Another chip company has entered the door of the "trillion club". With soaring performance and higher expectations, Broadcom (AVGO. US) recently became the ninth U.S. company with a market capitalization of more than one trillion yuan," said the next Nvidia (NVDA. US)" also emerged.

Unlike NVIDIA, which specializes in general-purpose GPUs, Broadcom specializes in custom chips, namely ASICs. Wu Quan, founding partner of Huaxin Jintong Semiconductor Capital, told the Financial Associated Press reporter that ASIC architecture usually achieves higher speed and efficiency in artificial intelligence than GPU and FPGA architectures. In this context, Morgan Stanley expects the AI ASIC market to grow from $12 billion in 2024 to $30 billion in 2027, with a compound annual growth rate of 34%.

With the popularity of ASICs, there are also many companies in the A-share market whose stock prices have risen due to the involvement of ASIC concepts. It is worth noting that most of the company's ASIC chip business is far from Broadcom's business direction, "although they are all ASIC chips, but the range of ASIC chips is very wide, we are mainly consumer electronics, smart home, and Broadcom's application direction is different." Rockchip (603893.SH) Securities Department told the Financial Associated Press reporter.

Broadcom has high expectations for ASICs

AI chip architectures mainly include GPU, ASIC, FPGA, etc., among which ASIC refers to integrated circuits designed and manufactured in response to specific user requirements and the needs of specific electronic systems.

Broadcom's high expectations for ASICs are an important reason why the ASIC concept has recently attracted market attention.

According to Broadcom's performance report released last Friday, Broadcom's AI-related revenue (data center Ethernet chip + AI ASIC) soared 220% year-on-year, reaching a total of $12.2 billion in the full fiscal year. Broadcom CEO Chen Fuyang also said at the results meeting that the company's artificial intelligence product sales will increase by 65% in the next fiscal quarter, far faster than the overall semiconductor industry growth of about 10%; At the same time, the market demand for ASICs is forecast to be between $60 billion and $90 billion in 2027.

For the AI industry, if the general-purpose GPU represented by NVIDIA GPU is "comprehensive", ASIC is more "precise breakthrough".

"The main demand for ASICs is that customers can perform specific AI tasks and optimize computing performance, while reducing unnecessary functions to achieve better energy efficiency. In addition, it can also achieve better cost control and customized needs. Qiu Peiwen, an analyst at TrendForce, told the Financial Associated Press.

Why ASIC will explode in demand at this point in time, Wu Quan told reporters that there are two main reasons:

First of all, it is due to the fact that large AI models are shifting from training to inference. "For example, the TPU of Google's partnership with Broadcom is mainly suitable for machine learning and deep learning, and its functions focus on inference, rather than other aspects such as training and rendering."

Barclays Research Report said that at present, the demand for AI inference computing will increase rapidly, and it is expected to account for more than 70% of the total computing demand of general artificial intelligence, and the demand for inference computing can even exceed the demand for training computing, reaching 4.5 times that of the latter.

Second, Wu Quan told the Financial Associated Press reporter that this is also related to the downstream players who want to get rid of their dependence on NVIDIA, "The demand for customized chips is basically big players, similar to Byte, Meta, Google, and they are worried about whether Nvidia chips can be supplied stably for a long time." By choosing ASIC, the downstream has stronger control over the product and can enhance the right to speak in procurement. ”

In a research report published on the 15th, Morgan Stanley believes that ASIC is expected to gradually strive for more market share from Nvidia GPUs by virtue of targeted optimization and cost advantages. In this case, Nvidia's stock price fell for several consecutive trading days.

However, on December 17, Citi analysts Atif Malik and Papa Sylla issued a report in support of Nvidia, reiterating that "these two chips will coexist", and predicting that by 2028, the total market size (TAM) of AI accelerators will reach $380 billion, of which AI GPUs will dominate, accounting for 75% of the share, and ASICs will account for only 25%.

"Customized chips have high requirements for use scenarios, and only a small number of key large customers will have them." An AI practitioner told reporters. Qiu Peiwen further told reporters that ASICs are not mass-produced like standard products, and the initial development cost is high, and after the development is completed, it is usually only implemented for specific customers or applications, and the flexibility is low.

Domestic ASICs are rising

With Broadcom's expectations rising, the A-share market has also ushered in a new narrative, and the stock prices of many companies that have said they design ASIC chips have risen sharply.

As of the close of trading on December 18, the share price of Cambrian-U (688256.SH), an ASIC concept stock, rose more than 8% on the same day, hitting a record high; Allwinner Technology (300458.SZ) shares rose 9.24%; The share prices of Rockchip Micro and Fuman Micro (Rights Protection) (300671.SZ) both rose by more than 8%; The shares of Hillstone Network Technology (688030.SH) and VeriSilicon (688521.SH) both rose by more than 5%.

It should be noted that although they are all ASIC companies, most of the ASIC targets are not AI chips, but SOC and other directions, and it is difficult to benefit from this wave of AI ASICs.

For example, Rockchip has publicly stated that the company can "provide ASIC chips designed for specific functions, including interface conversion, wireless connectivity, MCU chips, etc." However, the staff of Rockchip Micro Securities Department told reporters that unlike the ASIC customized by Broadcom for customers for large model inference training, the downstream of the company's products is mainly IOT, "In fact, this industry is very detailed, and the gap between each is relatively large." In addition, the ASIC security chip developed by Hillstone Network Technology and the downstream of Fumanwei are more consumer electronics, communication equipment, Internet of Things, etc.

In fact, ASICs are a very common architecture in the semiconductor industry.

According to Everbright Securities (Rights Protection), WSTS divides ASIC types into dedicated analog circuits, dedicated logic circuits, digital signal processors, etc., and the ASIC market size will reach US$248.57 billion in 2023, a year-on-year increase of 3.4%, accounting for 47.2% of the semiconductor market.

Many domestic AI chip manufacturers have indeed chosen the direction of ASIC, such as Cambrian, Yuntian Lifei-U (688343.SH) and other AI chip companies are all ASICs. But Wu Quan told reporters that this does not mean that domestic products currently have the opportunity to stand shoulder to shoulder with Broadcom.

"Compared with chip design in general-purpose fields such as CPU, GPU, and FPGA, it may be easier to design ASICs in a single domain, which opens up a field for various design startups, including some domestic AI chip manufacturers. However, due to the small market and insufficient product strength, few highly specialized ASICs have been widely commercialized in China. ”

Taking Cambrian, which has a market value of more than 250 billion yuan, as an example, in the first half of this year, Cambrian's revenue was only about 65 million yuan. Although the company's single-quarter revenue in Q3 reached 121 million yuan, which greatly increased the total revenue in the first three quarters, it also included nearly 59.19 million yuan of non-recurring profit and loss items.

In this case, other segments driven by ASIC may be more concerned.

Guotai Junan believes that Broadcom's optimistic expectations for the ASIC market show the determination of cloud factories to continue to invest in AI infrastructure on a large scale in the next 2-3 years, which will drive the continuous prosperity of upstream optical modules, switches, PCBs, high-speed cables and other industrial chains.

At present, the domestic optical module leaders include Accelink Technology (002281.SZ), Huagong Technology (000988.SZ), etc., Ruijie Network (301165.SZ), Unisplendour (000938.SZ), etc., and Zhongji Innolight (300308.SZ), Xinyisheng (300502.SZ), Luxshare Precision (002475.SZ), etc.