← All stories

Anthropic's IPO filing: what an operations team should read

Anthropic filed to go public on October 26. The $80.7B loss is mostly accounting. What matters if you build on Claude is who pays the bills, and how fast they can leave.

This story is also available in Español

Anthropic's IPO filing: what an operations team should read

Key takeaways

  • Anthropic filed an S-1 on October 26 to list on Nasdaq as ANT, reporting $16.3B in revenue for the first half of 2026 against $4.6B in all of 2025.
  • The $80.7B first-half net loss is mostly non-cash accounting charges, so it is not the number to watch.
  • About a fifth of revenue came from two customers, and a large buyer like Replit is reported to have moved spend to OpenAI over price.
  • Access to a model can change for reasons your contract doesn't cover, as the June suspension of Fable 5 and Mythos 5 showed.
  • Prompts, tests and vendor names kept portable, plus a second model in shadow mode, turn a vendor switch from a project into a config change.

Anthropic filed to go public on October 26, and the headline number is an $80.7 billion loss. Most of that is accounting. If you run operations on Claude, the lines worth your time are further down: who pays the bills, and how easily they can leave.

What happened

On October 26, Anthropic filed an S-1 to list Class A shares on Nasdaq under the ticker ANT, according to Yahoo Finance's report on the filing. Revenue for the first six months of 2026 was $16.3 billion. For all of 2025 it was $4.6 billion, so in half a year the company booked more than three times what it did in the whole year before.

The net loss for those six months was $80.7 billion. Fortune says the vast bulk comes from changes in the fair value of convertible notes and other non-cash charges, and argues investors will likely look past it. We would too. It is an accounting line, not a cash drain.

The cash side is compute. Anthropic spent $9 billion on AI infrastructure in the first half, against $7.3 billion in all of 2025. Yahoo also reports commitments of upward of $430 billion in cloud computing deals, including with Amazon and Google.

MetricFigure
Revenue, first half 2026$16.3B
Revenue, all of 2025$4.6B
Net loss, first half 2026$80.7B, mostly non-cash
AI infrastructure spend, first half 2026$9B
Share of revenue from two customersAbout one-fifth
Figures as reported by Yahoo Finance and Fortune from the S-1.
Revenue for all of 2025 against the first six months of 2026. Source: Yahoo Finance.

That last row matters more to a buyer than the rest. Roughly a fifth of revenue came from just two customers, Yahoo reports. Fortune's coverage makes the same point from the other side: customers can leave at any time.

How the dependency runs both ways. Figures as reported by Yahoo Finance and Fortune.

What it means for an operations team

The vendor isn't going away

Booking more than three times last year's revenue in six months is not a company about to run out of road. If your agents run on Claude, the filing is no reason to rip anything out. Keep building.

It does change how to read the relationship. A company with upward of $430 billion in compute commitments needs its capacity full and its biggest customers happy. That can work in your favor on price and support. It can also mean pricing and limits get tuned to hit quarterly numbers. That is our read, not something the filing says.

Big buyers move on price

The day before the filing, The Information reported, via Runtimewire, that Replit moved most of its AI spend from Anthropic to OpenAI. The reported terms: a discount of more than 30% on a contract worth over $100 million a year. None of the three companies has confirmed them. Replit's CEO has cited cost, performance and partnership as the reasons OpenAI became its preferred provider.

At that size, switching vendors is a negotiation. At yours, it is an engineering task, and how painful it is depends on choices you make now.

Access can change faster than a contract

In June, Anthropic pulled Fable 5 and Mythos 5 after a US government order, as Tom's Guide reported. Fable 5 came back on July 1, per Android Authority. Whatever you think of the order, the lesson for a buyer is plain: whether you can use a model depends on things your contract doesn't cover.

If that had been your support agent, the question wouldn't have been whether Claude is a good model. It would have been what your customers saw on the first day of the suspension, and whether anyone had the other model ready.

What to do this month

None of this needs a rewrite. It needs five habits, and you can start them on a single workflow.

A switch usually fails in boring places: prompts tuned to one model's habits, tool-call formats, output length, rate limits. Finding those on a calm Tuesday costs an afternoon. Finding them during an outage costs a lot more.

  1. Keep prompts, tool definitions and test cases in your own repository, not only inside a vendor's console.
  2. Put the vendor and model name in configuration, not in the code that calls it.
  3. Write 20 real cases for your most important workflow, with the answer you would accept for each. That is your switching test.
  4. Run a second model on those cases in shadow mode: same input, output compared but not used.
  5. Ask each vendor, in writing, what notice you get before a price or access change.

Shadow mode doesn't need much volume. Your 20 cases, run once a week, are enough to tell you when a second model is close and when it stops being close.

You don't have to switch. The point is that you could, and that your vendor knows it.

Sources

Is your operation tied to one model?We map which workflows depend on a single vendor and what it takes to make each one swappable.

Map my operation

Frequently asked questions

Is Anthropic in financial trouble?
Not on the numbers reported. Revenue for the first half of 2026 was $16.3B against $4.6B in all of 2025. The $80.7B net loss is mostly non-cash charges. The risks to watch are customer concentration and large compute commitments.
When will Anthropic's shares start trading?
No date is confirmed. The coverage we read gives no price range, share count or valuation, and one headline suggests trading could start as soon as November 16.
Should I stop building on Claude because of the IPO?
No. The filing is no reason to remove anything that works. It is a reason to keep prompts, tests and model names portable so a vendor change is a configuration change.
What does it mean to make a model swappable?
It means the model is the cheapest part of your workflow to replace. Keep prompts and tool definitions in your own repository, put the model name in configuration, and keep a test set you can run against a second model.
KytoWe design and build AI automations and business operating systems. Agency results + Academy sovereignty.

More stories

A black and white office facade at dusk with one window still lit in mint

Find the work your team could hand off.

Book 30 minutes. We map the workflows worth automating first and estimate the impact before you commit.