If Claude runs in production at your company, Anthropic's S-1 is the first audited document that tells you which parts of your contract carry the risk. The filing shows a quarterly operating profit and more than $130 billion in cash, so the exposure is in your terms. Nearly half of revenue arrives through AWS, Google Cloud and Azure marketplaces, so the binding terms may be the cloud provider's. Over 80% of recent revenue comes from models less than six months old, so the version you validated will be retired on someone else's schedule. Two 2026 government actions have already switched models off or put Claude under a federal supply-chain designation. Pin versions, put capacity in writing, and keep a second model qualified for regulated and government-adjacent work.
Anthropic, PBC filed its Form S-1 with the SEC on October 26, 2026 for a listing on the Nasdaq Global Select Market under the ticker "ANT", with Goldman Sachs and Morgan Stanley as lead bookrunners. The share count and price range are still blank. The full registration statement runs to hundreds of pages. The figures below come from the filing as reported by the outlets linked in each sentence.
What Does Anthropic's S-1 Show About Revenue and Profit?
Anthropic booked $16.3 billion of revenue in the first half of 2026 and its first quarterly operating profit. Data Studios' breakdown of the filing puts H1 2026 revenue at $16,286 million, against $1,163 million a year earlier, with an H1 operating loss of $693 million and compute and infrastructure spending of $8,945 million. That compute line alone exceeds the $7,326 million Anthropic spent on compute in all of 2025.
The quarterly split is what matters. ZeroHedge's read of the prospectus reports a Q1 2026 GAAP operating loss of $1.7 billion followed by a Q2 operating profit of $974 million, an 8% margin. Revenue per dollar of compute rose from $0.40 in Q1 2025 to $2.08 in Q2 2026. The $80.7 billion H1 net loss is mostly non-cash: $72.4 billion of it came from fair-value changes on convertible notes, derivatives and SAFEs.
The customer base is wide. As of September 30, 2026, the filing counts more than 11,900 customers paying $100,000 or more a year, more than 1,900 above $1 million and about 250 above $10 million, per Data Studios. Cash and short-term investments stood at $131.7 billion on June 30.
Steel-man the bull case and it holds up. A vendor with that cash balance, a positive operating quarter and four-digit counts of seven-figure customers will not vanish. The WSJ-reported draft said two customers made up about 20% of first-half revenue; ZeroHedge reports the public filing says no single customer exceeded 10% in the first nine months. The contract questions sit in the next four sections.
Why Does 48% Marketplace Revenue Change How You Contract?
Nearly half of Anthropic's revenue is sold through a hyperscaler, so for many buyers the binding terms are the cloud provider's. Sales through the AWS, Google Cloud and Azure marketplaces made up 48% of H1 2026 revenue, up from 11% in 2023, and those platforms account for 58% of net receivables, according to ZeroHedge's summary of the filing. Anthropic books that revenue gross, as principal. Channel partner fees rose from $120 million in H1 2025 to $3.2 billion in H1 2026.
The practical consequence is lifecycle. Anthropic's own model deprecation page says its dates apply to Anthropic-operated platforms, and that partner-operated platforms, Amazon Bedrock and Google-managed Claude on Google Cloud, "set their own retirement schedules." AWS says the same thing from its side: the Bedrock model lifecycle policy, which covers models launched on Bedrock on or after September 7, 2026, offers a Legacy period of either 6 months or 45 days before end of life, notes that "for Amazon Bedrock usage, only the dates on the model card apply," and warns that once a model goes Legacy, existing customers "may lose access after 15 days of inactivity."
So a team calling Claude on Amazon Bedrock, on Google Vertex AI and on Microsoft Foundry can be on three different retirement clocks for the same model. Our Bedrock vs Vertex AI vs Foundry comparison argued that your existing cloud commit should decide where you buy. The S-1 adds a second test: which route gives you the longer, written notice period.
Rate limits follow the same split. Anthropic's rate limit documentation states that all published limits "represent maximum allowed usage, not guaranteed minimums," and that organizations on Claude Platform on AWS cannot use the self-serve tier increase flow. If you need throughput at a set level, it has to be in an order form.
Do Anthropic's Compute Commitments Affect Your Capacity?
They can, because the filing shows Anthropic still renting capacity faster than any one supplier will provide it. The filing says Amazon, Google and Microsoft increasingly see Anthropic as a competitor and have not been willing to supply all the capacity it needs. The public filing lists the result: commitments of at least $114.5 billion to Google Cloud through 2033, at least $110 billion to Amazon through 2036, at least $44.9 billion to Microsoft through 2033, about $161.2 billion of Broadcom leases for Google TPU hardware, and Nvidia GPU capacity from Nscale, Lambda and SpaceXAI, per ZeroHedge's tally, which adds the named contracts to roughly $637 billion. The S-1 itself leaves the aggregate total blank.
One line deserves a buyer's attention. The SpaceXAI agreement (the former X.AI, now a SpaceX subsidiary) covers up to $120.5 billion of Nvidia GPU compute into 2029, with initial non-cancelable periods that end in January and March 2027, after which either side can terminate on 90 days' notice, per the same report. A supplier that can walk on a quarter's notice is a capacity risk Anthropic carries, and it flows to you as rate limits and queueing if it ever bites.
Your lever is the order form. If a workload has a throughput floor, buy reserved capacity and read the break-even first: our provisioned throughput analysis found annual reservations on Azure and Google pay off at 84% and 75% utilization.
Over 80% of Revenue From New Models Means Fast Deprecation
The S-1 confirms that Anthropic's revenue moves to each new model within months, which makes model retirement a routine event for your applications. The filing states that models released in the six months ended September 30, 2026 accounted for "over 80% of our revenue", as ZeroHedge quotes it. A vendor whose income follows the newest model has every reason to free capacity from the old ones.
The deprecation history shows the cadence. Anthropic's deprecation page commits to "at least 60 days' notice" before retiring a publicly released model. Claude Sonnet 4.5 was deprecated on September 30, 2026 and retires on November 30, 2026. Claude Opus 4.1 got notice on June 5 and was retired on August 5. Re-running a regulated evaluation, a model risk review and a change ticket inside sixty days is tight for most enterprises.
A model-version pin is a contract term that names the exact model ID your application runs on and obliges the vendor to keep serving it, or give a stated notice period, before it changes. Without one, the default terms govern. We saw what that looks like downstream when Replit's DPA let it switch its agent's default model without notice.
What Have Government Actions Already Done to Claude Customers?
Twice in 2026 a U.S. government action has reached through Anthropic to its customers, and the S-1 records both. In June, Commerce Department export restrictions forced Anthropic to disable Fable 5 and Mythos 5 for all customers until the restrictions were lifted, per the filing as summarized by ZeroHedge. The controls came from a June 12 letter, and Commerce Secretary Howard Lutnick said on June 30 that they were withdrawn, Fox Business reported. Access returned under new jailbreak classifiers, which Anthropic expected to flag more benign coding and debugging requests, according to CyberScoop. We covered the 19-day outage and the resilience playbook at the time.
The second action is still live. On September 25, 2026, the D.C. Circuit upheld the Department of War's designation of Anthropic as a supply chain risk under 41 U.S.C. § 4713, Taft's law bulletin explains. Taft notes that no current order bars contractors from using Anthropic, and that one would reach an existing contract only through an amendment or modification. Its advice to contractors: inventory where Anthropic products are used, keep that use separate from DoW contract performance, and prepare a transition plan. The S-1 puts government agencies at about 1% of Anthropic's revenue and warns that the dispute could affect customers that sell to the U.S. government, per ZeroHedge. That background goes back to the Pentagon's exclusion of Anthropic in May.
If your company holds defense contracts, the designation is a present-tense planning item, whatever happens to Anthropic's petition for rehearing.
What Should Claude Buyers Put in the Contract?
Use the S-1 as a checklist for your renewal: every risk factor that touches a customer maps to a clause you can ask for.
This Week:
- Export a model inventory. The Claude Console usage export breaks usage down by API key and model, per Anthropic, and Bedrock returns a
modelLifecyclefield for each model. List every Claude model ID you call, on which platform, and its earliest retirement date on that platform. - Flag anything on Claude Sonnet 4.5. It retires on the Claude API on November 30, 2026; check the Bedrock or Vertex model card if you buy it there, because the date can differ.
- If you sell to the Department of War, ask your contracts lead to map which programs touch Claude, directly or through a vendor's product, and to watch SAM.gov for a FASCSA order.
Before Renewal:
- Ask for a version pin with a notice period longer than 60 days for every model behind a regulated or customer-facing workflow, and put the number in the order form.
- Choose the contracting route on notice terms as well as commit drawdown. Compare Anthropic direct, Bedrock's 6-month or 45-day Legacy period, and Google's managed schedule, in writing, for your specific models.
- Turn any throughput you depend on into a committed capacity figure. Published rate limits are ceilings.
- Qualify a second model for each regulated or government-adjacent workload, and route through a gateway so the switch is configuration. Our guide to LLM gateways for multi-provider control covers the options, and this caching analysis explains why you keep the fallback cold and the primary cached.
- Ask what happens to your service if a model is disabled by government order, and get the answer as a credit or termination right.
The Bottom Line
The prospectus moves Anthropic from a vendor you assessed on demos and press reports to one with audited numbers and a risk-factor section you can quote in procurement. Those numbers describe a company that has found an operating profit, sells half its volume through other people's clouds, and depends on a compute supply chain with some short exits. Buyers ran the same exercise with the public cloud providers a decade ago, and the leverage went to customers who negotiated notice periods, capacity terms and an exit before they needed them. The comparison with OpenAI, whose reported run rate we found overstated, gets easier once both companies file quarterly. Your renewal does not need to wait for that.
Pull the model inventory this week, and bring the version-pin clause to your next Anthropic or hyperscaler call.
Continue Reading
- Bedrock vs Vertex AI vs Foundry: Your Cloud Commit Decides
- Replit's DPA Doesn't Require Notice for OpenAI Agent Switch
- 19 Days Dark: How a Shutdown Broke Enterprise AI's Vendor Myth
- Annual Azure and Google AI reservations break even at 84% and 75%
- OpenAI Run Rate Near $50B Is $20B Below Reported $70B
- Best LLM Gateways for Cost Control: Self-Host First