1.TSMC cut off the supply,16 and 14nm and below processes are strictly limited
TSMC has recently imposed a series of strict supply restrictions on integrated circuit (IC) design companies in mainland China, especially for 16/14nm process products.
The decision is closely related to the latest export control regulations issued by the US Department of Commerce’s Bureau of Industry and Security (BIS). According to this new regulation, from January 31, 2025, if related products of processes 16/14nm and below are not packaged under “approved OSAT” in the BIS white list, and TSMC does not receive a certified signed copy of the packaging plant, these products will be suspended from shipment.
TSMC will manufacture for chip designers that are already on the BIS white list. For chip designers that are not on the BIS white list (including mainland and foreign companies), an application must be submitted to the US Department of Commerce, or the final package must be forwarded to the approved OSAT list for seal testing. If the final sealed OSAT is not included in the white list, TSMC will suspend the shipment.
According to the latest BIS list, 33 IC design companies have been approved, and these companies are from well-known Western semiconductor companies.
As DeepSeek releases efficient AI models, such as DeepSeek-V3 and DeepSeek-R1, end customers will focus more on the rationality of AI infrastructure and reduce their reliance on hardware such as Gpus in favor of efficient software computing models, according to the latest research from Jibang Consulting. At the same time, CSPS are likely to expand their use of their own ASIC infrastructure to reduce costs. It is expected that after 2025, the demand for GPU chips in the AI industry will change, and China’s AI market will focus on the development of independent AI chips and software optimization to adapt to the changes in the international situation, meet the needs of domestic data center construction, and promote the diversification and commercialization of AI applications.
2.Reuters: OpenAI will complete the design of its first self-developed AI chip this year for mass production next year
The ChatGPT developer will finalize the design of its first in-house chip in the coming months and plans to send it to TSMC for manufacturing, sources told Reuters. The process of sending the first design into a wafer foundry for manufacturing is called a “flow sheet.” OpenAI and TSMC declined to comment. The latest news indicates that OpenAI is on track to reach its goal of mass production at TSMC by 2026. Typical streamers cost tens of millions of dollars and take about six months to produce a finished chip, unless OpenAI pays more to speed up manufacturing.
OpenAI in the United States has the following cooperation in self-developed AI chips:
- Cooperation with Broadcom: As a leading global semiconductor company, Broadcom has rich experience in chip design. OpenAI has been working with Broadcom for several months to focus on the development of inference chips. Broadcom has helped OpenAI adapt the chip design to manufacturing needs and has optimized the chip’s design elements to transfer information faster between chips and systems, which is critical for many chips to work together in large-scale AI systems.
- Working with TSMC: As the world’s largest semiconductor foundry, TSMC is a key partner for OpenAI to manufacture its chips. OpenAI’s first self-developed chip was preordered with TSMC’s A16 process to manufacture the AI chip, which was built specifically for Sora video applications. OpenAI, which has identified manufacturing capabilities with TSMC through Broadcom, plans to build its first custom chip in 2026, though the timeline is subject to change.
- Partnering with AMD: OpenAI plans to use AMD chips through Microsoft’s Azure cloud service and diversify its training and inference workloads with AMD’s new MI300X chip, further reducing its reliance on Nvidia.
3. Foundry growth slows to 20% by 2025, with AI and advanced processes as key drivers
According to Counterpoint Research’s forecast, the growth rate of chip foundry in 2025 could reach 20 percent, mainly led by TSMC and smaller competitors that have caught the AI wave. The forecast shows a slight slowdown in growth from last year. The foundry segment of the chip industry grew 22 percent in 2024, mainly benefiting from a rebound from a downturn in 2023, the analyst house said.
Counterpoint analyst Adam Chang told EE Times, a sister platform to EBusiness: “We expect overall foundry utilisation to be around 80 per cent in 2025, with advanced nodes utilising higher than mature nodes. Driven by China’s localization efforts, demand from mature node foundries in the country is expected to be stronger than their non-Chinese counterparts.”
Chang also noted that the industry utilization rate of advanced nodes (5/4 nm and 3 nm) is expected to remain above 90 percent as TSMC continues to benefit from high-end smartphone demand and a surge in AI-related orders from hyperscale companies. By hyperscale, we mean companies like Amazon, Microsoft and Google that offer a wide range of cloud computing and data services.
According to Counterpoint’s forecast, after 2025, the foundry industry is expected to maintain steady growth, from 2025 to 2028, the compound annual growth rate will slow to 13%-15%.
This long-term growth is mainly due to the development of advanced node technologies at 3 nm, 2 nm and below, as well as the accelerated adoption of advanced packaging technologies such as CoWoS and 3D integration, the report said. These technological advancements will be the main drivers of industry growth in the next three to five years, driven mainly by the growing demand for high-performance computing and artificial intelligence applications. Counterpoint believes TSMC will continue to lead the industry, using its technological strengths to shape industry trends.
TSMC has more than 60 percent of the global contract manufacturing market, followed by Samsung and Intel. TSMC is expected to spend between US $38 billion and US $42 billion on capital expenditure in 2025, up from US $29.8 billion last year.
Chip foundries will continue to dominate semiconductor equipment purchases, according to forecasts from SEMI, an industry group. The foundry industry is expected to add capacity at an annual rate of 10.9 percent this year, rising from 11.3 million wafers per month in 2024 to a record 12.6 million wafers in 2025, according to the group.
4. Smic’s 2024 revenue rose nearly 30% to exceed US $8 billion for the first time, and the company’s 2025 revenue guidance outperforms its peers
Smic released unaudited results for the fourth quarter of 2024 after the market closed on Feb 11.
The announcement showed that SMIC’s sales revenue for the fourth quarter of 2024 was about 15.917 billion yuan, an increase of 31 percent year-on-year and 1.7 percent quarter-on-quarter. For the year as a whole, SMIC’s sales revenue in 2024 was 57.796 billion yuan ($8.03 billion), up 27.7 percent year on year in 2023, and the first time SMIC’s annual revenue exceeded $8 billion.
In 2024, SMIC’s average capacity utilization rate was 85.6 percent. However, as SMIC co-CEO Zhao Haijun previously expected, due to the new capacity in 2024Q4 and the time required to verify the corresponding capacity, the capacity utilization rate and shipment volume were affected in the quarter.
The announcement shows that its monthly production capacity from the end of the third quarter of 2024 884,200 pieces equivalent to 8 inches of standard logic, further increased to the end of the fourth quarter of 2024 947,600 pieces.
Capacity utilization declined to 85.5% in Q4 from 90.4% in Q3; Q4 sold 1.99 million wafers equivalent to 8-inch standard logic, down about 6.1% from the previous quarter
As a result of the new plant opening, SMIC’s capital expenditures in the fourth quarter of 2024 were US $1.66 billion, an increase of approximately US $480 million from Q3. For the full year 2024, SMIC’s capital expenditures were approximately US $7.33 billion.
Corresponding capital depreciation and amortisation also increased, of which 2024 Q4 depreciation and amortisation was US $849 million, up 2.2 per cent sequentially. Depreciation and amortisation for full year 2024 increased 21.3 per cent year-over-year.
On the revenue side, by region, the importance of the China market to SMIC has further increased. The revenue share of China in Q4 2024 further increased to 89.1% from 86.4% in Q3. On the other hand, the market share of the United States and Eurasia decreased.
22/28nm, 40nm process is less affected, while 55nm and above process is relatively significant. Due to the requirements of the chip design industry for process maturity, capacity supply stability, etc., larger fabs do have significant advantages in order acquisition.
The analysis believes that AI will become an important pull factor for wafer foundry demand in 2025.
5. DeepSeek hit the screen, China’s big model stirred overseas AI lake
According to public information, DeepSeek, a Chinese AI startup, was founded in May 2023 and is a large model startup. Only half a year after its founding, DeepSeek launched DeepSeek Coder, a free commercial, fully open source big model of code. In May 2024, the company made a name for itself with the release of DeepSeek V2, an open-source model that reduced the cost of inference by nearly a hundredfold.
On December 27, 2024, DeepSeek launched its open source model, DeepSeek V3.
At that time, DeepSeek-V3 ranked seventh among all models and first among open source models in the foreign large model ranking Arena. Moreover, DeepSeek-V3 was the most cost-effective model among the global top 10.
On January 20, 2025, DeepSeek officially opened source the R1 inference model. Performance alignment OpenAI-o1, the official version of DeepSeek-R1 uses reinforcement learning technology on a large scale in the post-training stage, which greatly improves the reasoning ability of the model in the case of very little labeled data. In math, code, natural language reasoning and other tasks, the performance is comparable to OpenAI o1.
DeepSeek’s R1 release is now widely believed to mark an important turning point in inference modeling research, which until now had been an important area of industrial research, but lacked a seminal paper, just as AlphaGo used reinforcement learning to play countless games of Go and optimize its strategy to win, DeepSeek is using the same approach to boost its capabilities, so 2025 could be the first year of reinforcement learning.
On January 24, the DeepSeek R1 benchmark has risen to third place in the overall category of large models at Arena, where it is tied for first place with OpenAI o1 in the StyleCtrl category. Its Arena score reached 1357 points, slightly surpassing OpenAI o1’s 1352 points.
DeepSeek-V3 completed the training of the 671 billion parameter model using only 2,048 H800 Gpus at a cost of only $5.576 million, far less than the training cost of other top models.
As a reference point, researchers at Stanford University and Epoch AI published a study in the middle of last year suggesting that the largest models will cost more than $1 billion to train by 2027. In addition, third-party research firm Gartner predicts that hyperscale companies like Google, Microsoft, and AWS will spend as much as $500 billion on AI servers alone by 2028.
Therefore, many companies believe that DeepSeek’s low cost means that the demand for computing power investment in large models may tilt away from training lateral reasoning, that is, the demand for reasoning computing power will become the main driving force in the future. The traditional advantages of hardware vendors such as Nvidia are more concentrated on the training side, which may have an impact on their market position and strategic layout.
Yann Lecun, chief AI scientist at Meta, commented that the launch of DeepSeek-R1 does not mean that Chinese companies are surpassing US companies in the field of AI, but rather that open source large models are surpassing closed sources.
