The Surprising Part Isn't the Chips — It's the Business Model Shift
Amazon has been building its own AI chips for years. What's new, according to reporting from TechCrunch, is that AWS is in talks to sell those chips directly to other data center operators — companies that aren't AWS customers in the traditional cloud sense, but that run their own infrastructure.
That's a meaningful distinction. Until now, Amazon's custom silicon — Trainium (designed for model training) and Inferentia (designed for inference, meaning running a trained model) — has been available only as part of AWS's own cloud services. You could rent compute time on a Trainium instance; you couldn't buy the chip and rack it yourself. The reported talks would change that.
What Jassy Actually Said
CEO Andy Jassy has described this as a $50 billion opportunity for Amazon. It's worth being precise about what that claim does and doesn't mean. A $50 billion figure attached to a market opportunity is not the same as a $50 billion revenue forecast — it's a characterization of the addressable market, and those estimates are notoriously elastic. The data center chip market is genuinely large; Nvidia's data center segment alone generated over $47 billion in revenue in fiscal year 2024. But capturing share in that market requires more than a competitive chip.
The Nvidia Problem Is Mostly Software
Nvidia's dominance in AI compute isn't purely about hardware performance. It's substantially about CUDA — the proprietary programming framework that developers have built on for nearly two decades. Switching away from Nvidia means rewriting or recompiling workloads, retraining engineering teams, and accepting some degree of performance uncertainty on new silicon.
Amazon's chips have shown competitive performance on specific workloads in AWS's own benchmarks, but independent, apples-to-apples comparisons against Nvidia's H100 and B200 GPUs are limited. That gap between internal claims and externally validated benchmarks matters a lot if Amazon is asking data center operators to make long-term infrastructure bets.
Why This Is Still Worth Watching
That said, there are real reasons large data center operators might be interested. Nvidia's pricing power is substantial, and any credible alternative creates negotiating leverage even if it never becomes the primary supplier. Hyperscalers like Google (with TPUs) and Microsoft (with Maia) have made similar internal-to-external pivots, with mixed but real results.
Amazon also has a distribution advantage that pure chip startups lack: existing relationships with enterprises and cloud customers who already trust AWS infrastructure decisions.
The talks are ongoing, and no deals have been announced. The $50 billion framing is aspirational. But the strategic logic — turning a cost-reduction tool into a revenue line — is coherent, and the timing, as customers push back on Nvidia's pricing, is not accidental.