The Quiet Rise of Amazon's Silicon Empire
While Nvidia commands the headlines with H100 wait times and B200 launch events, Amazon has been methodically building a parallel AI chip business that is now genuinely large. AWS's custom silicon portfolio — anchored by Trainium for training and Inferentia for inference, alongside the Graviton general-purpose line — crossed a $20 billion annual revenue run rate in Q1 2026 and is growing faster than triple digits year-over-year.
To put that in context: this is a business that barely registered in public financial disclosures three years ago. Today it carries $225 billion in cumulative revenue commitments from long-term Trainium contracts.
- 📈 $20B+ — AWS custom silicon annualized run rate, Q1 2026
- 💰 $225B — Total Trainium revenue commitments on the books
- ⚡ ~30% — Trainium2's price-performance advantage over comparable GPUs
Trainium3 Is Already Gone
Amazon launched its third-generation Trainium chip in early 2026, and it is already "essentially sold out." The who's-who of AI is behind the demand:
- OpenAI locked in roughly 2 gigawatts of Trainium capacity to power its frontier models — one of the largest single chip commitments in industry history.
- Anthropic signed for up to 5 gigawatts of current and future Trainium chips.
- Meta and Uber have also secured multi-year commitments.
Trainium2, the predecessor, delivered approximately 30 percent better price-performance than comparable Nvidia GPUs and sold out entirely. The pattern has repeated with Trainium3 at even larger scale.
The Big Strategic Shift: Direct Sales to Outside Data Centers
Here is where the story gets structurally interesting. For its entire history, Amazon has kept Trainium chips AWS-exclusive — you could only access them through cloud instances, not buy hardware outright. That may be about to change.
Amazon CEO Andy Jassy disclosed in April 2026 that the company is in early talks to sell Trainium racks directly to third-party data center operators. Jassy called such external sales "highly likely" within the next few years, though no contracts have been signed yet.
If it happens, any organization building its own AI infrastructure — rather than renting capacity from a cloud provider — would gain a credible alternative to Nvidia H-series GPUs for the first time.
| Chip | Status | Use Case |
|---|---|---|
| Trainium2 | Sold out | AI model training |
| Trainium3 | Nearly sold out | Large-scale AI training |
| Inferentia2 | Available | Low-latency inference |
| Graviton4 | Generally available | General-purpose compute |
What This Means for Nvidia
Nvidia isn't going anywhere in the near term. Its CUDA software ecosystem, the density of developer tooling, and the H200/B200 performance leadership still make it the default for most AI workloads. But the structural dynamics are shifting.
When the world's largest AI spenders — OpenAI and Anthropic together representing the top of the model-training pyramid — are committing gigawatts of capacity to a non-Nvidia chip, the signal is clear: alternatives are viable at the frontier, not just for cost-optimization in commodity inference.
If Amazon's direct-sales talks progress into actual deals, the competitive surface expands beyond cloud pricing wars into the physical datacenter market — territory Nvidia has held nearly unchallenged.
- AWS custom silicon crossed $20B annualized run rate in Q1 2026 with triple-digit YoY growth
- Trainium2 sold out; Trainium3 is nearly fully subscribed months after launch
- $225B in cumulative Trainium revenue commitments from OpenAI, Anthropic, Meta, Uber, and others
- CEO Andy Jassy says direct sales of Trainium to outside data centers are "highly likely"
- Trainium2 delivers ~30% better price-performance than comparable Nvidia GPUs