Google is moving to expand supply of its Tensor Processing Units, a push that analysts say is intensifying competition among global tech giants over South Korea's artificial intelligence ecosystem. As the Korean government and companies scale up AI infrastructure investment and accelerate the commercialization of physical AI, Google is following Nvidia and AMD in widening its contact with the domestic AI ecosystem and stepping up efforts to secure talent.
Google has shifted its TPU business strategy this year, according to information technology industry officials and foreign media reports on the 22nd. The company had previously installed TPUs within its own cloud infrastructure to run internal services such as search and Gemini, or leased them to cloud customers. But on an April conference call, Google formally announced it would also supply TPUs to external data centers outside the Google Cloud Platform. The move follows expectations that computing demand for AI agents and physical AI technology will surge.
TPUs drew attention across the AI industry after being deployed in large volumes to train Gemini 3.0, released in October last year. Google, once seen as a latecomer in AI, unveiled a frontier-class AI model without relying on GPUs. TPUs are application-specific integrated circuits built for particular types of computation, setting them apart from graphics processing units, which are general-purpose AI chips. Their main advantage is greater power efficiency than GPUs. The latest TPU v8 delivers about three times the training performance per unit of power consumption of its predecessor, and about 1.8 times the inference performance.
External demand is already materializing. Anthropic, the developer of Claude, plans to expand its TPU deployment from 1 gigawatt this year to 5 gigawatts next year. Investment bank Morgan Stanley said in a recent report that Google's direct TPU sales would grow to $84 billion, or about 116 trillion won, in 2027 and $108 billion in 2028.
Industry officials believe Google has picked Korea as one of its main TPU business bases. Under three mega-projects announced by the government, a total of 550 trillion won in private investment is being pursued to build 8.4 gigawatts of AI data centers by 2029. According to S&P Global Ratings, Korea is projected to invest about 1,200 trillion won in data center construction through 2035. At that pace, it would become the second-largest data center market in the Asia-Pacific region after China.
Korea's diverse manufacturing ecosystem is seen as a particular draw, giving the country potential as a base for commercializing physical AI. That is why Nvidia Chief Executive Jensen Huang visited Korea twice, in October last year and June this year, to discuss various forms of cooperation. During his June visit, he cited three factors — a culture of rapid technology adoption, a geopolitically neutral position and a world-class industrial base — and stressed that "Korea can stand at the global center of the physical AI space." After Chief Executive Lisa Su visited Korea in March, AMD began pursuing the construction and demonstration of open AI computing infrastructure that combines its central processing units and GPUs with domestically developed neural processing units. Nvidia and AMD have also decided to set up AI research centers in Korea.
With Google now joining in, Korea is showing signs of rapidly emerging as a battleground for AI infrastructure among big tech firms. Google is seen as pursuing a strategy of broadening its contact with Korean industry, academia and research institutes by leading with full-stack AI technology, including not only TPU-based AI infrastructure but also AI models and security technology. The aim is to secure Korea's strong AI talent while generating AI chip revenue at the same time. "The lead time for Nvidia GPUs — the period until final delivery — runs as long as 30 weeks, and the bottleneck has not been resolved," an AI industry official said. "Google is taking aggressive business steps to exploit that gap."
For Korean industry, academia and research institutes, the interests align, as the range of options for AI computing resources widens. "Since we have to prepare even for the era of physical AI, the more computing resources the better," said Choi Byung-ho, a professor at Korea University's AI research institute. "For now, industry and academia are accustomed to using GPUs, so the key question is how much AI research output will come out of TPUs."
The prospect of expanded cooperation with major Korean AI companies is also gaining weight. Kakao signed a strategic AI partnership with Google in February this year, saying it would further advance its on-device AI services and actively adopt TPUs in developing its own AI. LG AI Research is also using TPUs as part of its computing resources in developing Exaone, its proprietary AI model.
Experts expect cooperation between U.S. big tech firms and Korea's AI ecosystem to continue over the long term. "From the U.S. perspective, apart from its adversary China, Korea is the only country worth partnering with in the physical AI field, so AI cooperation between the United States and Korea has no choice but to strengthen," Choi said.
