Today for AI
HOT RADAR
computeHEAT 0.0°

AI CURATED HOT STORY · 10/12/2026

Microsoft Unveils Maia 300 Chip: Achieves 40% Higher Token Throughput per Megawatt via TSMC N2P Multi-Chiplet Architecture, Set to Replace Nvidia in 2027

5 reports archived0 independent sourcesupdated 10/12/2026, 23:24:03
Synthesis & Latest Updates

Microsoft CEO Satya Nadella officially announced the third-generation Maia 300 AI inference chip, scheduled for full deployment by 2027 to replace Nvidia and AMD hardware as Azure's core infrastructure. Built on TSMC's N2P 2nm process with a multi-chiplet design, it claims 40% higher token throughput per megawatt compared to mainstream GPUs, aiming to significantly reduce large-scale inference costs by addressing data movement bottlenecks.

Technically, Maia 300 features 288GB of HBM4 memory across 8 stacks, delivering 18TB/s bandwidth and 13.8 PFLOPS FP4 compute at 950W. Optimized for long-context reasoning and agentic tasks, it integrates 96 accelerators per rack, achieving 7x better token processing efficiency than its predecessor, Maia 200. The latter is already supporting Copilot with 1.7x better energy efficiency than standard GPUs, validating Microsoft's custom silicon strategy.

This move represents Microsoft's biggest bet on self-developed chips, seeking independence from third-party GPUs by building gigawatt-scale autonomous inference clusters. As production nears in 2027, Maia 300 will reshape cloud hardware competition, driving AI infrastructure toward specialized high-efficiency inference and impacting developer ecosystems and supply chains.

LATEST/Microsoft confirmed that Maia 300 will enter production in 2027, initially targeting internal Azure services like Copilot and Teams before expanding to enterprise clients. Current Maia 200 deployments in the US, Europe, and Australia demonstrate 1.7x better energy efficiency than competitors, laying the groundwork for next-gen adoption.

HEAT TRENDHourly heat curve

54 fully observed hours
Now
0.0
Peak
65.210-13
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 5 reports · Latest first
  2. Hacker News AIT2·48 pts
    • Azure Maia 300 emphasizes system-level co-design from chip to datacenter rather than just hardware upgrades.
    • The architecture is specifically optimized for arithmetic intensity profiles of agentic workloads like long-context and tool use.
    • Core goal is maximizing intelligence per watt/dollar by minimizing data movement within fixed power and cooling constraints.
  4. Rohan Paul (@rohanpaul_ai)T2·86 pts
    • Microsoft plans to fully transition to its in-house Maia 300 as the primary inference chip by 2027, directly replacing NVIDIA and AMD solutions.
    • The Maia 300 utilizes a multi-chiplet architecture focused on solving data movement bottlenecks, claiming 40% better energy efficiency than leading GPU systems.
    • Designed specifically for long-context, low-latency chat, and multi-step agent tasks, it delivers up to 7x higher token throughput per rack compared to the previous generation.
  5. The Verge AIT2·78 pts
    • Maia 300 is positioned as Microsoft's primary AI inference system rather than a general-purpose training chip.
    • Key metric shows a 40% improvement in energy efficiency (tokens per megawatt) over leading GPU systems, significantly reducing large-scale inference costs.
    • The architecture is specifically optimized for long-context reasoning, low-latency conversational experiences, and multi-step agentic tasks.
    • Production deployment in Azure is expected to begin in 2027, reflecting the timeline from experimental silicon to scaled infrastructure.