Microsoft plans to expand Maia chip deployment starting in 2027, with Maia 300 possibly exceeding million units
BIBIBI
AT A GLANCE
Microsoft says it plans to expand Maia chip deployment, with Maia 300 possibly exceeding 1000000 units next year.
Article
Microsoft aims to make its self-developed chips the AI mainstay for inference, with Maia 300 possibly exceeding million units deployed next year
Beating AI News Flash, Microsoft announced that it plans, starting in 2027, to expand deployment of its self-developed Maia chips to several gigawatts, making them the primary inference platform for its AI products. According to Reuters, Microsoft may deploy more than 1000000 Maia 300 next year, further reducing its reliance on NVIDIA GPU.
The previous-generation Maia 200 is currently running in Microsoft Copilot, supporting OpenAI, Microsoft's own, and open-source models. The new-generation Maia 300 is optimized for long-context inference and multi-step Agent, with the compute chiplets using TSMC's 2nm process; each chip is equipped with 288GB HBM4 of memory, and memory bandwidth reaches 18TB/s。
Maia 300 a single rack houses 96 chips, with power consumption of approximately 167kW, twice the chip count of the previous-generation rack. Microsoft claims that the new rack's Token throughput reaches up to Maia 200 at 7 times the rate, and that the number of Tokens generated per megawatt of power is GPU higher than leading 40%systems.
Original link https://m.theblockbeats.info/flash/371498
Key points
01
Microsoft announced plans to expand deployment of its self-developed Maia chips to several gigawatts starting in 2027.
02
According to Reuters, Microsoft may deploy more than 1000000 Maia 300。
03
Maia 200 is already running in Microsoft Copilot, supporting OpenAI, Microsoft's own, and open-source models.
04
Maia 300 optimized for long-context inference and multi-step Agent tasks, using TSMC's 2nm process, with each chip equipped with 288GB HBM4 of memory and memory bandwidth of 18TB/s。
05
Maia 300 a single rack houses 96 chips, with power consumption of approximately 167kW。
06
Microsoft claims that the new rack's Token throughput reaches up to Maia 200 at 7 times the rate, and that the number of Tokens generated per megawatt of power is higher than leading GPU systems by 40%。
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Microsoft is advancing its self-developed AI Maia chips for AI inference, meaning the computation used when models generate responses. Maia 300 is designed for longer text and multi-step Agent tasks, but some details concerning deployment volume, performance, and specifications remain plans or claims made solely by Microsoft.
Why it matters to readers
If the plans and performance metrics are validated, Microsoft may increase its in-house AI inference capacity and partially reduce its reliance on NVIDIA GPU; however, the actual deployment scale and results cannot currently be confirmed.
This indicates that large-scale AI services rely not only on general-purpose GPU but are also attempting to design specialized chips to support inference for models at greater scale.
Risks and unknowns
The body does not provide the original Microsoft announcement or Reuters article, so the information source requires further verification.
The deployment figure of more than 1000000 units is a possible future plan and has not yet been confirmed as achievable.
Maia 300 Performance, power consumption, and process information for Maia primarily comes from Microsoft's claims; actual performance has not been independently verified in the body.
The body does not specify how much AI inference work Maia will ultimately handle, nor does it confirm the actual impact on purchases from NVIDIA GPU.
Related Developments
Loading event timeline…
Related concepts
AI inference
This term is not in the glossary yet. Browse related concepts in the glossary.
