Anthropic's confidential IPO prospectus, leaked this morning and reported by Reuters, reveals a scale of vendor commitment I have not seen before. Anthropic expects to spend at least $518 billion over a decade on AI infrastructure with six partners, and about 80% of that sum is non-cancelable or payable regardless of usage. The company plans to spend at least $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft over the next seven to 10 years. If actual spending falls short of the Google commitment, Anthropic must pay the difference, and similar terms apply to Amazon. A separate $161.2 billion in Broadcom-related equipment leases is largely non-cancelable. The vendor relationship has changed. You do not cancel when the bill is too high. You pay whether you use the capacity or not.
I do not commit $111 billion to a cloud vendor. I run a construction ERP at coenconstruction.com, an estimating SaaS at estimate.pro, and a permit-drawing platform that makes four thousand Claude API calls per month. The construction ERP calls OpenAI for invoice OCR. The estimator calls Anthropic for line-item extraction. The permit tool calls Gemini for site-plan markup and Claude for zoning-code summaries. Every integration depends on an API I do not control, but every contract allows me to stop paying when I stop calling. I can switch from Claude Opus 5 to Opus 5.5. I can switch from Opus to Gemini. I can switch from Gemini to a self-hosted model if the bill exceeds the value. Anthropic cannot. The vendor can raise prices, change the API, or deprecate the model, and the $111 billion commitment to Google does not reduce.
The lock-in multiplies with the infrastructure spend
Revenue grew 12-fold in 2025 to nearly $4.6 billion, even as the company lost more than $8 billion on an operating basis. The AI lab spent $7.33 billion on compute and infrastructure last year, a threefold surge from 2024, accounting for more than half of its $12.65 billion in total operating expenses. The construction ERP runs on a Cloudflare Worker with a D1 database. The monthly bill is $47. The estimator runs on Vercel with a Postgres database. The monthly bill is $83. The permit platform runs on another Worker. The bill is $52. Total infrastructure cost across three platforms is $182 per month, or $2,184 per year. If Cloudflare raises Worker pricing by 50%, I can migrate to AWS Lambda in two days. If Vercel changes the deployment API, I can move to Netlify or Render. The migration takes time. It does not take seven years and it does not require me to pay the original vendor after I leave.
Anthropic cannot migrate. The company committed to pay Google between April 2026 and July 2033 and Amazon between May 2026 and April 2036—if actual spend falls short, the company must pay Google the difference, and similar terms apply to Amazon. The commitment with Microsoft of $31.4 billion between November 2026 and May 2033 is non-cancelable except in the event of Microsoft's uncured material breach. The vendor has the leverage. Google can raise API prices. Amazon can deprecate an S3 endpoint. Microsoft can change the Azure authentication flow. Anthropic still pays. The company is locked into three cloud vendors, one chip vendor, and two additional compute partners for the next seven to ten years. If Google's API changes, Anthropic rewrites the integration or the training run breaks. If Amazon raises storage prices, the bill goes up and the contract does not allow Anthropic to move the data. The integration is temporary. The payment is not.
The non-cancelable clause does not protect the platform
The construction ERP has five dependencies: OpenAI for invoices, Twilio for SMS, Plaid for bank data, Stripe for subscriptions, and Cloudflare for the Workers runtime. The estimator has six: Claude for blueprints, Mapbox for project maps, an SMTP relay, Stripe, Vercel for hosting, and a separate Postgres instance. The permit tool has seven: Gemini and Claude for documents, Mapbox for site plans, two Cloudflare Workers, a D1 database, and a review gate that calls a third Worker. Every dependency is a vendor I do not control. Twilio can raise SMS prices from $0.0079 to $0.0095 per message. OpenAI can deprecate gpt-5.6 and release gpt-6-sol with a different JSON output schema. Stripe can add a required webhook signature. When the vendor changes the API, I change the code or the platform breaks. When the vendor raises prices, I pay more or I migrate. I do not sign a ten-year non-cancelable contract that requires payment regardless of whether I call the API.
About 80% of the $518 billion is non-cancelable or requires payment regardless of usage, and Anthropic is telling investors the commitments are necessary because access to computing power is becoming the key constraint on AI development. The constraint is real. The contract does not make the constraint go away. It moves the risk from the vendor to the customer. Google gets paid whether Anthropic uses the capacity or not. Amazon gets paid whether the model trains successfully or not. Microsoft gets paid whether Azure's API works or not. The vendor can shut down the API, change the authentication method, or deprecate the endpoint, and the non-cancelable clause does not stop them.
The contract locks the payment schedule. It does not lock the API.
The construction ERP makes twelve thousand OpenAI API calls per month. If OpenAI raises prices, the bill goes up. If OpenAI deprecates the model, I rewrite the prompt and redeploy. If OpenAI shuts down the API, I migrate to Claude or Gemini in three days and the platform stays online. Anthropic committed $252 billion to Google, Amazon, and Microsoft. The vendor can cut prices. The vendor can also raise them, and the contract still requires payment. The infrastructure commitment is non-cancelable. The model it trains is not.