Google's artificial intelligence (AI) model "Gemini 3" has drawn attention to Google's Tensor Processing Unit (TPU) after reports that it did not use Nvidia's graphics processing unit (GPU), which dominates the AI Semiconductor market, while delivering performance that threatens OpenAI's "ChatGPT." Some say TPUs can replace Nvidia GPUs more efficiently and at lower expense.
Jensen Huang, Nvidia's chief executive officer (CEO), appeared to take note and made an unusual remark checking Google's TPU. On the 25th (local time), he said on his official account on the social media (SNS) platform X that "Google has made great progress in AI," while emphasizing that GPUs remain ahead of Google's TPUs in technology. He also expressed confidence that even if Google builds AI models with TPUs, it will still need Nvidia GPUs.
◇ Can Gemini 3, which stunned the AI industry, work without GPUs
There have been attempts in the past to challenge Nvidia, which has monopolized the AI Semiconductor market. Not only Google but also big tech corporations such as Microsoft (MS), Tesla and Amazon have researched in-house chip designs for a long time to develop alternatives to GPUs. Some of these have actually been used for AI training models and, in certain areas, have shown better efficiency than GPUs.
So will there be a seismic shift in Nvidia's GPU monopoly structure? In short, experts say that is not possible for now. First, some analysts say the claim that Google's Gemini 3 was built only with TPUs without GPUs is flawed.
A senior official at domestic AI fabless companies said, "The reason Gemini 3 can be trained on TPUs is because Google has long trained data and frameworks based on Nvidia's GPUs," adding, "All the experience over the past several years—AI research and models, optimization techniques, distributed training—has been built on GPUs, and Gemini 3 optimized that data foundation for TPUs." The official said it is hard to accept the claim that Gemini 3 was developed without GPUs.
It is necessary to understand that GPUs and TPUs are conceptually and technically completely different hardware. A GPU is a general-purpose processor, a jack-of-all-trades that can handle images, video, simulations, and large language model (LLM) training and inference. In particular, because the software ecosystem needed for AI service development has formed around GPUs, the GPU is indispensable hardware for building AI infrastructure. However, prices are high, and issues like power efficiency mean purchase and operating expense are excessively heavy. There has also been steady criticism that using GPUs for all simple AI services is "over-spec" infrastructure relative to the service.
Google started developing TPUs in the same context. Google also bases a substantial portion of its data center training models on GPUs. But as GPU prices have skyrocketed and data center investment expense have become excessive, TPUs emerged after much deliberation. Early TPUs were developed optimized for YouTube, Gmail and Google Search. In other words, they were created with the aim of developing the most expense-efficient chips suited to Google's services.
Lee Byung-hoon, a professor in the Department of Electrical Engineering at Pohang University of Science and Technology POSTECH, interpreted the competition between Nvidia and Google's TPUs not as a "monopoly collapse" but as a "division of roles in the process of market expansion." Lee said, "Until now, GPUs have handled almost all AI training, but as AI applications become more segmented, specialized chips like NPUs (Neural Processing Unit (NPU)) and TPUs tailored to specific tasks will inevitably become more efficient," adding, "Work that GPUs used to do alone will gradually be partially replaced by specialized chips." He said investment is shifting toward mixing GPUs, TPUs and NPUs to make AI infrastructure investment, which had been concentrated on GPUs, more expense-efficient.
◇ Broadcom emerges as a "broker" for expense savings
Broadcom is playing the role of a "middleman" that reduces investment costs for corporations like Google and Meta. In reality, Google and Meta lack the capability to design and develop AI Semiconductors and even handle foundry (contract chip manufacturing). They cannot be compared at all with companies like Nvidia that employ thousands of specialized semiconductor design personnel. Google's team operating TPU development is also known to be a small group of fewer than 100 people.
Lee said, "Broadcom is not a company that makes engines like NPUs itself to compete, but a company that provides an XPU platform that can host various vendors' NPUs—an outer 'car shell,' so to speak," adding, "When NPU companies like Rebellions and Furiosa bring the engine, Broadcom builds the chassis that can mount that engine." He continued, "Just as TSMC sticks to manufacturing without competing with its customers, Broadcom's stance is, 'I won't compete with you. Instead I'll build it cheaply and well.'"
In this trend, he said, "Ultimately there will be a matchup between an XPU alliance centered on Broadcom and Nvidia," while emphasizing this is not a zero-sum game. Lee said, "The AI Semiconductor market is growing from 100 to 1,000 right now, so even if 20%–30% is taken, Nvidia will not immediately collapse," adding, "The market will move from an era when Nvidia monopolized with GPUs to one where the Broadcom alliance and various NPU companies grow the market together."
This competition is also an opportunity for Korean corporations to expand their market. Kim Jeong-ho, a professor in the School of Electrical Engineering at KAIST, said, "Nvidia GPUs are highly versatile thanks to the CUDA ecosystem, so they are hard to replace immediately," adding, "TPUs take a different path by optimizing for specific models. This is the starting point for diversified competition as the AI market matures."
He added, "Whether it's GPUs or TPUs, high-bandwidth memory (HBM) is essential to run high-performance models," noting, "Only Samsung Electronics, SK hynix and Micron can supply HBM, so this is actually an opportunity for domestic corporations."