Publication date: 12 October 2026
Microsoft plans to make its in-house Maia chips the primary inference system for its products and services, scaling the technology to multi-gigawatt deployments from 2027. The move is more than another semiconductor launch. It is an attempt to shift bargaining power inside the artificial-intelligence supply chain.
The company says its Maia 200 system is already serving production workloads including Microsoft Copilot and its own model-development teams, with GitHub Copilot and Microsoft Foundry due to follow. The third-generation Maia 300 is designed for long-context reasoning, memory-intensive models and sustained multi-step agent workloads.
Microsoft claims Maia 200 can deliver up to 1.7 times more tokens per watt for selected built-in services than leading graphics-processor infrastructure. It expects Maia 300 to improve tokens per megawatt by 40 per cent relative to leading GPU-based systems. These are company performance claims and will need independent testing across comparable workloads, but they reveal the metric that now matters most: useful AI output per unit of scarce power.
Inference changes the economics
Training advanced models receives most public attention because it requires enormous clusters of accelerators. Inference, however, is the recurring work performed every time a user asks a model to answer, reason or act. As AI becomes embedded in search, software development, office tools and healthcare workflows, inference can become a larger and more predictable source of cost.
Designing a chip specifically around Microsoft’s own services allows hardware, models, networking and software to be optimised together. General-purpose accelerators remain valuable because they can run a wide variety of workloads, but a cloud provider with sufficient scale can justify custom systems for tasks it performs continuously.
This vertical integration mirrors earlier changes in cloud computing, where the largest providers began designing processors, networking equipment and storage systems. Control of more layers can improve efficiency and reduce exposure to component shortages. It can also make a platform harder to leave if customers become dependent on features that work best on proprietary infrastructure.
Nvidia dependence will not disappear
Microsoft’s expansion does not mean it can simply replace Nvidia or AMD. Frontier training, specialised workloads and customer demand still require a diverse accelerator fleet. Software ecosystems, developer tools and established performance are powerful advantages for incumbent chip suppliers.
The more realistic shift is from dependence on one dominant supplier toward a portfolio. Microsoft can use its own silicon for high-volume internal inference while reserving external GPUs for workloads where flexibility or performance justifies the cost. That gives the company more leverage in procurement and more control over capacity planning.
The strategy also reduces the risk that another company’s product timetable determines when Microsoft can launch or expand an AI service. Owning the architecture creates faster feedback between product teams and infrastructure engineers.
Power is now a design constraint
Multi-gigawatt deployment brings a broader responsibility. Efficiency gains can reduce electricity used for each unit of computation, but total demand may still rise if lower costs encourage much greater use. This rebound effect means tokens per watt should be reported alongside total energy consumption, grid impact and carbon intensity.
Custom chips also concentrate operational risk. A design flaw, manufacturing delay or software limitation can affect many services at once. Microsoft will need supplier diversity, testing and fallback capacity even as it moves more workloads onto Maia.
For enterprise customers, the change may bring lower costs and improved performance, but it should also prompt questions about portability. Buyers need to understand whether models and applications can move between hardware environments, how pricing reflects efficiency savings and what happens if capacity becomes constrained.
The AI infrastructure contest is no longer only about who can buy the most chips. It is becoming a contest over who can design the entire system, secure enough power and turn hardware efficiency into reliable services. Microsoft’s Maia plan shows that cloud platforms increasingly view silicon as a core strategic capability rather than a commodity purchased from someone else.