Analysis
Amazon has hired Robert Hundt, a Google distinguished engineer who served as the original software leader for Google's Tensor Processing Unit program, to work on software for AWS's Trainium chips. Hundt joins Amazon's chip team under the same distinguished engineer title he held at Google -- a senior, individually-recruited hire rather than part of a broader team lift-out.
The move targets Trainium's most persistent weakness. Nvidia's dominance in AI training and inference isn't just about raw chip performance -- it's substantially about CUDA, the software ecosystem developers have built years of tooling and muscle memory around. AWS's Trainium and Google's own TPUs have both struggled to get developers to port workloads away from that ecosystem, even when the underlying silicon is competitive or cheaper. Bringing in the person who helped build TPU's software layer from the ground up is a direct bet that software expertise, not just chip design, is the bottleneck Amazon needs to solve.
“Google has the most mature program with TPUs; Amazon's Trainium and Microsoft's Maia chips are both earlier in their software maturity curves.”
This fits a broader 2026 pattern: every major hyperscaler is now running a custom-silicon program explicitly to reduce Nvidia dependency and capture more margin on AI compute rather than paying Nvidia's take on every training run. Google has the most mature program with TPUs; Amazon's Trainium and Microsoft's Maia chips are both earlier in their software maturity curves. Talent moves like this one are becoming a proxy battle for the custom-silicon race, much like model researchers were the proxy battle for the frontier-lab race in 2023 and 2024.
What to watch: whether Hundt's hire translates into measurable Trainium software improvements within the next few quarters, and whether this triggers further senior TPU-team departures to AWS or Microsoft as the custom-silicon talent war intensifies.
