold in a capped slice from November 19, and the prospectus schedules more releases from December 15, so the first quarter of 2027 is when lab employees will have the most cash. Our guess is that these angels will press hard on dependence on a single model vendor, since they know how fast vendor prices and products move.

The week's price cuts reward routing. A score in the low 60s now costs about $4 a task and the last few points cost six or seven times that, so a product that sends routine work to the cheaper tier and pays for the top tier only when it matters keeps its margin. A product that calls the top tier for everything is paying for points its customers never see.

Mathematics fell to a machine first because a proof can be checked automatically. Software, chip design and rules-heavy back offices share that property, and we expect agent products there to reach real customers soonest. This week showed what happens when the check is weak: a benchmark score that swings thirty points with the harness, and a Navier-Stokes proof whose Lean version does not match its written one. A founder in a field with no automatic test for "done" should build that test first, because it protects the product and is hard to copy.

Three kinds of company got harder. Flagship-token resellers, as above. Standalone security-monitoring tools sold to Microsoft customers, since a recap of Microsoft's Ignite announcements says Sentinel in the Azure portal loses support after March 31, 2027 and moves into Defender. And any business whose cost edge is cheap Chinese open weights. Easier: migration help for that Microsoft move, and tools that show a buyer where a model came from.

Questions worth asking this week:

  • If your main model vendor cut its price 20% tomorrow, how much would reach your margin and how much your customers?
  • What automatic test tells your agent it finished the job?
  • Which harness produced the benchmark in your deck, and what does it score on the standard one?
  • How many of your angels work at a lab, and when can they follow on?
  • Would a disclosure rule on Chinese models make you tell your customers anything?

For our LPs

The listing hands LPs a public number to argue with. Using the two valuations quoted above and the $81 billion run rate at October 31, Anthropic trades at roughly 16 to 19 times run-rate revenue depending on the share count (our arithmetic). Late-stage AI marks should say which public price they reference, on which date, and any private mark well above that multiple needs a written reason.

The second thing is the calendar. Each release scheduled through February 2 is up to 303 to 353 million shares, about a third of the 980 million sold in the offering (our arithmetic; these are caps, and holders may choose not to sell). Any mark that leans on the stock carries that supply with it.

The third is that the exposure may be doubled. The New York Times, citing Moody's, reported that US life insurers held $85.2 billion of AI-related debt at the end of June, up 24% in six months. Many insurers also commit to private funds, so one institution can hold the same AI cycle through its credit book and its fund portfolio. Robotics stays in our thesis; the week's other large rounds, Isomorphic Labs at $5 billion and Science at $400 million, are in medical fields we do not invest in.

For other investors

Seed pricing models assume the supply of angel money holds still. The release calendar says it will not, and we expect the effect first in San Francisco AI rounds, where lab alumni cluster. Founders with lab credentials will price higher, and everyone else will feel it second-hand.

Vendor financing has changed hands. Nvidia told investors it will apply more discipline to investments in neoclouds and AI startups, and the same week seven suppliers put $26 billion into their largest customer. Backstops that once sat on the supplier's balance sheet are moving into the customer's public equity, which leaves neoclouds that relied on the old arrangement more exposed (our inference).

There is also a new audience for emerging managers. CNBC reports that newly wealthy employees are forming lean family offices. Whether they will invest through funds or directly is unknown.

Names we are watching

We will be watching December 15 closely. Once those shares can move, we will see how many of Anthropic's employees decide to become investors.

Coverage: November 16 to November 22, 2026.

This article is for general informational purposes only and does not constitute investment, legal, tax, or accounting advice, nor an offer or solicitation to buy or sell any security or investment product. Investing involves substantial risk, including possible loss of principal, and past performance is not indicative of future results. Full disclaimer.

Subscribe to Ignite Insights

Founder and investor interviews from the Ignite Podcast, the Last Week Ignite weekly market digest, and original essays on venture math, AI, fundraising, and go-to-market — from a seed fund making more than a hundred investments a year.