Google TPU v6 and AWS Trainium 2 capture growing share of inference workloads — hyperscaler-internal silicon now standard for the largest deployments
Google's TPU v6 generation is expanding capacity rapidly through 2026, with third-party Google Cloud customers now able to provision TPU v6 at competitive per-token economics. AWS Trainium 2 has captured material share of inference workloads on Amazon's own platform and is opening to third-party customers. The pattern is the same on both clouds: hyperscaler-internal silicon now serves the largest inference deployments, with NVIDIA reserved for training and specialty workloads.
The economics are what drive the shift. Hyperscaler-internal silicon (TPU, Trainium, Inferentia) is optimized for the workloads the hyperscaler itself runs — Google's search-and-ads inference patterns, Amazon's Bedrock and Alexa inference patterns — and the per-token economics on those workloads beat NVIDIA on the same hardware-class comparison. For external customers running similar workloads, the price advantage transfers. The TPU v6 and Trainium 2 generations represent the first time hyperscaler-internal silicon is open enough and mature enough that third-party customers can adopt it without bespoke integration work.
The strategic consequence is that the compute moat fragments into hyperscaler-aligned segments. Customers on GCP increasingly default to TPU for inference (with Vertex AI managing the integration). Customers on AWS default to Trainium and Inferentia (with Bedrock as the routing layer). Customers on Azure remain NVIDIA-heavy because Microsoft hasn't yet shipped competitive in-house inference silicon, but the rumored Maia 200-series for late 2026 is the planned answer. By end of 2026, expect cross-hyperscaler customer mobility to be governed less by raw model capability and more by which silicon-stack the customer's workload runs best on.
S&P Global — AMD AI chips 2026 data center growth → · Yahoo Finance — AMD reveals new AI PC chips CES 2026 → · Clarifai — GPU Shortages 2026 →