Anthropic reportedly struggles to secure enough cloud capacity, and the hyperscalers that supply it increasingly see it as a competitor. TechCrunch's brief draws on a leaked S-1 filing, so its figures are not yet confirmed. For a founder who builds on Claude, the practical question is not Anthropic's balance sheet. It's whether capacity limits and vendor concentration could slow your product, raise your costs, or cut you off.
What Has Been Reported
The capacity squeeze is the core claim. TechCrunch says Anthropic is having trouble getting enough capacity from AWS, Google and Microsoft, even as its infrastructure spending grows. It also says the hyperscalers view Anthropic as a rival rather than only a customer.
The filing figures conflict across outlets. TechCrunch cites a first-half net loss of $80.7 billion and first-half revenue of $16.3 billion. SiliconANGLE, citing Reuters and the Financial Times, reports 2025 revenue of about $4.6 billion and a 2025 net loss of $42 billion. The periods don't match, so don't combine them. TechCrunch's "more $750 billion" cloud figure also doesn't appear in SiliconANGLE's account, which cites $518 billion in planned infrastructure spending over a decade.
Compute Limits Reach Your API Bill
Limits depend on your usage tier. Anthropic's rate limit documentation sets three tiers, Start, Build and Scale, with monthly spend caps of $500, $1,000 and $200,000. New organizations can start in an Evaluation tier with lower limits until they build usage history.
Two different limits can stop your traffic. A spend cap pauses API use until the next month, and the response carries no retry-after header, so retrying will not help. A rate limit returns a 429 that may include a retry-after value. Your code should handle these two cases differently.
Limits are maximums, not guarantees. The documentation says all limits "represent maximum allowed usage, not guaranteed minimums." Short bursts can exceed a per-minute limit even when your daily total is fine. The docs also warn that sharp jumps in usage can trigger acceleration limits, so ramp traffic gradually.
Caching raises your effective throughput. For most current models, cached input tokens don't count toward input-token-per-minute limits. The documentation gives an example with an 80 percent cache hit rate. Check your cache hit rate on the usage page before you request a higher tier.
Cloud marketplaces change billing, not limits. The docs say Claude Platform on AWS uses the same rate limits, but it bills through AWS Marketplace and doesn't offer the same tier-increase flow.
When Your Cloud Provider Is Also a Rival
Your host may compete with your model vendor. SiliconANGLE reports that nearly half of Anthropic's cloud budget goes to AWS, Microsoft and Google. Those same companies build their own models and chips, so the relationship has two sides.
Commentary sees a lock-in risk. Aragon Research analyst Jim Lundy argues that Anthropic has traded independence for guaranteed compute. He also says customers should watch for indirect cloud lock-in and for how funding arrangements affect pricing and model availability. This is an analyst's view, and the claims he cites are not all sourced.
Check your own dependencies. List every layer your product depends on: the model provider, the cloud that hosts it, and any marketplace that bills you. If one company sits in two of those layers, plan for a conflict.
Don't Depend on One Model Vendor
Concentration is a risk for vendors too. TechCrunch says two customers account for about a fifth of first-half sales. I couldn't confirm that figure, but the principle applies to you. If one model provider sets your pricing, limits and availability, a single capacity squeeze can hit your whole product.
Build a fallback path. Put model calls behind one internal interface, so you can switch providers or route simpler tasks to a cheaper model. Test that path regularly, not only during an outage.
Read Your Own Contracts Before the Liability Does
The filing warns about harmful model behavior. SiliconANGLE quotes the filing's warning that the company's models "could further increase the risk that our models cause harm." TechCrunch reports that the filing also warns damages could exceed Anthropic's insurance coverage. I couldn't verify that insurance claim.
Check your own coverage. Ask your insurer whether an AI-related incident is covered, and read the exclusions. Then check the indemnity clauses in your customer contracts, since they decide who pays when a model's output causes a loss.
Read Leaked Numbers Carefully
Leaks are early and partial. A leaked draft filing can change before it's public, and outlets pick different figures. Treat any single number as a lead, not a fact.
Match periods and sources. Before you repeat a figure, check whether it covers a quarter, a half-year or a full year, and whether the outlet cited the filing or a summary.
Recommendation Blocks
Plan for limits if: - Your product sends bursts of traffic, or your usage is growing quickly - Your users would notice a pause in the feature Diversify your model layer if: - One model provider handles most of your product's core calls - You have no tested fallback for an outage or a pricing change Review your contracts if: - Your product gives advice, generates content or takes actions on behalf of customers - Your insurance doesn't mention AI-related incidents
A Decision Framework
Step 1: Map your dependencies. List your model provider, your cloud host and any marketplace, and note any company that appears more than once.
Step 2: Read your limits. Check your tier, spend cap and cache hit rate, and set alerts before you reach them.
Step 3: Handle both errors. Treat spend-cap errors and rate-limit errors differently in your code.
Step 4: Build a fallback. Route calls through one interface, and test a second provider on a real workload.
Step 5: Review coverage and contracts. Confirm your insurance and indemnity terms before you sell into a regulated or high-stakes market.
Bottom Line
Anthropic's reported capacity squeeze matters most through the limits, pricing and availability it can affect in your product. The filing figures are early and conflicting, so use them as warnings, not benchmarks. The durable lessons are practical: know your tier and spend cap, avoid depending on one company in two layers of your stack, and keep a tested fallback.
This isn't financial or legal advice. Talk to your insurer and counsel before you change your contracts.
