Big Manufacturers Are Ditching GPUs

Nvidia's AI accelerator H200 is in short supply, highlighting a severe supply shortage in the global AI infrastructure market. Orders for graphics processing units (GPUs) are expected to be as high as 2 million this year, but only about 700,000 GPUs are currently available. If GPUs are not supplied on time, data center construction progress may be affected, which will prompt big tech companies to seek the use of application-specific integrated circuits (ASICs) as an alternative.

On the 5th, market research firm TrendForce said that it is expected that the demand for self-developed ASICs by cloud service companies in the artificial intelligence server market this year will exceed the growth rate of demand for GPUs. Cloud company ASICs are expected to grow at 44.6%, up from GPUs' 16.1%. This is interpreted as a structural signal that limited GPU supply is accelerating the adoption of ASICs.

Despite the versatility, excellent performance, and rapid deployment of AI infrastructure, the industry expects its supply chain risks to peak this year. GPU production processes, packaging, and high-bandwidth memory (HBM) are interconnected, so bottlenecks in any one link can lead to overall supply disruptions. It is particularly worth noting that the market demand for Nvidia's artificial intelligence semiconductors has surged, but the production capacity of Nvidia and TSMC, the world's largest wafer foundry, cannot meet even 50% of the demand.

TSMC is expanding advanced packaging process lines that are essential for producing AI accelerators, such as CoWoS (chip packaging on a wafer substrate). However, as capacity expansion investment takes time, there will inevitably be a gap between the rapid growth of orders and actual shipments this year.

Against this backdrop, ASIC chips - led by Google's Tensor Processing Unit (TPU) last year - are gaining traction. ASIC chips have higher initial development costs because they are designed for specific AI workloads, but they offer advantages in terms of energy efficiency, performance, and total cost of ownership (TCO) in the long run. Google has already handled a significant portion of AI training and inference tasks internally through TPUs, and Amazon is also reshaping its cloud cost structure with specialized chips such as Trainium and Inferentia.

ASIC institutional users of the AI accelerator market are expected to maintain a compound annual growth rate (CAGR) of about 28% through 2030, according to market research firm Mordor Intelligence in a semiconductor market report. Another market research firm, Credence Research, also predicts that the generative artificial intelligence ASIC market will grow from about $24.9 billion in 2024 to about $186.7 billion in 2032, with an average annual growth rate of about 28.6%.

The industry generally believes that this year will be a key turning point in the growth of the ASIC market. An Amazon Web Services (AWS) executive said: "This supply shortage is a short-term phenomenon, but it will have a long-term impact on decision-making. He added: "From the perspective of large technology companies, GPUs are no longer a stable 'basic commodity' but have become a strategic asset that is vulnerable to external factors." Therefore, reducing GPU dependence and increasing the proportion of ASICs in new data center investment plans are gradually becoming a viable alternative.