Microsoft on Monday unveiled the Maia 200, its second-generation in-house artificial intelligence chip, intensifying the battle among tech giants to break free from Nvidia’s hardware stranglehold. The new accelerator goes live this week at a data centre in Iowa, with another site in Arizona to follow soon.
Engineers designed the Maia 200 specifically for AI inference workloads, delivering three times the FP4 performance of Amazon’s latest Trainium processor and Alphabet’s TPU lineup. Built by Taiwan Semiconductor Manufacturing using cutting-edge 3-nanometer technology, it packs high-bandwidth memory, though an older version than Nvidia’s upcoming “Vera Rubin” chips.
Targeting Nvidia’s software fortress
Microsoft isn’t stopping at silicon. The company now offers Triton, an open-source programming tool co-developed with OpenAI, to let developers code for Maia 200 without Nvidia’s CUDA expertise. Triton lets non-experts match pro-level GPU performance, slashing the switching costs that trap coders in Nvidia’s ecosystem.
This software push comes as Nvidia holds about 85% of the AI accelerator market, with Wall Street viewing CUDA as its ultimate edge. Major players like Meta Platforms are already testing Google’s TPUs to close that gap.
Cloud titans forge their own chips
The Maia 200 marks the second wave of Microsoft’s custom AI hardware after the 2023 Maia 100 debut, delayed by OpenAI tweaks and team changes until full 2026 production. It echoes moves by rivals to pack SRAM for speedy chatbot handling, akin to Cerebras’ $10 billion OpenAI pact and Groq’s $20 billion licensing deal with Nvidia.
Google draws Nvidia customers like Meta, while Amazon pushes Trainium. These custom chips aim to slash costs and speed for massive AI training, reshaping the trillion-dollar cloud wars.
