Microsoft Reportedly Plans to "Significantly" Boost Production of Next-Gen AI Chips, Negotiating with Taiwan Semiconductor for Delivery of Over 300,000 Units in 2027

Microsoft is betting on breakthroughs in self-developed chips in an attempt to break free from its heavy reliance on NVIDIA.

According to a Monday report by tech media outlet The Information, two insiders revealed that Microsoft plans to release its new generation of self-developed AI chip, the Maia 300, this fall, with a possible public debut as early as next month. The company is currently negotiating a capacity contract with Taiwan Semiconductor for the delivery of over 300,000 chips in 2027—a quantum leap compared to the mere tens of thousands of units produced for the current-generation Maia 200. Meanwhile, Microsoft is actively negotiating with major cloud customers such as Anthropic regarding the use of Maia chips.

The core driver behind this expansion plan is cost advantage. Last month, Microsoft disclosed to investors that the operating costs of the Maia 200 chip were 30% to 40% lower than those of NVIDIA's cutting-edge chips when running OpenAI and Microsoft's own models; according to insiders, the Maia 300, optimized specifically for Microsoft's models, performs even better. Even if it fails to secure major external customers, Microsoft's backup plan is to shift more internal AI workloads to Maia while continuing to rent out high-priced NVIDIA chips to Azure cloud customers.

However, Microsoft still lags significantly behind Google and Amazon in the race for self-developed chips. Google's TPU and Amazon's Trainium have already been adopted by multiple major customers, with Google even beginning to sell TPUs to external clients. Currently, the Maia 200 is used only by Microsoft itself and, as of the end of last month, was operating in just two data centers within the United States.

Slow Start for Maia 200 Limits Expansion Ambitions

The deployment progress of the Maia 200 has fallen significantly short of expectations. The chip was delayed last year after early tests failed to meet internal targets, and it has since been deployed in only a few Microsoft data centers.

Microsoft CEO Satya Nadella stated in June this year that two data centers had been put into operation and that plans were in place to expand to more centers, including overseas deployments. However, according to one insider, as of the end of last month, the situation remained unchanged, with operations limited to the two data centers in the United States and no further progress made.

In contrast, competitors enjoy a significant scale advantage.

According to estimates by Morgan Stanley, Google plans to produce over 3 million TPU chips this year, with output reaching 5 million units next year. The order of over 300,000 Maia 300 chips currently under negotiation by Microsoft represents a significant gap compared to these figures.

Betting on Maia 300 with Targets Aimed at the Million-Unit Level

Despite this, Microsoft has not scaled back its ambitions.

The report states that Microsoft ultimately hopes to secure capacity for over one million Maia 300 chips, but the actual scale remains uncertain due to component supply constraints and ongoing capacity negotiations with Taiwan Semiconductor.

Andrew Wall, General Manager of Microsoft Azure Maia, declined to comment on specific production plans in a statement but indicated that Microsoft ultimately aims to produce Maia chips measured in gigawatts.

He said, "Microsoft continues to invest in custom silicon as part of our long-term AI infrastructure strategy. We expect Azure Maia deployments to support AI workload demands measured in gigawatts." Typically, data centers requiring several gigawatts of power can house millions of AI chips.

Regarding potential customers, The Information previously reported that Anthropic has been in negotiations with Microsoft for several months regarding the future use of Maia chips. Microsoft's appeal mainly stems from the cost side: internal company tests show that the Maia 300 offers particularly outstanding cost-effectiveness when running Microsoft's own models.

Strategic Drive to Break Free from NVIDIA

Behind the push for self-developed chips lies a core strategic concern for Nadella that has persisted for years. In an email sent in 2022, later made public through legal proceedings, Nadella wrote, "We are now just a thin layer on top of NVIDIA, with all intellectual property held by OpenAI," and mentioned that a certain business unit at Microsoft "would lose $4 billion next year."

According to one insider, Nadella was referring to the high costs of running OpenAI models on Azure—at the time, Microsoft could control neither the chip costs nor the AI models themselves. This situation marked the starting point for the subsequent dual-track strategy of developing self-owned chips and proprietary models.

Currently, Microsoft still relies primarily on NVIDIA chips to run most of its AI software, including the MAI model powering Copilot. The commercialization path for Maia involves gradually migrating internal workloads to self-developed chips and ultimately persuading external customers to follow suit with more competitive pricing.

Cobalt CPU Emerges Strongly, Offering an Alternative Path for Self-Developed Silicon

Notably, compared to the rocky progress of Maia, Microsoft's other line of self-developed chips—the Cobalt central processing unit (CPU)—has recently demonstrated stronger market traction. Last week, Microsoft announced that major customers such as OpenAI and Adobe have adopted Cobalt in over 25 data centers globally.

Cobalt falls into the traditional CPU category, differing from NVIDIA GPUs and AI-specific chips like Maia. However, against the backdrop of exploding demand for AI infrastructure, CPU demand has also surged significantly. The phased success of Cobalt provides a certain endorsement for Microsoft's overall self-developed silicon strategy and builds a foundation of customer trust for the subsequent promotion of Maia.