Jun.
Click the blue text to follow us
2025.06
Source: Content from Nomura.
According to reports, following in Google’s footsteps, Meta is fully entering the ASIC (Application-Specific Integrated Circuit) race. According to the latest research report by Nomura Securities analyst Anne Lee and her team, Meta’s ambitions in the AI server space are rapidly escalating. Their proprietary ASIC server project, MTIA, is expected to achieve significant breakthroughs by 2026, potentially challenging Nvidia’s long-standing market dominance.
The report cites the latest supply chain information indicating that Meta plans to launch millions of high-performance AI ASIC chips (1 to 1.5 million) between the end of 2025 and 2026. Meanwhile, cloud service giants like Google and AWS are also accelerating the deployment of their self-developed ASICs. This indicates a quiet shift in the competitive landscape of the AI server market.
The ASIC Chip Boom: Shipments May Exceed Nvidia Next Year
The report shows that Nvidia currently holds over 80% of the AI server market value, while ASIC AI servers account for only 8-11%.
However, from a shipment perspective, the situation is changing. By 2025, Google’s TPU shipments are expected to reach 1.5 to 2 million units, while Amazon’s AWS Trainium 2 ASIC is estimated at around 1.4 to 1.5 million units, and Nvidia’s AI GPU supply will exceed 5 to 6 million units.
Supply chain surveys indicate that the total shipments of Google and AWS’s AI TPU/ASIC have already reached 40-60% of Nvidia’s AI GPU shipments.
As Meta begins large-scale deployment of its self-developed ASIC solutions in 2026, Microsoft is expected to start large-scale deployment in 2027, with total ASIC shipments likely to surpass Nvidia’s GPU shipments at some point in 2026.
Meta’s MTIA Ambitions: Surpassing Nvidia’s Rubin Specifications
Meta’s MTIA project is one of the most anticipated cases in the current ASIC wave.
Supply chain data shows that Meta plans to launch its first ASIC chip, MTIA T-V1, in the fourth quarter of 2025. It is designed by Broadcom and features a complex motherboard architecture (36-layer high-spec PCB) and employs a hybrid cooling technology. Manufacturers responsible for assembly include Celestica and Quanta.
By mid-2026, the MTIA T-V1.5 will undergo further upgrades, with the chip area doubling, exceeding Nvidia’s next-generation GPU Rubin specifications, and its computational density will be directly comparable to Nvidia’s GB200 system. The MTIA T-V2 in 2027 may introduce larger-scale CoWoS packaging and high-power (170KW) rack designs.
However, Meta’s ambitions are not without risks.
Follow and star with non-networkeefocusScan to reply with the keyword “GPU, IoT, Automotive Electronics“ for industry groups and to receive reports
The report notes that, according to supply chain estimates, Meta aims to achieve shipments of 1 to 1.5 million ASICs by the end of 2025 to 2026, but the current CoWoS wafer allocation can only support 300,000 to 400,000 units, which may delay plans due to capacity bottlenecks. Not to mention the technical challenges of large-size CoWoS packaging and the time required for system debugging (Nvidia’s similar systems 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 Secure
Nvidia will certainly not sit idly by.
At the 2025 COMPUTEX conference, Nvidia introduced NVLink Fusion technology, opening its proprietary interconnect protocol, allowing third-party CPUs or xPUs to seamlessly connect with its own AI GPUs. This semi-custom architecture seems to be a compromise, but it is Nvidia’s strategy to consolidate its market share in cloud AI computing.
The report indicates that data shows Nvidia still leads in chip computational density (computational power per unit area) and interconnect technology (NVLink), making it difficult for ASICs to catch up with its performance in the short term. Additionally, Nvidia’s CUDA ecosystem remains the preferred choice for enterprise AI solutions, which is a barrier that ASICs find hard to replicate.
For investors, Nvidia’s technological moat remains deep, but whether its high-profit model will be forced to adjust under cost pressures from other CSPs is worth ongoing attention.
In summary, according to Nomura Securities’ report:
1.
Meta’s MTIA Plan:
-
Meta expects to launch 1 to 1.5 million AI ASIC chips (MTIA T-V1) between the end of 2025 and 2026, designed by Broadcom, featuring high-spec 36-layer PCB and hybrid cooling technology.
-
By mid-2026, the MTIA T-V1.5 chip area will double, with computational density approaching Nvidia’s GB200 system.
-
In 2027, the MTIA T-V2 will adopt larger-scale CoWoS packaging and high-power (170KW) rack designs.
2.
The Rise of the ASIC Market:
-
Nvidia currently holds over 80% of the AI server market, while ASICs account for only 8-11%.
-
By 2025, Google’s TPU shipments are estimated at 1.5 to 2 million units, and AWS Trainium 2 at around 1.4 to 1.5 million units, totaling 40-60% of Nvidia’s GPU shipments (5 to 6 million units).
-
In 2026, as Meta and Microsoft deploy on a large scale, ASIC shipments are expected to surpass Nvidia GPUs.
3.
Challenges and Risks:
-
Meta’s ASIC plans are constrained by CoWoS wafer capacity (only supporting 300,000 to 400,000 units), which may delay progress.
-
The technical challenges of large-size CoWoS packaging and system debugging (which takes 6-9 months) add uncertainty.
-
If Meta, AWS, and other cloud service providers accelerate deployment, high-end materials and components may face shortages, driving up costs.
4.
Nvidia’s Advantages:
-
Nvidia strengthens its market position through NVLink Fusion technology, opening interconnect protocols.
-
Its chip computational density, NVLink interconnect technology, and CUDA ecosystem remain leading, making it difficult for ASICs 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 ASICs Shake Nvidia?
Many may wonder why there is such interest in ASICs. Because NVIDIA is the emperor of this era, and ASICs are challenging the emperor.
However, I cannot deny that designing and manufacturing ASICs takes 2 to 3 years, and regardless of how good the specifications are at the design stage, they will be outdated in two years. He also mentioned that due to delays with Blackwell, many ASIC projects have emerged, but as Blackwell’s supply eases, many of these projects have been paused.
This statement makes a lot of sense. If Nvidia’s roadmap is accurate, then a large number of Nvidia chips will be released every year, and the two-year cycle of ASICs seems difficult to beat them.
However, this is not necessarily the complete answer. For example, how to prevent the delays experienced by Blackwell from reoccurring with Rubin?
Moreover, regardless of how low Nvidia’s total cost of ownership (TCO) is, will large tech companies continue to accept Nvidia’s high profit margins? I still feel it is difficult to shake off the feeling of dissatisfaction that large tech companies have with Nvidia’s profit margins.
Even so, I am increasingly inclined to view Nvidia positively.
1.
Dominating the AI Chip Market
Nvidia is the largest consumer of HBM, purchasing the most HBM from SK Hynix and Micron. According to my personal estimates, about 70% of SK Hynix’s HBM is sold to Nvidia. Additionally, due to its strong relationship with TSMC, Nvidia has the largest allocation of CoWoS capacity.
What does this mean? It means they can sell AI chips faster and in greater quantities 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 game of chicken; if you stop, you lose.
Conversely, this means that if large tech companies try to train using ASICs, having to customize software for them and work within the limited supply and performance of ASICs, they are losers compared to Nvidia.
In other words, this is a battle that Nvidia has already won from the start, when the foundations of AI technology began to be built on Nvidia’s chips and CUDA software. (Of course, this is assuming Nvidia continues to develop.)
2.
The Reality of Sovereign AI
Sovereign AI is similar.
I actually think ASICs are 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 to develop their own AI chips and create independent AI, rather than relying on American companies like Nvidia.
However, even in places like South Korea and Taiwan, which have rich chip design technology, nurturing their own AI chips is an extremely daunting and difficult policy.
Especially in countries with fixed terms, there is a strong tendency to formulate policies based on short-term results.
This means that very few countries will embark on the true path of sovereign AI, which involves the long development time of proprietary chips and uncertain outcomes. Instead, most will purchase Nvidia chips and create imitations of ChatGPT.
Unless a country, like China or Russia, restricts the sale of Nvidia chips, most countries will ultimately purchase Nvidia’s chips. That is because this is the easiest way to see results in the shortest time.
In this sense, sovereign AI may be seen as a “Nvidia Sovereignty Policy”—replicating American AI running on American chips.
*Disclaimer: This article is original by the author. The content reflects the author’s personal views, and Semiconductor Industry Observation reproduces it only to convey a different perspective, not representing Semiconductor Industry Observation’s endorsement or support of this view. If there are any objections, please contact Semiconductor Industry Observation.
END
Welcome to leave comments and exchange ideas in the comment area!
Industry Community
