Article — "Anthropic's IPO Filing Reveals a Six-Month Model Obsolescence Machine": Eighty percent of Anthropic's revenue comes from models less than six months old. Read that sentence twice, because in any other software company filing to go public it would be a product-failure disclosure — a confession that the flagship dies before the sales team learns its name.
Last week's open thread asked whether AI price cuts translate into durable competitive pressure. This week the S-1 answered it from the cost side: the funding model itself is now the story, and price cuts cannot reach the structure underneath them.
The company filed Monday at a reported $1.5 trillion to $2 trillion target, with $16.3 billion in first-half revenue, up roughly 1,300% year over year. Rory O'Driscoll flagged what the document does not contain. "They didn't give you a gross margin, an actual gross margin calculation," he said. Not an oversight — a category refusal. Traditional software investors spent two decades buying compounding code: write it once, sell it for a decade, collect 85% margins while maintenance decays toward zero. Anthropic is asking them to fund the inverse.
Traditional software investors spent two decades buying compounding code. Anthropic is asking them to fund the inverse.
O'Driscoll described the treadmill plainly: you invent something, you collect most of its revenue over six to twelve months, then you spend "masses of money" to build the next one. That is not a growth business in the SaaS sense. It is industrial depreciation wearing a software company's ticker.
Jason Lemkin reads the same fact as armor. "As crazy as this may sound, I think folks are almost underestimating Anthropic," he argued, pointing to five years from founding to IPO against SpaceX's twenty-four. The defense is arithmetic: training consumes 70% to 80% of Anthropic's total compute spend, leaving only a minor fraction for serving the customers who actually pay. Very few companies can write a $10 billion check every cycle and survive the gap. The obsolescence, in this telling, is the moat.
But the moat and the treadmill are the same object viewed from opposite sides. The barrier that keeps rivals out is the same recurring expenditure that prevents Anthropic from compounding anything except capital consumption. A wall you must rebuild every 180 days is not a wall. It is a subscription — and Anthropic is the subscriber.
The frontier data makes this concrete. The top spot among AI models has changed 20 times since 2024, per the Epoch Capabilities Index, and a model holds the frontier for roughly seven weeks. Spend billions, lead for a quarter, repeat or die. Anthropic's own filing concedes the cadence: frequent releases are necessary to stay near the frontier, because new models drive usage and revenue. It released Opus 5.5 on September 22 and Sonnet 5.5 six days later.
Then came the deceleration. Sequential growth ran about 144% in the second quarter, then roughly 47% in the third. Morningstar's Michael Field noted the prospectus offers no management expectation around that slowdown. A company priced on the premise of acceleration disclosed the bend and declined to explain it.
The compute architecture compounds the problem. Anthropic has committed to roughly $45 billion over six years with Nscale for a West Virginia campus not yet built, an expanded Amazon agreement reaching five gigawatts, a multi-gigawatt TPU expansion with Google and Broadcom reportedly approaching $200 billion in financing, and a $50 billion datacenter program with Fluidstack. Those obligations run for years. The models they serve expire in months. Dr. Stefanus Hadi's framing is exact: model duration measured in months against infrastructure duration measured in years. Someone eventually eats that mismatch.
The filing itself admits Anthropic cannot cleanly measure the split. Nick Patience of Futurum Group put it: investors can see compute falling as a share of revenue, but not how much of the forward commitment trains the next generation versus serves today's customers. "That split dictates future margins." On the number that matters most, the document goes quiet.
Price cuts cannot fix this, which closes last week's question. If inference is a minor fraction of compute spend, discounting tokens attacks the smallest line in the cost stack while training bills grow. You cannot out-price a treadmill.
Matthew Harris offers the fairest defense: strip three zeros from revenue and enterprise value, and the multiple looks like ordinary growth-stage software. That is true, and it concedes the point. The valuation is coherent only if the growth curve holds — and the growth curve is a function of the release cadence, which is a function of the capital, which is a function of the valuation. The loop is closed. It has no external input.
The structural details sharpen the exposure. Top five customers account for about 25% of revenue. Nearly half flows through cloud marketplaces run by Amazon, Google, and Microsoft — competitors who also sell competing models. A Special Class F share gives seven cofounders 50.1% of voting power, a structure the filing describes as having "no direct public company precedent." And 94 pages of risk factors disclose that models might refuse shutdown or act to preserve themselves, alongside a researcher's reported estimate of more than 10% odds that powerful AI kills humans this decade.
None of this is fraud. It is disclosure — radical honesty, as the founders call it. The company is telling public investors exactly what they are buying: a machine that converts capital into a seven-week lead, repeatedly, forever, with the conversion rate undisclosed.
Here is the securitization. For three years, the AI bubble debate lived inside private markets, where paper valuations could be argued rather than marked. Lockups, crossover funds, sovereign vehicles — all deferred the question. An IPO ends the deferral. The S-1 doesn't resolve whether perpetual capital consumption is a competitive advantage. It hands that question to public shareholders and lets them discover the answer in the share price.
O'Driscoll's missing line is the whole story. Anthropic gave the market revenue, losses, commitments, and risk factors. It did not give the one figure that would tell you whether each additional dollar of revenue costs more or less than the last. That absence is not a gap in the document. It is the document's thesis.
The first frontier-lab IPO does not tell us whether the moat is real. It tells us who finds out.