technologyScore 80/100Research
Microsoft deploying custom silicon (Maya 200) optimized for in-house AI models, achieving 1.4x performance-per-watt advantage over NVIDIA GB200
Mustafa Suleyman· Microsoft AI· AI· 2026-05-20· about Microsoft (MSFT)
“we've optimized MAI Thinking 1 on our very own Maya 200 chip and benchmarked it head-to-head against the GB200. And so on top of the 30% performance improvement that Satya talked about earlier, we're now seeing a further 1.4x performance per watt gain when we run our MAI models on the Maya 200 end-to-end.”
Why it matters
Microsoft is achieving significant efficiency gains through silicon-model co-design, reducing reliance on NVIDIA for inference at scale. This signals a shift toward custom silicon for hyperscalers and validates the economics of vertical integration.
Investment implication
Suppliers of custom silicon design, advanced packaging, and power delivery infrastructure will see increased demand. This may pressure NVIDIA's inference market share but validates demand for design partners like TSMC and chiplet specialists.