Vault
research

Anthropic Subsidy Stress Test — What Happens If the Compute Discount Goes Away

Created

Anthropic Subsidy Stress Test — What Happens If the Compute Discount Goes Away

Builds-on: anthropic-unit-economics-and-the-power-user-loss Related: ai-infrastructure-endgame-indicators Related: the-efficiency-counterthesis Related: hormuz-to-ai-repricing-causal-chain

The Question

The prior doc established that Anthropic's GM trajectory rests on AWS and Google selling compute below market. This deeper cut asks the next obvious question: how much compute is being given up by the hyperscalers, on what terms, and what does Anthropic's P&L look like if the subsidy normalizes.

The honest finding up front: the single most important variable is not publicly known. It will be revealed in the IPO S-1 in October 2026. Anyone modeling unit economics before then is guessing on the load-bearing assumption.

Where the Money Actually Flows: The AWS Side

Cumulative Amazon investment in Anthropic as of April 2026:

Amazon's books carry Anthropic at $16.0B preferred equity (Sep 30, 2025 10-Q) plus $5.3B in convertible notes (with $13.8B fair value at year-end 2024). Q3 2025 alone produced a $9.5B pre-tax non-operating gain from Anthropic mark-ups.

The headline-grabbing number is the $100B Anthropic commitment to AWS over 10 years. The Anthropic and Amazon press releases describe it as "spending more than $100 billion over the next ten years on AWS technologies." Neither party has disclosed whether this is:

  1. Take-or-pay (Anthropic must pay $100B regardless of usage)
  2. Minimum revenue commitment with downside outs
  3. Capacity-denominated (5 GW of compute regardless of dollar value)
  4. Aspirational (run-rate forecast, not contractually binding)

Futurum and Global Data Center Hub both flag this gap explicitly. Until the S-1, this is the most important unknown about Anthropic's economics. If take-or-pay, Anthropic is structurally locked and AWS holds long-term pricing power. If aspirational, Anthropic retains real switching leverage.

The capacity figure is firmer: up to 5 GW spanning Graviton + Trainium 2 → Trainium 4. Project Rainier already activated with ~500K Trainium2 chips; SemiAnalysis reports >1.3 GW under construction with nearly 1M T2 chips total. ~1 GW of T2/T3 online by end of 2026. Per-chip pricing is not disclosed — SemiAnalysis estimates Trainium2 effective committed cost at ~$0.50/chip-hour vs $2–5/hr for reserved H100.

Where the Money Actually Flows: The Google Side

Investment trajectory:

Compute scale:

The strategic signal worth noting: Google escalated commitments despite Gemini's strength. The naive read would be that Google should pull back as its own model competes more directly. Instead, the trajectory is the opposite — deeper, longer, larger. The interpretation is that TPU economics-of-scale require a frontier customer beyond Google itself; without Anthropic, TPU loses one of its few external proof points.

No public friction, no signs of pull-back. The trajectory is monotonically deeper since 2023.

Quantifying the Subsidy

If SemiAnalysis is right that the blended Trainium/TPU rate is 30–60% below Nvidia reserved at comparable throughput, then on a $7B 2026 inference budget the implicit subsidy is ~$2.1–4.2B/year at the market-rate counterfactual.

Across a 3-year horizon (2026–2028) on a roughly doubling compute base, total subsidy is in the $10–25B range. Wide because we don't know:

Anthropic's disclosed inference gross margin moved from 38% (2024) to 70%+ paid-tier-only in 2025. Including free users drops it to ~38%. Management cut overall GM guidance from 50% → 40% in Jan 2026 citing inference 23% above plan.

The 70%+ paid-tier figure is the number SemiAnalysis projects forward to 2027–28. If the subsidy normalizes meaningfully before then, the 70%+ projection breaks.

Conditions Under Which the Subsidy Dissipates

Four levers matter, ranked by current probability:

Lever 1: Trainium / TPU "proved" by external demand. Public Trainium customers besides Anthropic now include Databricks, Decart, poolside, Ricoh, Karakuri, SplashMusic, Arcee AI. TechCrunch reports Apple and OpenAI have at least evaluated or used Trainium. Trainium3 ramping early 2026 with ~1.6–1.7M unit volume target (Alchip sell-side). When external demand fills capacity, AWS no longer needs Anthropic as the anchor customer to validate the platform — and the strategic rationale for below-market pricing weakens.

Lever 2: Nvidia closes the price gap. Blackwell and Rubin economics are already narrowing the cost differential. If H100/B200 reserved hits $1–1.50/hr, Trainium's ~$0.50 advantage shrinks from a 4–10x edge to a 2–3x edge. The subsidy still exists but its strategic justification weakens because the alternative isn't punishingly expensive anymore.

Lever 3: Amazon's competitive ambitions reorient. Nova exists and is being updated, but Bedrock's strategy is "supermarket / optionality" rather than pushing Nova exclusively. The April 2026 deal deepens integration ("removing friction between Claude on Bedrock and Anthropic native tooling"), which suggests Amazon currently views Anthropic as the prestige anchor rather than a rival to displace. Could change. Not currently changing.

Lever 4: Google deprioritizes TPU subsidy as Gemini takes priority. Naive expectation. Reality is the opposite — the April 2026 $40B commitment is the deepest expression of Google's subsidy. Hard to see this dissipating in the modeled window.

The Microsoft–OpenAI Precedent

This is the only directly comparable arms-length subsidy structure. The trajectory is informative:

The OpenAI–Microsoft subsidy effectively lasted ~5 years before the lab gained enough leverage to renegotiate. Anthropic is currently ~3 years into its AWS relationship. The April 2026 deepening resets the clock at favorable terms — but only if the commitment minimums are real and two-way enforceable.

The structural point: as labs scale, they accumulate negotiating leverage, and exclusive subsidies don't survive the renegotiation cycle. Whether Anthropic's subsidy is "renegotiated" up (better terms for Anthropic) or normalized away (better terms for AWS as Anthropic scale increases the cost of subsidizing) depends on whose alternatives have improved more by the next cycle.

Lock-In and Switching Costs

Anthropic runs across TPU + Trainium + Nvidia simultaneously — explicitly multi-platform. CUDA→ROCm porting demos using Claude Code (IBM Mixture of Experts podcast, Techstrong reports) suggest the technical barrier to migration has fallen materially. But production-scale optimizations (kernels, attention implementations, KV-cache layouts) are still substantial sunk cost.

Practical read: Anthropic could migrate, but the economics work only if Azure or Oracle offered better-than-current pricing — unlikely while Microsoft is OpenAI-first and Oracle is OpenAI's $300B partner. Anthropic's switching options are realistically: AWS (current), Google (current), or build proprietary infrastructure (decade-scale).

The $100B commitment's enforceability against Anthropic depends on contract structure not publicly disclosed. Until S-1, this is the lock-in math nobody can model.

Stress Test Scenarios

Scenario A: Gradual 50% subsidy unwind, 2027–28. At a $7B+ inference base, ~$1–2B/year of margin compression. Pushes blended GM from ~40% toward high-20s before mitigations. Survivable for IPO but the "GM expansion to 70%+ by 2028" narrative breaks. IPO bankers price in the haircut.

Scenario B: AWS reprices Anthropic to standard customer (sudden). Most severe. If Anthropic is paying $0.50/chip-hour and standard rate is $2.50, repricing quadruples inference COGS. The $100B commitment becomes a ceiling on Amazon's pricing power only if it's denominated in capacity. The public language ("$100B in spending") suggests dollar-denominated, which would actually let AWS shrink delivered capacity if rates rise. Worst-case for Anthropic; best-case for AWS.

Scenario C: Google deprioritizes TPU subsidy. Anthropic loses redundancy. Capacity flexibility collapses. Anthropic becomes effectively a single-supplier customer of AWS. Negotiating leverage in the next contract cycle drops materially. AWS extracts more of the surplus. Slow grind, not a cliff.

Scenario D: Nvidia closes the price gap. Blackwell/Rubin narrows the moat. If H100 reserved hits $1–1.50/hr, Trainium's 4–10x edge becomes 2–3x. The subsidy still exists but its strategic rationale weakens. Most likely scenario in 2027–28. Slowly compresses the asymmetry without immediate P&L damage.

What the S-1 Will and Won't Reveal

For an October 2026 S-1, SEC requires related-party transaction disclosure (Item 404), revenue concentration, and material contracts.

Likely disclosed:

Likely not disclosed:

Amazon already discloses Anthropic only in aggregate. Anthropic will likely follow the same pattern. The expected outcome: bankers price in a "subsidy normalization" haircut to forward GM, probably 500–1,000 bps below SemiAnalysis's projected trajectory.

Comparable Public-Markets Precedent

There isn't a clean SaaS comp for compute-subsidy-dependent unit economics. Snowflake and Databricks paid market AWS rates. Palantir's government contracts are different. The closest analog is the cyclical chip-foundry capacity-commitment model (TSMC pre-paid wafer commitments by NVIDIA / AMD) — but those don't carry equity-linked structure.

This is novel territory for IPO disclosure. The market hasn't had to price a frontier-AI lab whose 70%+ gross margin trajectory depends on hyperscalers selling compute below market. October 2026 will be the first real data point.

What Is and Isn't Knowable

Knowable: total deal sizes, capacity in GW, equity investment values (Amazon's 10-K), public Trainium customer list, Anthropic's GM guidance, broad SemiAnalysis price estimates.

Not publicly knowable until S-1:

  1. Take-or-pay vs aspirational structure of the $100B commitment
  2. Actual per-chip-hour rates
  3. Whether commitments are dollar- or capacity-denominated
  4. Two-way enforceability
  5. Most-favored-nation clauses
  6. Anthropic's actual revenue concentration on AWS Marketplace / Bedrock vs direct

Item (1) is the load-bearing variable. The whole margin thesis hinges on it.

Bottom Line

The compute subsidy is real, large ($2–4B/year implicit), and currently growing rather than dissipating. The trajectory through 2026 is monotonically deeper integration with both hyperscalers, not pull-back. April 2026 was the year the subsidy expanded — Amazon to $33B potential, Google to $40B potential.

The Microsoft–OpenAI precedent suggests subsidies of this type last ~5 years before lab leverage forces renegotiation. Anthropic is ~3 years in. The most likely 2027–28 path is gradual normalization (Scenario D / A combination) rather than cliff (Scenario B). That path compresses GM by 500–1,000 bps from the SemiAnalysis trajectory but doesn't break the IPO.

The cliff scenarios (B, C) are tail risks that depend on hyperscaler strategic reorientation. Currently no signals point that way.

The single highest-value piece of information that doesn't exist yet is the contract structure. October 2026 S-1 is when the modeling stops being guesswork.

Sources

Backlinks