What Anthropic's IPO shows us about the future of software development

Anthropic, the company behind Claude, is heading for an IPO, and the draft prospectus obtained by Reuters in late September is probably the most revealing document ever published about the real economics of artificial intelligence. Not because of the grand plans, but because of the numbers.

On one side, revenue growth that would impress anyone: around US$ 4.6 billion in 2025, with 2026 quarters accelerating at a pace the company describes as multiplication, not addition. On the other, operating expenses of almost US$ 13 billion in that same 2025, a reported net loss in the tens of billions and, the most striking item in the document, commitments of more than US$ 500 billion in compute over the next decade, mostly non-cancellable, with Google, Amazon, Microsoft, Broadcom and Nvidia.

The cold reading is simple: growing revenue, a brutal need for cash and no clear path to profitability inside the document itself. The company talks about improving margins and positive adjusted quarters, but the obligations it has taken on are two orders of magnitude larger than current revenue. The IPO, which reportedly aims to raise around US$ 100 billion, is not a celebration of maturity. It is the next funding round of an operation that consumes capital at an unprecedented scale.

And this is not an Anthropic peculiarity. It is a portrait of the entire sector: its direct competitors burn cash at the same rate, with the difference that they have not yet had to show their numbers in this level of detail.

The liquidity tap is not eternal

While capital is abundant, the math works. Funds, big techs and now the public market finance the growth, and the price charged to the end user can remain artificially low. But the history of the technology market has no record of a liquidity tap that did not dry up at some point, whether through interest rates, disappointing results or a simple rotation of investment theses.

When that happens, and the prospectus itself admits the dependence on future fundraising, the pressure lands on the only variable the company truly controls: price. With hundreds of billions in non-cancellable obligations, there is no elegant plan B. Repricing stops being a hypothesis and becomes a schedule.

The playbook we have seen before

None of this is new. The playbook of technology companies financed by abundant capital is always the same, and it works in three stages: use the money raised to reduce user acquisition cost as much as possible, offering the service below its real cost; wait for companies and people to build their processes around the solution, until it becomes infrastructure; and then, with the base captured and switching costs high, raise prices.

It happened with the ride-hailing race, with streaming and, more quietly, with the cloud. The difference is that stage one has never been as expensive as it is in AI: every model response has a real marginal compute cost, and that cost is being subsidized to accelerate adoption. Whoever pays twenty dollars a month today for a tool that performs knowledge work is, in all likelihood, paying less than it costs to operate.

For the user, this is an excellent moment: sophisticated service at a subsidized price. The mistake is not in taking advantage of it. It is in building an entire operation assuming that today's price is the permanent price.

The question every technology leader needs to ask

From a business resilience standpoint, the question Anthropic's prospectus puts on every technology leader's table is a single one: how dependent are my processes on AI platforms?

This is not about avoiding AI, which at this point would be as sensible as avoiding the cloud. It is about being able to answer concrete questions. If the cost per call of the model I use triples, which processes become unviable? Which critical flows depend on a single model provider, with no abstraction layer that allows switching? What has been automated with AI and no longer has the manual procedure documented, nor the team that knew how to run it? How much of the roadmap assumes productivity tools at current prices?

It is the same exercise we proposed when discussing software resilience and vendor dependency, now applied to the newest and least mature layer of the infrastructure. The difference is that, with AI, the risk is not just a one-day outage. It is a permanent change in the economics that justified the adoption decision.

The counterpoint deserves registering: optimists argue that inference costs fall consistently with each generation of hardware and models, and that scale may close the gap before liquidity dries up. Maybe. But business resilience is not built on the optimistic scenario. It is built on the question "and if not?".

What we do about it in practice

At Espresso Labs, AI is already part of our development cycle and of the systems we deliver. Precisely because of that, we treat platform dependency as an architecture decision, not a detail: abstraction layers between application and model, the ability to switch between providers and open models, critical processes documented before being automated and cost-per-transaction math done at market price, not at promotional price.

Anthropic's IPO will be read as a milestone of the AI era, and it probably will be one. But for those who build and operate software, the lesson is in the prospectus footnotes: the most transformative infrastructure of the decade has not yet discovered what it costs. Design your systems knowing that.