Amazon

Every other name in DeadRisk’s mag7 tier is answering one question: build the frontier model, or rent someone else’s. Amazon’s answer was both, priced as an investment rather than a strategy. In a seven-week stretch from late February, Amazon committed up to $75 billion in equity to the two labs that also buy its cloud capacity: up to $50 billion into OpenAI, up to $25 billion more into Anthropic on top of $8 billion already there. Each check came bundled with a chip order — OpenAI for roughly 2 gigawatts of Amazon’s own Trainium silicon, Anthropic for up to 5 gigawatts of capacity. Meanwhile, the unit built to develop a rival model was cut back on three fronts: its AGI Lab’s head left in February, the lab closed in July, and Amazon was reported to be deprecating its flagship models the same month.

A SemiAnalysis report from September 2025 named this shape before it fully arrived: AWS was losing share to Microsoft Azure and Google Cloud, and the fix on the table was Anthropic as an “anchor customer” for Trainium — Amazon’s silicon, running someone else’s models. DeadRisk carries Amazon inside The AI Capex Supercycle as the counterparty that makes the neocloud trade legible from the other side: CoreWeave, Nebius, and Applied Digital sell raw compute to the same labs. Amazon does too — but it also buys equity in the customer.

The relationship started as a compute sale, not an investment. In November 2025, OpenAI signed a seven-year, $38 billion deal to buy AWS capacity — its first contract with the cloud leader, running on Nvidia GB200 and GB300 chips, not Trainium. About six weeks later, Amazon was in talks to invest $10B-plus in OpenAI, this time with Trainium attached. The figure climbed to up to $50 billion by late February, $35 billion of which could hinge on an IPO or an AGI milestone, and on February 27 Amazon formalized $15B upfront, $35B more once certain conditions were met, with OpenAI committing to roughly 2 gigawatts of Trainium capacity through AWS.

Three weeks before that formal deal, Amazon’s Q4 earnings had already shown the cost: a $200 billion 2026 capex guide, above the Street’s $146.6 billion estimate, and AMZN fell more than 10% after hours. Three days before the deal was formalized, David Luan, Amazon’s AGI lab head, left less than two years after joining through an acqui-hire — an exit the market shrugged off.

Amazon didn’t stop at OpenAI. On April 20, Amazon agreed to invest up to $25B more in Anthropic, on top of $8 billion already invested; Anthropic secured up to 5 gigawatts of capacity and pledged to spend $100 billion-plus on AWS over ten years. Ten days earlier, Anthropic had shown it wasn’t relying on Amazon alone: CoreWeave announced its own multiyear deal to supply it with Nvidia chips. By June 29, Amazon was weighing OpenAI’s models alongside Nova to cut costs, after Anthropic raised the price of the models running inside Amazon’s own products — a cost tension between Amazon’s role as customer and its role as investor.

Then, across eight days in late July, four separate reports landed in quick succession. On July 23, The Information reported that Amazon had closed its San Francisco AGI Lab that week, part of layoffs in the unit Luan had led. On July 28, Business Insider reported the Nova deprecation, Amazon shifting resources to something it called Frontier Model Research. On July 30, the Financial Times reported one instance of runaway AI spending: $1.8 million spent on Claude for a menial cataloging task, an overrun that took months to detect. On July 30, Amazon reported AWS revenue up 37% to $42.2 billion, its fastest growth since 2021, and raised capex guidance to $220 billion; the next day AMZN closed up 15.32%, $235.50 to $271.58, the largest one-day gain in our two-year price archive. On July 31, Amazon’s quarterly filing showed it had paid the remaining $21.3 billion, completing the $50 billion OpenAI investment; the Financial Times put Amazon’s stake at roughly 5%.

TEXXR’s regime-break detector marks 2026Q1, the quarter Amazon formalized the OpenAI investment, as a structural break. Most of that break is TEXXR’s, not Amazon’s. The corpus changed how it extracts edges that quarter: Amazon’s count rose from 153 to 960, Google’s from 265 to 959, Microsoft’s from 165 to 903, and a “competitor” label appeared that no earlier quarter carries. Only 105 of Amazon’s 960 edges carry a label at all, so the quarter’s predicate mix and semantic drift cannot be set against 2025’s. Nor can its ranking of most-connected names: OpenAI tops Amazon’s 2026Q1 list, but the 14 OpenAI edges we could read from that quarter all come from one 18 March report on Microsoft weighing legal action. The articles make the cleaner case. In each quarter of 2024 and 2025, one to four TEXXR headlines named both Amazon and OpenAI; in 2026Q1, fifteen did, seven of them on the investment.

Amazon’s chips business tracks a related but separate trend — the commercial scale of its own silicon, not the funding shift above. In Andy Jassy’s April 2026 shareholder letter, Amazon’s chips business was generating over $20 billion a year. By the Q2 release, just under four months later, Amazon disclosed a run rate exceeding $25 billion — both figures stated as floors, not precise levels, so only the direction is a fact: up. The record also shows Amazon isn’t content keeping Trainium captive: in June, AI chief Peter DeSantis said Amazon is in talks to sell Trainium externally — competing with Nvidia for capacity it now reserves for its own equity partners.

Positioning: how much of AMZN is sold short, on the record brokers file twice a month. It lags by design — read it as where the crowd stood, not where it stands.

Order flow is signed before it becomes revenue — the earliest of the three views. From the capital-flow ledger.

Signals in the record, not trade ideas:

• Whether “Frontier Model Research” ships. Business Insider’s sources said in July it is building a new flagship model expected to debut at this year’s re:Invent.
• Whether OpenAI and Anthropic models displace Nova inside Amazon’s own products, after Amazon weighed using OpenAI’s models alongside Nova in June over Anthropic’s pricing.
• The AWS contract, as filed. Anthropic’s S-1 puts its AWS commitment at $110 billion from May 2026 to April 2036, non-cancelable, with a true-up payment if usage falls short; the April pledge said $100 billion-plus. Sales through cloud providers’ marketplaces, Amazon’s among them, were 48% of Anthropic’s first-half revenue, booked gross. Whether that share falls as Anthropic sells direct is one read on how much of the lab’s growth runs through Amazon’s storefront.
• The chips run rate’s next data point. It moved from over $20B in April to over $25B in July; the next earnings report, due 2026-10-29, will show whether that pace held.
• Whether Amazon begins selling Trainium to outside data centers, the move DeSantis floated in June — a shift from captive demand toward open competition with Nvidia.
• Internal AI-spend controls, after the FT’s report of a $1.8M Claude overrun that took months to catch — an anomaly, or a pattern the next report repeats.

Amazon lives or dies by the same constraint as these names — small equal-weighted groups, one constraint each. Not investment products: see all categories.