Anthropic's S-1 Under the Microscope: Compute Efficiency, Loss Accounting, and the Case for Going Public

Ball opens with what he calls one of the most important numbers in the filing. Anthropic's compute spend on serving models rose 2.3 times over the past year, while the revenue generated over the same period rose 14 times. Taken together, he says, that implies roughly a 7x efficiency gain in serving inference.

He describes the business as almost accidentally profitable, and points to another figure from the S-1 he found striking: $141 billion of capital raised to date, inclusive of equity, convertible instruments and SAFEs, with a cash balance of $134 billion at the time. He contrasts that with the debt market, where he says a request to raise multi-billion-dollar debt for GPUs currently draws no bids.

Ball argues the decision to go public is in part about credit. Debt providers, he says, will underwrite Anthropic's creditworthiness and will expect full financials, and companies have historically been reluctant to open their books to that level of scrutiny. A public listing can lead to higher credit ratings and therefore lower borrowing rates.

He also notes that private capital pools have grown large enough to fund very big raises without a public offering, which makes the listing a choice of efficiency rather than necessity. Liquidity for employees and investors is a further reason, since tender offers are limited in size while a public company can offer a more open path to selling, subject to lockup terms that now vary by company.

Ball says management teams want everyday Americans to be able to participate in the equity upside of companies like Anthropic. He describes private markets as a small circle, limited to qualified purchasers, accredited investors, large funds, pensions and endowments, while doctors, dentists and ordinary households are excluded.

He warns that this imbalance could produce social unrest and, eventually, voter-driven regulation, contrasting capitalism's tendency toward centralization with democracy's tendency toward decentralization. He also notes that equity sits junior to debt in the capital structure, which makes a large equity cushion useful to lenders.

Ball points to the S-1 breakdown showing that only 27% of Anthropic's compute today goes toward serving models that generate revenue, with the remainder going to internal research, training, safety and alignment work and other non-serving uses. He says the company attributes its improved efficiency to improvement and investment loops described in the filing.

To illustrate the scale of the gain, Ball gives an example: a megawatt that once monetized at $10 to $15 million could now monetize at $70 to $100 million. He contrasts this with the traditional software model, in which companies burn cash early and become profitable later.

Ball says cash is the ultimate measure, and asks whether the cash balance is rising or falling. He acknowledges a standard critique: under GAAP, non-cash stock-based compensation is added back in free cash flow, which can overstate profitability, because employees would not join without equity.

He argues the cost should be treated as real, and even then Anthropic's free cash flow margin is still about 8%. By contrast, he says most mature software companies are free cash flow neutral or negative once stock-based compensation is counted as cash. He also cites a statistic that half of the Fortune 500 use Anthropic, while its spend represents only about 0.1% of their expenses.

Ball says much of the commentary about OpenAI and Anthropic is hyperbolic and boils down to three questions: can Anthropic stay on the frontier, is there a durable business to be built there, and how much will it cost. The second and third are connected, since the question becomes whether revenue can cover the cost of remaining at the frontier.

The S-1 shows that in the three months ending in September, 80% of revenue came from models six months old or newer. Ball expects frontier tokens to capture the largest share of economic value, while open-weight and cheap tokens will capture more of total token volume.

Ball describes Anthropic as seeking compute from every possible source, including building its own chip initiative and staffing a compute team, and buying TPUs. He notes that the hyperscalers are building their own chips too, with TSMC, Nvidia and ASML as the winners in the supply chain.

He says opposition to data centers in the United States has grown, citing noise, air pollution, water use, electricity bills and fears about AI risk. He suggests that the coming midterm elections could shape regulation. Ball argues this resistance acts as a natural governor on overbuilding, so the industry is less likely to face a glut like the dark fiber of the early 2000s, saying: "We have no dark GPUs today."

Ball says there is a limit to how far inference efficiency can be pushed, since it is bounded by physics, unless the chip itself changes. He suggests the company probably started out unoptimized and is now more optimized than two years ago.

On the question of whether frontier and open-weight models are converging, Ball says the more relevant comparison is between Anthropic's internal frontier and its public models. He says Anthropic has far more powerful models developed that have not been released, and points to an OpenAI researcher's chart showing internal models far outperforming released ones on math problems. He attributes the gap to delays in release due to safety testing and alignment work.

Ball applies a price-and-quantity logic to models: price has fallen sharply, quantity has risen sharply, and he expects the same with models and tokens. He gives the example that a binary classifier may not need a language model at all and can use a cheaper decision model, and expects further efficiencies of that kind.

He recalls that AWS's S3 and EC2 saw about 99% price reductions from launch, yet their business grew exponentially. He cites the quarter's reports, with Google Cloud growing over 100%, Azure almost 50%, and Amazon about 39 to 40%.

Ball points to a visible summer slump in Anthropic's revenue. He suggests it was partly because OpenAI's model was briefly ahead, and that many users started using it. He also notes that Google released new models and that open-weight models such as Kimi K3, the GLM series and DeepSeek became much more powerful quickly.

Ball says the Opus 5.5 and Sonnet 5.5 releases shifted that picture, with the run rate rising sharply in a short period. He says third-party data suggests revenue reaccelerated, and he would not be surprised to see that show up in October, the first month of the fourth quarter.

Ball describes Anthropic as having a clear focus on enterprise and coding, which he calls a deliberate choice. He says Anthropic has no decision model, has not released a standalone voice API, and has never released images. He contrasts this with OpenAI, which he says competes simultaneously with Nvidia, Google, Meta and Anthropic across several lines of business, and which shut down Sora.

He says that taking on several giants at once is a tall task, and that the stakes are higher for a trillion-dollar company. He credits Anthropic with discipline and says it has not been enamored with the next thing.

Ball describes a compute innovator's dilemma: an AI lab with many model families must decide what to serve, and demand for one product can squeeze compute for another. He says that much of the highest-return economic work has been in coding, which makes it hard to prioritize other areas. He calls coding the best generalizing model and framework, since it applies to other knowledge work.

Ball argues that Anthropic and OpenAI are chasing trillion-dollar opportunities, leaving smaller markets for startups and new labs, where a company could exit for $10 billion and still be a strong outcome.

Ball says he once hypothesized that labs would stop releasing models and instead build applications on their best unreleased models. He notes that some leading application companies are moving to in-house their own models to control their own destiny.

His investment thesis is that a strong application company needs both a strong product organization and a strong research organization, and that the two should be aligned so research improvements feed product improvements. He expects a much bigger push by both Anthropic and OpenAI into vertical applications, though he says the actions of Anthropic so far show no serious intent in that direction.

Ball compares product organization inside large companies to classic enterprise management: sales teams sell what already works, and new SKUs are harder to push. He says the company's focus has been a strength, and that OpenAI faces a tough task competing across several businesses at once.

Ball says the balance sheet shows no debt, with convertible instruments included in the $141 billion of capital raised. He says the reported net loss of about $80.7 billion in the first six months of 2026 is often cited as evidence of a poor business, but he argues it is misguided.

Ball explains the mechanism: as the liability tied to convertibles rose to $108 billion, assets did not fall, so equity had to fall, and because net income flows to retained earnings, the rising value of the liability shows up as a loss on the income statement. He says about $72 billion of the loss comes from appreciation of convertibles held by Amazon and Google, and about $9 billion from separate SAFEs and notes. He concludes that 100% of the loss in the first six months is just the value appreciation of those converts, saying "100% of the loss in the first six months is just the value appreciation of their converts."

Ball says that if Anthropic's S-1 had shown a doomy financial profile, the whole AI ecosystem would have reason for worry, since the company's returns affect everyone downstream. He says that as an intellectually curious student of finance, he finds the balance sheet and cash flow statement as informative as the income statement.

He reviews the gross margin, which he says is improving rapidly, while noting that the figures exclude cloud partner revenue share, which he understands sits in sales and marketing rather than cost of revenue. He puts the share of revenue from cloud partners at about 48%, with roughly $3 billion of related sales and marketing expense.

Ball says multi-billion-dollar GPU debt deals are drawing no bids, a situation he compares to an IPO being pulled. He asks why there is 100 billion in equity but not 100 billion in debt.

He explains that credit funds look through to the offtake customer, because that is who pays the bills. Many neocloud data centers are being built for Anthropic and OpenAI, and credit providers feel full after lending to Coreweave, Nebius and others. He expects debt markets to go through a period of indigestion as lenders ask whether projects are on time and revenue is flowing back, and notes that project delays can come from many sources, including local community opposition.

Ball works through a rough estimate. He notes that a gigawatt can cost roughly 50 to 70 billion dollars, and that both OpenAI and Anthropic have said they want to reach about five gigawatts by the end of the year. OpenAI's Stargate plan targeted 30 gigawatts by 2030.

He assumes both labs go from zero to 30 gigawatts over five years, adding 60 gigawatts, and that about 40% of that becomes total capacity, giving roughly 100 to 120 gigawatts. At around $50 billion per gigawatt, he estimates about $5 trillion over five years, or roughly $1 trillion a year. He says that amounts to a nontrivial percentage of real GDP, "That's like five% plus or minus of GDP." He says this would require very deep and open global debt markets.

Ball says that in a highly uncertain world, he falls back on core beliefs: the world will consume more models, more power, more compute, networking and storage, and it will need far more data centers to serve the intelligence being built. He looks for businesses that are direct derivatives of those beliefs.

He mentions a chip company still in stealth, American Terawatt, which aims to build a new industrial grid for industrial offtake, and compute-related work. He gives an example of raising utilization from 30 to 40% to 60%, which effectively doubles usable GPUs.

Ball describes Invest America as an investment account funded by the US government at birth, to which anyone can contribute, and which cannot be touched until age 18. He says the account lets children see investments appreciate over time and participate in the upside of the country.

He notes that the law has passed and each newborn receives $1,000. He mentions donations including a reported pledge of around $6 billion from Michael Dell. The current focus, he says, has shifted to getting parents to claim accounts, since money is sitting in accounts for newborns and must be claimed.