Story perspectives

Microsoft’s Maia 300 AI chips deliver 40% more tokens

By Drooid · Oct 12, 2026, 6:16 PM UTC · How we work

Summary

Microsoft Unveils Maia300

Microsoft unveils Maia 300, a new AI inference accelerator.
Production deployments start in 2027, scaling the Azure Maia family to multi-gigawatt levels.
Microsoft says Maia 300 yields 40% more tokens per megawatt than leading GPU systems.
A rack of 96 accelerators offers up to seven times Maia 200’s token throughput.
Each chip integrates eight HBM4 stacks, delivering 18 TB/s bandwidth within a 950 W TDP.

Summary

Amazon Announces AI Silicon

Microsoft expects to spend over $175 billion in capital expenditures this year.
Finance chief Amy Hood says in-house silicon has boosted margins over the past decade.
CEO Satya Nadella frames the chip rollout as a move to lessen Nvidia dependence.
TSMC is slated to produce more than 300,000 Maia 300 chips for 2027, according to The Information.
Amazon, AMD and Google also announced custom AI silicon, intensifying competition.

History

Historical Context

Google began building custom AI silicon in 2013 when Amir Salek was hired, and under Norman Jouppi’s lead the first Tensor Processing Unit entered production in May 2016, and later generations delivered about 45 tera-operations per second with far higher performance-per-watt than contemporary CPUs and GPUs. The TPU’s rapid success spurred other firms, and Microsoft’s 2027 Maia 300 rollout follows the same in-house-chip strategy.

1 of 3

Microsoft Unveils Maia300

Microsoft unveils Maia 300, a new AI inference accelerator.
Production deployments start in 2027, scaling the Azure Maia family to multi-gigawatt levels.
Microsoft says Maia 300 yields 40% more tokens per megawatt than leading GPU systems.
A rack of 96 accelerators offers up to seven times Maia 200’s token throughput.
Each chip integrates eight HBM4 stacks, delivering 18 TB/s bandwidth within a 950 W TDP.

1 / 3