Tech giants face real challenges in "de-NVIDIA" as Microsoft's AI chips are delayed until mass production in 2026

Despite investing huge amounts of money and assembling large teams, technology departments such as Microsoft and Google are facing severe technical and talent challenges in their pursuit of developing AI chips to challenge Nvidia's market dominance.

On Friday, media reports citing informed sources indicated that the design cycle of Microsoft's latest generation AI chip far exceeded expectations, making it difficult for its performance to compete with similar products from Nvidia. Its chip, Maia 100, is currently only used for internal testing.

At the same time, Google's work in the chip sector has also been hit by talent loss. According to direct insiders, when Google collaborated with MediaTek to develop the next generation Tensor Processing Unit (TPU), key members of MediaTek's core team responsible for critical network technology switched to Nvidia.

For investors, these events further demonstrate Nvidia's absolute leading position in the hardware field. Although self-developed chips are a necessary path to reduce costs, the short-term and performance gaps during the development process mean that large tech companies will still be important customers of Nvidia in the short term.

Development Progress Severely Delayed, Design Changes Causing Project Obstacles

Microsoft made a high-profile announcement in 2023 regarding its flagship self-developed AI chip, Maia 100, which received public praise from OpenAI's Altman, stating that it paved the way for reducing model operational costs. However, several insiders and former Microsoft employees admitted that there is a gap between reality and promotion. The chip has yet to provide computing power for any AI services, with its main use still involving internal testing.

The root of the problem lies in its outdated design. The development of Maia 100 began in 2019, with the design goal at that time focused on solving image processing issues rather than serving the currently dominant generative AI market. After the explosive industry growth triggered by OpenAI's ChatGPT, the architecture of this chip can no longer meet new demands, rendering it "ill-timed" and difficult to play a role in practical applications.

To catch up with technological trends, Microsoft has developed an ambitious roadmap for subsequent chips. According to insiders, this plan is set to be reported, with three successor chips codenamed Braga, Braga-R, and Xylia scheduled for deployment from 2025 to 2027.

However, Microsoft's next-generation AI chip codenamed Braga is facing at least a six-month delay, pushing its mass production from 2025 to 2026. They stated that when it finally goes into mass production next year, its performance is expected to be far below that of Nvidia's Blackwell chip, which will be released at the end of 2024.

During the development process, Microsoft requested design changes for Braga to integrate new features brought by OpenAI. These changes caused the chip to become unstable during simulation runs. Although significant design changes occurred, Microsoft refused to postpone the deadline for completing the design by the end of the year, putting immense pressure on the team, which resulted in a loss of one-fifth of its members.

Insiders indicated that Microsoft's AI chips will not be able to compete with Nvidia's products in terms of performance at least until Maia 300 is released. Meanwhile, the company has canceled its plans to develop AI training chips at the beginning of 2024, with all current chip efforts focused on inference applications

Google Faces Talent Drain, Nvidia Takes Initiative

Microsoft's predicament is not an isolated case. Google has also encountered difficulties while collaborating with MediaTek in Taiwan to design the next-generation TPU. According to a source with direct knowledge, key members of the MediaTek team responsible for the critical networking technology of the TPU (which enables multiple chips to work together) have joined Nvidia.

With the urgent growth of self-developed chip plans from clients, Nvidia President Jensen Huang clearly stated to reporters at last month's developer conference that most custom chip projects pushed by tech companies will ultimately be abandoned. He countered, "What’s the point if the application-specific integrated circuit you develop is not better than the chips you can buy?"

In addition to publicly expressing confidence, Nvidia has also adopted a proactive defense strategy. According to insiders, to make it difficult for clients to replace their self-developed chips with other products, Nvidia has set corresponding behavioral goals for its newly released flagship AI system GB200. The intention is to continuously enhance its market leadership through ongoing technological leadership