Microsoft is preparing to make its own AI silicon a much bigger part of the cloud infrastructure behind Copilot and other services. Its third-generation custom accelerator, Azure Maia 300, is designed for AI inference and is scheduled to enter production deployment in 2027. Microsoft says the system could deliver 40% more tokens per megawatt than leading GPU-based systems, while a full rack could provide up to seven times the token throughput of its predecessor, Maia 200.
The chip was announced on October 12 as part of Microsoft’s plan to scale its Azure Maia platform to multi-gigawatt deployments from 2027. The company says Maia 300 is built for long-context reasoning, fast conversational responses and multi-step AI-agent tasks, where the ability to serve model outputs efficiently can matter as much as raw computing power. The figures are Microsoft’s own projections, not independently verified benchmark results.
What Maia 300 brings
Microsoft is treating Maia 300 as a complete AI-computing system rather than a standalone processor. Its technical announcement describes a multi-chiplet design that separates compute, memory-controller and input/output functions. The compute chiplets use TSMC’s N2P process, while the wider system is tuned to reduce unnecessary data movement between processors, memory and networking.
Microsoft lists several headline specifications:
- 40% more tokens per megawatt: the company’s expected improvement over what it calls leading GPU-based systems.
- Up to 7x rack-scale token throughput: Microsoft’s comparison with Maia 200 in a full 96-accelerator configuration.
- 18.0 TB/s of peak memory bandwidth: provided by eight stacks of HBM4 memory, according to Microsoft.
- Up to 950 watts thermal design power: the stated chip-level thermal design figure.
- 96 accelerators per rack: a Maia 300 rack is specified at about 167 kilowatts, with 27.6 TB of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth across its high-speed scale-up domain.
The architecture reflects a key constraint in modern AI: generating answers from a trained model, known as inference, can be limited by memory bandwidth and the movement of data as well as by the processor itself. Microsoft says Maia 300’s dataflow design and software optimisations are intended to keep the compute, memory and network working together more efficiently. Its Maia 300 technical overview provides additional imagery and system details.
Why Microsoft is building its own AI chips
Large cloud providers increasingly want control over more of the hardware that runs their AI services. Custom silicon can be tailored to a provider’s own models and workloads, potentially improving throughput, supply planning and operating economics. It can also reduce dependence on buying every accelerator from outside suppliers, including Nvidia, whose GPUs remain central to the AI industry.
A Reuters report said Microsoft could deploy more than one million Maia 300 chips in 2027, although Microsoft’s public announcement states a plan for multi-gigawatt deployments rather than a confirmed unit shipment target. That would be a substantial expansion of the company’s in-house silicon effort. It does not mean Microsoft will stop using external GPUs: Azure serves many different customers and workloads, and Microsoft has not announced that Maia 300 will replace its full range of third-party accelerators.
The efficiency headline also needs context. Microsoft has not named the specific GPU systems behind its 40% comparison, so readers cannot independently reproduce it from the public announcement alone. Reuters’ report notes that a Maia 300 rack is expected to draw about 167 kW with 96 accelerators, compared with roughly 142 kW for Nvidia’s latest GB300 rack with 72 GPUs. Those configurations differ, and rack power figures by themselves do not establish which system delivers more useful output per unit of power for a comparable workload. Microsoft’s metric is a claim worth watching, not yet a neutral, apples-to-apples verdict on the hardware market.
What could change for Copilot and cloud customers?
Microsoft says Maia 300 is intended to become the primary inference system for its products and services as deployments scale. Its previous-generation Maia 200 is already handling production workloads, including Microsoft Copilot, with the company planning further use across services such as GitHub Copilot and Microsoft Foundry. More efficient inference could help Microsoft serve more AI requests within a fixed power envelope, or give it room to support larger and more complex workloads.
That does not automatically guarantee faster responses, lower subscription prices or new features for every customer. Those outcomes will depend on deployment scale, software optimisation, the models being served and how Microsoft allocates capacity. Maia 300 is infrastructure for Microsoft’s services, not a consumer chip that PC buyers can purchase separately.
What it means for India
For Indian businesses and consumers, the relevance is indirect but practical: the economics of AI services they use through Microsoft’s cloud may be affected as the company improves its own infrastructure. However, Microsoft has not announced a Maia 300 deployment schedule for Indian data centres, so there is no confirmed India-specific launch date or pricing impact. The company’s existing local cloud footprint, including the India South Central region in Hyderabad, should not be taken as confirmation that this particular accelerator will be installed there.
India is also pursuing its own computing capacity through initiatives such as the government’s plan to secure long-term access to 10,000 AI GPUs. Maia 300 is a reminder that AI capacity is not only about how many chips a country or cloud provider can obtain: memory, networking, power availability and the performance of the whole system all influence how much useful AI work can be delivered.
The main test begins in 2027
Maia 300 is a significant step in Microsoft’s effort to design AI infrastructure around its own services, but the most important evidence will arrive with production deployments. Independent comparisons, real-world performance, power use under comparable workloads and the pace of rollout will show whether the promised efficiency gains translate into better cloud economics. Until then, the 40% figure is Microsoft’s target, while the broader shift toward custom AI chips is the more immediate story.
