we examine the competitiveness of Google's TPUs and Nvidia GPUs, which Google and Broadcom have spent a decade developing. we also analyze the performance innovations brought about by Zemini 3.0, changes in the global AI accelerator market, and the impact on domestic semiconductors, including the expanding demand for Samsung and SK Hynix HBMs.

google's recently unveiled Gemini 3.0 is said to have surpassed ChatGPT in terms of AI performance, and attention has been focused on Google's TPUs used for training. the rise of TPUs has tipped the balance in the AI infrastructure market, which has been dominated by Nvidia. In fact, Google and Broadcom stocks have soared on the back of Gemini 3.0, which utilizes Google TPUs, while Nvidia's stock price has declined slightly. Let's take a look at what Google TPUs mean and what's to come.

GPU vs TPU: Concepts and differences

Even if you don't know much about AI and computers, you've probably heard the term "Graphics Processing Unit" (GPU) at least once. Originally developed for game graphics and video editing, GPUs are strong at parallel computation, which makes them efficient for AI training. nvidia holds more than 95% of the global AI accelerator market thanks to its GPU performance. Tensor Processing Units (TPUs), on the other hand, are AI-specific chips (ASICs) that Google and semiconductor designer Broadcom have been working on for more than a decade. As the name suggests, they are optimized for tensor computation and have an excellent power-to-performance ratio for deep learning computations.

  • GPU: A general-purpose semiconductor developed for gaming and graphics. they are highly parallel but power-hungry, leading to global shortages.

  • TPU: A chip dedicated to AI developed by Google and Broadcom. specialized for deep learning computations, it is energy efficient and consumes less power per performance.

to use a simple analogy, it's like the difference between a versatile SUV and an AI-specific sports car GPUs are the SUV that can do a lot of different things, while TPUs are the sports car that runs on a dedicated, optimized circuit.

the rise of Google TPUs

TPUs were unveiled by Google in 2016 and became famous when they were used in AlphaGo's Lee Sedol match. originally developed to speed up its search service, its strength in AI computation has led to a wide range of uses. In the past, Google has only offered TPUs in select clouds, but this year it announced that it would begin selling them externally, due to the performance of its seventh-generation TPUs (codenamed "Ironwood"). In fact, AI startup Enthropic has announced that it will utilize up to 1 million TPUs in Google's cloud, meaning that TPUs are now a technology that anyone can use.

google TPUs also played a crucial role in the development of Gemini 3.0. we didn't use Nvidia GPUs for training Zemyni 3.0 at all, but instead used Google TPUs to perform the computations. This resulted in a higher power-to-performance ratio, allowing us to do more computations with the same amount of power, which significantly improved the performance of the model. Of course, in the service phase, we are using GPUs and TPUs in parallel to ensure both compatibility and efficiency.

The AI chip race: nvidia vs Google

up until now, AI semiconductors have been the domain of Nvidia. nvidia's GPUs are so powerful that almost all AI data centers use them, and countries including South Korea have signed large-scale GPU supply contracts, such as a deal to supply 260,000 Nvidia GPUs to South Korean companies by 2030. but the rise of Google TPUs is a game changer.

the market is already talking about an "anti-Nvidia" movement. meta is considering TPUs for its data center, which is scheduled to go live in 2027, and Amazon, MS, and Apple are developing their own AI chips. while the Google-Broadcom alliance is moving underwater, Nvidia has emphasized its dominance in its official announcement, even while praising Google. While the GPU throne may not be toppled in the short term, the mere fact that a new competitor has emerged is shaking up the market.

impact on the domestic semiconductor industry

this AI semiconductor tectonic shift is also an opportunity for the domestic semiconductor industry. As TPUs require high-bandwidth memory (HBM) for large-scale computations, memory companies such as Samsung Electronics and SK Hynix will benefit. in fact, Google and Broadcom's stock prices have risen significantly since the announcement of Gemini 3.0, while Nvidia's has fallen slightly. memory-related stocks such as Samsung Electronics and SK Hynix have also risen, signaling a shift in the semiconductor market sentiment.

conclusion and outlook

the arrival of Google's TPUs marks a new inflection point in the AI infrastructure market. with Google's TPUs challenging Nvidia's dominance in GPUs, the AI chip race is set to intensify. in the short term, Nvidia's dominance will continue, but in the mid- to long-term, the AI semiconductor war is likely to be a two-way street. in Korea, this could create another opportunity for memory semiconductor companies. How do you think the AI infrastructure market will change in the future?