Custom AI Chips Challenge NVIDIA Dominance as Tech Giants Enter

Custom AI Chips Challenge NVIDIA Dominance as Tech Giants Enter

Google, Meta, Amazon Develop Specialized Semiconductors as AI Industry Prioritizes Power Efficiency

Like Windows in the PC era, an OS (operating system) leading the AI (artificial intelligence) age.’

This sentence defines NVIDIA’s GPU (graphics processing unit). GPUs, which process vast amounts of data simultaneously, have become essential in the AI era, and NVIDIA, which first introduced GPUs to the market, has risen as the AI industry’s dominant player. NVIDIA’s market share in the GPU-based AI chip sector reaches 90%. Priced at $30,000–$40,000 (approximately 40 million–50 million Korean won) per unit, GPUs are expensive, and even those with sufficient funds struggle to secure them. Thanks to this, NVIDIA has become the world’s most valuable company.

However, major tech firms are recently developing custom chips (ASICs) or diversifying semiconductor suppliers, signaling cracks in NVIDIA’s AI empire. ASICs, specialized for specific purposes, offer advantages over NVIDIA GPUs in power efficiency and cost. The shift in AI development paradigms—from “training,” which requires massive computing power, to “inference,” which demands relatively less—is also weakening NVIDIA’s monopoly. Unlike training, inference benefits more from power-efficient custom chips.

◇Custom AI Chips to Replace NVIDIA

When Google unveiled its AI model “Gemini 3,” its custom chip TPU (tensor processing unit) garnered as much attention as the model itself. TPUs are high-performance semiconductors Google developed around a decade ago to support its AI initiatives. Google designs the TPU’s basic architecture, while U.S. chip designer Broadcom and Taiwan’s MediaTek handle the physical design. These chips incorporate HBM (high-bandwidth memory) from SK Hynix, Samsung Electronics, and Micron. Taiwan Semiconductor Manufacturing Company (TSMC) then manufactures the final product. Specialized for AI, TPUs outperform GPUs in certain tasks and consume less power, reducing operational costs. AI startup Anthropic plans to use up to 1 million TPUs for its AI model development, and Meta is reportedly introducing Google’s TPUs into its data centers.

OpenAI, partnering with Broadcom, aims to produce its own chips by late next year. This is due to the massive chip demand for its “Stargate” project, which involves building data centers with a $500 billion investment. Meta is developing its own AI chip, “MTIA,” for AI development and services. Amazon Web Services (AWS) operates its own AI data centers equipped with 500,000 “Trainium2” chips, serving major clients like Anthropic and Databricks. Chinese companies, including Alibaba and Baidu, are also training AI models with self-developed semiconductors to reduce reliance on NVIDIA.

◇AI Ecosystem Likely to Shift

The move away from NVIDIA is driven largely by economic factors. Custom chips are cheaper and more power-efficient, offering operational advantages. According to Morgan Stanley, installing 24,000 of NVIDIA’s latest Blackwell GPUs costs $852 million (approximately 1.2 trillion Korean won), while the same scale of Google TPUs costs $99 million (approximately 145 billion Korean won). The emergence of affordable chips could alleviate concerns about an AI bubble fueled by excessive investments in AI infrastructure.

The paradigm shift from training to inference in AI is also influencing the trend. In the early stages of AI model development, “training” vast datasets required high-performance NVIDIA GPUs. However, the “inference” phase—delivering services based on trained AI—does not demand the same level of performance. This has increased demand for power-efficient, lightweight semiconductors like TPUs and NPUs (neural processing units). A source from the tech industry stated, “Many companies currently use both NVIDIA GPUs and custom chips from other firms, but the proportion of NVIDIA GPUs is expected to decline.” Still, evaluations overwhelmingly acknowledge that NVIDIA’s GPU performance remains superior to other custom chips.

The AI ecosystem, once centered on NVIDIA, is also poised for change. Currently, TSMC manufactures chips designed by NVIDIA, a structure that has solidified. As big tech firms collaborate with design companies to produce their own chips, firms like Broadcom are emerging as competitors.

☞CPU, GPU, TPU, NPU

The CPU (central processing unit), the computer’s basic “brain,” is like a master chef who skillfully prepares Korean, Japanese, and Chinese cuisines. However, handling everything alone takes time. In contrast, the GPU (graphics processing unit) is akin to 1,000 fast but less skilled part-timers working simultaneously to quickly produce specific dishes. This explains why GPUs are favored in the AI era, which requires repetitive data processing and learning. However, employing 1,000 workers increases costs (electricity) and space requirements. The TPU (tensor processing unit), developed by Google for AI, is comparable to a specialized machine that excels at one task (e.g., making dumplings). It requires fewer workers than a GPU but still needs a large facility. The NPU (neural processing unit), modeled after the human brain, is small, lightweight, and energy-efficient, making it ideal for smartphones and home appliances.

READ the latest news shaping the AI Chips market at AI Chips News

Custom AI Chips Challenge NVIDIA Dominance as Tech Giants Enter, source