Amazon (Anthropic stake + Trainium)
Andy Jassy (CEO); Anthropic stake is a corporate strategic investment, compute via AWS.
Amazon mirrors Alphabet's playbook on its own silicon: an $8B (and growing) Anthropic stake paired with AWS Trainium chips. Anthropic is Amazon's anchor Trainium customer — Project Rainier, a ~$8B cluster of nearly 500,000 Trainium2 chips, was the world's largest AI cluster at launch, and Anthropic has committed >$100B to AWS over ten years for up to 5GW of new capacity. The stake exists to validate and fill Trainium capacity against Nvidia.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
- Private marks — Material value sits in private marks that have not been exit-tested.
Top holdings
$8B in to date; April 20, 2026 agreement to invest up to an additional $25B as part of the infrastructure deal.
>$100B / 10yr commitment — Anthropic uses >1M Trainium2 chips; Project Rainier (~500k Trainium2, ~$8B) was world's largest AI cluster at launch. Up to 5GW new capacity; ~1GW Trainium2/3 online by end-2026.
Recent moves
April 20, 2026: agreed to invest up to an additional $25B in Anthropic (on top of $8B prior); Anthropic commits >$100B to AWS over ten years for up to 5GW new compute. Project Rainier (~500k Trainium2) completed/activated; ~1GW Trainium2+Trainium3 coming online by end-2026. Sources: CNBC, About Amazon, Anthropic, DCD.
Loading disclosed positions…
Our take
The same vertically-integrated thesis as Google but on Amazon's custom silicon: own Anthropic equity, then have Anthropic burn it on Trainium and AWS — proving Amazon's chips can train a frontier model without Nvidia. Edge: Trainium validation is strategically worth more than the equity mark. Caveats: minority stake, no control; the >$100B AWS commitment is a related-party flywheel (Amazon funds the customer that fills its clusters), the same circularity critique aimed at Nvidia; and Anthropic now multi-sources compute (5GW on Google TPUs too), so AWS exclusivity is gone.
The strategic logic is stronger than the accounting optics, and the two are being conflated.
An anchor customer contractually bound to 5GW is exactly what a merchant-silicon program needs — you cannot bootstrap a CUDA competitor without a workload that stresses it. But Q1 2026 net income of $30.3B included $16.8B of pre-tax gains from the Anthropic stake in non-operating income, against $23.9B of operating income: more than half the headline profit is a mark on an unlisted private company whose latest price Amazon helped set by joining the round. Simultaneously TTM free cash flow fell to $1.2B from $25.9B, with $147.3B of TTM property-and-equipment purchases (up $59.3B YoY). The honest read: reported earnings are flattered by a self-referential mark at precisely the moment cash generation is consumed by the buildout the mark is meant to justify. Two different businesses, one number — and only one of them is cash.
- 2026-07Cash generation is not merely squeezed: TTM free cash flow is -$7.6B (from +$18.2B a year earlier), TTM property-and-equipment purchases $169B (+64%), Q2 capex $54.2B, FY2026 capex guidance $220B from $200B.
Thesis
Amazon's Anthropic position is not an investment portfolio — it is customer financing for a silicon program. Amazon has put $13B of equity into Anthropic ($8B in 2023–24, plus $5B closed 20 Apr 2026 at a ~$380B valuation), with up to $20B more milestone-gated. In return Anthropic committed >$100B over ten years to AWS and up to 5GW of Trainium capacity. The equity is the inducement; the payoff is an anchor tenant large enough to prove Trainium can train a frontier model and to amortise the buildout against something other than Nvidia's price list.
Corporate strategic investment, not a fund. Take a minority, non-controlling stake in an anchor AI lab and convert it into a decade-long AWS commitment (>$100B, up to 5GW, Trainium2 through Trainium4); use that committed load to justify custom-silicon capex and cut Nvidia dependence. Project Rainier (~500k Trainium2) was the proof build; ~1GW of Trainium2+3 due by end-2026. Structurally: vendor financing with an equity kicker, run off the balance sheet.
- 2026-06Amazon's invested total is now $18B — $8B of convertible notes plus two $5B preferred tranches (Series G and H). Anthropic's Series H closed 28 May 2026 at $965B post-money, two rounds above the ~$380B April mark.
Assessment
- The compute commitment is contractual and dated (>$100B / 10yr, ~1GW Trainium2+3 by end-2026) — harder to walk back than a memorandum of intent.
- Trainium is not purely Anthropic-fed: Jassy disclosed $225B of committed multi-year chip demand on the Q1 2026 call, including ~2GW from OpenAI starting 2027.
- Amazon monetises the relationship three ways — AWS revenue, chip margin, equity mark — so the thesis does not depend on any single leg.
- Minority and non-controlling, which keeps Anthropic's independence credible to its own customers and avoids the antitrust surface a control stake would create.
- The circularity is real: Amazon funds the customer that fills Amazon's clusters, then marks up the stake on a round Amazon participated in. That is the same critique aimed at Nvidia's vendor financing.
- $16.8B of a $30.3B quarter is a paper mark on a private position. Strip it and the AI-profit narrative looks materially different from the headline.
- Disclosure is thin. Amazon publishes the gain but no discrete carrying value, ownership percentage, or Anthropic-attributable share of AWS revenue — organic and financed demand cannot be separated.
- AWS is 'primary,' not exclusive. Anthropic secured ~1M Google TPUs and >1GW from Google and serves Claude on Azure Foundry. Marginal training can migrate on price-performance.
- $20B of the $25B is milestone-gated and the milestones are undisclosed. Headline commitment and committed capital are being read as the same figure by most coverage.
- 2026-07Q2 2026 magnifies the point: net income $62.6B included a $53.4B pre-tax Anthropic mark against $27.5B of operating income — about 85% of headline profit is a private mark, not the ~55% Q1 showed.
- 2026-07Amazon now equity-funds both frontier labs: it completed a $50B investment in OpenAI ($28.7B funded by 30 Jun) and became exclusive third-party cloud for OpenAI's Frontier program.
Record
There is no standalone return series, and there cannot be one — a private, unlisted stake has no independent price. What is disclosed is the periodic non-operating mark: $16.8B pre-tax in Q1 2026, off a cumulative $13B invested. That implies a large paper multiple, but the multiple derives from a valuation Amazon helped establish, not a realised exit or an arm's-length clearing price. Attribution is therefore unresolvable. What is measurable is the strategic leg: AWS grew 28% YoY to $37.6B in Q1 2026 with $14.2B of segment operating income, and Jassy put the chips business (Trainium, Graviton, Nitro) at a ~$20B annualised pace with $225B of committed multi-year demand — the Trainium-only split is not disclosed. The equity gain is not measurable that way: it reverses on any down-round and is only truly priced by a liquidity event. Treat the $16.8B as an accounting entry, not evidence of investing skill.
Risks & fit
- Mark-to-model risk: an Anthropic down-round or a compressed AI-private-valuation regime reverses the gain through the income statement with no cash offset.
- Capex/FCF risk: $147.3B TTM capex against $1.2B TTM FCF leaves little cushion if AI demand decelerates before the assets depreciate.
- Concentration risk: Trainium's proof point remains one equity-funded anchor that has itself diversified to Google TPUs — though OpenAI's ~2GW blunts this.
- Circularity risk: if regulators or auditors reframe funded-customer revenue as contra-revenue rather than sales, both the AWS growth rate and the mark are affected.
- Backlog-conversion risk: the $225B commitment book and the $20B contingent tranche are commitments, not shipped capacity or recognised revenue.
Breaks if non-Anthropic Trainium revenue stalls — if the ~$20B chips run-rate proves predominantly Anthropic consumption funded by Amazon's own equity, or if OpenAI's ~2GW and the $225B commitment book fail to convert into shipped capacity. Conversely, the accounting critique is falsified if Anthropic lists and clears at or above the ~$380B mark in an arm's-length market, or if Amazon discloses Trainium revenue ex-Anthropic and it holds up. Watch: conversion of the $225B backlog into disclosed revenue, and a third-party-led Anthropic round that reprices the stake without Amazon in it.
Useful for understanding how hyperscaler AI economics are constructed — where strategic investment, vendor financing and reported earnings blur. Relevant to parsing Amazon's income statement (separating operating results from the Anthropic mark), assessing custom-silicon challenges to Nvidia, or the circular-financing question spanning Amazon, Alphabet, Microsoft, Nvidia and Oracle. Not a way to access Anthropic — the stake is not separable.
Not a fund — no management fee or carry. Capital is deployed from Amazon's balance sheet; returns accrue to AMZN shareholders via AWS revenue, chip margin, and non-operating marks. No fee load applies.