Trainium 3 Demand Spike
Signals show AWS Trainium 3 demand accelerating fast and projected to be sold out by 3Q26, a capacity squeeze that’s already lifting AWS margins. The note also flags Marvell as a key supplier for Trainium ASICs — a supply‑chain node worth watching. (x.com)
Trainium3 powers Trn3 UltraServers that can house up to 144 Trainium3 chips per UltraServer rack and deliver roughly 362 FP8 petaflops, with AWS claiming ~4.4x more compute and ~4x better energy efficiency versus Trainium2. (datacenterdynamics.com) AWS executives said the company has deployed more than 1 million Trainium chips across its fleet and that Anthropic’s Project Rainier already runs on roughly 500,000 Trainium2 devices. (techcrunch.com) Amazon disclosed that its in‑house chip programs (Trainium plus Graviton) are running at an annual revenue run rate north of $10 billion, and management outlined a $200 billion capital‑spend plan tied to AI and infrastructure build‑out. (techcrunch.com) (geekwire.com) Marvell’s public release describes a five‑year “multi‑generational” collaboration with AWS that explicitly covers supply of custom AI products, optical DSPs, active electrical cable DSPs, PCIe retimers, DCI optical modules and Ethernet switching silicon. (marvell.com) Bank analysts reporting on meetings with Marvell say company management told investors purchase orders for the Trainium3 XPU ASIC program are in place for 2026, a point JPMorgan highlighted when defending Marvell’s AI outlook. (investing.com) Marvell reported record Q3 FY2026 revenue of $2.075 billion and management has cited rapid data‑center traction and accelerating custom‑ASIC contribution—Marvell said custom revenue scaled materially in fiscal 2026 and guided to stronger data‑center growth ahead. (investor.marvell.com) (marketbeat.com)
