The $895 Billion Compute Book: Mapping Anthropic’s Supplier Stack

Most AI Power Players stories start with who sells the chips or who builds the campuses. Anthropic’s Amendment No. 1 to Form S-1, filed with the U.S. Securities and Exchange Commission on 2 November 2026, flips the camera. It puts a single number on what one frontier lab says it has already committed to spend on compute and infrastructure — and then breaks that commitment across a multi-vendor stack.

This article stays inside that filing. It does not assess any offering, forecast a listing outcome, or offer investment advice. Every dollar figure below is taken from the S-1/A unless noted as an AndBeyond.Tech sum of disclosed line items.

The headline number — and what it includes

As of and subsequent to 30 September 2026, Anthropic states that it has entered into purchase commitments and lease agreements — including short-term commitments and leases not yet commenced — with an aggregate committed spend of approximately $895.4 billion, “substantially all of which is related to compute and infrastructure.”

Timing, as disclosed:

• About $447.5 billion expected to be incurred before 31 December 2030.
• Substantially all of the remainder expected between 2031 and 2045.
• About $69.8 billion due in the 12 months following 30 September 2026.

The company also warns, in risk-factor language, that some compute arrangements involve minimum purchase commitments, take-or-pay provisions, and other terms that obligate it to pay for resources regardless of actual utilization. That is the difference between a shopping list and a balance-sheet-shaped obligation.

Multi-chip by design

Anthropic describes its compute strategy as a single fungible pool drawn from multiple partners, with workloads steered to wherever the marginal unit is most productive. It trains and serves Claude “natively across four major chip families”: AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, using a proprietary orchestration layer.

That architecture is the industrial logic behind the $895.4 billion. Diversification is not a press-release virtue; it is how the company says it avoids being trapped on one accelerator roadmap while still signing take-or-pay scale with each supplier.

The named supplier book (contractual figures in the S-1/A)

The filing does not present one tidy pie chart. It describes individual infrastructure arrangements. The largest disclosed dollar commitments include:

AndBeyond.Tech sum of the seven lines above: about $643.8 billion. That is less than $895.4 billion — as expected. The headline aggregate also includes other purchase commitments, short-term commitments, leases not yet commenced, and arrangements not broken out at the same level of detail in the passages above. The S-1/A’s contractual-obligations table, for example, presents a narrower subset (purchase commitments and commenced leases) than the $895.4 billion narrative total; Readers should rely on the company’s own reconciliation language, not third-party bridges.

Outside the $895.4 billion — still disclosed, not yet fixed

The filing separately estimates future payments under additional compute arrangements that lack a fixed dollar commitment, including further TPU-related capacity, at roughly $125 billion to $150 billion (and, in related passages, broader estimates of approximately $150 billion to $175 billion for certain arrangements). Anthropic states these amounts are not included in the contractual-obligations table in the same way as fixed commitments.

On AMD, the S-1/A describes a Hardware Supply Agreement (effective July 2026) to purchase, lease or access AMD-based Helios rack-scale systems with Instinct MI450 series GPUs. The agreement does not specify a fixed dollar commitment. AMD has committed to purchase, at Anthropic’s election, up to $5.0 billion of Anthropic Class N common stock if delivery and deployment milestones are met — equity flowing the other direction from the AMD–OpenAI/Meta warrant pattern. Treat AMD volumes as commercial intent under conditions, not as part of the $895.4 billion fixed book.

Liquidity against the near-term bill

The same prospectus puts cash resources next to the obligation calendar. As of 30 September 2026, Anthropic reports:

• $95.6 billion of cash and cash equivalents
• $36.3 billion of short-term investments
• $141.1 billion of capital raised in aggregate to date

Cash plus short-term investments total about $131.9 billion (AndBeyond.Tech sum), against about $69.8 billion of commitments due in the following twelve months. The remainder of the compute book stretches to 2045; the filing’s risk factors tie it to continued access to capital, customer growth and supplier delivery. This is arithmetic on disclosed figures, not an assessment of the company’s financial position.

This article does not analyze dilution, offering size or valuation. It only notes that the S-1/A itself pairs the $895.4 billion disclosure with a large cash buffer — and with risk factors about take-or-pay rigidity if utilization lags.

What the revenue side shows (same filing)

For context on the business funding this stack, the S-1/A reports revenue of $17.0 billion for the three months ended 30 September 2026, and describes rapid growth from commercial launch in the first quarter of 2023. Nine-month 2026 results and GAAP losses appear in the financial statements; this article does not restate every P&L line or interpret profitability. The Power Players takeaway is narrower: Anthropic is already both a large buyer of compute and a large seller of AI services, which is the commercial basis on which counterparties underwrite multi-year infrastructure.

Claude is also distributed through the three major clouds — AWS, Google Cloud and Microsoft Azure — under arrangements in which Anthropic says it keeps direct customer contracts and pays the CSP a fee. That go-to-market detail matters because the same companies appear on the supply side of the compute book. Equity investments, cloud resale and take-or-pay hosting can sit in one relationship, which the filing flags as a source of potential conflicts.

Fixed, contingent and cancelable — three risk shapes

Not every dollar in the $895.4 billion behaves the same way if demand or delivery slips.

Fixed take-or-pay (Google, Amazon, Microsoft). These CSP hosting deals use language of “at least” spend “regardless of actual usage,” with true-ups if Anthropic underspends. Cancellation is generally limited to uncured material breach (and, for Microsoft, undelivered tranches). Once capacity is accepted, the economic risk of idle GPUs sits largely with Anthropic.

Lease-shaped TPU infrastructure (Broadcom). The ~$161.2 billion of rent (excluding interest) is tied to equipment becoming ready for service. Timing risk is operational — when halls light up — more than optional cancel-for-convenience.

Cancelable or financing-conditioned GPU capacity (SpaceXSI, Nscale, Lambda). SpaceXSI’s “up to $120.5 billion” becomes cancelable on 90 days’ notice once the initial non-cancelable periods end in early 2027. Nscale’s “up to ~$44.6 billion” accrues only as capacity is accepted and depends on qualifying financing. Lambda’s ~$41.3 billion includes a large block terminable if a facility lease is not secured on time. These lines can shrink without Anthropic “breaking” a take-or-pay.

No fixed dollar (AMD Helios; additional TPU estimates). Strategic and potentially large, but deliberately outside the fixed aggregate until dollar minimums exist.

Reading the compute book as one undifferentiated $895 billion therefore overstates how similar the contracts are. The Power Players map is really a map of obligation shapes.

How to read a compute book without overreading it

Three discipline rules follow from the text:

1. “At least” and “up to” are different. Google, Amazon and Microsoft are framed as minimum spend regardless of usage. SpaceXSI and Nscale are framed as “up to,” with cancelability or financing conditions that can shrink the realized total.
2. Delivery risk sits with suppliers and developers. Nscale’s qualifying-financing covenant and Lambda’s facility-lease contingency are explicit reminders that a signed commitment is not energized megawatts.
3. Chip family ≠ cash commitment. Naming Trainium, TPU, Nvidia and AMD as supported architectures is not the same as equal dollar weights. Broadcom/TPU lease economics and Google hosting dominate the disclosed fixed dollars; AMD appears strategically important but without a fixed spend line in the $895.4 billion total.

Why this matters for the AI Power Players map

Hyperscaler capex tells you what cloud companies are building. Dirt-to-model campus stories tell you who is trying to own land and power. Anthropic’s compute book tells you something else: how a frontier lab finances and diversifies scarcity.

The Power Players on the other side of these contracts — Amazon, Google, Microsoft, Broadcom, SpaceXSI (SpaceX’s AI unit), Nscale, Lambda, AMD — are not interchangeable. Some sell cloud with true-ups. Some sell leased accelerator infrastructure. Some sell cancelable GPU capacity. One (AMD) sells racks and may buy Anthropic stock if milestones clear. The $895.4 billion is less a single bet than a portfolio of differently shaped dependencies.

Bottom line

Anthropic’s S-1/A of 2 November 2026 is one of the clearest primary-source maps yet of a frontier lab’s supplier stack. Approximately $895.4 billion of committed compute and infrastructure spend, about half of it expected before 2031, spans take-or-pay cloud deals, TPU equipment leases, Nvidia-powered capacity from multiple hosts, and a multi-chip software strategy that includes AMD without yet booking AMD as a fixed dollar minimum. The number is large because the obligations are contractual. Whether every gigawatt arrives on time is a delivery question the filing itself refuses to paper over.

Related reading

• Equity for Silicon: How AMD Turned OpenAI and Meta Into Stakeholders
• Capex Arms Race: Mapping Hyperscaler AI Spend and What It Buys
• Vertical Integration from Dirt to Model
• Beyond Stargate: The New Gigawatt Map of AI Factories

Images: original illustrations generated for AndBeyond.Tech; no logos or real people depicted.