Amazon's custom silicon business just crossed a $20 billion annual run rate, growing over 100% year over year. CEO Andy Jassy now ranks it among the top three datacenter chip companies globally—but the technical story goes deeper than revenue.
The portfolio spans three distinct workloads: Graviton for general compute, Trainium for AI training, and Nitro for security offload. Trainium2 delivers 30% better price-performance than comparable GPUs and is fully subscribed. Trainium3, shipping since early 2026, adds another 30-40% improvement and is nearly sold out. Trainium4 reservations already exceed $225 billion in commitments, with broad availability 18 months out.
What matters for engineers: OpenAI committed roughly two gigawatts of Trainium capacity through AWS, ramping in 2027. Anthropic locked in up to five gigawatts across current and future generations. Meta signed on for tens of millions of Graviton cores to handle agentic AI workloads. Jassy explained the reasoning: as AI shifts from answering questions to taking actions, post-training and inference scale pulls heavily on CPUs. Graviton delivers up to 40% better price-performance than x86, and 98% of the top 1,000 EC2 customers now use it.
AWS deployed over 2.1 million AI chips in the past year—more than half Trainium, paired with over a million Nvidia GPUs. This heterogeneous strategy gives customers options. Power, not chips, is the binding constraint. Trainium at scale saves tens of billions in capex annually and adds hundreds of basis points to operating margins versus third-party silicon.
The architectural takeaway: vertical integration at hyperscale shifts the cost curve. AWS AI revenue run rate tops $15 billion—260x the $58 million from its first three years. Bedrock processed more tokens in Q1 than all prior years combined. Customer spending there rose 170% quarter over quarter.
Risks remain: execution on Trainium4, power constraints, and continued Nvidia partnerships. But the $225 billion in chip commitments signals a structural shift. AWS's backlog sits at $364 billion. Jassy floated selling Trainium to third parties. If that happens, the $50 billion standalone run rate he described becomes a floor, not a ceiling.
Chips aren't a hedge anymore. They're the engine.
Source: Webpronews