When SpaceX publicly filed its S-1 on May 20, the prospectus disclosed in full what had been leaked in pieces since early May: Anthropic has agreed to pay xAI $1.25 billion per month through May 2029 — with reduced rates the first two months as Colossus ramps — for dedicated access to all 300 megawatts of Colossus 1 in Memphis, including more than 220,000 NVIDIA H100, H200, and GB200 GPUs. The full-term value is around $45 billion. Either party can terminate with 90 days’ notice.
That’s the largest publicly disclosed compute commitment in AI history, and it landed inside an S-1 that almost nobody in enterprise procurement has read.
The story everyone is writing is about Elon, xAI’s underutilization, and the orbital-compute footnote at the back. The story nobody is writing — and the one your CFO or CIO actually needs — is the procurement read. Specifically: given this disclosure, what do you change about your Q3 Claude renewal, your refresh-risk model, and your vendor-concentration posture before the end of June?
Below is the procurement decision table the SERP doesn’t have. Six scenarios, six honest scores, three buyer-type verdicts, and the three questions to ask your Anthropic enterprise rep before signing anything in Q3.
The two facts to anchor every conversation
Before the table, two non-obvious facts the popular coverage has buried.
One. The S-1 explicitly frames this as “monetizing unused compute capacity.” That’s xAI’s framing, written into a regulated disclosure document. It confirms the worst-case interpretation of the underutilization commentary that’s been circulating since early May. The deeper read for the buyer side: when SpaceX’s own filing concedes the counterparty needed this revenue to clear pre-IPO numbers, the durability of the contract is governed less by Anthropic’s compute hunger and more by whether xAI’s next opportunity is a better one. That math changes every quarter.
Two. Either party can terminate with 90 days’ notice. The clause is in the S-1. The most-quoted post on X this week came from @ThierryBorgeat: “The biggest customer is also the most fragile… legally cancelable with three months’ warning. This is the kind of risk factor that gets buried on page 47 and becomes the entire story two earnings calls later.” For any enterprise that bet Q3 production capacity on Claude, a 90-day window between “everything is fine” and “we have a structural compute interruption” is now a real refresh-risk you should price into your contract terms.
With those two anchors in place, the table.
The 6-scenario × 7-dimension Q3 Claude renewal decision table
| Scenario | (1) Capacity guarantee | (2) 24-mo $ cost vs status quo | (3) Refresh-risk if 90-day clause fires | (4) Lock-in risk | (5) Exit cost | (6) Multi-model resilience | (7) Disclosure-readability for your audit committee |
|---|---|---|---|---|---|---|---|
| A. Renew 3-yr enterprise commit now at current Tier-1 pricing | Strong (as worded; see verdict) | Lowest (locked rate, no Sept OpenAI-IPO benchmark) | High exposure; no contractual buffer | High — single-vendor | High | Low — depth-first on Claude | Hardest to defend if Tier-1 doesn’t include explicit Colossus-failure language |
| B. Renew 1-year + reopen post-Sept (post-OpenAI IPO) | Medium | Slightly higher Q3 (premium for short term) | Moderate — buys 6-9 months optionality | Medium | Medium | Medium | Cleanest “we waited for full disclosure” defense |
| C. Negotiate Tier-1+ with explicit Colossus-failure clause | Strongest | Highest in Q3 (margin for explicit guarantee) | Lowest — contractual offset | High but documented | High | Low | Strongest with audit committee — explicit risk acknowledgment |
| D. Hedge with Bedrock or Vertex backup of 30-40% workload | n/a (multi-vendor) | Highest engineering cost; flat $ on inference | Lowest — workload-level failover | Lowest | Lowest | Highest | Strong if multi-model architecture is documented |
| E. Move to direct Anthropic API + absorb Jun-15 billing split | Self-managed | Lower than Tier-1; you absorb price volatility | Moderate — same vendor exposure | Medium | Medium | Medium | OK — direct billing relationship has cleaner audit trail |
| F. Re-architect for multi-model now (full portability) | Lowest single-vendor; highest portfolio | Highest 24-month engineering cost; lowest ongoing inference | Lowest — fully fungible | Lowest | Lowest | Highest | Best long-term posture for any regulated industry |
A few cells need a sentence of unpacking before the verdicts.
Cell (1) capacity guarantee. Anthropic’s May 6 announcement of the xAI compute partnership on its own page implies guaranteed access; the S-1 specifies the 90-day termination clause. Tier-1 enterprise commits with Anthropic do carry capacity language today, but in most existing Tier-1 contracts that language predates the Colossus disclosure. Your guarantee is only as strong as the explicit Colossus-failure clause your counsel adds in renewal — Scenario C is the only scenario that addresses this directly.
Cell (3) refresh-risk if 90-day clause fires. Reasonable worst case: xAI gets a better offer from another buyer (OpenAI, Meta, a sovereign-AI program); xAI invokes the 90-day clause; Anthropic has 90 days to source replacement capacity in a market where every hyperscaler is already capacity-pinched. The likely Anthropic mitigation: route more inference to AVP1 (the Akamai partnership announced May 10 — V2 already covered), pull forward TPU capacity from the April Anthropic-Google-Broadcom 3.5 GW deal, and ration Tier-1 customers down to contracted minimums during the bridge. If you’ve built around current Tier-1 limits, the 90-day window is when your latency budget breaks.
Cell (7) disclosure-readability for your audit committee. Every regulated industry CIO will face the question from their audit committee in Q3 or Q4: “What’s our exposure if Anthropic’s compute supplier walks?” Scenarios A and E both make that conversation harder than necessary. Scenario C makes it easier because the explicit Colossus-failure clause documents that the risk was considered and priced. Scenarios D and F make it easier because multi-vendor architecture itself answers the question.
Three buyer-type verdicts
Different industries should weight these dimensions very differently. Here are three honest cuts.
If you’re a regulated financial-services or healthcare CIO
Scenario C or D favored, with Scenario B as a defensible bridge. Two reasons. First, the audit-committee-readability row matters most for you — your model-risk management framework and your board will both ask “what’s our exposure,” and the answer “we negotiated explicit Colossus-failure language” or “we route 40% to Bedrock as failover” is much cleaner than “we renewed Tier-1 at the pre-disclosure price.” Second, regulated industries pay a 90-day-interruption tax that’s much higher than tech-startup-land — your service-level commitments to your own customers don’t have a Claude-dependency caveat.
Scenario A is defensible only if your renewal contract is amended in writing with explicit Colossus-failure language. Without that amendment, the existing capacity language is governed by ordinary force-majeure terms, and “vendor’s supplier walked” is contested territory under most force-majeure interpretations.
If you’re a high-growth-tech VP-platform with engineering depth
Scenario E or F, often hybrid. Two reasons. You can absorb a model swap faster than a regulated industry can — your code-base size and your engineering culture favor portability. The engineering cost of Scenario F (full re-architecture for multi-model) is real, but for any team with an existing model-routing layer or a Vercel AI Gateway / OpenRouter pattern in production, the marginal cost is smaller than your CFO assumes.
The configuration that usually wins here is direct API + 30% workload on a second vendor (Bedrock or Vertex), with the routing layer engineered to absorb a 90-day primary-vendor interruption. That looks like Scenario E + partial D.
If you’re a CFO at a mid-market enterprise (300-2,500 employees)
Scenario B is the conservative default. Two reasons. The Sept timing maps to OpenAI’s confidential S-1 (filed May 22, public expected by Q3) — once OpenAI is public, you have a second set of disclosed unit economics to triangulate against. And the 1-year commit + reopen optionality keeps your hands free during the period when the most material AI-vendor financial information will become public.
Scenario A only makes sense at this size if the price discount on a 3-year commit is meaningfully above your cost of capital — and right now, with Anthropic’s Q3 run-rate above $30 billion and the public disclosure of a $15 billion/year compute obligation, Anthropic has weaker incentive to discount aggressively than it had in early 2026.
Three questions to ask your Anthropic enterprise rep before signing
“Given the Colossus disclosure, what explicit capacity-guarantee language are you willing to add to our renewal? Specifically — if xAI invokes the 90-day clause, what’s our contractual position on continued Tier-1 access during the bridge period?” Reps will offer verbal reassurance. Side-letter language is what makes the reassurance real. If they won’t put it in writing, weigh that signal heavily.
“What’s your contractual liquidated-damages or service-level credit if Tier-1 capacity is throttled by more than X% for more than Y days during a vendor-side compute interruption?” The point isn’t to actually collect damages. The point is to surface — early — how much real risk Anthropic itself prices into the timeline. If the answer is “we don’t offer service-level credits at any level,” that’s a procurement signal worth pricing into your renewal premium.
“What’s your contingency posture? Specifically, what percentage of your inference capacity is on AVP1 (Akamai), TPU (Google), and Colossus respectively, and how does that mix change if you need to re-route?” This question is the most important one and the hardest to get a real answer to. Push for it anyway. Even a partial answer — “we don’t disclose mix percentages, but we have written contingency to shift X workloads” — is more diagnostic than the standard sales-team boilerplate.
What this means for you
If you’re renewing a Claude enterprise contract in Q3 2026
Push to Scenario B (1-year + reopen) unless you can get explicit Colossus-failure language. The 90-day clause is too material to renew over for 3 years without contract-level acknowledgment.
If you’re choosing between Claude API direct and a reseller bundle (Bedrock, Vertex)
The reseller bundles now look comparatively safer than they did 3 months ago. AWS Bedrock’s Claude offering inherits Anthropic’s compute supply but also includes AWS’s own infrastructure abstraction — meaning your AWS-side SLA is independent of the Anthropic-xAI contract. For mid-market buyers who don’t want to negotiate side-letter language directly with Anthropic, Bedrock just got more attractive.
If you’re at a startup that depends on a single Claude model for production
Test the failover path now. Spend a weekend in Q3 wiring your inference layer to a backup provider (Bedrock, Vertex, Together, Replicate, OpenRouter) and run shadow traffic for 30 days. You’ll learn whether your code is actually portable or whether you’ve been carrying hidden Claude-specific assumptions. The Q3 timing is right because the OpenAI IPO will land somewhere in the same window, and that’s when multi-model strategy goes from “nice to have” to “expected of any AI-platform team.”
If you’re investing in or covering Anthropic for an investor newsletter
The disclosure forces a unit-economics reset. Anthropic’s $30+ billion run-rate paying $15 billion/year to xAI puts gross-margin pressure squarely on a level that the Where’s Your Ed At “Profitability Swindle” critique (V2 covered May 22) had already flagged. The new question is whether Anthropic’s net-margin trajectory survives the compute commitment through 2027, and whether the October IPO target window remains stable. The 90-day termination clause is a feature here, not a bug — it’s the only structural protection Anthropic has if the Colossus economics deteriorate.
If you’re a CISO at any of the above
The compute-supplier disclosure changes the threat model for vendor-concentration risk. Add a new line to your Q3 vendor-risk review: “Critical AI vendor’s compute supplier is publicly disclosed as terminable on 90 days. Document our exposure and our 90-day mitigation runbook.” That’s not a panic; that’s good hygiene under disclosed conditions.
What this analysis can’t tell you
It can’t tell you whether the 90-day clause will ever fire. Most multi-billion compute contracts run their full term. The base rate of catastrophic-termination events in this market is genuinely low. The procurement question is not “will it fire” but “given non-zero probability, what’s the contractual offset.”
It can’t price the orbital-compute footnote. The S-1’s mention of Anthropic-SpaceX joint interest in multi-gigawatt orbital AI capacity is genuinely the most interesting line in the disclosure for 2030-2032, but the commercial window is too far out to inform your Q3 renewal. Watch for follow-up announcements, particularly any Starship-cadence-tied capacity commitments after the SpaceX IPO closes in June.
It can’t substitute for your enterprise account team’s read. Anthropic’s enterprise reps have material context the S-1 doesn’t surface. Schedule the call regardless of which scenario you favor.
It can’t price the geopolitical layer. Memphis Tennessee is not a regulated-supply geography in the conventional sense, but the energy-supply pressure and the Tennessee state energy-review posture could both shift in the next 12 months. The 90-day clause is the proximal risk; the regional regulatory risk is the structural one.
The bottom line
The S-1 disclosure didn’t change how Claude works; it changed what your renewal contract should say. Most enterprise buyers won’t notice for two quarters and will renew on pre-disclosure pricing under pre-disclosure terms. The teams that read the S-1, modeled the six scenarios, and negotiated explicit Colossus-failure language into their renewal will look back in 2027 and find they bought genuine optionality at a small Q3 premium.
If you’re building the AI-procurement and vendor-management muscle on your team — covering compute-supply-chain risk, multi-model architecture, contract language, and the audit-committee conversations that turn S-1 disclosures into actual procurement decisions — our Enterprise AI Rollout Playbook and AI Consulting & Advisory courses walk through the working frameworks.
Sources
- Anthropic inks deal to use all of SpaceX’s Colossus 1 compute capacity (Wall Street Journal)
- Anthropic will pay xAI $1.25B per month for compute (TechCrunch, May 20)
- Anthropic is paying $15 billion a year for access to Elon Musk’s data centers (The Verge, May 22)
- Anthropic is paying SpaceX $15 billion per year (Axios)
- SpaceX IPO Anthropic paying AI compute (Business Insider)
- SpaceX S-1 reveals Anthropic paying $1.25B/month for AI compute (The Deep Dive)
- New Compute Partnership with Anthropic (xAI, May 6)
- Anthropic to use all of SpaceX-xAI’s Colossus 1 data center compute (Data Center Dynamics)
- Notes on the xAI/Anthropic data center deal (Simon Willison, May 7)
- SpaceX S-1 filing on SEC EDGAR (May 20)
