Summary of the Survey on Supply and Demand of AI Chips in 2025
Topic: Supply and demand of AI chip market
With the rapid development of artificial intelligence technology, the AI chip market has become a focal point in the global tech industry. This interview is conducted by LookWhole Insight with experts from the AI chip industry. It covers multiple dimensions of the market, including concerns from the demand side, positive market changes, trends in computing power demand, and the current state and challenges of global AI infrastructure, aiming to reflect the true landscape of the AI chip market. - LookWhole Insight
1. Q: From the demand side, what were the main concerns in the past six months, especially in Q1 of 2025?
A: The main concerns included: global economic uncertainty leading enterprises to reduce AI investment projections; rising computing power requirements due to new models; and the below-expectation capital expenditures from companies like Alibaba and Tencent, coupled with export control policies that could impact the implementation prospects of domestic strategic plans.
2. Q: What positive developments have occurred in the industry since May?
A: Although global economic uncertainty remains high, diplomatic and commercial negotiations between governments have been progressing, improving market expectations for outcomes. In particular, agreements starting in the Middle East, such as the UAE relaxing its export limit to 500,000 units per year, and Jensen Huang's (NVIDIA CEO) negotiations with European governments on sovereign AI data center initiatives, are seen as positive signs.
3. Q: How do the scope and complexity of AI applications impact the growth of computing power demand?
A: The scope and user base of AI are expanding, and the complexity of AI is increasing. For instance, not only has model training become more complex, but growth in post-training and test-time inference, along with more intuitive user experiences like chain-of-thought reasoning, are also driving up computing power demand.
4. Q: What data reflects this trend in computing power demand?
A: Specific data points include OpenAI increasing demand via collaboration with Google despite delays with Microsoft. Google’s API call volume surged to 480 billion in a short period. Also, the launch of “Doubao” (likely an AI product) saw a significant rise in daily usage, indicating an explosion in demand driven by the increased complexity of new inference models.
5. Q: What are the current issues and developments in China’s domestic AI infrastructure?
A: Under the H20 restrictions, the RTX Pro series is expected to begin deliveries, partially addressing China’s computing power needs. While there are concerns that RTX Pro supports inference but not training, in reality, training capacity is currently sufficient and not growing exponentially—existing infrastructure can still be reused. Inference, however, is under pressure due to a rise in application numbers and users per application. The RTX Pro may have some limitations in inference efficiency, but overall, it serves as an adequate transitional product.
6. Q: What are the expectations for RTX Pro shipments and the market reaction after major players receive certification?
A: Shipment expectations for RTX Pro have risen from several hundred thousand units to the million-unit level. More precise data will be available after deliveries begin in June. Certification by major players will improve visibility. Although the growth rate of computing power may slow due to technological maturity, the growth trend remains intact, and we remain confident in 2026 investments.
7. Q: What are the current market concerns and timing strategies for China’s computing power infrastructure?
A: There are concerns regarding domestic computing power capacity, and investors are waiting for concrete shipment data of RTX Pro and Chinese chips in the second half of the year before making deployment decisions. This timing may be approaching, offering a potential opportunity for computing-related industry chain segments and IDC companies in China.
8. Q: What is the long-term trend of AI demand?
A: Long-term demand is driven by strategies including pre-training, post-training, and test-time usage, as well as the evolution from generative AI to AI agents to physical AI. Despite short-term headwinds, we remain confident in long-term demand growth.
9. Q: What are the global server shipment volumes and AI server share projections?
A: In 2025, global server shipments are projected to reach 11.61 million units, a 7% YoY increase. AI servers will account for 15% of this, with a 30% YoY increase. General-purpose servers are also expected to grow about 10%. At the chip level, NVIDIA is expected to account for 60% of unit volume and nearly 70% of total market value. Google and AWS are projected to account for 15% and 12% respectively. AMD and Intel currently have relatively smaller shares.
10. Q: What are the main market concerns regarding NVIDIA?
A: Concerns center on NVIDIA’s reduced investment in TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) technology, raising doubts about demand sufficiency in 2025–2026. Opinions on CoWoS are divided—there was a surge in H2 2024, but it returned to normal levels in Q1 2025. Export controls have affected allocation, and there are varying views on CoWoS capacity and ramp-up scale. However, no major strategic shift has occurred, and NVIDIA is expected to ship 360K–370K units in 2025, with no clear signs of demand reduction for the B-series.
11. Q: How is the market responding to NVIDIA's GB-series products like GB200 and GB300 NVL72?
A: Based on Jensen Huang’s remark about a 1,000-unit weekly output, the annual volume could reach nearly 30,000 units, and possibly exceed 35,000 considering ramp-up in Q3 and Q4. Microsoft is the main customer, with a 40% share; Google has 20%, and others like Oracle, Amazon, and Meta hold the rest. With new OpenAI-Google agreements and growing demand in the Middle East, tech giants are expected to increase their share by 2026.
12. Q: What updates are there in NVIDIA's product roadmap?
A: Jensen Huang mentioned that the next-gen GB300 has started shipping in Q2 at the chip level, with server-level small-batch production starting in Q3 and mass production in Q4. GB300 NVL72 will feature 12 HBM stacks with higher capacity. The Rubin product is expected in H2 2026, transitioning from HBM3/HBM3e to HBM4, though storage capacity may fall short of market expectations. The Rubin Ultra NVL576 is due in 2027 with upgraded HBM specs and packaging. The Feynman architecture will launch in 2028.
13. Q: What is your view on the ASIC market?
A: While overseas ASIC players like Broadcom and Marvell have strong stock performance, uncertainties remain around customer project implementation beyond Google and competition from Taiwanese firms. Nevertheless, the value ASIC represents for the industry is acknowledged. Google, as a leading AI ecosystem player with strong hardware capabilities, offers competitive products. We hold a positive view on TPU demand, which may benefit Chinese companies in ASIC, PCB, and switching networks.
14. Q: What are your views on the future of China’s computing power ecosystem?
A: There are signs of marginal improvement and growth potential on both supply and demand sides. Though hard to quantify in H1—especially in the realm of physical AI—the overall trend remains positive. We are optimistic about the sustainability of computing power demand. Globally, hardware stocks have rebounded to near historical highs, suggesting possible short-term volatility. In China, segments like IDC companies have faced pressure due to slow capex from leading firms and unmet market expectations. However, with improved chip supply in H2, China’s computing power industry chain could see new opportunities.
In addition, the following are more detailed key questions and insights on China's domestic AI chip industry in 2025 collected through expert interviews and market research. These questions and answers focus on key aspects such as capacity, ecosystem development, technology gaps and business models
Summary of Key Questions and Answers on China's Domestic AI Chip Industry in 2025
|
No. |
Question |
Answer |
|
1 |
Which generation of HBM is used in 910C? Where is it sourced from? |
It uses 3rd-generation HBM, mainly Samsung chips in the past, procured via distributors. A major customer stocked up last year. Currently based on spot purchases. |
|
2 |
How is the market feedback for the 384-card cluster cabinet product? |
Market acceptance is moderate. Due to cooling and interconnect complexities, customer adoption is slow; most are still in the testing phase. Promotion takes time. |
|
3 |
What is the current monthly wafer volume for 7nm AI chips? |
Currently about 5,000 wafers/month, expected to expand to 8,000–10,000 wafers/month in the second half of the year. |
|
4 |
Approximately how many AI chips can be produced from one wafer? |
Theoretically 80–90 chips per wafer, but actual yield is around 20–30 chips per wafer. |
|
5 |
With a prepayment of RMB 100 million, how many wafers can be purchased? |
At RMB 70,000/wafer and a 50% prepayment rate, RMB 100 million can procure around 3,000 wafers (equivalent to a total value of RMB 600 million). |
|
6 |
Can AI chip inventory be directly recognized as revenue on a 1:1 basis? |
No, it must go through packaging, testing, and other processes. Only final boards can be recognized as revenue. |
|
7 |
What are the shipment targets for major domestic AI chips in 2024? |
Huawei targets 700,000–1,000,000 units for the year, with 910C aiming for over 400,000. Cambricon expects 200,000–300,000 units, with strong growth. |
|
8 |
Can SMIC support chip shipments for companies like Cambricon? |
Yes. To ship 200,000 units annually, Cambricon needs about 1,250 wafers/month. SMIC’s current and planned capacity can support this. |
|
9 |
How will the domestic AI card ecosystem evolve? |
A landscape dominated by Huawei and 5–6 key vendors is expected. Huawei has a leadership edge but won’t be monopolistic. |
|
10 |
What is the estimated cost to develop a new generation AI card? |
Requires over 200 staff, with annual personnel costs of RMB 200 million. Adding EDA/IP/licensing/fab costs, total investment is several hundred million RMB. |
|
11 |
Is there a large performance gap between domestic chips and TSMC chips? |
Performance gap is small; the main differences are in power efficiency and yield. Domestic chips are price-competitive if performance suffices. |
|
12 |
What is the current yield of 910C at SMIC? |
Around 30% currently, due to dual-die packaging and large chip area. Improvements still in progress. |
|
13 |
Can DeepSeek R2 run on domestic AI chips? |
Yes, it supports R1 training and inference. Inference performance is close to H20. Compatibility is improving on platforms like 910C and 590. |
|
14 |
Do internet companies developing their own ASICs threaten general chip vendors? |
Limited short-term impact. Some partner with chipmakers to co-develop ASICs, which helps bind customers. Vendors need to balance confidentiality and resource allocation. |
Source: LookWhole Insight, Expert interviews, 2025
About US
LookWhole Insight is a global leader in data analytics and market research, offering deep insights into industries, economies, and consumer behavior across the world. We deliver comprehensive data and analysis on thousands of products and services, making us the first choice for organizations pursuing growth and exploring untapped, blue ocean markets.
Our offerings include syndicated research reports, customized research solutions, and strategic consulting services. The LookWhole Insight database is trusted by prestigious academic institutions and Fortune 500 companies alike, providing a robust foundation to navigate both global and regional business environments. Our data spans 26 industries across 35 key economies, backed by thousands of metrics and detailed analyses.
As an independent provider of global business intelligence, we empower clients with market analysis and consumer insights that range from local to global, and from tactical to strategic. Our research solutions guide critical decisions on when, where, and how to scale your business with confidence.
Email: market@lookwhole.com
www.lookwholeinsight.com
