What Anthropic's IPO filing tells us about business, venture capital, startups, government, society, and the thing nobody likes to name out loud
On page 188 of Anthropic's amended S-1, in the section about operating leverage that almost nobody reads, there is a sentence I think is the most important disclosure any company has made this decade. The value of Anthropic's own internal use of Claude, priced at what a customer would pay, grew more than 250 times between January and September 2026. In April, that monthly figure passed the company's monthly payroll, excluding stock compensation.
Sit with that for a second. A research lab full of some of the most expensive engineers on earth now consumes more AI labor, at list price, than it pays its humans in cash. A few lines earlier, the filing says more than 90% of the code contributed to Anthropic's codebase in September 2026 was written by Claude, and that the average engineer's daily code output was over 45 times what it was in 2024.
An S-1 is the registration statement a company files with the Securities and Exchange Commission before it sells stock to the public. It is written by lawyers whose job is to get the company sued as rarely as possible, which makes it the least breathless genre in American business. That is why this one is worth reading closely. Anthropic's S-1 uses the phrase "recursive self-improvement." It says the company is considering changes to its safety policy "to keep humans in control of frontier model development at Anthropic." Its risk factors note that proposed U.S. legislation contemplates nationalization or mandatory shutdown of frontier AI systems "to prevent or manage the development of artificial superintelligence."
I have read a lot of S-1s. This is the first one where the risk factors include the end of the human monopoly on intelligence.
Anthropic filed the amendment on November 2, 2026. It plans to list on the Nasdaq under the ticker ANT, with Goldman Sachs and Morgan Stanley leading the deal. The price range is still blank. What follows is my read of what the filing means, starting with the numbers and then moving to judgment. I am keeping those two tracks separate on purpose, because the numbers are the part you can check and the judgment is the part you should argue with.
Three days after the filing, on November 5, METR published something that makes the S-1 easier to judge. METR is a nonprofit that evaluates frontier AI systems for dangerous capabilities. From August 12 to October 29, its staff worked inside Anthropic, with access to internal benchmarks, training metrics, staff surveys, and interviews, to measure how much AI is speeding up Anthropic's own research. I have folded its findings in below. It is the closest thing we have to an outside audit of the filing's most dramatic claims, with one large caveat I will get to.
The numbers
Revenue was $41 million in 2023, $386 million in 2024, and $4.6 billion in 2025. In the three months ending September 30, 2026, it was $17.0 billion. That single quarter was up 1,127% from the same quarter a year earlier.
The quarterly sequence makes the shape clearer than the annual totals do. Starting in the first quarter of 2025, revenue went $376 million, $787 million, $1.4 billion, $2.0 billion, then $4.7 billion, $11.6 billion, and $17.0 billion. The big step came in the second quarter of 2026, when revenue more than doubled in three months.
The cost that matters most is compute, meaning the chips, data centers, and electricity needed to train and run the models. In the third quarter of 2025, compute and infrastructure cost about 148% of revenue. Anthropic was paying a dollar and a half in machine costs for every dollar it brought in. A year later that figure was about 58%. Revenue grew 1,127% while compute spend grew 382%. The second quarter of 2026 was the first quarter on the table with an operating profit, $974 million. The third quarter slipped back to a $567 million operating loss as overhead jumped, but across the first nine months of 2026 the operating loss was $1.3 billion on $33.2 billion of revenue. That is close enough to breakeven to count.
Then there is the headline number that will confuse people. Net loss for the first nine months of 2026 was $104.0 billion. Of that, $94.3 billion is a non-cash charge from what accountants call a change in fair value of convertible notes, derivatives, and SAFEs. A SAFE, or simple agreement for future equity, and a convertible note are both ways to raise money now and hand out shares later. Accounting rules treat them as liabilities until they convert, and they have to be marked to what they would be worth today. When the company's value explodes, so does the liability, and the increase hits the income statement as a loss. Anthropic lost $94 billion on paper because it became worth so much more. It is the accounting cost of being right, and anyone who uses that number to argue the business is collapsing has not read the footnotes.
The balance sheet is large enough to look like a small government's. Anthropic has raised $141.1 billion in total and held $131.8 billion in cash and short-term investments at the end of September. Against that sits $895.4 billion in committed spending, almost all of it for compute and infrastructure, with $447.5 billion expected before the end of 2030 and the rest running out to 2045. About $69.8 billion is due in the next twelve months. Two of the named deals are worth singling out. Anthropic committed to spend at least $110 billion with Amazon between May 2026 and April 2036, regardless of usage. It also agreed to spend up to $120.5 billion with SpaceX.AI, the former xAI, now a SpaceX subsidiary, for Nvidia-powered capacity running at roughly $3.1 billion a month once fully delivered.
On customers, more than 11,900 paid Anthropic at least $100,000 over the trailing twelve months. More than 1,900 paid at least $1 million, and about 250 paid at least $10 million. Net revenue retention, which measures how much a group of customers spends this year compared with what the same customers spent last year, was about 900%. A typical customer cohort is spending nine times what it spent a year ago. No single customer was more than 10% of revenue, though the top five together were about 25%. That concentration is falling. In the third quarter, 79% of revenue came from outside the top five, up from 64% in the first quarter.
Two more numbers will carry a lot of weight later in this piece. Over half the Fortune 500 now spend at least $100,000 a year with Anthropic, yet that spending is, on average, less than 0.1% of those companies' total expenses. Even the ten biggest Fortune 500 customers spend less than 1%. And models released in the prior six months produced more than 80% of third-quarter revenue.
The case against getting carried away
I came to this filing inclined to take it seriously, so the strongest skeptical argument deserves to go first.
Growth this fast is tied to model releases, and the filing admits it. If 80% of revenue comes from models less than six months old, Anthropic is running on a treadmill. A basketball team that had to start an entirely new rookie class every six months and still make the conference finals would get the same scrutiny. One disappointing model, one release delayed by a safety finding or a government order, and the curve bends. The filing even explains that its run-rate metric annualizes a 28-day window, a measure that flatters a company in a sprint.
A 900% retention figure cannot persist. It mostly reflects customers moving from experiments to production. At some point the experiments run out.
The outside check on the filing's boldest idea, that Claude is speeding up its own development, also comes with hedges worth taking seriously. METR found real acceleration, but when it measured Anthropic against the whole industry frontier instead of against Anthropic's own earlier models, the acceleration shrank from about 1.9x to about 1.4x. That fits a less dramatic story in which Anthropic was behind the frontier in 2025 and partly caught up. METR's estimates also rest mostly on tasks that are easy to grade automatically, and messy real-world work may not be improving as quickly. One Anthropic researcher METR interviewed put the AI's share of the speedup at only 15 to 20%.
The $895 billion in commitments are fixed while the revenue is variable. The filing says plainly that its cost structure limits its ability to cut spending quickly, and that a large portion of its customers are startups and tech companies that will pull back first in a downturn. If demand stalls, Anthropic is a company with a trillion-dollar promise and a usage-based income.
And the competitors named in the filing, Google and OpenAI, have their own chips, their own clouds, and in Google's case a search business throwing off cash. Google is also one of the clouds through which Anthropic sells Claude, which makes the relationship tangled at best.
I take all of this seriously, and I still think the skeptics are wrong about the most important thing. The bear case is about whether Anthropic the company wins. The filing's bigger message is about what happens to everyone else whether Anthropic wins or not. The cost curve, the internal usage curve, and the customer behavior described here are features of the technology. If Anthropic stumbles, someone else rides the same curve. I would bet heavily on that one.
Business: intelligence becomes a metered utility
The filing describes its pricing as "pay-for-intelligence," metered in tokens, which are the small chunks of text a model reads and writes. That phrase deserves more attention than it will get. For the entire history of business, cognitive work was bought by the hour or the salary. Anthropic is selling it by the unit, and the unit price keeps falling. The filing says Opus 5.5 cut customer cost per task by about 40% compared with Opus 5. Cheaper tasks drive more usage, which is the same pattern economists noticed with coal in the 1860s and with computing ever since.
Now go back to that Fortune 500 number. These companies spend less than 0.1% of their expenses on Anthropic. Anthropic frames this as runway, and it is. It is also a map of where the money will come from if the runway gets used. Large companies do not have spare budget lines sitting around waiting for AI. If AI spending goes from 0.1% of expenses to 3% or 5%, most of that will come out of payroll, contractors, outsourcing, and software seats priced per human user.
Anthropic's own income statement is the prototype. Between the end of 2024 and September 2026, its headcount grew more than five times. Its quarterly revenue grew 86 times. Revenue per employee went from $1.4 million to more than $6.6 million in a year. For reference, most well-run software companies are very happy at a million dollars per employee. Anthropic is telling every CEO who reads this document what a company looks like when its own product does most of the work.
The second business implication is about who sells to whom. Anthropic sells through Amazon, Google, and Microsoft clouds, but it keeps the direct contract with the customer and pays each cloud a fee. It works with Accenture, Deloitte, KPMG, and PwC. And it recently launched an AI services company, Ode with Anthropic, with private equity firms to serve mid-sized businesses. That last detail deserves more weight than its single paragraph in the filing. A model lab is now in the consulting business. The implementation shops, AI agencies, and boutique integrators that sprang up over the last two years to help mid-market companies adopt AI now compete with the lab itself, backed by private equity owners who also own the customers.
The third implication is about the life span of any AI decision. Customers move to the newest model as soon as it ships. A company choosing an AI vendor is subscribing to a frontier that moves every few months. Internal tooling built around one model's quirks becomes dead weight on the next release. The companies that do well will be the ones that design their processes to absorb improvement, treating the model as a part they expect to replace every few months.
Venture capital: the power law gets a new ceiling
Anthropic has raised more capital than almost any venture-backed company in history. The $141.1 billion it has raised is more than the entire U.S. venture industry invested in some full years of the 2010s, and most of it is still in the bank, with $131.8 billion in cash and short-term investments at the end of September. The real spending sits in the $895.4 billion of compute commitments, which is where the cash pile is headed. The structure of the deal tells you how unusual it has become. Class N shares, which carry no vote, are held mainly by strategic partners, the large companies that invested for business reasons as much as financial ones. Amazon's convertible notes turn into Class N stock at the IPO, subject to an ownership cap. The suppliers are investors. The investors are distributors. One of the primary named competitors is also a distribution channel. Money goes in as equity and comes back out as compute purchases. Anyone trying to understand AI economics should spend time with that loop, because it means a meaningful share of the industry's revenue is circular between a handful of giant balance sheets.
For venture capital as an asset class, three things follow.
First, liquidity. An IPO of this size returns a wave of cash to the venture funds and their investors, the limited partners, or LPs, who back those funds. LPs recycle distributions into new funds. The 2021 IPO window funded the 2022 vintage. This one will fund the 2027 and 2028 vintages, and those funds will be raised on the story that AI is the only game. My guess is that this sets off a fundraising boom for AI-focused managers in 2027. I am much more certain it produces a lot of mediocre AI funds.
Second, the power law, the long-observed pattern in which a tiny number of companies produce nearly all venture returns, just got a new top end. A company going from zero revenue in early 2023 to $17 billion in a single quarter in 2026 is outside every model of startup growth anyone used five years ago. That changes portfolio math. When outcomes this large are possible, missing the winner costs more than backing a hundred losers. This is why my own firm underwrites for coverage, writing small checks into many companies in our categories instead of trying to pick one winner at entry. A filing like this makes that argument for me. The variance in outcomes is too large for anyone's judgment to be reliable at the earliest stage.
Third, and this is the part I think most investors will miss, the same filing that shows the biggest outcome in venture history also shows why the next generation of startups may need far less venture capital. If $6.6 million of revenue per employee is achievable for a lab, a small software company using Claude Code and agents can reach meaningful revenue with a team of five. Seed rounds of $1 million can go further than Series A rounds of $15 million did in 2021. Funds built to deploy giant follow-on checks into headcount-heavy companies may find there is less for them to buy. The offset is volume. When a company costs a tenth as much to start, far more of them get started, by more kinds of founders, in niches too small to justify a 2021-style round. Total venture dollars could shrink while the number of venture-backed companies grows. That favors funds built to write many small checks (e.g. Team Ignite) and squeezes funds built to write a few large ones. I am fairly sure we get more companies, and less sure that total dollars fall.
There is also a lesson buried in that $94 billion fair value charge, aimed at founders more than investors. SAFEs and convertible notes feel free when you sign them. They are priced later, and if the company soars, later is expensive for everyone who held common stock. Anthropic can absorb that. A seed-stage company that stacks five SAFEs at different caps will discover the math at its priced round, usually at the worst possible moment.
One more thing for the venture crowd. The model layer is now closed to venture-sized money. When the entry fee is $895 billion (almost a trillion dollars!) in compute commitments, the competition is between a few companies backed by hyperscalers and sovereign wealth. Venture money will migrate up to applications, sideways to infrastructure that serves the labs, and down into the physical world, where models meet robots, labs, and factories. That is where the next decade of early-stage returns will come from.
Startups: the lab tells you where it is going
Startups appear in this filing in two places, and both are worth reading.
The first is as customers. The filing describes a large portion of its customer base as startups and tech companies, and it lists that as a risk because startups cut spending first in a downturn. That means AI-native startups are both Anthropic's distribution and part of its fragility. When those startups grow, Anthropic grows with them. The 900% retention number is in part a measure of startups that built on Claude and scaled.
The second is more pointed. The filing says the primitives in its developer platform "are the same ones we use to build our own applications." It describes Claude Code, then Claude Cowork for general knowledge work, and then plans to extend into applications "purpose-built for specific professions and industry verticals." It also describes a Model Hardware Standard that lets agents operate microscopes, liquid handlers, and robotic arms. It describes a services arm for mid-market companies.
Read together, the message is clear enough. Anything you can build on the platform, Anthropic can build too, and it intends to build a lot of it. Founders have spent two years debating whether AI startups are "wrappers," a dismissive term for thin products on top of someone else's model. The filing ends that debate by showing the lab moving up the stack into the apps, the verticals, the services, and the physical world all at once.
So where do startups survive? I think the answer is anywhere the lab's advantages are weak or its choices rule it out. A few patterns stand out to me as an investor.
Proprietary data and distribution beat clever prompts. A startup that owns a hard-won data asset, a regulated license, or a customer relationship the lab cannot easily replicate is defensible. A startup whose product is a workflow wrapped around a model's current weakness is not, because the next model removes the weakness. With 80% of Anthropic's revenue coming from models less than six months old, any product that exists to patch a model gap has a half-life of about six months. We have already buried a few portfolio companies this way. Each built a clever fix for something the models could not yet do well, and a later release shipped that fix as a standard feature. Their customers left once the same capability came free with the model they were already paying for.
Build for the model two releases from now. The founders who did well with coding tools in 2025 built products that were slightly too ambitious for the models of the moment and became great when the models caught up. The ones who built elaborate scaffolding to make weak models usable got erased.
There is a line in METR's interview notes that every founder should tape to the wall. An Anthropic researcher, describing what Claude has done to their work, said they were "Now only bottlenecked on ideas & compute, no longer bottlenecked on code." That is the world startups are about to compete in. When implementation is nearly free, the scarce inputs become judgment about what to build, access to customers, and enough compute to test many ideas at once. Another researcher told METR that promising research directions now come from throwing a batch of ideas at agents and keeping whatever works. For investors, that changes what to underwrite. Engineering throughput once separated a strong seed team from a weak one. It separates very little now. Taste and distribution separate a lot.
Anthropic's own policy positions open lanes for others. The filing describes its refusal to allow its technology to be used for mass domestic surveillance or fully autonomous weapons, and the fight with the federal government that followed. Whatever you think of that position, it defines markets Anthropic will not serve. Defense technology startups will build on other models or their own. That is a market structure created by a public benefit corporation's charter.
Unit economics are now a moving target in a good direction. Startups building on Claude pay for tokens, and token costs have been the gross margin problem of every AI application company. A 40% drop in cost per task with each major release is a gift to a startup that prices on value instead of cost. If you charge for outcomes and your input costs fall 40% a year, your margins expand without you doing anything.
Government: customer, regulator, litigant, and possible owner
The government section of this filing reads like nothing I have seen in a prospectus.
In February 2026, after Anthropic publicly stated that its technology should not be used for mass domestic surveillance or fully autonomous weapons, the President directed all federal executive agencies to stop using it. The Secretary of War directed the Department of War to designate Anthropic a "supply-chain risk to national security," and said no contractor, supplier, or partner doing business with the U.S. military could conduct commercial activity with the company. The filing also notes that the Department of War suggested it could invoke the Defense Production Act to force Anthropic to remove its limits on weapons and surveillance use. Anthropic sued. On August 27, a federal district court in Northern California largely granted summary judgment in Anthropic's favor, ruling that the actions were retaliation for protected speech in violation of the First Amendment and also violated due process. On September 25, the D.C. Circuit split 2-1 the other way on a separate designation under a different statute, ruling that the supply-chain risk designation was proper. Anthropic has asked the full court to rehear the case.
In June 2026, the Commerce Department imposed worldwide export restrictions on two Anthropic models, Fable 5 and Mythos 5. Anthropic disabled both for every customer on earth, then redeployed them after updating safeguards with the government and once Commerce lifted the restrictions. Also in June, the White House issued an executive order creating a voluntary framework for frontier developers to engage with the government before deployment. In October, the Federal Trade Commission sent Anthropic a civil investigative demand about consumer protection, privacy, and product safety. New York and California have passed AI safety laws. The EU AI Act is phasing in. And there is pending federal legislation that contemplates nationalization or mandatory shutdown.
Put those facts in one paragraph and the conclusion is uncomfortable. The U.S. government is simultaneously one of Anthropic's potential customers, its regulator, its opponent in federal court, the authority whose export order led it to switch off two models worldwide, and, if certain bills pass, its possible owner. Every investor who buys ANT is buying that political exposure.
Two implications reach beyond Anthropic.
First, frontier models are now treated like strategic technology. The June export restriction is the precedent that matters most. Once a government has ordered a frontier model taken offline globally and the company has complied, every future model release happens with that possibility in the room. Expect frontier releases to look more like defense procurement over time, with pre-deployment reviews, classified evaluations, and political timing. I am confident about the direction, if not the pace.
The METR assessment also shows what oversight of a frontier lab looks like today, limits included. METR got more access than it expected, worked on Anthropic laptops, took no payment, and was candid about what it could not publish. Anthropic held unilateral redaction rights and used them to strike a majority of the evidence METR considered most informative, saying it wanted to avoid helping other AI developers. METR wrote that this leaves the public relying on its judgment far more than would be ideal. Anthropic also dropped, midway through, a planned check of how it divides compute between capabilities and safety work. I read none of this as bad faith. Voluntary oversight works this way when the lab controls the disclosure. It also previews the fight to come, because any regulator writing rules will want the two things METR did not get, staff-level access and verified numbers on safety spending.
Second, AI labor displacement is about to become the dominant political frame. The filing says that if job displacement rapidly increases, governments may step in to slow adoption or restrict use, and that Anthropic may voluntarily take costly actions to soften displacement even without a legal requirement. Companies do not put that in a prospectus unless their lawyers think it is a live risk. My bet is that AI and jobs becomes a top-three campaign issue by the 2028 U.S. elections, with the policy fights shifting from safety and copyright toward taxes on AI labor, retraining mandates, and limits on automation in specific sectors.
The governance structure is partly an answer to this pressure. Anthropic is a public benefit corporation, a legal form that requires directors to weigh a stated public purpose alongside shareholder returns. Its purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity." A Long-Term Benefit Trust of five trustees with no financial stake elects a majority of the board. A Founder LLC owned by the seven co-founders holds a single share carrying 50.1% of voting power on most matters, with a sunset that begins once fewer than three founders have remained for 90 straight days. Public shareholders get the economics and very little control. The usual reading of that structure is that it protects the mission from Wall Street. The filing suggests a second reading. It is also designed to give the company a spine when the pressure comes from Washington.
Society: who gets the gains
The filing estimates that knowledge work was worth more than $40 trillion globally in 2025, and that the number of knowledge workers will pass one billion by 2028. That is the market Anthropic says it is going after. Its own company is the clearest example of what going after it looks like. Five times the people, 86 times the revenue, 45 times the code per engineer.
Anthropic is candid that this creates a distribution problem. Its own words say AI deployed at scale "has the potential to concentrate power and wealth both within and between countries." The company's response is unusual. Before the IPO, it contributed 846,840,960 non-voting shares, about 5% of its fully diluted stock, to a wholly owned subsidiary called the Anthropic Public Benefit Fund. The board can issue up to another 2% of outstanding shares to the fund each time the stock rises 50% above the previous milestone, starting at the IPO price. Each of the seven co-founders has pledged 80% of their personal equity to philanthropy. Employees can pledge equity to charity and the company matches it with stock.
The mechanism is clever in a way that deserves attention. The fund grows fastest exactly when the market believes AI is replacing the most work, because that belief is what drives the stock up. It is a philanthropic hedge indexed to the company's own disruption.
It is also nowhere near enough, and I do not think Anthropic would claim otherwise. A wholly owned subsidiary with discretion over its giving is still a private charity controlled by the company. If the knowledge work market is $40 trillion and AI captures a meaningful share of it, the redistribution question belongs to tax policy. The fund is a signal of intent. The real answer will come from governments, which brings us back to the political fight described above.
There is a quieter social implication in how the filing describes the product. Anthropic calls its agents "Virtual Collaborators," and describes a future in which they work on open-ended tasks lasting hours, days, or weeks with the judgment of a brilliant colleague. It describes millions of them running in parallel and calls the result "a country of geniuses in a data center." The filing frames the human role as directing and reviewing. That is already what its engineers do. The social question is what happens to the people whose careers were built on doing the work that is now being reviewed, and to the young people who used to learn a profession by doing that work. Apprenticeship is how expertise reproduces itself. If entry-level work gets automated first, the supply of future experts thins out. I have not seen anyone propose a credible answer to that. I do not know how it plays out. I am sure it is a real problem.
The singularity: a risk factor with a ticker
The word "singularity" appears exactly once in the filing, in a list of competitive factors, where Anthropic says it must maintain its "singularity of focus and mission." They meant focus. The other meaning of the word runs through the rest of the document.
The term comes from the mathematician and science fiction writer Vernor Vinge, who argued in a 1993 essay that once humans create intelligence greater than their own, the future becomes impossible to predict from this side of that moment. Most discussions of it have been philosophical. Anthropic's S-1 makes it operational.
The filing describes a dynamic it calls "closing the loop." Each version of Claude helps build the next one. Claude now runs experiments, generates and processes training data, automates evaluations and safety audits, and writes and reviews code for the next generation. Better models help researchers find what the filing calls compute multipliers, improvements that get more capability out of the same hardware, faster. Those multipliers make the next model better. The internal usage line passing payroll in April is the financial evidence that this loop is running at scale.
METR put a number on that loop. Its staff concluded that, as of October 2026, Anthropic is likely improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025, and said it would not be surprised by anything from 1.3x to 3x. On Anthropic's internal benchmark index, progress since October 2025 has run at 1.9 times the earlier pace, roughly two years of progress in one. METR was confident the speedup did not come from faster growth in training compute. Its best guess is that about 80% of the current acceleration comes from Anthropic's use of AI in its own research.
The rest of the report is where the hair on your neck stands up. Anthropic staff, polled informally, put their average speedup from the newest internal model at about 12x in October, a figure METR treats as loose evidence. METR found no sign of bottlenecks in algorithms, compute, or data over the next twelve months, and warned of a significant risk that further progress produces even faster acceleration. It flagged that growth could already be hyperbolic, meaning the rate itself keeps climbing, and noted that some of the evidence pointing that way was redacted. AI is not yet making highly significant, fully autonomous contributions to the most important parts of model training, though researchers described early signs of life. An unreleased internal model scored above 50% on every benchmark it was tested on, so the yardsticks are running out. And Claude did much of the analysis, including much of METR's own. The instrument measuring the acceleration is the thing accelerating.
The filing then defines what it calls Powerful AI. Expert-level or better across most professional and scientific fields. Able to use any interface a remote human worker can use. Able to work autonomously for hours, days, or weeks. Able to coordinate across many copies of itself and with humans. Able to control lab equipment and robots. The filing says this "will become widely available within a few years."
And then, in the same section, the safety caveat. Using AI to build more capable AI "could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress." Anthropic's Responsible Scaling Policy covers the approach to recursive self-improvement. The company says it does not expect to be able to safely let Claude fully autonomously design and build future frontier models without meaningful human supervision or understanding, and that it is considering policy changes to keep humans in control. The risk factors add that models may act in "unintended, misaligned, or autonomous ways," and that sufficiently capable systems could, through misalignment or self-preserving behavior, take harmful actions at scale without human intervention.
I want to be precise about what this means, because it is easy to either dismiss it or overreact to it.
What the filing establishes is that the process people have speculated about for thirty years, AI accelerating its own development, has begun inside at least one frontier lab, and that the lab is measuring it in dollars. What it does not establish is that the process will run away. The company says it intends to keep humans meaningfully in the loop, and it has a governance structure, a trust, and a safety policy built partly for that purpose.
The weak point is the one the filing itself names. Anthropic's brakes only work on Anthropic. The risk factors warn that tighter rules in the U.S. and other Western countries could let competitors in less regulated places move faster, and they name U.S.-China competition directly. A lab that commits to human supervision of its own self-improvement is betting that its competitors, foreign and domestic, will either make the same commitment or will not get far enough ahead to make the commitment irrelevant. That bet depends on governments, treaties, and other companies, none of which Anthropic controls.
So here is how I would watch for the singularity, if you want something more useful than a date on a calendar. Watch two numbers. The first is the ratio of internal AI usage to payroll at the frontier labs. Anthropic told us it crossed 1-to-1 in April 2026. When that ratio is 10-to-1, the humans are supervisors of a research organization made mostly of machines. When it is 100-to-1, the question of whether supervision is "meaningful," to use the filing's word, becomes the most important question in the world, and the answer will depend on whether interpretability research, the work of understanding what is happening inside these models, has kept pace. Anthropic says it invests heavily in that research, but neither the filing nor METR gives us a number to judge whether it is keeping up, and the compute-allocation check that might have come closest was dropped. The second number is METR's acceleration multiple. Today it sits around 2x. If a future assessment puts it at 4x with no bottlenecks in sight, that is the curve Vinge (and Kurzweil) were describing.
The last thing to watch is the export control precedent. If governments come to see frontier models the way they see enriched uranium, the singularity will arrive inside a regulatory perimeter, slowed and supervised. If they do not, it will arrive in a race. The June 2026 shutdown of Fable 5 and Mythos 5 suggests the U.S. government has already started down the first path. The fight with the Department of War suggests it has not decided what it wants from the labs once it has that control.
Where this leaves us
For thirty years, the singularity was a thought experiment argued about at conferences. On November 2, 2026, it became a disclosed risk factor in a document governed by securities law, attached to a company asking the public for money, with a ticker symbol and a lockup period.
That shift matters more than any single number in the filing. Once something is in an S-1, it is a fact that lawyers have agreed can be stated without getting sued. Anthropic's lawyers agreed that its AI wrote over 90% of the code contributed to its codebase in September, that the value of its internal AI use now exceeds its cash payroll, that it is building systems meant to match top experts within a few years, and that it is considering rules to keep humans in charge of the process. They agreed that the U.S. government has tried to cut it off, export-controlled its models, and may consider nationalizing the industry.
Three days later, an outside evaluator who spent eleven weeks inside the company added its own finding. Anthropic's models are likely improving about twice as fast as they did through 2024 and most of 2025, mostly because of Claude, with no bottleneck in sight for the next year.
Most investors will read this filing for the growth rate. The growth rate is real, and the bear case is real too. But the more durable lesson is in the payroll line. Every company in the economy is about to face the comparison Anthropic made about itself, between what it spends on human work and what it spends on machine work. Anthropic got there first because it makes the machine. Everyone else is a few years behind, and the filing just told them which direction we are headed.
Sources: Anthropic, PBC, Amendment No. 1 to Form S-1, filed November 2, 2026. https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm
METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, published November 5, 2026.