Anthropic's Six-Month Product

In a podcast published in August 2025, Dario Amodei, Anthropic's chief executive, explained the model business with a made-up example. A lab spends $100 million to train a model in 2023. The model earns $200 million the next year. Meanwhile the lab spends $1 billion on the next model, then $10 billion on the one after. The first two models each earn back twice their cost, and the company loses more money every year.

His punchline: "If you consider each model to be a company, the model that was trained in 2023 was profitable." It's a good idea with one soft spot. A company needs a lifespan. How long does each of these companies live?

On October 26, 2026, Anthropic filed its S-1 to go public. The filing gives a strong hint. Not long, at least when you count the revenue.

The hint sits in one sentence of the filing's discussion of results: model versions released in the six months to September 30, 2026 "accounted for over 80% of our revenue" in the third quarter.

Now the second clock. The computing power those models run on is bought in advance. The majority comes under multi-year commitments "that we typically procure 12 to 18 months in advance of availability."

Put the two together and you have the business in one picture: a product that earns most of its money in its first six months, made in a factory you have to order a year ahead.

Most coverage will go to the $80.7 billion net loss for the first half of 2026. That number is mostly paper. $72.4 billion of it is the rising value of convertible notes and similar instruments, which the filing says rose mainly because its valuation went up. The two clocks are the more useful story for anyone who sets strategy, because they explain the company's choices.

A software company can think of research and development as an investment, because the product it builds sells for years. Anthropic can't count on years. So the real cost of selling today's model includes building tomorrow's.

The filing shows how much that changes the picture. Count only the computing power that serves customers, and the first half of 2026 shows a gross margin of about 85%, by our arithmetic on Note 8. Count all of it, including what trains the next model and runs research, and 45% is left. Since 2025 the income statement leads with that second view, all compute on one line. The filing describes serving customers and building models as one cycle, each "a direct input that informs the other."

So far, revenue has caught up with the bill. Compute and infrastructure cost 251% of revenue in the first quarter of 2025 and 48% in the second quarter of 2026. The filing is plain about the other direction: as new compute comes online, its cost lands before the revenue it will support, and "we expect operating margin and Adjusted Operating Margin may compress in those periods."

And the size of the total bet? The filing says Anthropic has entered into purchase commitments and leases "with an aggregate committed spend of more than $", and leaves the figure blank, like many numbers in this preliminary draft. That blank is the number to read when an amendment fills it in.

When customers chase the newest version, the safest product to own is the one that replaces yours. The S-1 says enterprises "consistently utilize our latest and most capable models as soon as they are released." So Anthropic keeps replacing its own models and making each one cheaper to use: Claude Opus 5.5 "reduced customer costs per task by approximately 40% compared to Claude Opus 5." The filing says the reduction came partly from a lower price and partly from efficiency.

The other half of the rule is speed. "Each version of a Claude model assists in building the next," the filing says. The company calls this closing the loop. On a clock this fast, the lasting advantage is less any one model than the machine that makes the next one.

Anthropic isn't alone. Nvidia, one of the chip makers whose hardware Anthropic runs on, keeps a similar clock. Its latest annual report says it is "bringing new advanced architectures on a one-year product cadence." It describes the mismatch from its side of the table too: "Long manufacturing lead times and uncertain supply and capacity availability, combined with a failure to estimate customer demand accurately, has led and could lead to mismatches between supply and demand."

Nvidia commits early too. It "significantly increased" its supply and capacity commitments "from $119 billion last quarter to $279 billion as of July 26, 2026 to meet future demand." And it names the price of its own speed: "reduced demand for current generation architectures when customers anticipate transitions." Some buyers wait for the next chip. The faster the clock, the more today's product competes with tomorrow's.

Two companies at two points on the same supply chain, playing the same game. The product clock runs faster than the capacity clock, and the gap between them is where the risk sits.

Most of our readers sit on the other side of this trade. They buy these models, or advise companies that do. The two clocks matter to them too.

If most of a supplier's revenue comes from versions less than six months old, the model you evaluated in the spring is probably not the model your teams will use in the fall. Plan for that. Write contracts around the work you need done, not a named model. Rerun your tests every time a new version ships, because the cost of the same task can move a long way in one release: Opus 5.5 cut it by about 40%, the filing says. And keep the work you build on top portable enough to move between versions without starting over.

The deeper point is about where the value sits. A supplier on a clock this fast needs you to stay through every replacement. Your bargaining power comes from being able to leave.

One sentence is carrying a lot of weight here. The 80% figure covers one quarter, in a year of extraordinary growth: revenue rose 1,368% in the second quarter from a year earlier. In a boom like that, new models would dominate any quarter even if the old ones kept selling.

That's fair, and it points to the real asset. The models turn over. So far the customers have stayed and spent more: net revenue retention, which leaves out Pro and Max subscribers, was about 900%. That is a net figure, and the filing warns that many top customers can "reduce or in some cases cease spending at any time and for any reason." Still, on the evidence so far, the product expires and the relationship compounds.

Amodei named the deeper risk himself. In his example the run ends one of two ways: "a large, very profitable business," or "the models will stop getting better" and the last, biggest round of spending buys nothing. The S-1 can't tell you which. It does show one quarter of operating profit, the only one in its table: $974 million, earned while the next model was still being paid for.

Check two clocks. How long does your product stay current, and how far ahead do you have to commit the capacity to make it? When the product clock is faster, the rest follows. Count the next product as a cost of selling this one. Make your own product obsolete before a rival does. And hold on to the customer, because the customer is the thing most likely to outlast the product.

Any business that ships new versions faster than it can build new capacity lives with some version of this. The question for yours is simple: which clock is faster?