Hyperscalers Shrink Merchant Accelerator Market
Tenstorrent
Hyperscalers shrink the market above Tenstorrent by taking the biggest, most predictable AI buyers off the open market and onto their own chips. AWS, Google, and Microsoft now package Trainium, TPU, and Maia directly inside their clouds, so a customer already running workloads there can buy training and inference as a cloud instance instead of buying merchant accelerators from an outside chip vendor. That makes Tenstorrent strongest where buyers want infrastructure control, open software, or deployment outside the big clouds.
-
AWS and Google are not just designing chips, they are turning those chips into default cloud products. Trainium is sold through EC2 and SageMaker with AWS claiming lower training cost and better price performance versus GPU instances, while TPU v6e is positioned for training, fine tuning, and serving inside Google Cloud.
-
This matters most at the top of the market. The largest model builders and enterprise AI deployments already spend billions inside hyperscaler clouds, so every workload that moves to Trainium, TPU, or Maia is one less workload available to merchant accelerator vendors. Tenstorrent itself identifies AWS Trainium and Google TPUs as its most structurally threatening indirect competitors.
-
The closest analog is SiFive, which sells the picks for custom silicon rather than the finished accelerator. As more companies choose bespoke chips, value can move from merchant chip sales toward IP, design enablement, compiler software, and sovereign or private deployments, which is where Tenstorrent tries to differentiate with an open ML stack and non hyperscaler partnerships.
The next phase of AI infrastructure will split more clearly in two. Hyperscalers will absorb a growing share of mainstream cloud demand with in house silicon, while companies like Tenstorrent will have to win the segments that do not want to be locked into AWS, Google, or Azure, especially sovereign AI, private clusters, and buyers that value open architecture over cloud bundling.