Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Infrastructure

Wall Street Is Raising $60 Billion to Fund Chips for Anthropic. As a Customer, I Read It as a Memo About My Rate Limits.

October 7, 2026 6 min read

On Friday, October 2, Bloomberg reported that Broadcom's Wall Street syndicate was starting to gather $60 billion of fresh AI chip financing "to benefit Anthropic PBC and other companies." Bank of America, Citigroup and Morgan Stanley were poised to send syndication letters for a $42 billion Class A senior-secured tranche, which Broadcom effectively backstops through what Bloomberg calls residual value support. Blackstone is leading an $18 billion Class B junior tranche and committing $9 billion from its own funds. Broadcom declined to comment.

Most coverage read it as a credit-market story: is the AI build-out still financeable? Fair question, not mine. I build a product on Claude. My question is narrower. What does the way my supplier's compute gets paid for tell me about the capacity, prices and limits I will get?

This is not investment advice. It is supplier due diligence.

The plumbing, from the filings

In this structure, Anthropic does not buy the chips. Broadcom's quarterly report, filed with the SEC on September 10, describes the mechanism without naming the customer. A financial partner buys AI racks built on Broadcom's custom accelerators and leases them to "a customer" over five-year terms. Broadcom backstops the lease: if the customer defaults, it covers the gap between 85 percent of what is owed and what the racks resell for. For the first deal, launched in June with $35 billion led by Apollo, that maximum exposure is about $29 billion. Apollo's release says that tranche funds more than one gigawatt of compute for Anthropic, and that the platform aims at more than 20 gigawatts for frontier labs through 2028.

The same 10-Q has a paragraph worth quoting in full:

"In connection with lease agreements that enable access to compute capacity, our customer may, under certain circumstances and if needed, issue to us convertible promissory notes up to an aggregate principal amount of $42 billion. The notes are subject to various conditions, including the requirement to use the notes or proceeds from the notes, if any, solely for the customer's obligations under the lease agreements. As of August 2, 2026, no notes have been issued to us."

The 10-Q does not name the customer. Anthropic's IPO prospectus does, according to Reuters, which reported on October 1 that Broadcom agreed to lend Anthropic up to $42 billion, enough to cover about a third of a $125.2 billion, five-year lease of TPU computing capacity. Reuters places it in the context of the deal Anthropic announced on April 6: multiple gigawatts of next-generation TPU capacity with Google and Broadcom, coming online from 2027. Broadcom's filing that day put the part Anthropic accesses through Broadcom at about 3.5 gigawatts and added that "the parties are in discussions with certain operational and financial partners."

Bloomberg's report does not say which chips the new $60 billion pays for, so I will not either. But the shape is clear: the money for Anthropic's next wave of compute is being raised now, by banks, against long leases.

Capacity: bought long, sold short

The sentence I keep rereading is in that April filing: "The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success."

That is the model. Anthropic commits to years of compute so the capacity exists in 2027. The lenders get leases, hardware as collateral and a guarantee, in writing.

Now look at what a customer gets in writing. Anthropic's rate limit documentation says that "all limits described here represent maximum allowed usage, not guaranteed minimums." The service tiers page opens with a warning that Priority Tier capacity commitments "are no longer available for purchase" and that anyone who needs guaranteed capacity should contact sales. Even existing commitments do not cover Opus 5.5 or Sonnet 5.5, the current models.

None of this is scandalous. Utilities buy long and sell short. But it means the capacity being financed now is an obligation Anthropic carries and a promise I do not hold. My share arrives as a per-minute bucket, on a best-effort basis.

Price: fixed costs want full racks

Five-year leases are fixed costs, and fixed costs want the racks busy. Broadcom describes its guarantees as low-probability liabilities "supported by the strong profitability trajectory of the leading frontier AI labs."

My reading, not a forecast: a supplier in that position competes on volume. It cuts the price of the tokens agents burn most, sells commitments to large accounts and has no reason to price the long tail out. That fits September, when Opus 5.5 launched at $4 and $20 per million tokens, 20 percent below Opus 5, with cache reads 60 percent cheaper.

The other side comes from Anthropic itself. Its prospectus, per Reuters, warns that Broadcom's decisions around pricing and hardware could affect its ability to procure enough computing infrastructure. Part of what Claude costs to run is now negotiated with a company that is also a prospective lender. I cannot price that. I can know it is there.

Concentration: diverse overall, not in this block

Anthropic is not a one-chip shop. Its April announcement says it trains and runs Claude on AWS Trainium, Google TPUs and Nvidia GPUs, that Amazon remains its primary cloud provider, and that Claude is sold on Bedrock, Vertex AI and Foundry.

The concentration sits inside this block. Broadcom co-develops the TPUs with Google, has its name on the financing, backstops the senior debt and may lend $42 billion directly. Per Reuters, Anthropic is expected to become its largest compute customer in 2027, and Anthropic's own filing calls the arrangement a source of "potential conflicts of interest." Robert Leitao of Rothschild & Co told Reuters: "It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened."

I wrote about the wider supplier picture in Anthropic Has Three Single Points of Failure on October 3. This point is narrower. The concentration I can act on is not Anthropic's. It is mine: one model vendor behind my product.

What I check on Monday

Four things, all cheap.

Split your errors. A 429 means my organization hit its own limit, or tripped the acceleration limits the docs describe after a sharp increase in usage. A 529 means the API is overloaded, which the docs say happens with "high traffic across all users." Their ratio is the only capacity signal a small customer gets before a press release, so I log them separately, next to the remaining-tokens headers the API returns on every response.

Ramp, do not jump. If a launch can double traffic overnight, the documentation is explicit: ramp up gradually and keep usage consistent. For the API, a launch is a capacity event.

Treat caching as capacity, not only cost. For most current models, cache reads do not count toward the input tokens per minute limit. The docs' own example: a 2 million limit with an 80 percent cache hit rate processes 10 million input tokens a minute.

Put capacity in the contract if you are big enough to sign one. The Priority Tier that is no longer on sale makes a ready checklist: committed input and output tokens per minute, a named model version, a duration, an uptime target (Priority Tier targets 99.5 percent) and what happens on overflow. If you are not that big, your contract is a second route, the same model through another cloud or another vendor, tested on real traffic.

The people financing Claude's 2027 compute have leases and a guarantee in writing. I have a token bucket. That is fine, as long as I build like it.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk