This website uses cookies

Read our Privacy policy and Terms of use for more information.

Hey {{first_name|Investor}} -

Anthropic's contracts with Broadcom, Google, Amazon and Microsoft add up to about $431 billion of lease payments and minimum cloud spend over their full terms, with narrow exits, and some of those suppliers have agreed to lend it money to pay them. And all Class A shareholders together, IPO buyers included, get at most 49.9% of the vote.

On Wednesday I covered the business for Insiders: $11.6 billion of revenue in the June quarter, a first operating profit, and about $2.3 billion of first-half operating cash once you strip out working capital and capital spending (my sensitivity check). Today is what that business has promised to pay, and what's left for a Class A share.

All figures are from the S-1 unless I say otherwise.

The compute commitments

This is the part of the filing I'd spend the most time on. The total and the year-by-year payment table are blank, but the named contracts aren't.

Counterparty

What

Amount

Term

Binding exposure / exit terms

Broadcom

5-year equipment leases, TPU-based compute

~$161.2B in rent, before interest

Capacity from 2026 and 2027

Non-cancelable except on default

Google

Cloud compute

At least ~$114.5B

Apr 2026 to Jul 2033

Minimum spend; shortfall paid as a true-up; exit only on Google's uncured breach

Amazon

Cloud compute

At least $110B

May 2026 to Apr 2036

Minimum spend with true-up; exit only on Amazon's uncured breach

Microsoft

Cloud compute

At least $44.9B

Nov 2026 to May 2033

Minimum spend; exit on Microsoft's uncured breach or for undelivered tranches

SpaceXAI (SpaceX's AI unit, formerly xAI)

NVIDIA GPU compute

Up to $120.5B, ~$3.1B a month at full delivery

Into 2029

Either side can cancel on 90 days' notice after early periods end in Jan and Mar 2027

Nscale

NVIDIA GPU compute

Up to ~$44.6B

Into 2033

Fees only on delivery; depends on Nscale getting financing (Reuters, Oct 22: Nscale pushed its own IPO to later in the year)

Lambda

NVIDIA GPU compute

~$41.3B

Into 2034

Either side can end $34.9B of it if Lambda doesn't secure the facility lease in time

Google and other TPU providers

More TPU compute (outside the commitments table)

Estimated $125B to $150B

2028 to 2031

Depends on capacity and deployment terms

AMD

Helios racks, MI450 GPUs (outside the table)

Over $20B if the full first tranche is taken

From 2027

Non-cancelable except on AMD's breach

These are different kinds of promises, so I'd sort them before adding anything up (my grouping and arithmetic):

  • Minimum spend with narrow exits: Broadcom, Google, Amazon, Microsoft. About $431 billion. Mostly the supplier's uncured breach; Microsoft's can also end tranche by tranche if capacity isn't delivered.

  • Capacity that's capped, cancelable or conditional: SpaceXAI, Nscale, Lambda. Up to about $206 billion.

  • Estimates outside the table: the extra TPU compute and AMD. Roughly $145 billion to $170 billion or more.

The filing's own totals are bigger and cut differently. At June 30, Anthropic reported about $265.6 billion of fixed or minimum commitments under its compute and infrastructure agreements, and about $321.4 billion of leases that hadn't started yet, some subject to conditions. After June 30 it signed new or amended agreements with about $245 billion more of minimum payments, starting between 2026 and 2036, plus short-term agreements with about $21 billion of minimum non-cancelable payments. Those use different dates and categories, and I haven't reconciled them to the named contracts, so they don't stack on my groups. My $431 billion is a subtotal of named contracts, and it isn't Anthropic's total.

One signed deal the S-1 doesn't name: Akamai announced on September 24 a seven-year, $11.6 billion commitment from Anthropic for CPU cloud capacity, which could grow by up to $9 billion (Akamai release and 8-K). It's outside my groups.

For scale: Reuters reported on October 14 that September-quarter revenue was $16.8 billion, and that times 4 is about $67 billion. SpaceXAI alone, at $3.1 billion a month, is about $37 billion a year once fully delivered, which Anthropic expects in December. Without the payment schedule, that comparison tells you size. It doesn't tell you the funding gap in any given year.

The dependence may run both ways. SpaceX's 10-Q shows an unnamed "Customer B" at 34.5% of its $14.6 billion of September-quarter revenue, about $5 billion (my arithmetic). Bloomberg reported on October 29, citing one person familiar with the matter, that Customer B is Anthropic. That's a single-source report, and I haven't seen either company confirm it.

More may be coming. Bloomberg and Reuters reported on October 19 that Anthropic and Nscale were in talks on a second deal of about $45 billion, for another 460-megawatt building at the same West Virginia campus. Those are reported talks, so they're outside my groups above. Nscale has said in its own IPO filing that it hasn't yet lined up financing for the first Anthropic contract, as reported by The Next Web.

On the supplier side, the biggest dollars point at Google's TPU chips: the Broadcom leases, the Google cloud contract and the TPU estimate come to $400 billion or more. NVIDIA's chips reach Anthropic mostly through GPU clouds: SpaceXAI, Nscale and Lambda. Rent paid under the Broadcom leases is Anthropic's cost, and the filing doesn't say how much lands as Broadcom revenue.

The rest of this issue is for Insiders.

Below the line: how Broadcom, Amazon, Google and AMD finance what Anthropic owes them, and why Blackstone changed its credit view. What a Class A share gets, including who holds the other 50.1% of the vote. The government, safety and copyright risks. The reported price. And what I'm watching before the range.

logo

This is where it gets interesting.

Insiders get the full research — the emerging patterns, the second-order effects, and the trends I think matter before they're obvious

Unlock the full post

A subscription gets you:

  • Weekly deep dives on trends before they're consensus
  • Operator-level research and frameworks behind every thesis
  • Full archive of every past issue and report