Microsoft Unveils Maia 300 AI Inference Chip for Multi-Gigawatt Azure Deployments
The AI compute build-out sees Microsoft pushing custom silicon for inference, while major chip and power deals reshape global infrastructure and supply chains.
The story
The AI hardware landscape is seeing a significant push towards custom silicon, with Microsoft unveiling its Maia 300 AI inference chip at the OCP Global Summit in San Jose. Designed to be the primary inference system for Microsoft products and services, Maia 300 is projected to deliver 40% more tokens per megawatt than existing GPU-based systems. Microsoft plans multi-gigawatt Azure deployments of the chip beginning in 2027. This move signifies a broader industry trend among hyperscalers to develop in-house processors to reduce reliance on third-party GPU providers like Nvidia and AMD, optimize costs, and enhance performance per watt.
Concurrently, the foundational infrastructure supporting AI is undergoing substantial shifts. Nvidia agreed to invest $5 billion in Intel, with Intel expected to manufacture custom chips for Nvidia's AI platforms, indicating a deepening collaboration in chip production. Additionally, AMD and OpenAI announced a 6 GW deployment agreement for AMD Instinct GPUs, with the first 1 GW planned for late 2026, while OpenAI and Broadcom plan to deploy 10 GW of OpenAI-designed accelerators by 2029. These large-scale deals highlight the immense and growing demand for AI compute capacity and the diverse strategies companies are employing to secure it, from custom silicon to strategic partnerships.
Silicon
Maia 300
Maker: Microsoft
What: Third-generation custom AI inference chip for long-context reasoning, promising 40% more tokens per megawatt.
For Whom: Microsoft's internal AI services and Azure cloud platform.
GPU D-Series (formerly Crescent Island)
Maker: Intel
What: AI inference accelerator with 1.3 PFLOPS of FP8 compute and 1.5 TB/s memory bandwidth, designed for air-cooled enterprise servers.
For Whom: Large language models, long-context inference, multi-tenant AI services, and robotics.
Structera X (second-generation)
Maker: Marvell
What: Six-channel CXL 3.2 memory expanders improving compute and memory utilization across AI systems.
For Whom: Hyperscaler customers like Alibaba and Samsung for AI infrastructure.
The build-out
AI infrastructure expansion
Who: AM Intelligence, NTT Data
Scale: 20,000 more Nvidia Rubin GPUs; India nearing 2GW data center capacity
Where: India and Malaysia
US Data Center Portfolio Expansion
Who: Gorilla Technology Group
Scale: Acquiring first US data center (6 MW current, potential 36 MW), targeting over 100 MW within 12-18 months.
Where: United States
Supply & policy signals
US AI data center startups raised $33.5 billion by early October 2026.
Implication: Investment is accelerating into technologies that integrate energy supply with computing infrastructure, reflecting power availability as a critical constraint for AI data center expansion.
Nvidia's $5 billion investment into Intel for custom chip manufacturing.
Implication: Indicates a strategic move by Nvidia to diversify and secure advanced chip production capacity through partnerships.
OpenAI's large-scale deployment agreements for 6 GW of AMD Instinct GPUs and 10 GW of OpenAI-designed accelerators with Broadcom.
Implication: Signifies an aggressive long-term commitment by major AI developers to secure vast amounts of compute, intensifying demand across the silicon supply chain.
What we'll be watching
- Microsoft Ignite 2026 conference for further details on Maia 300 and Azure strategy.
- ASML's Q3 2026 earnings report on October 14, followed by TSMC's full Q3 earnings on October 15.
- Deployment of Microsoft's Maia 300 chips in Azure, beginning in 2027.
- Initial 1 GW deployment of AMD Instinct GPUs by OpenAI in the second half of 2026.
- Broader U.S. data center acquisitions and infrastructure expansions by Gorilla Technology Group over the next 12-18 months.
Reporting + analyst voices: grounded via Google Search at publish time.