Google TPUs account for $282.5 billion, or 43.9% by our arithmetic, of the $643.8 billion in compute contracts with stated dollar amounts in the amended S-1 registration statement Anthropic filed with the SEC on 2 November. Nvidia GPUs account for $251.3 billion (39.0%) and Amazon for $110.0 billion (17.1%), grouping each contract by the chip platforms the filing itself lists.
The two blocs carry different terms. Every TPU dollar sits in contracts the filing describes as non-cancelable. Of the $251.3 billion tied to Nvidia GPUs, $206.4 billion, or 82% by our arithmetic, can be cancelled on 90 days’ notice after early 2027 or accrues fees only as capacity is delivered.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on paying for usage instead of fixed commitments. It teaches the general idea only and does not discuss Anthropic’s contracts.
The key insight: Anthropic’s amended S-1 names seven compute contracts with stated amounts worth $643.8 billion. By our arithmetic, Google TPUs are $282.5 billion of it, or 43.9%, all in contracts the filing calls non-cancelable, while $206.4 billion of the $251.3 billion tied to Nvidia GPUs, or 82%, can be cancelled on notice after early 2027 or is billed only as capacity is delivered.
Broadcom, $161.2 billion. Anthropic has five-year master equipment leases with Broadcom and certain financing entities. The filing says rent over their life equals “approximately $161.2 billion excluding interest” and that the leases “provide us with access to TPU-based compute capacity”.
Google, $121.3 billion. Under a CSP hosting agreement amended between May and October 2026, Anthropic “committed to spend at least approximately $121.3 billion between April 2026 and July 2033, regardless of actual usage”.
Amazon, $110.0 billion. An April 2026 agreement commits “at least $110.0 billion between May 2026 and April 2036, regardless of actual usage”. The filing says the Amazon collaboration provides access to “AI accelerators, including custom-designed Trainium chips”.
SpaceXAI, up to $120.5 billion; Microsoft, at least $44.9 billion; Nscale, up to about $44.6 billion; and Lambda, about $41.3 billion. The filing lists its “arrangements with Microsoft, SpaceXAI, Nscale, Lambda, and other cloud providers” as its route to “compute capacity powered by NVIDIA GPUs”.
Anthropic’s seven compute contracts with stated amounts in its S-1/A, $643.8B in total, grouped by chip family: Google TPUs $282.5B (43.9%), Nvidia GPUs $251.3B (39.0%), Amazon including Trainium $110.0B (17.1%). Ochre marks the $206.4B that can be cancelled on 90 days’ notice after early 2027 or is billed only as delivered. Data: Anthropic S-1/A, 2 Nov 2026; our arithmetic.
What Was Posted on X
$GOOGL $AVGO $AMZN $MSFT $NVDA $SPCX $AMD $ANT
google’s TPUs are 44% of the $644bn of named compute deals in anthropic’s amended S-1, and each deal leaves the hardware risk with a different party.
1. broadcom, $161.2bn of TPUs on five-year leases in two SPVs, so anthropic carries that risk, with broadcom’s $42bn convertible facility funding the lease fees.
2. google, $121.3bn, up from $114.5bn in the first filing and payable regardless of usage, with google carrying the hardware.
3. spacex, up to $120.5bn of nvidia capacity, about $3.1bn a month once delivered and terminable by either side on 90 days’ notice from january and march 2027.
4. amazon, at least $110bn of trainium to april 2036, take-or-pay.
5. microsoft, at least $44.9bn on azure.
6. nscale, up to $44.6bn, and lambda, $41.3bn, where the neoclouds and their lenders carry it.
7. AMD, more than $20bn if all the initial capacity is bought, which sits outside the $644bn.
by chip that’s $282.5bn of google TPUs, $251.3bn of nvidia GPUs and $110bn of trainium.
it all sits inside about $895bn of commitments, about $69.8bn of it due within twelve months, while third-quarter revenue was $16.96bn against about $9.86bn of compute.
broadcom is the sharpest case to me. it’s lending at 9% against hardware whose replacement it’s also designing, and its residual support pays out only if anthropic defaults and the racks resell for less than 72.5% of what’s owed.
anthropic’s IPO, if and when that happens, is when the public stock buyers would join the end of that chain. not any time soon.
A post by @nodiligence at 12:58 UTC on 8 November that sums the named compute contracts in Anthropic’s amended S-1 by chip, putting Google TPUs at 44% of $644bn. The amounts in this piece are taken from the filing itself, and our arithmetic on them matches the post’s totals.
Fixed Terms on TPUs, Flexible Terms on GPUs
The filing says the Broadcom leases cannot be cancelled by either party during the lease period, except by the lessor in the case of a default. The Google and Amazon commitments are owed regardless of usage, with a true-up if spending falls short, and are non-cancelable except for the counterparty’s uncured material breach.
The GPU side is looser. After initial periods ending in January and March 2027, each SpaceXAI agreement “may be cancelled by either party upon 90 days’ advance written notice”. Nscale and Lambda fees accrue “only as capacity is delivered to and accepted by us”, and for Lambda’s August 2026 agreement, about $34.9 billion of the total, either party may terminate if Lambda does not secure the facility lease in time. Microsoft’s $44.9 billion is the exception: non-cancelable except for breach or undelivered tranches.
Adding SpaceXAI ($120.5 billion), Nscale ($44.6 billion) and Lambda ($41.3 billion) gives $206.4 billion, which is 82% of the $251.3 billion Nvidia bloc. That is our arithmetic on the filing’s figures; the SpaceXAI, Nscale and Lambda amounts are maximums or depend on delivery.
The TPU Share Grows With the Filing’s Own Estimates
Two amounts sit outside the $643.8 billion. The filing says Anthropic also “committed to procure additional TPU compute, without a fixed dollar commitment”, expected between 2028 and 2031, and estimates those payments at “approximately $125.0 billion to $150.0 billion”. Its AMD agreement has no fixed dollar commitment, but Anthropic expects spending “to exceed $20.0 billion” if it buys the full initial capacity.
Adding both estimates, with AMD at $20 billion, TPUs would be $407.5 billion to $432.5 billion of $788.8 billion to $813.8 billion, or about 52% to 53%, by our arithmetic. Because the AMD figure is a floor, those shares are upper bounds. Our earlier piece on how Anthropic’s $895 billion of commitments compare with Meta’s first reported the additional TPU estimate.
The Business Lever: A Committed Base and a Flexible Layer
The filing describes a “multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs” that lets Anthropic “orchestrate workloads to optimize for price and performance”. As we read the contract terms, the TPU and Amazon contracts form the committed base, while most of the Nvidia capacity bought through SpaceXAI and the neoclouds can be cancelled or paid for only as it arrives.
The split matters for who carries the risk if demand changes. SpaceXAI, Nscale and Lambda are the parties whose revenue depends on delivery or on notice periods, as our coverage of the SpaceX bill reaching $3.1 billion a month and Apollo’s $75 billion of chip deals has followed. Google and Amazon hold commitments the filing says are owed regardless of usage, and Broadcom holds leases the filing says neither party can cancel during the lease period except the lessor on default.
As we read the filing, Anthropic has put its TPU capacity, through Broadcom leases and a Googlecloud contract, entirely on non-cancelable terms, and kept most of its Nvidia capacity on terms that can be cancelled on notice or are paid only on delivery.
As we read the terms, the neoclouds and SpaceXAI carry more of the delivery risk than Google, Broadcom or Amazon do, because their fees depend on delivery or on notice periods.
As we read the estimates, the TPU share would rise to about half of named compute if the filing’s $125.0 billion to $150.0 billion of additional TPU compute is delivered, by our arithmetic.
Anthropic, amended S-1, 2 November 2026
“multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs”
Three Implications
FOR GOOGLE AND BROADCOM The filing says the Google commitment is owed regardless of actual usage, and that the Broadcom leases cannot be cancelled during the lease period except by the lessor on default.
FOR NVIDIA CAPACITY PROVIDERS The filing says each SpaceXAI agreement may be cancelled by either party on 90 days’ notice after initial periods ending in January and March 2027, and that Nscale and Lambda fees accrue only as capacity is delivered and accepted.
FOR THE NUMBERS TO WATCH The filing estimates $125.0 billion to $150.0 billion of additional TPU compute delivered between 2028 and 2031, and expects AMD spending to exceed $20.0 billion if it buys the full initial capacity; neither is in its contractual obligations table.
The essay describes “Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”
As we read it, Anthropic’s filing puts dollar amounts on two of those categories: $282.5 billion of hyperscaler-custom TPUs and $110.0 billion with Amazon, which includes Trainium, against $251.3 billion of Nvidia’s generalist GPUs, by our arithmetic. The essay does not discuss Anthropic’s contracts.
What Is Not Established
The filing does not give the chip mix inside the Microsoft or Amazon contracts; we group Microsoft with Nvidia and Amazon with Trainium because that is how the filing lists its chip platforms. It does not give prices per chip, the power behind each contract, or how much has been spent so far. The seven contracts sum to $643.8 billion, about 72% of the roughly $895.4 billion headline commitment by our arithmetic, and the filing does not map each contract into that total.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic’s S-1/A names seven compute contracts with stated amounts worth $643.8 billion. By our arithmetic, Google TPUs are $282.5 billion of it, or 43.9%, all on terms the filing calls non-cancelable, while 82% of the $251.3 billion tied to Nvidia GPUs can be cancelled on notice after early 2027 or is billed only as delivered.
A note on sourcing. Contract amounts, dates, terms and quotations are from Anthropic’s amended S-1 registration statement filed with the SEC on 2 November 2026. The grouping by chip family follows the filing’s own list of chip platforms; sums and shares are our arithmetic. A post on X by @nodiligence first summed the seven contracts by chip on 8 November.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
