Source: Wall Street News
Author: Zhang Yaqi.

With the rapid development of generative AI applications, whether AI ASIC can become a viable alternative to$NVIDIA (NVDA.US)$GPUs has been a global hot topic. On the 15th, Morgan Stanley released a Research Report titled "AI ASIC 2.0: Potential Winners," suggesting that ASICs, with their targeted optimization and cost advantages, are expected to gradually capture more market share from NVIDIA's GPUs.

Morgan Stanley expects the market size of AI ASIC to grow from 12 billion dollars in 2024 to 30 billion dollars in 2027, with a compound annual growth rate of 34%.

In this situation, NVIDIA, with its advantages in training large language models, will continue to dominate.$Broadcom (AVGO.US)$Alchip and $Socionext (6526.JP)$are considered Bullish. $Cadence Design Systems (CDNS.US)$ 、$Taiwan Semiconductor (TSM.US)$And its supply chain partners (ASE, KYEC, etc.) will benefit from the rapid growth in ASIC design and manufacturing.

Morgan Stanley stated that the rise of ASIC does not mean the decline of GPUs. On the contrary, these two technologies will coexist for a long time, providing optimal solutions for different demand scenarios.

Will ASIC become a strong competitor to NVIDIA?

With the rapid development of generative AI applications, global AI computing demand is experiencing explosive growth. Reports predict that under the basic scenario, the market size of Cloud Computing Service AI Semiconductors will reach 238 billion USD by 2027, and under optimistic scenarios, it may even reach 405 billion USD.

In this field, ASICs, with their targeted optimization and cost advantages, are expected to gradually seize more market share from NVIDIA GPUs.

Morgan Stanley predicts that the market size of AI ASICs will grow from 12 billion USD in 2024 to 30 billion USD in 2027, with a compound annual growth rate of 34%.

Although NVIDIA's AI GPU performance is outstanding, Morgan Stanley believes that Cloud Computing Service providers such as $Alphabet-C (GOOG.US)$ 、$Amazon (AMZN.US)$and$Microsoft (MSFT.US)$The company is still actively promoting ASIC design. The main driving forces behind this are two.

First, it is to optimize internal workloads. By developing custom chips, CSPs can more efficiently meet their internal AI inference and training needs.

Secondly, it is about better cost-effectiveness. The report points out that although NVIDIA's GPUs have strong computing performance, their hardware prices are high, especially during AI training processes. In contrast, ASICs have a lower unit cost, particularly after large-scale adoption.

For example, Amazon's Trainium chip is about 30% to 40% cheaper than NVIDIA's H100 GPU for inference tasks. Google is also continuously optimizing its TPU series, with the latest TPU v6 achieving a 67% improvement in energy efficiency over the previous generation.

Morgan Stanley emphasizes that although NVIDIA's GPUs remain the first choice for most CSPs, in the coming years, as ASIC design becomes increasingly mature, these cloud giants may gain greater bargaining power in procurement negotiations through self-developed ASICs.

Winners and Losers: Who Will Dominate the Future Market?

Morgan Stanley outlined the Global ASIC supply chain in the report and identified six potential winners:

• AI GPU: NVIDIA will continue to dominate the market, particularly in large-scale language model training, where its solutions remain the best choice.
  

• ASIC Suppliers: $Broadcom (AVGO.US)$ , Alchip (World Leading Electronics) and $Socionext (6526.JP)$ are seen as potential stocks in the ASIC market. Among them, Alchip is expected to significantly increase its market share by 2026 due to its deep collaboration with Amazon Web Services.
  

• Electronic Design Automation Tools: $Cadence Design Systems (CDNS.US)$ Expected to achieve structural growth.
  

• Foundries: $Taiwan Semiconductor (TSM.US)$ And its supply chain partners (such as ASE, KYEC, etc.) will benefit from the rapid growth of ASIC design and manufacturing.
  

• Testing services:$Advantest (6857.JP)$It is a leader in AI Chip testing, and its focus on testing AI devices will bring significant growth.
  

• HBM: Samsung Electronics is the leading player in the non-NVIDIA HBM market and will benefit from the growth in ASIC demand.

In contrast, some traditional chip companies and foundries may face challenges. For example, AMD may lose more market share due to its failure to close the gap with NVIDIA in the AI GPU field. Additionally, foundries like UMC that lack support for advanced process nodes may also struggle to gain a foothold in the high-end AI chip market.

TCO analysis: Is ASIC really cost-effective?

Morgan Stanley compared the cost-effectiveness of ASIC and GPU in AI training and inference tasks using a TCO model. The results show that although NVIDIA's GPU has a significant performance advantage, the initial cost of ASIC is lower, making it particularly suitable for budget-constrained cloud service providers.

For example, under the same budget, AWS's Trainium 2 can complete inference tasks faster than NVIDIA's H100 GPU, with a cost-performance increase of 30-40%. Trainium 3 is scheduled to be launched in the second half of 2025, with a 2-fold increase in computing performance and a 40% improvement in energy efficiency.

However, the report also points out that NVIDIA remains competitive in TCO calculations due to its more mature system integration capabilities and stronger software ecosystem, especially in scenarios requiring flexibility to adapt to different AI tasks.

The Research Reports mention that the potential rise of quantum computing may impact the demand for AI Semiconductors, but currently, the applicability of quantum computing in AI inference is low, and it is unlikely to replace ASIC and GPU in the short term. Moreover, retired GPUs may also pose a threat to the ASIC market. Some cloud service providers may opt to reduce costs by using retired GPUs rather than investing in expensive ASIC.

Morgan Stanley concluded that the rise of ASIC does not mean the decline of GPU. On the contrary, both technologies will coexist in the long term, providing the best solutions for different demand scenarios.

In the future AI market, ASIC will compete for more share due to cost and energy efficiency advantages, while NVIDIA will continue to solidify its market position through its CSI Leading Technology Index.

Editor/rice