Microsoft has unveiled its Azure Maia 300 AI accelerator, promising a 96-chip configuration that delivers seven times the token throughput of its predecessor, with production deployments scheduled for 2027. The Redmond giant announced that its current-generation Azure Maia 200, revealed earlier this year, is already handling production workloads, while the successor represents what the company calls "a complete intelligent system" built for emerging AI workload types. Microsoft plans to scale the new chip to multi-gigawatt deployments starting next year.
The Azure Maia 300's full 96-accelerator setup will provide up to seven times the rack-scale token throughput of the Maia 200, while delivering 40 percent more tokens per megawatt compared to leading GPU-based systems, according to Microsoft. Each individual Maia 300 chip will include eight HBM4 memory stacks that provide up to 18 TB/s of peak memory bandwidth. The company envisions deploying the chips in rack-scale systems containing up to 96 accelerators, designed to support execution of large models and long-context workloads.
Scott Guthrie, Microsoft's EVP for Cloud and AI, described the chip as featuring an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads. The company states that Maia 300 is "built as a complete system, with software, silicon, physical rack infrastructure, networking, and datacenters co-designed to operate as a single platform to meet the demands of cloud-scale agentic inference." Microsoft insists this system-level approach is necessary for agentic workloads, which involve multi-step execution, long-context decoding and tool use, and "change the computational profile and arithmetic intensity of inference."
The announcement marks a shift in Microsoft's infrastructure strategy, as the company now controls more of its AI stack through in-house chips, rack-scale designs, and internally developed models. This growing focus on vertical integration stands in contrast to the company's 2024 "AI Access Principles," which emphasized partnerships across its technology stack and even criticized Google to European competition authorities for being "vertically integrated in a manner that provides it with strength and independence at every AI layer from chips to a thriving mobile app store." The report notes that Microsoft at the time said "everybody else must rely on partnerships to innovate and compete," yet now possesses its own AI chips and full-stack control.
Guthrie said the firm will share more details on its strategy at next month's Microsoft Ignite 2026 event in San Francisco, with multi-gigawatt Azure Maia deployments set to begin in 2027. The move positions Microsoft to handle what it describes as the new computational demands of agentic AI workloads, which require different processing characteristics than traditional inference tasks. The pivot from partnership-dependent infrastructure to proprietary system integration may reshape competitive dynamics as hyperscalers increasingly design purpose-built silicon rather than relying solely on third-party accelerators.