Microsoft Is Building Its Own AI Chips. Nvidia Should Pay Attention.

Microsoft wants to change the hardware powering its artificial intelligence services, and one of the biggest names in the chip industry has a reason to pay attention.

On October 12, the company unveiled Maia 300, its latest custom-designed artificial intelligence chip, alongside plans to expand deployment of its Maia computing systems to a multi-gigawatt scale beginning in 2027.

The announcement is more than another product launch. Microsoft is trying to gain greater control over the expensive infrastructure behind AI, at a time when demand for computing power is growing rapidly and companies are competing to make AI services faster, more efficient and cheaper to operate.

Nvidia has become one of the biggest beneficiaries of this demand. Its graphics processing units, or GPUs, are widely used to train AI models and run them after training. But Microsoft's latest move highlights a growing shift across the industry: some of the companies buying AI chips are also developing their own.

Microsoft's Maia 300 is part of that effort.

The company says the new system is designed to deliver 40% more AI output, measured in tokens per megawatt, than leading GPU-based systems. It also says a rack containing 96 Maia 300 accelerators can achieve up to seven times the token throughput of a comparable Maia 200 configuration.

Those are significant claims, although they are Microsoft's own performance estimates rather than independently verified results across every workload.

To understand why they matter, it helps to look at what AI chips actually do.

When someone asks Microsoft Copilot to summarise a document, help write code or answer a question, the AI model must perform calculations before generating a response. That process is known as inference.

Training an AI model requires enormous computing resources, but inference becomes particularly important once a model is serving millions of requests. Every conversation, generated image, coding suggestion or automated task consumes computing capacity.

Multiply those interactions across Microsoft's products and cloud services, and the costs become substantial.

Hardware efficiency therefore matters enormously. If a company can generate more AI output using the same amount of electricity, it may be able to serve more users without increasing power consumption at the same rate. Depending on hardware costs, utilisation, software and other operating expenses, that could also improve the economics of running its services.

This is where Microsoft's custom-chip strategy becomes interesting.

Rather than relying entirely on processors designed by outside suppliers, the company can build hardware around the specific AI workloads it expects to run. It can coordinate chip design with its software, networking equipment, memory systems and data centres.

That approach gives Microsoft more opportunities to optimise the entire system instead of treating the chip as an isolated component.

Maia 300 is designed for demanding AI workloads, including conversations that require quick responses, reasoning across long inputs and AI agents that perform multiple steps to complete a task. Those capabilities are increasingly important as technology companies develop assistants that can do more than simply respond to individual prompts.

Microsoft is also building on an existing foundation. Its Maia 200 chip is already running production workloads, including Microsoft Copilot. The company says Maia 200 delivers up to 1.7 times more tokens per watt for certain built-in AI capabilities compared with leading GPU infrastructure.

Maia 300 is expected to extend that approach, with production deployments scheduled to begin in 2027.

The implications extend beyond Microsoft's own products.

Nvidia's position in AI computing has been strengthened by its hardware, software ecosystem and relationships with cloud providers and AI developers. Its GPUs remain important to the industry, and Microsoft's announcement does not mean the company is about to stop buying them.

Instead, Microsoft is pursuing greater flexibility. By developing its own chips, it can potentially reduce its dependence on external suppliers, optimise costs for its internal workloads and gain more control over how its AI infrastructure evolves.

Other major technology companies are pursuing similar strategies. Google has its Tensor Processing Units, Amazon has developed its Trainium and Inferentia chips, and Meta has been investing in its own AI silicon. The broader pattern is clear: controlling the hardware is becoming an increasingly important part of competing in artificial intelligence.

However, designing a chip is only one part of the challenge.

Manufacturing advanced processors at scale requires specialised facilities, sophisticated packaging, high-bandwidth memory and a reliable supply of components. Even a promising design must prove itself in real-world deployments, where performance, reliability, software compatibility and operating costs all matter.

Microsoft also has to make sure developers and its own AI teams can use the hardware effectively. A powerful processor is less useful if the software running on it cannot take full advantage of its capabilities.

And there is a larger challenge hanging over the entire industry: electricity.

Modern AI data centres consume enormous amounts of power. As companies add more computing capacity, they must also secure energy supplies, manage heat and cooling, and build the infrastructure needed to keep their systems running. Improving the number of AI operations completed per unit of electricity can help, but it does not automatically eliminate the energy demands of expanding AI services.

Microsoft's multi-gigawatt deployment plans show the scale of infrastructure it expects to need. They also illustrate why efficiency has become a competitive issue rather than merely an engineering preference.

For now, Maia 300 is a promising strategic move, not proof that Microsoft has displaced Nvidia. Its real significance will become clearer as the chips enter production, performance claims are tested under practical workloads, and the company demonstrates how effectively it can deploy them at scale.

Still, the direction is unmistakable. The AI race is no longer only about who builds the most capable model. It is also about who can afford to run that model efficiently, serve more people and keep the infrastructure behind it under control.

Microsoft wants a bigger role in that part of the race. Building its own chips is one way to get there.