Tech

Microsoft has officially unveiled its latest piece of custom silicon , the Maia 200,

Published

5 months ago

January 27, 2026

Zoro

Microsoft unveils Maia 200, a custom AI chip with 100B+ transistors designed for high-speed inference, efficiency, and scaling complex AI models.

Building on the foundation of the Maia 100

, which debuted in 2023, the new 200 series has been re-architected to run complex AI models with greater speed and efficiency

. Packed with more than 100 billion transistors

, the chip represents a significant leap forward, delivering over 10 petaflops in 4-bit precision and roughly 5 petaflops of 8-bit performance—a substantial upgrade over its predecessor

For context, inference is the computational heavy lifting

required to actually run an AI model, as opposed to the initial training phase. As the AI sector matures, the soaring costs

associated with inference have become a critical line item for companies, sparking a race to optimize these processes

Microsoft aims to lead that charge with the Maia 200

, promising smoother operations and reduced power consumption

for AI businesses. The company claims that, practically speaking, “one Maia 200 node can effortlessly run today’s largest models, with plenty of headroom for even bigger models in the future

.”

This launch underscores a broader industry shift where tech giants are developing bespoke silicon to reduce reliance on Nvidia

, whose GPUs currently dominate the market. Google has long offered its Tensor Processing Units (TPUs) via the cloud

, while Amazon recently rolled out the Trainium3

, the latest iteration of its own AI accelerator, in December. By offloading workloads to these proprietary chips, companies can diversify their hardware usage and manage costs more effectively

Microsoft is positioning the Maia 200 directly against these rivals

. In a press statement released Monday, the company touted impressive benchmarks

, noting that Maia delivers triple the FP4 performance of Amazon’s third-generation Trainium chips and outperforms Google’s seventh-generation TPU in FP8 capabilities