Microsoft has announced Maia 300, its next-generation custom
artificial intelligence (AI) chip designed to run AI models across its data centres, as technology companies seek to control the cost of operating AI services. The company plans to begin large-scale deployments of its Azure Maia systems in 2027, with the chips expected to become the primary inference platform for its products and services.
Microsoft is not the only major technology company deploying custom chips for AI inference. From Google's Tensor Processing Units (TPUs) to Amazon's Trainium and OpenAI's planned inference chip, the industry is moving towards greater control over the hardware that powers AI services.
Why inference is becoming the focus
AI computing has two broad stages: training and inference. Training involves processing large amounts of data to build a model, while inference is the process of using that trained model to generate an answer, write code, summarise a document or complete another task.
As AI tools become available to more users, inference becomes a recurring operating expense. Every query, generated image, coding suggestion or interaction with an AI agent consumes computing resources. Services used by millions of people can therefore require substantial processing capacity even after their underlying models have been trained.
Not every inference task requires the same computing power. A complex reasoning request or a long-context AI agent may need substantial processing and memory, while a simpler classification or summarisation task may run efficiently on a less powerful, specialised processor. Companies are seeking hardware suited to particular workloads rather than using expensive, general-purpose accelerators for every request.
This distinction is important because a chip's value is determined not only by its peak computing power but also by how quickly and efficiently it can complete a task.
Cost, power and control over infrastructure
The economics of inference are a major reason for the shift. Custom chips can be designed around a company's models, software and expected workloads, potentially improving the amount of computing delivered for every rupee or dollar spent on hardware and electricity.
Power consumption is another consideration. Data centres must supply electricity and cooling to large numbers of processors operating continuously. Improving the number of AI tokens generated per unit of energy can help companies serve more requests within existing power constraints.
Custom silicon also gives companies greater control over hardware supply and infrastructure planning. Depending on their scale and chip design, companies may be able to reduce purchases of external accelerators, plan capacity more predictably and optimise the interaction between processors, memory, networking and software.
However, designing a chip does not automatically make it cheaper. Development costs, manufacturing yields, memory availability, software support and the expense of building compatible data-centre systems all affect the final budget.
Microsoft is scaling its Maia programme
Microsoft's Maia 300 builds on the company's earlier Maia processors. According to its announcement, Maia 200 is already serving production workloads, including Microsoft Copilot, and supports Microsoft's own models as well as OpenAI and open-source model families.
The company said the Maia 300 architecture is designed for long-context reasoning, memory-intensive processing and multi-step AI agent tasks. Its planned uses include Copilot, GitHub Copilot, Microsoft Foundry and Dragon Copilot.
Microsoft's strategy is to coordinate chip design with its cloud infrastructure and AI software. This could allow it to optimise the entire system for the tasks its services perform, rather than adapting every workload to a processor designed for a broader range of applications.
Building competitive AI hardware requires designing entire systems and optimising everything from memory and networking to software and power consumption, Rani Borkar, Microsoft's president of Azure Hardware Systems and Innovation, told Axios.
"We prioritize Maia for inference," Borkar said, describing it as the largest and fastest-growing opportunity for AI computing and an area where Microsoft can distinguish itself on cost and efficiency. "For us, it's not an either-or strategy," she said. "We want a broad ecosystem while investing deeply in our first-party silicon."
The planned deployment is also significant because Microsoft operates Azure, giving it a large internal market in which to use and refine its chips before expanding their role across its services.
Google and Amazon are taking similar approaches
Google has developed its Tensor Processing Units (TPUs) for years and uses them to support its AI models, including Gemini. In April 2026, Google announced its eighth-generation TPU systems, separating the design into TPU 8t for training and TPU 8i for inference.
According to Google, TPU 8i is designed for low-latency inference and the demands of AI agents that must reason and complete multiple steps. TPU 8t, meanwhile, targets the computing and memory requirements of training large models. Google also continues to offer Nvidia GPU instances to cloud customers, indicating that its own chips are part of a wider hardware portfolio rather than an outright replacement strategy.
Amazon is pursuing a similar strategy through its Trainium family. In July 2026, Amazon said its custom-chip business had crossed a $25 billion annual revenue run rate. The company said Trainium3 offers up to 40 per cent better price-performance than Trainium2, although such comparisons reflect Amazon's own reported results. Its Bedrock service also uses Trainium for much of its inference workload.
The developments suggest that custom silicon is becoming a core part of cloud infrastructure, particularly for companies operating AI services at scale. It can also give cloud providers a way to offer customers different computing options based on performance, cost and workload requirements.
OpenAI moves towards owning more of the AI stack
OpenAI, which has relied on external computing infrastructure to develop and operate its models, announced its first custom inference processor, Jalapeño, with Broadcom in June 2026.
According to OpenAI and Broadcom, the chip is designed around the requirements of current and future large language models, with deployment planned at gigawatt scale across data-centre partners over multiple generations. The companies said early testing showed improved performance per watt.
This move is crucial because OpenAI operates services such as ChatGPT, where inference demand grows with usage. Designing a processor around its models could give the company greater control over how its AI services are served and scaled. Its partnership with Broadcom also illustrates that developing a custom chip does not necessarily mean manufacturing it independently: chip design, fabrication, packaging and deployment involve a broader supplier ecosystem.
Nvidia's position is changing, not disappearing
The expansion of custom chips challenges Nvidia's position in parts of the AI market but does not mean its graphics processing units (GPUs) will become redundant. Nvidia offers a broad hardware and software ecosystem, including its CUDA platform, that supports a wide range of AI workloads. This flexibility remains important for companies training new models, experimenting with architectures or deploying systems across different environments.
As the technology industry moves towards a mix of processors, with companies choosing hardware according to the workload rather than relying on a single chip category, Nvidia is also moving deeper into the AI ecosystem, with its recent acquisition of AI repository platform Hugging Face being an example.
 "Beyond Nvidia: Why tech firms are turning to custom chips for AI inference")
Connect with us on WhatsApp