news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

There have been rumors that OpenAI's CEO Altman wants to make a big move in the field of AI and open a new company to develop and produce new AI chips.

Recently, research organization The Information revealed that OpenAI has been discussing the development of new chips with semiconductor design companies including Broadcom (AVGO.US), and hinted that the AI ​​startup is seeking to break away from its dependence on Nvidia (NVDA.US) and cultivate its own supply chain.

According to relevant reports, the discussions between OpenAI and Broadcom focus on Broadcom's potential role in the development of OpenAI's new chip, and the current discussion is still in the preliminary stage. Broadcom has not yet made any comment on this matter.

In addition to OpenAI, Microsoft (MSFT.US) and its main competitor Google (GOOG.US) and Anthropic, in which Google and Amazon (AMZN.US) have invested, and other artificial intelligence companies currently rely on Nvidia's AI chips to train and run large models. However, Nvidia's AI chips are expensive and in short supply, which may inspire various technology companies developing AI to find other ways.

Advanced Micro Devices (AMD.US) and Intel (INTC.US) have both launched their own AI chips to seize this demand opportunity, but their mass production schedules are still behind schedule and cannot meet the requirements of these technology giants to develop first.

This highlights the current situation of a widening gap between supply and demand in the AI ​​chip industry chain. So who will benefit the most in this situation? Undoubtedly, the upstream of the supply chain, and among them, the wafer foundry TSMC (TSM.US) is the first to bear the brunt.

TSMC: Very short supply! Really very short supply!

TSMC, which just announced its second-quarter 2024 results, achieved quarterly performance that exceeded its original set targets.

TSMC's revenue in the second quarter increased by 13.6% quarter-on-quarter (in NT dollars) and 10.3% in US dollars, mainly driven by strong demand for industry-leading 3nm and 5nm processes, which partially offset the impact of the seasonal decline in demand for smartphones;

TSMC's gross profit margin increased by 10 basis points quarter-on-quarter to 53.2% in the second quarter, slightly higher than its performance guidance, mainly because the overall capacity utilization rate was higher than expected.

The material cost of AI chips, such as H100, is mainly composed of three parts: logic chip, HBM memory chip, CoWoS package, in addition to PCB and other auxiliary materials. Since AI chips have extremely high requirements for advanced processes, in terms of mass production and technical capabilities, they are mainly supplied by TSMC. In addition, CoWos package is also exclusively provided by TSMC.

TSMC's management said at the earnings conference that in the past three months, it has found that demand for AI and high-end smartphones is very strong, which has led to an increase in the capacity utilization rate of its advanced 3nm and 5nm processes in the first half of 2024, and a decrease in inventory turnover days of 7 days to 83 days, mainly due to the increase in N3 wafer deliveries. 2024 is expected to be a year of strong growth for TSMC. Management also revised TSMC's full-year performance guidance upward, expecting its revenue in 2024 to increase slightly by more than 25% in US dollars.

When asked about the supply and demand of CoWoS, Wei Zhejia said: "I am also trying to achieve a balance between supply and demand, but I can't do it because the demand is too high. TSMC is already working hard to produce to meet customer needs and will continue to increase production, hoping to achieve a balance in 2025 or 2026. The current supply is still very tight, and this tension may not be relieved until 2026."

In view of the technological innovation brought about by the AI ​​investment boom, TSMC has stated that it will make every effort to catch up in technology.

Wei Zhejia revealed the progress of 2 nanometers at the performance conference, and said that N16 will be launched as its next nano chip technology.

Wei Zhejia pointed out: "The continued surge in AI-related demand supports the structural demand for energy-efficient computing power, and we expect that the output of 2nm technology will be higher than that of 3nm and 5nm technologies in the first two years."

Compared with N3E, 2nm technology can achieve a 10-15% speed increase at the same power, or a 25%-30% power increase at the same speed, and a full-node performance and power consumption advantage of more than 15% chip density.

Management revealed that the development of N2 technology is progressing smoothly, with equipment performance and yield rates being completed as planned or ahead of schedule, and it is possible to achieve mass production in 2025, with production ramp-up conditions similar to N3.

In addition, TSMC will also launch N2P as an extension of its N2 family, which will add two functions that can further improve performance by 5% at the same power, or increase power by 5% to 10% at the same speed, while supporting smartphones and high-performance computing applications. Mass production is planned for the second half of 2026.

TSMC's SPR is the industry's first power delivery solution that uses a back-side contact scheme to maintain gate density and device flexibility. Compared with N2P, N16 can increase speed by 8% to 10% at the same power, increase power by 15% to 20% at the same speed, and increase chip density by 7% to 10%. N16 is best suited for specific high-performance computing products with complex signal routing, power delivery, and intensive work. Mass production is expected in the second half of 2024.

TSMC believes that N2, N2P, N16 and derivative products will further consolidate its technological leadership to seize future development opportunities.

TSMC said it is working on solving the shortage of production capacity, including converting part of N5 capacity to N2. In addition, two overseas wafer fabs will be put into production next year, namely the first phase of the Arizona wafer fab and the first phase of the Kumamoto wafer fab. However, overseas wafer fabs may have a negative impact of 2 to 3 percentage points on its gross profit margin next year and in the next few years.

Summarize

The hot investment wave continues to attract capital investment, leading to the emergence of more new AI companies that rely on the empowerment of AI chips.

At the same time, there are continuous good news about technological breakthroughs in the development of AI. For example, NVIDIA will launch the "most powerful AI chip on the planet" to meet the needs of these end customers.

However, the high cost has led more and more AI companies to think about how to save money by building their own supply system. For example, OpenAI plans to ease its dependence on Nvidia through self-sufficiency.

Amid fierce market competition, NVIDIA and its peers inevitably rely on TSMC as a key foundry partner as they race to launch new products to gain market share, and as end customers build their own supply chains in pursuit of cost-effectiveness.

Therefore, regardless of the final outcome of this AI competition, TSMC will be the clear winner with comprehensive benefits. However, it also needs to be wary of the potential uncertainties that come with its fame.

Author: Mao Ting