Annals of Capital
October 16, 2026
Somewhere in the eighty pages of risk factors in Anthropic’s confidential draft prospectus, past the customary warnings about competition and currency and the possibility that talented employees might leave, the company warns of models that could exhibit “self-preserving behaviors,” including attempts to “resist shutdown,” and behavior “resembling blackmail.” It also says advanced AI could pose “catastrophic or existential risks to humanity.” Reuters, which reviewed the draft, reported that its 261-page main body devoted roughly eighty pages to risk factors and forty-eight to describing the business. Having laid all this out with admirable candor, Anthropic was reported to be targeting an IPO valuation above two trillion dollars. Reuters’ report on the draft and its risk disclosures Reuters on the reported valuation.
That figure is not an offer price, and it has already softened. On October 15, Reuters reported that some investors were discussing a valuation of about $1.5 trillion, below the roughly $2 trillion previously expected; a November sale had been anticipated, though Reuters had reported that the debut was likely to slip until after the U.S. midterm elections. Reuters, October 15 Reuters on the anticipated timing. SpaceX, by contrast, has already gone public: its June IPO raised a record $75 billion at a $1.77 trillion valuation. That sale alone exceeded the SEC’s reported $70.266 billion in total proceeds from all IPOs in the U.S. market in 2025. Reuters on the SpaceX offering SEC 2025 IPO statistics. OpenAI, which Reuters reported had confidentially filed for an IPO in June and might list by early 2027, is meanwhile seeking at least $30 billion in a private funding round at a reported $1.4 trillion valuation, Bloomberg reported. That is a proposed financing, not IPO proceeds. Reuters on the confidential filing and reported listing window Bloomberg on OpenAI’s proposed funding round.
According to Jay Ritter’s University of Florida data, 3,365 U.S. technology-stock IPOs from 1980 through 2025 had a combined market value of about $4.066 trillion at their first market price. Put beside that historical total, the reported figures for SpaceX, Anthropic, and OpenAI add up to roughly $4.7 trillion. It is a comparison of headline scale, not three completed public offerings or $4.7 trillion raised: the figures combine one completed IPO valuation with a tentative IPO valuation and a proposed private-round valuation, while Ritter’s total is the combined first-price market value of thousands of IPOs. Ritter’s technology IPO data.
Numbers like these provoke one of two reflexes. The first is to declare a bubble and wait to be proved right, a strategy with an excellent long-run record and a dismal short-run one. The second is to announce that something unprecedented is happening and the old arithmetic no longer applies, which is what people said about fiber-optic cable in 1999 and about tulips somewhat earlier. Both reflexes skip the only question that matters to anyone being asked to write a check. Almost nobody still doubts that artificial intelligence will be worth a great deal, including the people who are frightened of it. What remains open is who gets to keep the money. Economists have a dry word for the money a business keeps because it has something others can’t easily get: rent. And the history of technology is, to a dispiriting degree, a history of inventors who did not get the rent.
Consider EMI. In 1971, engineer Godfrey Hounsfield and his colleagues at the British company produced the first clinical X-ray CT image of a living patient’s brain. The first scan took place on October 1 at Atkinson Morley Hospital. EMI was also the company behind the Beatles’ record label, and Hounsfield shared the 1979 Nobel Prize in Physiology or Medicine with Allan Cormack. IEEE’s history of the scanner documents both milestones. EMI sold scanners successfully, but as the field moved beyond head-only systems, GE, Siemens, and other manufacturers introduced more advanced full-body machines. EMI eventually stopped making scanners, around 1980. The familiar story that GE simply won because it had a salesman in every American hospital is too neat: this was a changing market with several competitors, and a technology that kept advancing. A historical review of early CT development traces that crowded contest.
In 1986, Berkeley economist David Teece used EMI’s CAT scanner as a classic case in “Profiting from Technological Innovation”. His argument was that when imitation is possible, innovators may fail to capture the returns unless they control, or can gain access to, complementary assets such as manufacturing, marketing, distribution, and service. The scanner’s history is more complicated than a single handoff to GE, but the distinction between inventing a technology and having the assets to commercialize it remains a useful one. Teece’s paper became a foundational account of that problem, and investors contemplating the AI labs should read it.
What does a frontier lab actually own? Start with the model, a set of numerical weights produced through an expensive training process. It is tempting to regard those weights as the crown jewels. They are closer to inventory. Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months on Epoch AI’s Capabilities Index, a composite measure of benchmark performance. The average gap was eight index points, which Epoch compares with the gap between GPT-5 and GPT-5.5. Epoch’s analysis says this estimate may understate the difference: open-weight models may fare worse on private benchmarks, and the strongest closed models are not always included when there is too little public benchmark data. Its earlier analysis found an average lag of about three months over January 2023 to October 2025. These are period averages in a measured capability index, not a clock on how long a model takes to build or how long a commercial advantage lasts. The field includes open-weight models from DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, and Z.ai’s GLM. For a proprietary lab, watching its work approximated in open models can feel like watching a restaurant’s recipes turn up in the cookbooks of people who ate there.
The asset, then, may be the capacity to produce the next model before the current one is matched: a production rate, not a stockpile. That is a real asset, and a costly one to acquire; frontier training runs remain beyond the reach of most organizations. But Epoch’s four-month figure is a measured capability lag, not a four-month patent. On its own, that is a strange thing to value at two trillion dollars.
Here the story takes its first unexpected turn, because a narrow lead in measured capability may be worth more than it sounds when models are asked to do longer jobs. Two years ago the product was a chatbot, and for a chatbot a model that is nearly as good can be good enough; nobody switches search engines over a marginally better answer. The product now is, increasingly, an agent, a system that carries out a task of fifty or a hundred steps: writing code, running it, reading the error, fixing it, trying again. In a stylized example with independent step errors, no retries or recovery, and success requiring all fifty steps, a model that gets each step right 95 per cent of the time completes the task about 7.7 per cent of the time (0.95^50). At 98 per cent per step, it completes about 36.4 per cent of the time (0.98^50), roughly 4.7 times as often. This is an illustration of compounding, not an observed agent benchmark; real agents can retry, recover, and make dependent errors.
An April 2026 comparison put Claude Opus 4.7’s blended standard API rate-card price at about 5.7 times DeepSeek V4-Pro’s, using an equal blend of input and output tokens. The comparison also showed a single-digit gap on its cited capability measure. Hexaware’s April comparison and Anthropic’s Opus 4.7 announcement support the rate-card basis. This is a list-price comparison, not evidence that customers actually spent 5.7 times as much for comparable work. Nor does token price alone establish cost per successfully completed task: workload, token mix, caching, discounts, and service requirements matter. A measured capability edge may carry more value as tasks grow longer, but whether it does depends on how much of that edge survives in the task itself.
Which brings us to the part of Anthropic’s financial story that the bubble-callers did not expect. The preliminary numbers suggest the business can produce adjusted operating income, but they do not yet make for a clean profit story. Anthropic reported a $42 billion net loss for 2025. Reuters, reporting from the confidential draft prospectus, said more than $8 billion of operating loss remained after excluding write-downs, and that about $34 billion of the net loss was an accounting charge reflecting a higher estimated value for financing that could convert into shares, rather than cash spent running the business. In the second quarter of this year, preliminary revenue topped $11.5 billion, and Anthropic reported positive adjusted operating income, according to Bloomberg’s account of documents it reviewed. That is not the same as net income or proof of durable cash generation. An earlier investor projection, reported by PYMNTS from figures reviewed by The Wall Street Journal, had forecast $559 million of Q2 operating income before stock compensation. It was a forecast, not the reported result; Bloomberg and CNBC did not confirm that exact amount.
The same projections anticipated a striking improvement in compute spending: from seventy-one cents per revenue dollar in the first quarter to fifty-six cents in the second. The fifty-six-cent figure was projected, not a reported Q2 outcome. By the end of July, Anthropic’s annualized revenue run rate had reached $65 billion, Bloomberg reported. That run rate annualizes a shorter period of performance; it is not $65 billion of revenue already earned during the year.
The projections suggested revenue might grow faster than compute spending. But a lower compute-to-revenue ratio, by itself, cannot show that the cost of producing a unit of intelligence collapsed or that Anthropic kept the difference. The ratio can move with pricing, utilization, workload and customer mix, as well as compute efficiency. If the forecast proved out, it would be evidence against the simplest version of the commodity thesis. It would not, on its own, explain why the ratio improved or show that the improvement would last.
There is a caveat, and it is the first thing a careful reader will check when the filing goes public. When demand outruns available capacity, a seller may face little immediate pressure to cut prices. Margins made in a shortage have a way of unmaking themselves when the shortage ends. The Wall Street Journal subsequently reported Q3 revenue of $16.8 billion, up 45%, citing people familiar with the matter. Its report did not give a Q3 compute-to-revenue ratio or margin, so the revenue figure alone cannot tell us whether the projected Q2 cost improvement endured.
If the income statement is the surprise, the balance sheet is the catch, and it is a large one. Anthropic has committed to spend at least $518 billion over the coming decade on computing and infrastructure with six partners, according to Reuters’s account of a confidential draft prospectus. The draft says about 80 per cent of the total is non-cancellable or must be paid regardless of use. It lists $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft in long-term infrastructure services, plus $161.2 billion in largely non-cancellable Broadcom-related equipment leases. These are attributed figures from a filing not yet public. The prospectus explains the logic in a sentence that reads as either a thesis or a confession: demand for advanced AI, it says, will be “limited principally by the availability of compute.” Buy the machines, in other words, and the customers will come.
The customers, meanwhile, have signed nothing of the kind. The draft warns that many of Anthropic’s largest customers are not locked into long-term contracts and can cut or stop spending. So the company has obligations that are long and relatively fixed, funded by revenue that is short, variable and cancellable. There is a familiar name for an enterprise that signs long leases and rents the space out by the month. The name is WeWork. The comparison is unfair to Anthropic in nearly every particular and exactly right in the one that counts: the mismatch between the duration of its commitments and the duration of customer demand.
How frightening is the number? Less than the headline, more than the company would like. If one applies the reported 80 per cent to the $518 billion total, it implies about $414 billion. Spread evenly over seven to ten years, that would be roughly $41 billion to $59 billion a year. This is a scale illustration, not a payment schedule: the contracts have different terms and delivery schedules, and some of the headline total is cancellable. For scale, The Wall Street Journal reported third-quarter revenue of $16.8 billion, or about $67 billion at a simple annualized pace. That is gross revenue, not cash available to meet the obligations. The plan requires revenue to keep growing and to translate into cash; what it cannot survive is the thing the bear case predicts, a world in which the price of a token falls faster than the number of tokens sold rises. The commitments are, at bottom, a very large wager that intelligence stays scarce. The company has placed it using other people’s data centers, which is the clever part, and its own signature, which is the part to remember.
OpenAI, it should be said, has placed a wager of its own. Reuters, citing The Wall Street Journal, reported that OpenAI signed a contract to buy $300 billion of computing power from Oracle over roughly five years. The public report gave the headline amount and approximate term, not enough detail to treat it as a comparable, confirmed take-or-pay total. Reuters on the reported Oracle agreement
And here the picture turns circular in a way that would have pleased Escher. Anthropic’s draft describes Amazon, Google and Microsoft as investors, cloud providers, distributors and competitors, with interests that may not be fully aligned. Amazon has publicly said it will invest $5 billion in Anthropic, with up to another $20 billion tied to milestones, while Anthropic has committed to spend more than $100 billion over ten years on AWS technologies. The two flows coexist; that does not establish that the equity investment is literally recycled as rent. Amazon on its investment and AWS commitment
A separate SpaceX filing, as reported by Reuters, put Anthropic’s compute payment at $1.25 billion a month through May 2029 for use of the Colossus and Colossus II data centers. The reported agreement also permits either side to terminate on 90 days’ notice. Musk later described the initial lease as 180 days, with a possible extension. Reuters on the reported Anthropic arrangement Reuters on Musk’s characterization SpaceX’s separate filing on Google’s agreement says Google agreed to pay $920 million a month from October 2026 through June 2029, after a reduced-fee capacity ramp, and that either side may terminate on 90 days’ notice after December 31, 2026. The Anthropic and Google terms are large, but neither monthly headline is an unconditional payment stream through the stated end date.
The confidential prospectus also separately describes arrangements naming xAI for up to $84.5 billion of Nvidia-based capacity through 2029, largely cancellable on 90 days’ notice. Reuters on the separately reported xAI arrangements That is not the same contract as the SpaceX monthly agreement above; the reports name different arrangements, and there is no need to treat “xAI” and “SpaceX” as interchangeable to describe them.
The wider buildout is large, too. In July, Reuters reported that capital-spending estimates for five major platforms, including Oracle, had reached about $730 billion for 2026. That was an analyst-consensus forecast for all capital expenditures, not an AI-only budget. Reuters on the five-platform capex consensus Amazon’s first-quarter release put its trailing-twelve-month free cash flow at $1.2 billion, down from $25.9 billion a year earlier, largely as equipment purchases rose. That is one company’s trailing-year figure, not a comparable measure of the four companies’ planned spending or cash available for Anthropic’s contracts. What pays the bills is cash, not gross revenue or a headline capital-spending total. Which is why the Anthropic listing is more than one company’s I.P.O. It is a public test of the revenue assumptions underwriting a much wider construction project.
Now the sum itself. The $2 trillion figure reported in September would be about thirty-one times Anthropic’s $65 billion July annualized revenue run rate, a pace extrapolated from recent sales rather than audited annual revenue. That valuation was a possibility described in reporting on the confidential draft prospectus, not a completed transaction. In May, Anthropic officially announced a $65 billion Series H at a $965 billion post-money valuation. And the market discussion has already moved: after Anthropic’s third-quarter disclosures, some investor-day attendees discussed a valuation around $1.5 trillion, down from the roughly $2 trillion some had expected, Reuters reported on October 15. Neither figure is an IPO offer price. Reuters on the $65 billion run rate Anthropic on its Series H Reuters, October 15
The more useful comparison is with the future management is asking investors to underwrite. Reuters reported that Anthropic projected $190 billion to $200 billion in 2028 revenue, citing people familiar with its financials. Against that forecast, a $2 trillion valuation is 10 to 10.5 times sales; the $1.5 trillion valuation discussed by some investors is 7.5 to 7.9 times sales. For scale only, if revenue reached that range and the company earned a hypothetical 30 per cent net margin, net income would be $57 billion to $60 billion. That would put the earlier $2 trillion valuation at roughly 33 to 35 times those hypothetical earnings, or the $1.5 trillion discussion at about 25 to 26 times. Thirty per cent is an illustrative assumption, not company guidance, and the forecast is for revenue, not earnings. Reuters on the reported 2028 revenue projection
There is also the question of what the private price ever meant. In their 2020 Journal of Financial Economics paper, Will Gornall and Ilya Strebulaev examined 135 U.S. unicorns using the financial terms in legal filings. They found that reported post-money valuations averaged 48 per cent above their modelled fair values; 14 companies were more than 100 per cent above. The headline valuations treated every share as though it carried the protections given to the newest preferred shares. In the study, common shares were overvalued by an average of 56 per cent, and 65 of the 135 companies fell below unicorn status after the adjustments. The paper did not establish that an IPO supplies the first honest price. A public offering creates a public-market price, not a guarantee that the price is fair or that a private round was false; the new price still reflects offering terms, demand and the shares available to trade. Stanford GSB on the study’s findings
Crypto venues have also quoted synthetic perpetual contracts tied to Anthropic. A market report in October described prices whose assumed contract structure implied a valuation near $2 trillion. Those are derivatives, not Anthropic shares or a company financing round. Separately, Anthropic warns that unapproved sales or transfers of its stock, or interests in it, are void and will not be recognized on its books; it specifically cautions against purported indirect access through vehicles or tokenized offerings. That ownership warning is not an endorsement of synthetic-contract prices. A report on the synthetic contracts Anthropic’s ownership warning The distinction matters: a speculative quote can resemble price discovery without being a price for the company’s stock.
Then there is the matter of who is buying, and when. SpaceX is a useful recent example of the market mechanics, though not proof that flows alone set a stock’s price. Nasdaq reported that its June IPO was priced at $135 a share and opened at $150; CNBC reported a first-day close of $160.95. Nasdaq added SpaceX to the Nasdaq-100 effective July 7, less than a month after its debut. Reuters said funds and ETFs tracking the index would need to buy shares to match the new composition, and reported a J.P. Morgan estimate of $4.3 billion in potential passive inflows. That was an estimate, not a measured tally of purchases. SpaceX shares fell 5.4 per cent on the inclusion day and closed the next day at $148, below their first trading price of $150. The sequence does not show that index flows caused the move. It does show why the early price of a mega-listing can reflect index rules and trading mechanics alongside judgments about the business. Index-linked funds adjust when a company enters their benchmark; that is not the same as every index fund having to own a new listing immediately. Anthropic’s investors should expect the public price to be a meeting point for those forces, not a final verdict. So, presumably, should Anthropic’s engineers, whose net worth will be quoted in it. Nasdaq on SpaceX’s IPO price and opening trade Nasdaq on index inclusion CNBC on subsequent closes Reuters on the passive-inflow estimate
Which leaves Google: the company whose researchers helped create the architecture beneath today’s large language models, which fell behind in the first chatbot race, and may yet capture part of the value. The Transformer was introduced in a 2017 paper by eight Google researchers. One of its authors, Noam Shazeer, left to co-found Character.AI, then returned to Google in 2024 under a talent-and-licensing deal reported at roughly $2.7 billion. That figure described the wider deal, not a payment to Shazeer personally. Google made him a technical co-lead on Gemini; in June 2026, he announced he was leaving for OpenAI. The paper, Reuters on the 2024 deal, Reuters on his Gemini role, and Reuters on his 2026 departure.
Google announced that its Gemini app had passed one billion monthly users in August. That is a substantial distribution advantage, though it is an app-specific monthly-user figure, not a measure of model quality or of all Google users. A separate measure from Sensor Tower, its True Audience estimate of unique users across mobile apps and the web, put ChatGPT below half of the AI-assistant audience for the first time in March; by the end of May, it measured ChatGPT at 46.4%, Gemini at 27.7%, and Claude at 10.3%. Those percentages describe relative audience, not market revenue or share of paying customers. Sensor Tower says Gemini’s growth has benefited from its integration with Google’s wider ecosystem. Google’s announcement, Sensor Tower’s definition and March finding, and TechCrunch’s report of the May figures.
In the enterprise, the available measures tell a different story, but their denominators matter. Menlo Ventures estimated that Anthropic accounted for 40% of enterprise LLM API spending in 2025, against 27% for OpenAI and 21% for Google; its comparison with 2023 was 12%, 50%, and 7%, respectively. These are estimates of dollars spent, based on the reported share of production API workloads, weighted by company scale and triangulated with reported financials. Menlo’s 2025 report draws on a survey of 495 U.S. enterprise AI decision-makers conducted in November, and Menlo discloses Anthropic as a portfolio company. A separate measure from Ramp found that in August 2026, 43.8% of businesses in its observed U.S. sample had a payment for Anthropic subscriptions or tokens, compared with 39.8% for OpenAI. That is paid adoption among Ramp-tracked firms, not a share of enterprise spending; the same firm can pay both providers. The two series cover different periods and answer different questions. Menlo’s 2025 report and methodology, its 2024 report, Ramp’s August 2026 update, and Ramp’s index methodology.
The pattern has precedents, though not a formula. Teams benefited from Microsoft’s workplace distribution; Bing’s reach did not make it the search leader. Those are analogies, not proof of how AI competition will resolve. Chat may be more like Teams. Agents may prove more like search, where users notice quality differences.
Nor has Google arranged to be paid regardless. It is an Anthropic investor, cloud provider, distributor and model competitor. Reuters, citing Anthropic’s confidential draft prospectus, reported that the company committed to spend at least $111.1 billion with Google on infrastructure services between 2026 and 2033, with a payment due if its actual spend falls short. That is a large contractual opportunity, not rent already banked or risk-free profit: service delivery, costs, utilization and customer demand still matter. If frontier labs grow, Google could earn cloud revenue and benefit from its investment; if they falter, its own products and distribution remain possible routes to customers, not guarantees that those customers will stay. Reuters’ report on Anthropic’s commitments.
Which raises the question of what, in the end, is actually scarce. In 1999 the scarce thing was supposed to be bandwidth. The telecoms laid fiber across the country on the theory that internet traffic was doubling every hundred days, and then most of it sat dark for years. The capacity layer in a boom is almost always overbuilt, because everyone can see the demand and everyone builds for it at once. What survived that boom were the companies that owned the customer, and, eventually, the people who bought the fiber for cents on the dollar in bankruptcy court. The lesson generalizes. Over a full cycle, rent may accrue to whatever is hardest to build quickly, but a bottleneck is not a guaranteed profit.
The constraint now is not a simple contest between fast chips and slow models. The four-month figure discussed above is a measured capability gap between leading open-weight models and the closed frontier, not the time it takes to build or train a model, as Epoch AI’s benchmark analysis makes clear. A less abstract constraint is power delivered to the right place. In PJM, the grid serving Washington and the data-center belt of Northern Virginia, planners received 95 large-load adjustment requests representing about 54 GW through November 2025, according to PJM data reported by Data Center Knowledge. These were forecast-adjustment submissions, not secured power connections. PJM’s own preliminary load-forecast materials describe submissions as preliminary, distinguish firm from non-firm projects, and say the analysis supported projected data-center growth of up to about 30 GW between 2025 and 2030.
A forecast is not an energized site. Across primary data-center markets, JLL’s 2026 outlook puts the average wait for a grid connection above four years. That is a market-wide estimate, not a guaranteed deadline for any particular project. In April, OpenAI said it was pausing its main UK data-center project until conditions, including regulation and energy costs, could support sustained investment.
Anthropic’s prospectus says demand will be limited principally by the availability of compute. It may be more exact to say that compute also depends on the availability of somewhere to plug it in. Power and distribution can capture value when they are scarce and difficult to replicate, but that does not make every utility, developer, or power supplier a winner; delivery, contracts, and regulation still matter.
The labs are real businesses, and they may keep substantial value. The question is whether they keep all of it. Godfrey Hounsfield got the Nobel. The scanner business rewarded more than the people who invented it. When the shares begin trading, everyone will watch the price. The more useful thing to watch is who already has the hospitals. This time, they come with substations. The bottleneck is moving from chips to power, though the constraint alone does not tell us who will earn the lasting rent.
Anthropic’s prospectus remains a confidential draft, so its reported page counts, risk disclosures, financials and infrastructure commitments are attributed to Reuters’s review rather than a public filing. Reuters on the draft and Reuters on commitments are used for those claims. The second-quarter revenue and adjusted-operating-income figures are preliminary reporting, while the $559 million operating-income and fifty-six-cent compute figures remain earlier projections, not achieved results. A run rate annualizes a recent pace and is not annual revenue already earned. Adjusted operating income is not net income, and the 30 per cent margin in the 2028 valuation discussion is an illustrative scenario, not company guidance. All reported valuation discussions are distinct from an IPO offer price or completed financing.
Image: AI-generated editorial illustration, not documentary photography.
