Google is on track to manufacture more artificial intelligence accelerators than Nvidia sells to the entire market by 2028, according to new research from Fubon. The company plans to produce between 12 and 15 million ninth-generation TPUs (Tensor Processing Units), custom chips designed specifically for AI workloads, compared to Nvidia's projected 12.4 million data center AI GPUs in the same year.

What Are TPUs and Why Does Google Need So Many?

TPUs are custom-built processors that Google has been developing for roughly a decade to power its AI services. Unlike Nvidia's general-purpose graphics processors, TPUs are tailored to Google's specific software and data center needs. The ninth-generation TPUs will feature four compute chiplets, essentially smaller processing units stacked together to boost performance. This architectural choice signals Google's confidence in the design and represents a significant engineering undertaking.

The sheer volume Google plans to deploy underscores a fundamental shift in how major technology companies approach AI infrastructure. Rather than relying entirely on off-the-shelf hardware from vendors like Nvidia, Google is betting heavily on vertical integration, where companies design chips matched to their own workloads and software stacks. This approach gives Google tighter control over its AI capabilities and potentially lower costs at massive scale.

Why Would Google Need Intel's Help Alongside TSMC?

Manufacturing 12 to 15 million chips annually is an enormous undertaking. Fubon researchers believe that even TSMC, the world's leading chip manufacturer, cannot handle Google's entire demand alone. The company will likely need to tap Intel Foundry Services, Intel's contract manufacturing division, to meet its volume targets.

This partnership makes practical sense because of how modern chips are assembled. Google's TPUs use advanced packaging technologies to connect their four compute chiplets. Intel has developed proprietary packaging methods like EMIB (Embedded Multi-die Interconnect Bridge) that are incompatible with TSMC's competing technology. Reports suggest Google has already tested Intel's packaging capabilities and may order three million TPUs from Intel Foundry.

How to Understand the Competitive Implications for the AI Hardware Market

• Market Share Shift: If accurate, Google's 12 to 15 million TPUs would match or exceed Nvidia's 12.4 million projected GPU shipments in 2028, marking a dramatic rebalancing of AI hardware dominance away from a single vendor.
• Software Ecosystem Risk: Nvidia's real competitive advantage is not just its hardware but CUDA, its software platform that developers rely on to write AI applications. Google's TPUs depend on a rival software stack, which could eventually challenge Nvidia's lock on developer mindshare.
• Continued Demand Growth: Both companies can grow simultaneously because AI demand is expanding so rapidly that the market can absorb massive increases from multiple suppliers without saturation.
• Supply Chain Diversification: Google's reliance on both TSMC and Intel Foundry reduces its dependence on a single manufacturer and mirrors a broader industry trend toward geographic and vendor diversification in chip production.

The implications extend beyond raw unit counts. Google will likely become the world's largest consumer of AI accelerators, deploying more chips annually than any other organization. This positions the company to own what could become the world's most capable AI hardware fleet, giving it enormous computational resources for training and running advanced AI models.

Whether Google will use this overwhelming compute capacity primarily for its own services like Search, Gmail, and Cloud offerings, or will make significant capacity available to other companies and researchers, remains an open question. The answer will shape the competitive landscape for years to come.

For Nvidia, the milestone would not necessarily diminish its market position in the near term. The company's AI GPUs remain sold out, and demand continues to outpace supply. However, Nvidia should focus less on the volume of TPUs Google can deploy and more on the fact that Google's processors run software that rivals CUDA, potentially eroding Nvidia's most defensible competitive advantage.

Nvidia plans to release its Feynman and Feynman Ultra processors in 2029 to 2030, which could accelerate its own production capacity and maintain competitiveness. But the window for Nvidia to respond is narrowing as Google's vertical integration strategy matures and the company proves it can manufacture AI chips at scale.