Skip to main content

Google's Ambitious AI Chip Plan: 15 Million by 2028, A Direct Challenge to Nvidia

Google is making a bold move in the AI hardware arena. According to a market analysis report cited by Taiwanese media Jutai, the tech giant plans to deploy between 12 million and 15 million of its self-developed TPU v9 AI acceleration chips by 2028. That's a staggering number, especially when you consider that Nvidia's AI GPU shipments are projected to reach around 12.4 million units in the same year. If these figures hold, Google could actually surpass Nvidia in server AI accelerator shipments—a direct challenge to the reigning champion of computing power.

But what exactly is a TPU? It's an ASIC chip designed by Google specifically for AI training and inference. Unlike Nvidia's GPUs, which are sold commercially, TPUs have been kept under wraps, used internally and never publicly sold. This shift from renting Nvidia GPUs to developing in-house chips is a strategic move for Google to cut computing costs and reduce its reliance on a single supply chain. It's a classic case of "if you want something done right, do it yourself."

A 4 Compute Die Architecture and the Quest for Manufacturing Partners

The report also sheds light on the technical side of TPU v9. It's expected to adopt a 4 Compute Die structure, which will push the limits of advanced process and packaging technologies. This means Google will need to look beyond TSMC for manufacturing partners. Intel and Samsung foundry services are likely to see new business opportunities as Google adopts a multi-supplier strategy to meet its capacity needs.

This development reflects a deeper transformation in the industry: leading tech giants are moving away from simply buying general-purpose GPUs and instead developing customized AI chips. As Google prepares to build an independent computing infrastructure with millions of its own chips, it may be positioning itself as the strongest challenger to Nvidia's dominance in AI chips.

Key Points

  • Google plans to deploy 12-15 million TPU v9 AI chips by 2028, potentially surpassing Nvidia's projected 12.4 million AI GPU shipments.
  • TPU v9 will feature a 4 Compute Die architecture, requiring advanced manufacturing and packaging.
  • Google is expected to diversify its manufacturing partners, benefiting Intel and Samsung foundries.
  • This move underscores a broader industry trend toward custom AI chip development by major tech companies.