Anthropic filed Amendment No. 2 to its Form S-1 on November 10, 2026, updating how it describes Microsoft and Google infrastructure spend. The filing text raises stated Azure commitments to at least $47.2 billion and up to $49.2 billion through June 2033, and Google CSP commitments to at least about $121.3 billion through July 2033, in each case regardless of actual usage with limited exit rights.
If you build on Claude, this is not a product launch. It is IPO disclosure about where Anthropic has locked spend for the rest of the decade, beside Amazon, SpaceXAI, Lambda, and TPU option language already in the registration statement.
TL;DR
| Counterparty | Stated commitment (S-1/A-2 text) | Period | Cancelability (filing) |
|---|---|---|---|
| Microsoft | At least $47.2B, up to $49.2B | Nov 2026 – Jun 2033 | Non-cancelable except Microsoft uncured breach or undelivered tranches |
| Google CSP | At least ~$121.3B | Apr 2026 – Jul 2033 | Non-cancelable except Google uncured breach; shortfall true-ups |
| Google TPU add-on | ~$125B–$150B estimated (not in obligations table) | Delivery 2028–2031 | Payment mechanics vary by delivery method |
| Lambda Aug tranche | ~$34.9B of ~$41.3B stack | Into 2030s | Either party may exit if lease not timely secured |
What changed in Amendment 2
The October 26 filing already listed big-number hyperscaler tables. Amendment 2, dated November 10, 2026, reflects contract amendments through November on the Microsoft side and through October on Google, per the risk-factor prose.
The passages builders cite most often are verbatim in EDGAR:
- Microsoft: agreement from November 2025, amended February through November 2026, committing $47.2B minimum and $49.2B maximum between November 2026 and June 2033, regardless of actual usage.
- Google CSP: agreement from March 2026, amended May through October 2026, committing at least approximately $121.3 billion between April 2026 and July 2033, again regardless of actual usage, with true-up if spend falls short.
Secondary coverage from The Information summarized the day as roughly $10 billion of additional Microsoft and Google commitments versus earlier drafts. Always reconcile journalists against the SEC PDF before you put numbers in procurement decks.
Take-or-pay language and builder impact
Take-or-pay (or minimum spend regardless of usage) means Anthropic owes the floor even if a model generation shift moves work between hosts. For API customers the practical questions are narrower:
- Will my Claude tier slow down? Not because of this filing alone. Throughput still depends on accepted capacity, routing, and incident response — the same variables in our Fable 5.2 launch week coverage.
- Does this change token prices today? No announced list-price move accompanied Amendment 2. Economic pressure from $895B-scale obligations shows up in investor materials, not necessarily in next week''s invoice.
- Should I diversify hosts? If you already multi-home models across vendors, this filing is a reminder that Anthropic itself multi-homes at contract level. Teams running agent harnesses should keep portable evals and provider failover keys whether or not Anthropic exercises a neocloud exit clause.
The filing also repeats an estimated $125B–$150B band for additional TPU compute delivered 2028–2031, outside the contractual obligations table. That line matters for Google TPU vs Nvidia GPU mix, but it is an company estimate, not a fixed invoice.
Anthropic S-1 amendment cloud spend across hyperscaler data centers through 2033
How this sits next to Lambda and the IPO roadshow
On the same news cycle, outlets including Crypto Briefing relayed that Anthropic does not plan to terminate the $34.9B Lambda tranche after a reported lease miss. explainx.ai published the termination-right details in a dedicated post on Lambda lease language.
Think of the mosaic this way:
| Bucket | What Amendment 2 emphasizes | Builder takeaway |
|---|---|---|
| Hyperscaler (Microsoft, Google) | Amended floors, hard to cancel | Core Claude training and serving likely stay multi-cloud |
| Neocloud (Lambda) | Conditional exit if lease slips | Supply narrative risk, not an automatic API outage |
| IPO roadshow | Fresh numbers during pricing week | Watch effective date and later 10-Q tables for realized spend |
Our S-1 hub analysis walks the broader $895B commitments frame; this amendment is the Microsoft and Google delta slice.
Microsoft vs Google: what the filing does not split
The S-1 prose does not publish a per-model routing map. You cannot read Amendment 2 and learn whether Opus batch jobs sit on TPU vs Blackwell tonight.
What you can infer responsibly:
- Both deals are non-cancelable except for provider breach — stronger than Lambda''s lease trigger.
- Google''s paragraph couples CSP hosting with a TPU procurement option tree (host with Google, buy with financing, or third-party TPU lease). That matches Google''s public push on TPU variants and Anthropic''s historical TPU use.
- Microsoft''s amended tranche lands during the same quarter ChatGPT Enterprise hit AWS Marketplace, a reminder that frontier labs buy capacity from every hyperscaler even while they compete on models.
Finance vs engineering: who should read this
Investors read Amendment 2 for cash timing and concentration risk ahead of the roadshow launch.
Engineering leads should skim it for single-vendor assumptions in multi-year roadmaps. If your compliance packet says Claude runs only on one neocloud, the filing disagrees at the contract layer.
FinOps teams should note regardless of actual usage language. Anthropic''s COGS story can diverge from your utilization if you autoscale down while their minimums stay fixed — that does not directly bill you, but it can influence discount renewals described in our enterprise discount cap reporting.
Questions procurement should ask Anthropic or resellers
| Question | Why ask it now |
|---|---|
| Which public regions map to which host class for our Enterprise workspace? | Filing shows multi-host contracts, not your tenant placement |
| Do incident comms distinguish hyperscaler vs neocloud faults? | Lambda headlines can spook teams even when Azure routes cover slack |
| Are committed-use discounts tied to Anthropic''s host floors or your token chart? | Take-or-pay on their side does not automatically mean cheaper overage for you |
Document answers in your AI registry the same way you would for Bedrock tokenomics or Vertex access claims.
Related reading
- Anthropic Lambda $34.9B termination rights
- Amended S-1 Q3 revenue and commitments
- Anthropic IPO roadshow
- TSMC board strategic investments
- Microsoft Chinese transceiver ban supply risk
- Neocloud selloff and S-1 verification
Private placement and Google concurrent tranche (same amendment)
Amendment 2 is not only about rent. The filing also describes concurrent private placements with AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC Partners for $26.0 billion of Class N stock at the IPO price, and a Google option to purchase up to $10.0 billion more concurrently with the offering, which Google agreed to exercise in full per the text.
Those equity lines matter for capacity planning because they bind strategic suppliers to Anthropic''s cap table at pricing week. They do not tell you which GPU tranche powers your Claude Code session tonight, but they explain why compute disclosures arrive in one dense S-1 instead of scattered press releases.
Amazon and SpaceXAI lines still anchor the mosaic
Readers sometimes treat the $10B Microsoft and Google headline as if it replaced other hosts. The same registration family still lists ~$110B Amazon commitments and up to ~$120.5B SpaceXAI caps in indexed summaries such as MetaPrompt''s obligation table. Amendment 2 is an update, not a reset.
When you model vendor concentration risk for your org, copy the whole table into an internal wiki, not one counterparty column. Our norway grid pause coverage shows how physical power bottlenecks can matter as much as contract text.
Roadshow timing and what effective means
Anthropic''s newsroom listed a roadshow launch on November 10, 2026 beside model releases. Filings during roadshow weeks are read aloud on earnings-style podcasts within hours. For builders, effective date is the gating event for legal offers; until then, numbers can still change in Amendment 3.
If your enterprise procurement team is negotiating a Claude commit during IPO week, ask whether host amendments trigger MFN clauses on your side. The S-1 does not answer that; your paper does.
A short FAQ for platform engineers
Do I need to re-benchmark Fable after this filing? Not because of disclosure alone. Re-benchmark when release notes or incidents change routing.
Does Google''s $121.3B line mean Claude trains only on TPU? No. It is a spend floor on Google-hosted services, not a chip map.
Does Microsoft''s $49.2B cap imply Azure-only inference for Enterprise? No public tenant mapping was filed.
Should startups panic about neocloud exits? Lambda''s lease trigger is real but optional per reporting; hyperscaler floors look stickier in the same document.
Reading list for finance-curious developers
You do not need an MBA to read an S-1. Search the PDF for regardless of actual usage, true-up, non-cancelable, and material breach. Those four phrases tell you who eats idle capacity risk.
Pair the PDF with explainx.ai''s amended Q3 post for revenue scale, and with TSMC''s strategic investment for foundry alignment.
What we are watching next
| Event | Why it may matter to builders |
|---|---|
| Effective registration | Locks offering terms; watch for risk factor edits |
| First post-IPO 10-Q | Shows realized vs minimum cloud spend |
| Lambda lease status | Could convert optional exit into delivery delay comms |
| Google TPU option tranche delivery 2028–2031 | May shift batch job economics if TPU share rises |
Commitment figures follow Anthropic S-1/A-2 as filed November 10, 2026. Registration statements are preliminary until effective; verify the latest amendment on EDGAR before citing in compliance documents.