AI Chip Battle: NVIDIA vs. AMD, TSMC Dominance

In the current landscape of intense competition within the AI chip market, both NVIDIA and AMD have been actively vying for market share. The AMD MI300A series has successfully commenced mass production this quarter, garnering favorable adoption from clients. In response, NVIDIA is gearing up to launch an upgraded version of its AI chip to maintain competitiveness. TSMC, securing orders from both NVIDIA and AMD, emerges as a major beneficiary.

Industry projections indicate that the combined total shipments of AI chips from NVIDIA and AMD for the year will surpass one million units, potentially exceeding 1.5 million units. This surge is significantly bolstering TSMC's order momentum for its 5-nanometer and 3-nanometer advanced processes.

While TSMC traditionally refrains from public commentary on client dynamics and order trends, President Wei Zhejia acknowledged in a December forum that uncertainties persist in 2024 due to external factors such as high inflation and escalating costs. Nonetheless, the rapid development of AI applications is anticipated to yield opportunities throughout the year.

Observers note that the global AI boom, initiated in 2023, remains a focal point in 2024. A noteworthy shift from the previous year is the introduction of AMD's MI300 series, challenging NVIDIA's historical dominance in the high-performance computing (HPC) sector.

The AMD MI300A series, utilizing TSMC's 5-nanometer process for CPUs and GPUs and the 6-nanometer process for IO chips, incorporates advanced packaging technologies like TSMC's System Integration on Chip (SoIC) and CoWoS. Simultaneously, the standalone MI300X, without integrated CPU cores, is shipping. In comparison to NVIDIA's GH200 and H200 chips, AMD's new AI chips exhibit better-than-expected performance, competitive pricing, and an impressive cost-performance ratio, attracting adoption from system manufacturers.

Major cloud service providers such as Microsoft and Meta have been placing orders for AMD's MI300 series products for the past one to two years, urging Original Design Manufacturers (ODMs) to exclusively design AI servers using the MI300 series to diversify risks and reduce costs. Industry estimates indicate a demand of at least 400,000 units for the AMD MI300 series chips this year, with the potential to reach 600,000 units if TSMC can provide additional production capacity.

In response to AMD's competitive position, NVIDIA is proactively upgrading its product line. Amid persistent shortages of chips like H200 and GH200, NVIDIA is expected to introduce new products such as B100 and GB200, utilizing TSMC's 3-nanometer process, by the end of the year.

Market analysts are optimistic, projecting that NVIDIA's AI chip shipments will commence at least at one million units this year, representing a significant growth multiplier compared to 2023. With the initiation of AMD's MI300 series production, the total AI high-performance computing chip shipments from NVIDIA and AMD to TSMC are expected to exceed one million units, potentially reaching 1.5 million units. This, in turn, is anticipated to boost TSMC's utilization of advanced processes like 3-nanometer and 5-nanometer, driving the company's performance to rebound quarterly and return to a growth trajectory by the end of the year.