Microsoft CTO says he wants to swap most AMD and Nvidia GPUs for homemade chips

Original Article ↗ Hacker News Discussion ↗

Market power, pricing, and motives

  • Several see the announcement as an attempt to gain leverage over Nvidia’s pricing rather than a near-term technical shift.
  • Commenters note Nvidia’s “excess” profits and argue only hyperscalers threatening to go in-house can push prices down.
  • Some are cynical that this is also about sustaining the “AI growth” narrative for Wall Street in a broader tech-bubble pattern.

Vertical integration and big-tech playbook

  • Many compare this directly to Apple Silicon and Google TPUs, and to Amazon’s Graviton/Trainium: hyperscalers cutting out the middleman to save on per-wafer costs.
  • View that Microsoft is “late to the party,” but others say many hyperscaler silicon efforts actually began around 2018–2019, including inside Microsoft.
  • Discussion notes that these efforts are often based on ARM Neoverse cores plus custom accelerators, not fully custom CPU designs.

GPUs vs custom accelerators (training vs inference)

  • Broad agreement that inference (and much of training) is “embarrassingly parallel,” making custom ASICs and SoCs attractive.
  • Debate on whether an “inference-only” Nvidia chip is meaningfully distinct from a GPU; some cite TPUs, Groq, Tenstorrent, Etched as examples of more radical designs.
  • Several emphasize interconnects, memory bandwidth, and networking as the real bottlenecks and the hardest part to replicate, more than raw ALU performance.

Software ecosystem and CUDA moat

  • Strong consensus that Nvidia’s real advantage is CUDA and its mature tooling, not just hardware.
  • Some argue that for internal use you only need a small set of primitives (e.g., for transformers), so CUDA’s breadth matters less.
  • Others counter that developer inertia, ecosystem depth, and the scarcity of top-end engineers make it very costly to bet against CUDA at scale.

Microsoft’s credibility and impact on the ecosystem

  • Mixed views on Microsoft’s ability to execute: some point to prior in-house hardware (Catapult, Brainwave, Maia) and Azure systems; others call the company institutionally slow and see this as largely talk.
  • Concern that in-house chips across big tech could create hardware silos, limiting access for smaller players, though some hope it frees more Nvidia GPUs for consumers.