Aug.
Click the blue text to follow us
2025.08
With a positive outlook on the rapid growth of the AI training market, not only is the shipment of AI GPUs from NVIDIA or AMD expected to grow rapidly, but the training ASIC market is also expanding quickly under the push of the four major American CSPs, becoming a highlight for the further growth of AI servers.
As a result, institutional investors are more optimistic about the ASIC market for this year and next, expecting a CAGR of up to 70% for ASIC chip shipments from 2024 to 2026. Especially from this year to next, more ASIC cabinet designs will be launched.
Research institutions estimate that this year, the shipment of ASICs for AI training will grow by over 200% year-on-year, reaching up to 5 million units. In the overall AI server category, the shipment ratio of AI GPUs to ASICs will shift from 62:38 to 60:40 by 2026 as the proportion of ASICs increases.
DIGITIMES believes that self-developed ASIC accelerators can not only reduce energy consumption and help CSPs control costs and supply chains for AI infrastructure but also reduce their dependence on NVIDIA or international political factors. Through technological innovation and establishing market barriers, they can further ensure the competitiveness of CSPs’ AI application products and services.
Among the four major American CSPs, AWS is expected to launch the Teton 2 cabinet using Trainium 2/2.5 in the second half of the year, boosting its ASIC chip shipment growth by over 40%, while also driving the shipment momentum of its main assembly factory, Wistron, and its main supplier, Wistron.
As for Meta, it plans to gradually start mass production of the Minerva cabinet using its own MTIA chip from the second half of the year, with major assembly factories including Celestica and Taiwanese company Quanta benefiting.
What is ASIC? Why does it threaten NVIDIA?
What is an ASIC chip? Simply put, an ASIC chip is a customized chip for specific applications, while NVIDIA’s GPU is a general-purpose processor suitable for a wider range of functions.
Broadcom’s CEO, Hock Tan, predicts that by 2027, the sales of ASIC-type chips led by Broadcom will reach an astonishing $60 billion to $90 billion! This means that ASIC chips will occupy an increasingly important position in the AI chip market.
The rapid growth of ASIC chips produced by companies like Broadcom raises the question: does this mean NVIDIA’s GPUs will lose market share? The fluctuations in their stock prices show that many investors are concerned. What do experts and foreign investors think?
First, Dan Gallagher, a columnist for the Wall Street Journal, believes that the AI chip market is not a zero-sum game; ASICs and GPUs can coexist and share the prosperity of the AI industry.
Gallagher analyzes that while ASIC chips have advantages in specific application scenarios, GPUs still play an indispensable role in the AI field. Moreover, the versatility of GPUs allows them to handle various AI tasks, which is something ASICs lack. Therefore, the future AI chip market will present a diversified pattern where ASICs and GPUs shine in their respective fields.
Dot dot dotFollow and star Eefocus
Scan to reply“Chip” keyword for industry groups and to receive reports
Morgan Stanley: ASIC and GPU Will Coexist Long-Term
Additionally, Morgan Stanley recently released a report on the competitive landscape between ASICs and GPUs titled “AI ASIC 2.0: Potential Winners.” The report points out that while NVIDIA’s GPUs have a clear performance advantage, the lower initial cost of ASICs makes them more attractive to cloud computing service providers with limited budgets. Morgan Stanley expects the ASIC chip market for AI to grow from $12 billion in 2024 to $30 billion in 2027, with a compound annual growth rate of 34%.
However, Morgan Stanley also emphasizes that the rise of ASICs does not mean the decline of GPUs. On the contrary, these two technologies will coexist for a long time, providing optimal solutions for different needs within the AI industry. They remain optimistic that NVIDIA will continue to dominate the AI chip market, especially in large-scale language model training, where its solutions remain the best choice.
The report also mentions that Amazon’s Trainium chip costs about 30% to 40% less than NVIDIA’s H100 GPU for inference tasks. Google is also continuously optimizing its TPU series chips, with the latest TPU v6 improving energy efficiency by 67% compared to the previous generation. These data show that ASICs can save costs and improve efficiency for certain tasks.
Jensen Huang’s View: ASIC Has Its Value
So, how does NVIDIA’s CEO Jensen Huang view the rise of ASICs? Back in October this year, Huang clearly discussed the pros and cons of ASICs in a podcast hosted by Silicon Valley venture capitalist Brad Gerstner, while also explaining why it does not significantly impact NVIDIA’s strategy.
Huang admitted that ASICs do have their value. He said, “If you are very clear about what you want to use the chip for and that will not change in the future, you can build an ASIC.” However, he also emphasized the limitations of ASICs, stating that they lack flexibility. “A perfect ASIC performs excellently in certain tasks but poorly in others. Once the AI workload changes, it becomes useless, like creating a piece of software with fixed functions,” he stressed.
Huang believes that NVIDIA’s GPUs have the advantage of flexibility and versatility. He emphasized, “AI models change very quickly, and the effectiveness of GPUs comes from their ability to adapt to these changes. They are general-purpose devices that can perform matrix operations and can be programmed.” He further emphasized that GPUs can switch from one AI application to another, from one research area to another, and this capability is very valuable.
From Huang’s perspective, NVIDIA’s strategy is not to confront ASICs head-on but to leverage the versatility and flexibility of GPUs to continue innovating and expanding their application range in the AI field. He understands that AI is a constantly evolving field, and being able to respond flexibly to changes is crucial.
Huang: NVIDIA’s Strategy is an Ecosystem, Not a Single Product
Moreover, Huang has repeatedly emphasized the importance of NVIDIA’s “platform strategy.” He believes that NVIDIA’s success lies not only in hardware but also in its complete software ecosystem, which is key to NVIDIA’s competitiveness. Therefore, the rise of a new chip hardware technology is not enough to shake NVIDIA’s entire ecosystem.
In summary, the rise of ASICs does put some pressure on GPUs; however, foreign investors and experts believe it is not a zero-sum game. In the short term, it is more likely that both will grow together in the AI industry. It is recommended to continue monitoring the further developments of NVIDIA and Broadcom to grasp the overall situation.
END
Note: The cover image of this article is from freepik, created by the author, and publicly available media, all authorized.
Industry Community
