ASIC chip craze: shipment volume may exceed Nvidia next year

The report shows that Nvidia currently holds over 80% of the market value of AI servers, while ASIC AI servers only account for 8-11%.

However, from the perspective of shipment volume, the situation is changing. By 2025, Google's TPU shipments are expected to reach 150000 to 2 million units, Amazon AWS Trainium 2 ASIC will be around 140000 to 1.5 million units, and Nvidia's AI GPU supply will exceed 5-6 million units.

According to a supply chain survey, the total shipment volume of AI TPU/ASIC from Google and AWS has reached 40-60% of Nvidia's AI GPU shipment volume.

As Meta begins large-scale deployment of its self-developed ASIC solutions in 2026, Microsoft will begin large-scale deployment in 2027, and it is expected that the total ASIC shipments will surpass Nvidia GPU shipments at some point in 2026.

Meta's MTIA Ambition: Exceeding Nvidia's Rubin Specification

Meta's MTIA project is one of the most anticipated cases in the current ASIC wave.

According to supply chain data, Meta will launch its first ASIC chip MTIA T-V1 in the fourth quarter of 2025. It is designed by Broadcom, with a complex motherboard architecture (36 layer high specification PCB), and adopts a hybrid technology of liquid cooling and air cooling. The manufacturers responsible for assembly include Celestica and Quanta.

By mid-2026, MTIA T-V1.5 will undergo further upgrades, doubling its chip area and surpassing the specifications of Nvidia's next-generation GPU Rubin. Its computing density will directly approach Nvidia's GB200 system. The MTIA T-V2 in 2027 may bring larger scale CoWoS packaging and high-power (170KW) rack design.

However, Meta's ambition is not without risks.

The report points out that according to supply chain estimates, Meta's goal is to achieve 1 million to 1.5 million ASIC shipments by the end of 2025 to 2026, but currently CoWoS wafer allocation can only support 300000 to 400000 wafers, which may be postponed due to capacity bottlenecks. Not to mention the technical challenges of large-sized CoWoS packaging and the time required for system debugging (similar systems from Nvidia require 6 to 9 months of debugging time).

If Meta, AWS, and other CSPs accelerate deployment simultaneously, the high-end materials and components required for AI servers may face shortages, further driving up costs.

Nvidia's technological moat remains safe

Nvidia will definitely not sit idly by.

At the 2025 COMPUTEX conference, Nvidia launched NVLink Fusion technology, opening up its proprietary interconnect protocol that allows third-party CPUs or xPUs to seamlessly connect with its own AI GPU. This semi custom architecture may seem like a compromise, but it is Nvidia's strategy to consolidate its market share in cloud AI computing.

The report points out that data shows that Nvidia still leads in chip computing density (computing power per unit area) and interconnect technology (NVLink), which makes it difficult for ASICs to catch up with its performance in the short term. In addition, Nvidia's CUDA ecosystem remains the top choice for enterprise AI solutions, which is a barrier that ASICs find difficult to replicate.

For investors, Nvidia's technological moat is still deep, but whether its high profit model will be forced to adjust under the cost pressure of other CSPs (cloud service providers) is worth continuing to monitor.

in summary

According to a report by Nomura Securities:

one

Meta's MTIA Plan:

Meta expects to launch 1 million to 1.5 million AI ASIC chips (MTIA T-V1) by the end of 2025 to 2026, designed by Broadcom, using high specification 36 layer PCBs and hybrid cooling technology.

By mid-2026, the MTIA T-V1.5 chip area will double and the computing density will approach Nvidia's GB200 system.

In 2027, MTIA T-V2 will adopt a larger scale CoWoS package and high-power (170KW) rack design.

two

The Rise of ASIC Market:

Currently, Nvidia accounts for over 80% of the AI server market, while ASIC only accounts for 8-11%.

The estimated shipment volume of Google TPU in 2025 is 1.5 million to 2 million units, and AWS Trainium 2 is about 1.4 million to 1.5 million units, totaling 40-60% of Nvidia GPU shipments (5 million to 6 million units).

In 2026, with the large-scale deployment of Meta and Microsoft, ASIC shipments are expected to surpass Nvidia GPUs.

three

Challenges and Risks:

Meta's ASIC plan may be delayed due to CoWoS wafer capacity limitations (supporting only 300000 to 400000 wafers).

The challenges of large-size CoWoS packaging technology and system debugging (requiring 6-9 months) increase uncertainty.

If cloud service providers such as Meta and AWS accelerate deployment, there may be a shortage of advanced materials and components, which will drive up costs.

four

Advantages of Nvidia:

Nvidia strengthens its market position by opening up internet protocols through NVLink Fusion technology.

Its chip computing density, NVLink interconnect technology, and CUDA ecosystem are still leading, and ASIC is difficult to surpass in the short term.

Nvidia's technological moat is solid, but its high profit model may face cost pressures from cloud service providers.

Can ASIC shake Nvidia?

Many people may wonder why they are so interested in ASICs? Because NVIDIA is the emperor of this era, and ASIC is challenging the emperor.

However, I can deny that designing and manufacturing ASICs takes 2 to 3 years, and no matter how good the specifications are at the time of design, they will become outdated after two years. He also mentioned that due to Blackwell's delays, many ASIC projects have emerged, but with Blackwell's supply easing, a large part of these projects have been suspended.

This statement actually makes a lot of sense. If NVIDIA's roadmap is true, then a large number of NVIDIA chips are released every year, and two-year cycles of ASICs seem difficult to beat them.

However, this may not be the complete answer. For example, how to prevent the delay experienced by Blackwell from recurring on Rubin?

Furthermore, no matter how low NVIDIA's total cost of ownership (TCO) is, will large tech companies continue to accept NVIDIA's high profit margins? I still find it difficult to shake off the feeling of dissatisfaction with NVIDIA's gross profit margin among large technology companies.

Nevertheless, I am increasingly inclined to favor NVIDIA's viewpoint.

one

Dominating the AI chip market

NVIDIA is the largest consumer of HBM, purchasing the most HBM from SK Hynix and Micron. Based on my personal estimation, approximately 70% of SK Hynix's HBM was sold to NVIDIA. In addition, due to its strong relationship with TSMC, NVIDIA has the largest CoWoS capacity allocation.

What does this mean? This means they can sell AI chips faster and more than anyone else.

I fundamentally believe that AI is a race against time. Large tech companies are trying to develop AI that is faster and more powerful than their competitors. This is a chicken game, and if you stop, you lose.

On the contrary, this means that if large tech companies attempt to use ASICs for training, have to customize software for them, and work with limited supply and performance of ASICs, then they are losers compared to NVIDIA's situation.

In other words, this is a battle that NVIDIA won from the beginning, when the foundation of AI technology began to be built on NVIDIA's chips and CUDA software. (Of course, this is based on the assumption that NVIDIA is also continuing to develop.).

two

The Reality of Sovereign Artificial Intelligence

The same goes for Sovereign AI.

I actually think ASIC is more suitable for Sovereign AI.

From the perspective of each country, I have always believed that true sovereign AI means cultivating local fabless companies in each country, developing their own AI chips, and creating independent AI, rather than relying on American companies like NVIDIA.

However, even in places such as South Korea and Taiwan, China, China, where chip design technology is rich, cultivating their own AI chips is an extremely difficult and difficult policy.

Especially in countries with fixed terms, people strongly tend to formulate policies based on short-term outcomes.

This means that few countries will embark on the path of truly sovereign artificial intelligence, which involves long development times and uncertain outcomes for proprietary chips. On the contrary, most people will purchase NVIDIA chips and manufacture counterfeit ChatGPT products.

Unless a country, like China or Russia, restricts the sale of NVIDIA chips in the United States, most countries will eventually purchase NVIDIA chips. That's because it's the easiest way to see results in the shortest amount of time.

In this sense, sovereign AI may be seen as "NVIDIA's sovereign policy" - replicating American AI running on American chips.