AI - Microsoft is betting big on its own Azure Maia chips
Microsoft is going all-in on its own AI silicon, announcing plans to scale Azure Maia to multi-gigawatt deployments from 2027. That is a serious commitment, but it is not just about making another chip: Microsoft wants Maia to become the primary inference system for its own products and services.
Not dissimilar to others as Google has their own TPUs, AWS is trying something too, Meta has its own AI chips for internal use. Microsoft is just catching up, with its own chips and also its own models (which are still way behind the competition) 🤣
Everyone is trying to get less dependant on NVIDIA, I wonder why 🤣🤣🤣
Azure Maia 300 and Microsoft’s silicon bet
Official announcement
Microsoft to expand Azure Maia system to multi-gigawatt deployments – The Official Microsoft Blog
Microsoft describes Maia 300 as an inference system for long-context and multi-step workloads:
Maia 300 is a complete intelligent system designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks.
By improving efficiency across the system, we expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems. And in a full 96-accelerator configuration, it delivers up to 7x the rack-scale token throughput of Maia 200.
The plan builds on Maia 200, which Microsoft says is already serving production workloads, including Microsoft Copilot. GitHub Copilot and Microsoft Foundry are expected to follow by the end of the year. Microsoft says Maia 200 runs OpenAI models, its own MAI models and open-source model families. It also reports up to 1.7× more tokens per watt for built-in AI capabilities in Teams, Copilot and Cowork compared with leading GPU infrastructure.
The next generation, Maia 300, is designed for inference workloads such as long-context reasoning, latency-sensitive conversations and sustained multi-step agent tasks. Microsoft says Maia 300 will:
- Deliver 40% more inference tokens per megawatt than leading GPU-based systems.
- Provide up to 7× the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration.
- Combine eight HBM4 stacks, with 18 TB/s peak memory bandwidth, in a 950-watt chip.
- Begin production deployments in 2027.
Satya Nadella framed the announcement as a major commitment to Microsoft’s own silicon:
Today we are announcing our biggest bet yet on our own silicon.
Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services.
And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems.
The design is a system, not just a processor. Microsoft describes a multi-chiplet architecture that separates compute, memory-controller and I/O functions, alongside software-defined dataflow and high-bandwidth networking. In plain language, the aim is to keep data moving efficiently between the parts of the system instead of spending power and time shuffling it around.
Why build its own AI chips?
The obvious draw is efficiency. If Microsoft can deliver more inference from the same power budget, it can serve more requests from its data centres. That matters as electricity and data-centre capacity become constraints on how quickly AI services can grow. Custom chips may also give Microsoft more flexibility when demand is hard to predict and data centres take years to build.
There is also a strategic benefit in designing hardware alongside the models, software and services that will run on it. Microsoft can tune the full stack for its own workloads rather than relying on one general-purpose accelerator to be ideal for everything.
Not a GPU replacement
And this does not mean GPUs are disappearing from Microsoft’s infrastructure. The announcement is about making Maia the primary inference system for Microsoft’s own products and services; it does not say Microsoft will stop using Nvidia or AMD accelerators for other workloads. Hardware and software will also continue to change before Maia 300 starts production deployments in 2027.
Still, the investment is notable. Maia 200 is already running real Microsoft workloads, and Microsoft is now committing to much larger deployments of its own inference hardware. The interesting question is not whether custom silicon replaces every GPU, but how much of Microsoft’s own AI workload it can serve more efficiently 😊