--- title: "Gavin Baker Says Trainium Is This Year's TPU. The AWS Multiple Disagrees." url: "https://www.readplaza.com/articles/gavin-baker-says-trainium-is-this-years-tpu-the-aws-multiple-disagrees" type: "article" publisher: "Hardwired" category: "Company Coverage" published: "2026-05-28T14:41:00+00:00" updated: "2026-06-01T14:49:33.850183+00:00" reading_time_minutes: 4 tags: ["Semiconductors"] --- # Gavin Baker Says Trainium Is This Year's TPU. The AWS Multiple Disagrees. _Hardwired followed his call on $ALAB, the AI networking infrastructure play, and that position returned 190% as of today._ Gavin Baker manages $7 billion at Atreides Management. Hardwired followed his call on ALAB, the AI networking infrastructure play, and that position returned 190% as of today. If you're interested in the AI buildout and semiconductors story, he's a name you want to be familiar with. Three weeks ago Baker said Amazon's Trainium is the most underestimated chip in AI. His specific claim: "The significance of Trainium to 2026, especially after the substantial deployment of Trainium 3 in the second half of this year, is akin to the importance of TPU to 2025." When Google's TPU reached scale, it stopped appearing in cost and started appearing in margin. Baker is watching for the same thing at AWS. So are we. Incoming demand explosionAnthropic's run-rate revenue was $9 billion at the end of 2025. By April 2026 it had crossed $30 billion. That growth, in four months, runs on Amazon's chip. Over one million Trainium2 chips are training and serving Claude through Project Rainier. Scaled Trainium3 capacity comes online in the second half of this year. OpenAI's equivalent compute runs on Nvidia GPUs at market rates. Anthropic's runs at 50% lower cost per training run. Every model generation, that differential widens. The third business Amazon doesn't headline.Amazon's custom silicon stack, Graviton, Trainium, and Nitro, crossed a $20 billion annualized revenue run rate in Q1 2026. Jassy's own words on the Q1 call: "if that chip business sold externally the way Nvidia does, the run rate would be $50 billion. The other $30 billion doesn't appear as revenue. It flows directly into AWS margins." He said Trainium will deliver several hundred basis points of operating margin advantage for inference and save tens of billions in capex annually. AWS operating income came in at $14.16 billion for the quarter, beating consensus by $1.3 billion, with growth at 28% year-over-year, fastest in 15 quarters. The stack doesn't have a segment. Amazon doesn't break it out. A chip business running at $20 billion annualized, inside a cloud division growing at 28%, doesn't carry a chip multiple. It carries a cloud expense-line assumption. That's the gap. Bedrock, Amazon's own AI inference product, runs most of its inference workload on Trainium today. Amazon is its own best customer for the chip it's selling to 125,000 other organizations. Every inference call that routes through Bedrock on Trainium is margin that doesn't leave the building. 80% of Fortune 100 companies are on Bedrock. Amazon isn't its own best Trainium customer in isolation. It's running the chip at the center of how almost every major enterprise in the world now accesses AI. The backlogAmazon holds $225 billion in Trainium revenue commitments. Trainium3 is nearly fully subscribed. Much of Trainium4, still 18 months from broad availability, is already reserved. The demand is contracted two chip generations out. Total AWS backlog at Q1-end: $364 billion, excluding the Anthropic deal. Anthropic's commitment alone spans Trainium2 through Trainium4 and future generations, with $100 billion committed over ten years. Trainium3 is nearly fully subscribed. Much of Trainium4, still 18 months from broad availability, is already reserved. The demand is contracted two chip generations out. The bear case treats $200 billion in 2026 capex as a bet on unproven demand. The demand is contracted. The capex is building pre-sold infrastructure. The second compounding vectorAmazon has deployed one million robots across US fulfillment operations. Sequoia, its AI and computer vision inventory system, shortens distance traveled and reduces touches per package. Jassy called the current generation "a step change in efficiency" on the Q1 call. The retail margin pressure is real. What the bear model isn't running: a million-robot network getting incrementally more efficient on every shift, reducing the cost structure the bears are worried about. What would change thisTrainium 3 significantly underperforms Baker's expectation, measurable at Q3 earnings. The FTC proceeding results in structural AWS separation, well beyond current legal precedent. The Anthropic deal Q3. The second and third would be immediate and public. Until one of those three moves, $364 billion in committed backlog is the floor. Most Amazon coverage is running the retailer's model against the capex line. The thesis is on the silicon.