# Hyperscalers Move Inference to In-House Silicon

Large cloud operators will keep shifting their own inference workloads from merchant GPUs to custom accelerators, compressing Nvidia's share of inference spending while raising dependence on TSMC and HBM suppliers.

- Conviction: 40 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-10-13T00:00:00.000Z
- Last updated: 2026-10-13T06:34:40.436Z
- Canonical: https://polylog.news/ai/trends/hyperscaler-custom-inference-silicon
- Publisher: Polylog
- Affected regions: Global

## Recent evidence

- [confirms] Microsoft says its Maia chips will become its main inference platform at multi-gigawatt scale from 2027 (2026-10-13): Microsoft says its Maia chips will become its main inference platform at multi-gigawatt scale from 2027. The third-generation Maia 300 is built on TSMC N2P with eight HBM4 stacks, and Microsoft claims 40% more tokens per watt than leading GPU systems, a figure no one has independently measured. That commits inference spending to in-house silicon and deepens dependence on TSMC and HBM suppliers, as the thesis predicts.