Net revenue retention of about 900%. That is the number in Anthropic's IPO filing that should get every enterprise AI buyer's attention, more than the valuation and more than the revenue.
A recent piece on TechCrunch walked through what the S-1 reveals about how Anthropic keeps enterprise customers. I read it from the buyer's side of the table, because that is where most of my readers sit.
What the filing actually says
As of September 30, Anthropic had more than 11,900 customers spending at least $100,000 a year, more than 1,900 spending over $1 million, and about 250 spending $10 million or more. Revenue for the first half of 2026 was $16.3 billion.
The strategic number is this one. More than half of the customers spending at least $100,000 now use both the Claude platform and a first-party app like Claude Code or Cowork. At the end of last year, that share was about 30%.
And the apps change spending. In the third quarter, the top 5% of Claude Enterprise customers using Claude Code spent about 20 times as much per month as the average enterprise user who did not use the coding tool.
The ecosystem playbook, applied to AI
The pattern is familiar. Amazon Web Services started with compute, then added storage, databases and dozens of other services until leaving became a project nobody wanted to own. Uber started with rides and added food and freight.
Anthropic is running the same play across three layers: the models, the developer platform and the applications people use directly. A company starts with the API, rolls out Claude Code to engineers, then Cowork to the wider workforce. Each step is sold to a customer who already trusts the vendor, which is far cheaper than finding a new one.
The filing describes extra pressure behind that motion. An internal Applied AI team helps embed Claude inside customer stacks, and a separate venture with Blackstone and other private equity firms, called Ode with Anthropic, places forward-deployed engineers inside businesses to build bespoke tools. Anthropic also holds the direct contract when it sells through AWS, Google Cloud and Microsoft Azure, which lets it set pricing and own the relationship.
Where the switching cost really lives
I build on Claude in my own products, including Madison AI, so I am not writing this as a critic of the company. The products are good, which is exactly why the lock-in works.
But buyers should be clear about where the switching cost sits now. It used to be in the API. Changing a model behind an API call is a few days of work and some testing. Most serious teams already route between providers.
The apps are different. Once hundreds of employees build their daily habits, project memory, shared prompts, connectors and automations inside one vendor's application, the model underneath becomes almost irrelevant. You are no longer locked into a model. You are locked into a workplace.
The second-order effect is on negotiation. A 900% net revenue retention figure means existing customers keep expanding. Every expansion makes the next renewal harder to walk away from, so pricing power shifts to the vendor with each new team that adopts the app.
I covered the risk-factor side of the earlier filing in what Anthropic's prospectus means for buyers. This filing shows the other half: the growth engine is built on exactly the dependency those risk factors describe.
The vendor admits its own exposure
The filing is honest about the tension. Anthropic says the cloud providers that help sell Claude increasingly see it as a competitor, and that Amazon, Google and Microsoft may prioritise their own requirements or rival models they have invested in, which could make compute harder to secure.
It also says a significant reduction in spending by its largest customers could hurt the business. That sentence is worth reading twice. If your company is one of the largest customers, you have more bargaining power than you think, at least until the next team rolls out the app.
The filing also says Anthropic will spend less effort entrenching current products and more on new classes of products meant to dwarf the existing ones. For a buyer, that means the app you standardise on today may be repositioned in eighteen months.
What I would do as a buyer
First, separate the layers in your contracts and your architecture. Keep model access behind your own routing layer, so the API stays replaceable. I made the broader case in don't let one vendor own your agents, and the filing makes it more urgent.
Second, decide deliberately which work lives inside the vendor's apps and which lives in systems you own. Prompts, workflows and project knowledge should be exportable and documented outside the app.
Third, negotiate while you are expanding, not at renewal. Seat growth is your bargaining power. Use it for price protection, data portability terms and exit clauses.
If you would rather build the workflows inside your own architecture than inside a vendor app, difrnt.ai designs custom AI solutions end-to-end on top of whichever models fit.
Anthropic wrote a very clear plan for keeping its customers. Buyers should write one for keeping their options.