Microsoft Maia 300 will enter production deployments in 2027 as the company scales its custom AI inference platform to multi-gigawatt installations. Microsoft unveiled the third-generation accelerator on October 12, positioning it as the primary inference system for products including Copilot, GitHub and its Azure model-serving services.

 

The launch shifts Maia from a specialized internal program toward one of the largest custom-silicon deployments announced by a cloud provider. Reuters reported that Microsoft could install more than one million Maia 300 chips next year, although the company has not published a site-by-site schedule or total capacity target.

 

The Maia 300 system introduces five major changes:

  • TSMC N2P compute chiplets
  • Eight HBM4 memory stacks per accelerator
  • 96 accelerators in each 167-kilowatt rack
  • Superpods scaling to 12,288 accelerators
  • Production deployments beginning in 2027

 

Microsoft Maia 300 Moves Into Production in 2027

Microsoft designed Maia 300 for inference rather than model training. Its target workloads include long-context reasoning, conversational systems with tight latency requirements and agents that execute sustained, multi-step tasks. Those jobs spend substantial time moving model weights and cached context, making memory bandwidth and communication efficiency as important as raw arithmetic performance.

 

The company says Maia will become the primary inference platform behind its own products and services. Maia 200 already serves built-in features across Microsoft Teams, Copilot and Cowork, while the next generation is intended to extend custom silicon into GitHub coding workloads, Microsoft Foundry deployments, the MAI model team and Dragon Copilot for clinicians.

 

That rollout does not mean Microsoft is abandoning outside processors. Azure operates as a heterogeneous cloud and continues to deploy Nvidia and AMD hardware. Maia gives Microsoft another option for predictable, high-volume internal workloads where matching the chip, software stack, network and data center can reduce operating costs.

 

How Microsoft Maia 300 Chiplets and HBM4 Work

Maia 300 divides major functions across specialized chiplets instead of building every component on one process. Compute chiplets use TSMC’s N2P technology, while memory-controller and input-output functions sit on dies selected for bandwidth, connectivity and cost. Software manages locality so data remains close to the compute region using it.

 

Each accelerator carries eight HBM4 stacks with 288 gigabytes of capacity and 18 terabytes per second of peak bandwidth. Microsoft says that is 2.5 times Maia 200’s memory bandwidth, 33% more capacity and roughly twice the HBM bandwidth per watt.

 

The design also supports enhanced MXFP4 and storage formats below three bits. Lower-precision representations reduce the bytes required for model parameters and key-value caches, which can allow longer contexts and more concurrent sessions inside the same memory and power envelope. The tradeoff is that model quality must be validated for each workload after quantization.

 

Microsoft’s software-defined dataflow architecture coordinates local SRAM, HBM, direct-memory-access engines and the network. The goal is to reduce unnecessary movement rather than relying solely on more computing units, because transferring data across a large inference system consumes power and can leave arithmetic hardware idle.

 

The 96-Accelerator Rack and 12,288-Chip Superpod

A full Maia 300 rack combines 96 accelerators across 24 liquid-cooled trays within a 167-kilowatt power envelope. Together they provide 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate memory bandwidth, doubling the accelerator density of the previous Maia rack.

 

Microsoft connects the chips with a two-tier Ethernet-based fabric. Electrical links handle communication within a rack, while optical links join as many as 128 racks into a superpod of 12,288 accelerators. One communication model spans both levels, allowing software to place tensor, pipeline and expert-parallel work across the available hardware.

 

The network can move key-value cache state directly between prefill and decoding workers, avoiding a separate host-mediated path. Schedulers can also redirect work to spare trays when hardware becomes unavailable, trading some latency for higher utilization and resilience across the superpod.

 

Microsoft’s Efficiency Claims Face a Production Test

Microsoft says a 96-accelerator Maia 300 rack can deliver up to seven times Maia 200’s token throughput and more than 2.5 times its performance per watt. It also expects 40% more tokens per megawatt than leading GPU-based systems, but has not named the comparison systems or released reproducible benchmark results.

 

Reuters noted that a Maia 300 rack uses about 167 kilowatts, compared with roughly 142 kilowatts for an Nvidia GB300 rack containing 72 GPUs. The figures are not a direct performance comparison: the systems contain different accelerator counts, architectures and software stacks, and Microsoft’s metric depends on the models and serving conditions used.

 

Production deployment will therefore provide the decisive evidence. Utilization, reliability, model compatibility and delivered tokens per dollar matter more than peak specifications. If Microsoft can achieve its efficiency targets at multi-gigawatt scale, Maia 300 could lower the cost of serving agents and long-context models while reducing the company’s dependence on any single external accelerator supplier.

 

Background Reading