Something here is worth watching

The report

Second hyperscaler says its own inference accelerator now serves a majority of internal traffic

Reuters · 2026-07-19

What this says

• —A second large cloud provider states publicly that internally designed accelerators now carry the majority of its own inference traffic.
• —This is the specific development named as a kill criterion — not because one customer matters, but because it is evidence the software lock-in is escapable at scale.
• —It is a press report, not a disclosure. The claim is unaudited and the volumes are not quantified, so it raises the criterion to warning rather than breaching it.

Assumptions affected

Which parts of your reasoning this bears on, and how.

1. CUDA lock-in keeps customers on-platform once rivals close the gapThe report describes production inference, not experimentation, which is the case the assumption was written to exclude.Undermines

What would prove the thesis wrong

Where each criterion stood after reading this source.

1. 1A major hyperscaler moves production workloads to in-house siliconReported publicly by the customer, but not quantified and not confirmed in any filing.Near