Microsoft Scales Back Ambitions for AI Chips to Overcome Delays
The Takeaway
• Microsoft will focus on less ambitious chip designs through 2028
• Decision follows delays in in-house chip development
• Microsoft wants to reduce reliance on Nvidia for AI computing
Microsoft is revising its roadmap for its internally developed artificial intelligence server chips and will focus on releasing less ambitious designs through 2028, hoping to overcome problems that have caused delays in development, according to two people with direct knowledge of the situation.
Microsoft hopes that by scaling back some of the designs and pushing out the schedule for other AI chips it is already working on, it can develop them more easily. The hope is that when these chips are released over the next three years, they will still remain competitive with Nvidia’s AI chips.
Microsoft executives told engineers in its silicon team about the new plans in a meeting last week. The decision comes after Microsoft had to push back the release of its latest-generation AI chip, Maia 200, to 2026 from 2025, as The Information previously reported.
Like both Google and Amazon, Microsoft designs its own chips meant to power AI services such as OpenAI’s ChatGPT in hopes of creating an alternative to Nvidia’s chips, which dominate the market. Nvidia tightly controls the prices and supply of its own chips.
Microsoft last year was Nvidia’s largest customer by revenue, according to multiple Nvidia employees, and spends billions of dollars a year buying Nvidia AI chips for its Azure cloud service.
Microsoft launched its first AI chip, dubbed Maia 100, in 2024 and immediately began working on three successors—codenamed Braga, Braga-R and Clea—due for release in 2025, 2026 and 2027, respectively. However, both Braga and Clea were based on entirely new designs, making their development especially challenging.
The Braga chip’s design was only completed in June, missing a year-end deadline by around six months. Braga’s delay raised concerns internally within Microsoft that the chips due in 2026 and 2027 also could be delayed, making them even less competitive with Nvidia’s chips by the time they were released, the people said.
As a result, Microsoft executives told engineers last week that the company is considering developing an intermediary chip for release in 2027 that will sit between Braga and Braga-R in terms of performance. This chip will likely be known as Maia 280 and is still largely based on Braga’s design. However, the Maia 280 consists of at least two Braga chips linked together so that they can work together as a single, more powerful chip.
A Microsoft spokesperson didn’t comment directly on the development of its Maia chips but said the company “remains committed” to developing in-house hardware based on its customers’ and its own computing needs while continuing to work with “close silicon partners.”
Microsoft executives believe the Maia 280 approach will still deliver between 20% and 30% better performance in terms of performance per watt compared to the chips that Nvidia will release in 2027, one of the people said. The new roadmap allows Microsoft’s chip engineers to release new chips in 2027 and 2028 without the need to redesign Braga from scratch to get performance gains. The company plans to ramp up production with each new chip generation and eventually make hundreds of thousands of AI chips a year, according to two people with direct knowledge of the matter.
Braga-R, which will now be publicly called Maia 400, is expected to enter mass production in 2028. Braga-R will connect the brains of two chips prior to their full assembly at the so-called die level, allowing for even faster performance once assembled into a single chip. Braga-R will also incorporate faster, high-bandwidth memory, the people said. The trade off is that there is a higher potential for manufacturing defects using the die-to-die method.
That’s a different approach than is used by chip makers like Nvidia, which link two finished chips to increase performance. The new approach to improve on Braga’s performance reflects an acknowledgment by Microsoft’s leaders that designing a new chip from scratch each year wasn’t feasible.
Although Braga-R’s design is still based on Braga, Microsoft plans to utilize a so-called chiplet architecture, which means dividing the chip up into smaller components and functions that can be manufactured separately by outside parties and later recombined, three people said.
The chiplet architecture helps reduce manufacturing defects by keeping the surface area of each chiplet smaller and lowers the likelihood of defects occurring during production.
It also allows Microsoft to save money because it can outsource portions of the chip’s design, several people involved in the project said. The tradeoff is that chiplet-based chips don’t perform as well as designing the same functions into a one single chip.
The release of Microsoft’s third-generation AI chip, Clea, has been pushed beyond 2028, and its future remains unclear. Clea was intended to be the first AI chip that would be competitive with Nvidia’s chips in terms of performance per watt.
Microsoft’s revised roadmap has negative implications for Marvell, a specialized chip designer that Microsoft hired to work on some of the chiplets in Braga-R, according to two people with direct knowledge of the matter. Marvell’s shares surged last year due to partnerships with major tech firms including Amazon, which relies on Marvell to help build their AI chips, and Marvell had anticipated revenue from Microsoft much sooner, the people said. However, its stock price has fallen this year amid delays in chip projects from its customers, a slowing global economy and trade tensions between the U.S. and other countries.
Not all of Microsoft’s chip projects are facing problems. The company released a central processing unit for servers called Cobalt in 2024 as an alternative to chips made by Intel and AMD. CPUs are easier to design than graphical processing units, the main component in AI chips, and other companies such as Amazon have had similar success with building in-house CPUs as opposed to GPUs. Cobalt is used both internally by Microsoft to power services such as its Teams communication software, and the chip also is available for use to Azure customers. Unlike Maia, Cobalt already is generating revenue for the company, according to multiple people with direct knowledge of the matter.
The design for Cobalt’s successor, codenamed Kingsgate, was completed around March and also will utilize a chiplet-based design and faster high-bandwidth memory, according to two of the people.