Remember to star this public account ⭐️ to receive timely updates.
“Just as the Windows operating system led the personal computer era, it is now leading the era of artificial intelligence.”
This statement defines NVIDIA’s GPUs, or Graphics Processing Units. GPUs can process massive amounts of data simultaneously and have become indispensable products in the age of artificial intelligence. NVIDIA, the first to bring GPUs to market, has risen to become a leading company in the AI field, holding a staggering 90% share of the GPU-based AI chip market. Each GPU can cost between $30,000 to $40,000 (approximately 40 million to 50 million Korean won), making them prohibitively expensive even for those with substantial financial resources. As a result, NVIDIA has become the world’s most valuable company.
However, major tech giants have recently begun to develop their own Application-Specific Integrated Circuits (ASICs) or expand their semiconductor suppliers, signaling that NVIDIA’s dominance in the AI field is being challenged. ASICs are chips designed for specific purposes, offering better energy efficiency and cost advantages over NVIDIA’s GPUs. The shift in AI development from resource-intensive “training” to relatively resource-light “inference” is also undermining NVIDIA’s monopoly. Unlike training, inference is more suited to using energy-efficient dedicated chips.
Custom AI Chips Replacing NVIDIA
When Google released its AI model “Gemini 3,” its custom chip TPU (Tensor Processing Unit) also garnered significant attention. The TPU is a high-performance semiconductor developed by Google about a decade ago to support its AI development. Google is responsible for the TPU’s infrastructure design, while U.S. chip design company Broadcom and Taiwan’s MediaTek handle the chip’s physical design. The TPU integrates HBM (High Bandwidth Memory) from SK Hynix, Samsung Electronics, and Micron. The final product is manufactured by Taiwan’s TSMC. Since the TPU is specifically designed for AI, it outperforms GPUs in certain tasks while consuming less power, thereby reducing operational costs. AI startup Anthropic plans to use up to 1 million TPUs to develop its AI models, and reports indicate that Meta is also incorporating Google’s TPUs into its data centers.
OpenAI plans to collaborate with Broadcom to produce its own chips by the end of next year. This is due to the need for a large number of chips for the $500 billion “Star Gate” project, which involves building data centers. Meta has developed its own AI chip, “MTIA,” for AI development and services. Amazon Web Services (AWS) operates AI data centers equipped with 500,000 “Trainium2” chips, with major clients including Anthropic and Databricks. Chinese companies like Alibaba and Baidu are also using self-developed semiconductors to train AI models, aiming to reduce their reliance on NVIDIA.
The AI Ecosystem May Also Change
The shift away from NVIDIA chips is primarily driven by economic reasons. Custom chips are cheaper and more energy-efficient than GPUs, making them more advantageous in operations. According to Morgan Stanley, installing 24,000 of NVIDIA’s latest Blackwell GPUs would cost $852 million (approximately 1.2 trillion Korean won), while the cost of installing an equivalent number of Google’s TPUs is only $99 million (approximately 145 billion Korean won). The emergence of cheaper chips is expected to alleviate concerns about an AI bubble caused by recent over-investment in AI infrastructure.
The transition from training to inference in AI paradigms has also had an impact. In the early stages of AI model creation, “training” massive datasets is crucial, requiring a large number of NVIDIA’s high-performance GPUs. However, during the “inference” phase, which is based on already created AI, the performance level required is far less than that of GPUs. Therefore, semiconductor devices like TPUs and NPUs (Neural Processing Units), which have high energy efficiency and lightweight characteristics, are receiving increasing attention. A tech industry insider stated, “Many companies are currently using both NVIDIA GPUs and custom chips from other companies, but the proportion of NVIDIA GPUs is expected to gradually decline.” However, NVIDIA GPUs still dominate in performance evaluations compared to other custom chips.
The AI ecosystem centered around NVIDIA is also expected to change. Currently, TSMC’s foundry for NVIDIA chips is well-established. As large tech companies collaborate with chip design firms to produce their own chips, companies like Broadcom are emerging as new competitors.
CPU, GPU, TPU, NPU
The Central Processing Unit (CPU) is the basic brain of a computer, akin to a skilled chef capable of cooking various cuisines—Korean, Japanese, Chinese, and so on. However, since it handles all tasks alone, it tends to be time-consuming. In contrast, the Graphics Processing Unit (GPU), while slightly less skilled, can quickly prepare specific dishes, functioning like 1,000 “part-time workers” operating efficiently at the same time. This is why GPUs are highly regarded in the age of AI, as AI requires simple repetitive calculations and the ability to learn from vast amounts of data. Since GPUs require 1,000 “workers,” they are costly (in terms of power consumption) and take up significant space. The Tensor Processing Unit (TPU) is a high-performance semiconductor developed by Google for AI. Unlike CPUs and GPUs, TPUs are like specialized machines that excel at a specific task (like making dumplings). While TPUs do not require as many “part-time workers” as GPUs, they still need a large factory to operate. The Neural Processing Unit (NPU) is a semiconductor device that simulates the human brain. It is small, lightweight, and energy-efficient, making it highly effective for use in smartphones and home appliances.
(Source: Content compiled from Chosun)
*Disclaimer: This article is original to the author. The content reflects the author’s personal views, and Semiconductor Industry Observer reproduces it solely to convey a different perspective, not representing its endorsement or support of the views expressed. If there are any objections, please contact Semiconductor Industry Observer.
END
This is the 4240th issue shared by “Semiconductor Industry Observer”. Thank you for your attention.
Recommended Reading
★A Chip That Changed the World
★U.S. Secretary of Commerce: Huawei’s Chips Are Not That Advanced
★“ASML’s New Lithography Machine, Too Expensive!”
★Quietly Rising New Competitors to NVIDIA
★Chip Prices Plummet, All Blame Trump
★New Solutions Announced to Replace EUV Lithography
★Semiconductor Equipment Giants, Salaries Surge by 40%
★Foreign Media: The U.S. Will Propose Banning Software and Hardware Made in China for Cars
Star this account ⭐️ to receive timely updates, small account to prevent loss
Seeking likes
Seeking shares
Seeking recommendations
