Microsoft has announced its Azure Maia 300 custom AI accelerator, planned for 2027, built as a vertically integrated system spanning silicon, racks and software. For UK buyers already running or planning AI server capacity, it signals a shift in how Azure pricing and availability will be structured.
View the data behind this chart
| Maia 200 rack | Maia 300 rack | |
|---|---|---|
| Token throughput index | x1 | x7 |
Microsoft's next Azure AI chip moves from blueprint to roadmap
Microsoft says its first custom accelerator, Azure Maia 200, announced earlier this year, is now serving live production inference workloads on Azure. Its successor, Maia 300, is already slated for 2027 and has been described by Microsoft as "a complete intelligent system" designed around the demands of agentic AI inference.
Microsoft's EVP for Cloud and AI, Scott Guthrie, outlined the specs in a blog post: a full rack-scale configuration spans up to 96 accelerators, and Microsoft claims this delivers up to seven times the rack-scale token throughput of Maia 200, alongside 40 percent more tokens per megawatt than leading GPU-based systems. Each Maia 300 chip carries eight HBM4 memory stacks, offering up to 18 TB/s of peak memory bandwidth — aimed squarely at long-context, memory-intensive and decode-heavy workloads that typify agentic AI.
A reversal on vertical integration
What's notable here is the change in Microsoft's own stance. In its 2024 'AI Access Principles', Microsoft argued that partnerships across the technology stack were essential to competition, and even complained to EU regulators that Google had an unfair advantage by being vertically integrated 'at every AI layer from chips to a thriving mobile app store'.
Two years on, Microsoft is building exactly that: its own silicon in Maia, its own rack-scale system design, and its own MAI models running on top. For UK organisations with Azure as a strategic platform, this matters because capacity, pricing and feature rollout decisions are increasingly made inside a closed Microsoft stack rather than negotiated across multiple silicon vendors.
Custom inference chips vs Nvidia GPUs: what it means for workload fit
Maia 300's inference-first architecture puts it in direct contrast with the general-purpose rack-scale systems Nvidia has built for its GPU lines, and the Helion and Helios infrastructure AMD has more recently developed. Each vendor is tuning hardware, software and rack design around its own silicon — which means workload portability between platforms becomes a bigger planning question, not a smaller one.
UK teams currently sizing GPU estates via tools like our AI GPU sizing calculator, or comparing options across GPU accelerators and NVIDIA DGX systems, should treat Maia 300 as a genuinely distinct category rather than a drop-in GPU alternative — it's optimised specifically for decode-heavy, long-context agentic inference rather than general training or mixed workloads.
Azure capacity planning for 2027 budgets
Microsoft says it plans to scale Azure Maia to "multi-gigawatt deployments" starting in 2027. That is a meaningful capacity signal for UK enterprises and public-sector bodies planning multi-year Azure commitments: it points to substantial new inference capacity entering the pipeline, but also to Microsoft directing more of its own infrastructure investment toward proprietary silicon rather than third-party GPU procurement.
For budget owners, this reinforces the case for modelling AI infrastructure spend separately from general cloud compute. Teams using an IT finance calculator to plan 2027 AI budgets should factor in that pricing structures tied to Maia-based inference tiers may not map directly onto existing GPU-instance cost models, and early visibility will be limited until Microsoft publishes concrete SKUs.
What UK IT leaders should do before Ignite 2026
Microsoft has confirmed it will share more detail on its Maia strategy at Microsoft Ignite 2026 in San Francisco next month. Until then, concrete UK pricing, regional availability and instance types remain unconfirmed, so this is a planning window rather than a procurement deadline.
Practical steps worth taking now: map which current or planned workloads are genuinely decode-heavy or agentic in nature (where Maia 300's design advantage is claimed), review current server configuration choices against that workload profile, and keep broader infrastructure modernisation — including licensing shifts covered in our look at VMware alternatives — on the same roadmap, since AI capacity decisions rarely happen in isolation from the rest of the estate. Teams wanting a second opinion before committing budget can talk to a Servnet engineer.