Google is accelerating mass production of its in-house artificial intelligence accelerator, the Tensor Processing Unit (TPU). With computing demand surging as the AI agent market expands, attention is turning to whether the chip can crack Nvidia's dominance in graphics processing units.
TPUs began drawing attention across the AI industry after Google deployed them in large volumes to train Gemini 3.0, released last October, according to information technology industry sources and foreign media reports on the 24th. Google, long seen as a latecomer in AI, delivered a frontier-class model without relying on GPUs.
A TPU is an application-specific integrated circuit built for particular types of computation, which sets it apart from the GPU, a general-purpose AI chip. Its main advantage is greater power efficiency than GPUs. Google says the latest version, TPU v8, delivers about three times the training performance per watt of its predecessor and about 1.8 times the inference performance.
Broadcom, Google's U.S. manufacturing partner for the chip, disclosed the production status of the TPU. On the 2nd, in its third-quarter earnings announcement, Broadcom said it had begun production shipments of Google's next-generation inference chip, the TPU 8i, and planned to expand volume shipments in earnest in the fourth quarter. The training chip, the TPU 8t, code-named Sunfish, may enter mass production around the same time. A defining feature of TPU v8 is that it is split into separate chips for training and inference.
Google has paired that production push with an aggressive business strategy. Until now it had deployed TPUs mainly within its own cloud infrastructure, running internal services such as search and Gemini or leasing capacity to cloud customers. But on an April conference call, the company said it would also supply TPUs to outside data centers. Anthropic, the developer of Claude, is reported to be raising its TPU deployment from 1 gigawatt this year to 5 gigawatts next year. In a recent report, investment bank Morgan Stanley projected that Google's direct TPU sales would grow to $84 billion in 2027 and $108 billion in 2028. A semiconductor industry official said TPU adoption is rising as part of a set of alternatives aimed at improving AI computing efficiency, even if the chips cannot fully replace GPUs.
As Google pushes to expand TPU supply, its plans in South Korea are also drawing interest. The industry believes Google has designated the country as one of its main bases for the TPU business. S&P Global Ratings, the international credit rating agency, projected that about 1,200 trillion won (roughly $900 billion) will be invested in data center construction in South Korea through 2035.
South Korea's diverse manufacturing ecosystem is also cited as an attraction, making the country a potential base for commercializing physical AI. During a visit in June, Nvidia Chief Executive Jensen Huang pointed to three factors — a culture that adopts technology quickly, a geopolitically neutral position and a world-class industrial base — and said South Korea could stand at the global center of physical AI. AMD is working to build and demonstrate open AI computing infrastructure that combines its own central processing units and GPUs with domestically developed neural processing units.
Google appears to be leaning on a full-stack AI offering — AI models and security technology on top of TPU-based infrastructure — to widen its contacts with South Korean industry, academia and research institutes. The strategy looks designed to secure top AI talent in the country while generating AI chip revenue at the same time. For local industry, academia and research institutes, the interests align because their options for AI computing resources broaden. An industry official said lead times for Nvidia GPUs run as long as 30 weeks and the bottleneck persists, adding that in such conditions Google is likely to press ahead aggressively.
