In the downward cycle of the semiconductor industry over the past year, in addition to Nvidia achieving rapid performance growth as the dominant GPU player, some semiconductor giants that are not primarily focused on GPUs but are based in the AI custom chip market have also achieved steady growth in a low-key manner.
According to Gartner's statistics, in 2023, based on semiconductor sales, Nvidia entered the top 5 semiconductor (excluding third-party wafer foundries) camp for the first time with a year-on-year revenue growth rate of 56.4%; The other two companies in the top ten that can achieve growth in the downward cycle are Broadcom (+7.2%) and STMicroelectronics (+7.7%), but considering the income gap between the two companies, Broadcom's growth ability appears to be stronger.
In the eyes of the outside world, Broadcom is a company that is stronger than mergers and acquisitions, and in recent years, it has extended its acquisition arm to software, completing multiple major moves. One of Broadcom's strong businesses is the customization of AI chips, which supports its growth against the trend in response to the surging demand for AI in large models.
Marvell, another company with a relatively small overall market, has also seen a significant increase in related performance under the demand for customized AI chips. The latest financial report shows that the company's data center related business revenue increased by 54% year-on-year in the fourth quarter of 2023, with AI related revenue of approximately $200 million during the quarter.
Unlike GPUs that are more positioned as general-purpose graphics processors and large-scale parallel computing, custom AI chip ASICs are a more specialized category. Due to its high degree of customization, its potential in the early stages of industrial development with strong demand for computing may be relatively limited. However, as large models continue to develop, their importance and growth potential cannot be ignored.
Customized chips are in high demand
The commonality between Botong and Meiman Electronics is that they both have customized AI chips and related data exchange businesses, but due to their different historical backgrounds, each company has its own focus on overall development.
According to the financial report for the first quarter of the 2024 fiscal year (three natural months ending in February 2024), Broadcom achieved a revenue of 11.961 billion US dollars, a year-on-year increase of 34%; Net profit of 1.325 billion US dollars, a year-on-year increase of 17.2%; The gross profit margin is 75.4%.
Hock Tan, CEO and Chairman of Broadcom, stated that there are two major revenue growth drivers in the first quarter and the entire fiscal year of 2024. One is that the company recently completed the acquisition of Vmware, which led to an increase in revenue from the infrastructure of Broadcom Software as customers deployed Vmware's infrastructure; The second is the strong demand for network products in artificial intelligence data centers, as well as the demand for AI custom accelerators in large-scale data centers, driving growth in the semiconductor sector.
The financial report shows that its business mainly consists of two main parts: semiconductor solution revenue of 7.39 billion US dollars and software infrastructure revenue of 4.571 billion US dollars. In the semiconductor business, Broadcom's AI revenue during the quarter was approximately $2.3 billion, mainly driven by the demand for customized AI chips from two ultra large scale customers.
Although the company has not officially announced who the two major clients are, the industry believes that the more likely ones are Google and Meta, which have been promoting self-developed AI chips.
Broadcom expects that the unexpected growth of artificial intelligence in the fiscal year 2024 will be sufficient to offset the cyclical weakness of broadband and server storage, and will achieve medium to high single digit growth; It is expected that the semiconductor business will reach 30 billion US dollars this year, and AI related revenue will exceed 10 billion US dollars.
According to the financial report of Meiman Electronics, in the fourth quarter of 2024 (the three natural months ending in February 2024), it achieved a net revenue of 1.427 billion US dollars, a year-on-year increase of 1%; The gross profit is 664 million US dollars, with a GAAP gross profit margin of 46.6% and a non GAAP gross profit margin of 63.9%; Non GAAP's net profit is 402 million US dollars.
Marvell President and CEO Matt Murphy stated that artificial intelligence has driven strong growth in the data center terminal market, with a month on month growth of 38% and a year-on-year growth of 54%.
According to the performance meeting, the company began shipping two AI ASICs in the first quarter and plans to achieve significant growth in the second half of the fiscal year. In early March, Marvell announced that its first AI acceleration chip product on the 2nm platform would collaborate with TSMC. Prior to this, the two had already collaborated on the 5nm and 3nm platforms.
From this, it can be seen that the driving force behind the growth of these two customized AI chip businesses lies in the international technology giants considering the high cost of purchasing chips and deciding to reduce the difficulty of computing power deployment costs through self-developed chips.
Industry insiders told 21st Century Business Herald reporters that both GPU and Google's TPU chips, which are AI chips, have their own advantages and disadvantages. For example, the advantage of GPUs lies in their strong versatility and suitability for various computing tasks, but there are certain bottlenecks in their storage and communication, which is also the reason why the demand for high bandwidth storage of HBMs has been strong in the past two years; TPU belongs to a type of ASIC customized AI chip, which has the advantages of being designed specifically for deep learning and high computational efficiency, but the disadvantage is poor universality.
Not only these two major manufacturers, Nvidia also has signs of entering the custom chip market in the near future, which may cause some disruption to the development pattern of this market.
Zhang Yubin, senior analyst at the IT department of Sigmantell, analyzed to 21st Century Business Herald reporters that Nvidia has strong technical strength and market position in the GPU field, and ASIC design and GPU design are similar in some aspects, such as parallel processing and low-power design. By laying out ASICs, Nvidia can apply its technological strength in the GPU field to ASIC design, achieving technological collaboration and resource sharing. Meanwhile, the design and production of ASICs also require advanced semiconductor manufacturing processes and equipment, which can complement Nvidia's existing GPU production lines.
According to Lu Jing, Partner and Managing Director of Sullivan Greater China, unlike the flexibility of GPUs and FPGAs, customized ASICs cannot be changed once manufactured. Therefore, the high initial cost and long development cycle make it a high entry barrier. "Nvidia is preparing to enter the ASIC field with the aim of occupying a part of the explosive market for customized artificial intelligence chips and protecting itself from the influence of more and more companies seeking their product alternatives. Nvidia's entry into the ASIC field is likely to erode the market share of Broadcom and Meiman Electronics in the ASIC field."
Differentiated competition
Broadcom and Meiman Electronics are not traditional chip giants that focus on general-purpose computing capabilities. It's just that in the current era of expanding demand for high-speed computing, computing power is no longer sufficient. To achieve rapid data processing, computing power, storage power, and transportation capacity are indispensable.
"Transportation" involves the process of transmitting data. Broadcom and Marvell have deployed communication network related chip services, such as Ethernet switching chips, optical transmission DSP chips, etc.
China International Capital Corporation (CICC) research points out that AI has increased requirements for data center communication performance, which is expected to drive the expansion of the data center communication market in the long term; Broadcom has a leading market share in the upstream device field of multiple cloud network devices, such as switching chips, and is expected to fully benefit from the changes in the AI era.
Guojin Securities also believes that in the long run, due to the lower requirements for computing power in AI inference compared to training, with the expansion of AI applications in the future, non Nvidia AI chips will have a large market space in the inference end, driving the application of Ethernet in AI networking.
In addition, in recent years, Broadcom's mergers and acquisitions have continued to move towards software and cloud related activities. Yao Jiayang, a senior observer in the semiconductor industry, analyzed 21st Century Business Herald reporters, "My personal observation is that Broadcom's acquisition of software business in the past few years and Broadcom's ASIC business are two parallel lines, and it is unlikely to form a complementary relationship between CUDA and NVIDIA. The development strategies of the two companies are different."
He pointed out that Vmware, which was recently acquired, has deep cooperation with cloud computing vendors such as Google and Meta. The acquisition of software will greatly help increase the gross profit margin, profit margin, and EPS of Broadcom's financial performance. From this perspective, software acquisition will be an important direction for the operation of Broadcom Enterprises.
In addition, multiple chip giants have shown in their financial reports that in addition to AI training, more and more demand for AI reasoning is also being concentrated recently, which is expected to bring new development space.
Yao Jiayang told reporters that AI reasoning is indeed releasing a significant demand, which will bring development space for ASIC chips. Of course, whether GPU or ASIC has more development advantages depends on specific application demands. "For example, some chip manufacturers may make special optimizations for the facial recognition part to achieve lower power consumption or better performance, and can use ASIC chips for this function. However, GPUs can also be used for related development. Therefore, it depends on customer needs, whether to use GPUs to develop facial recognition or use ASIC chips to meet special application needs."
"I believe that ASIC and GPU will coexist and have their own needs," he continued. Currently, the development of AI big models is still in its early stages, so there will be different big models continuing to develop in different fields. So it is difficult to develop customized ASIC chips for specific fields in the early stages of industrial development; The universality of GPUs makes them relatively suitable for the current stage. "Of course, because Meta and Google have specific application scenarios, they will develop corresponding ASICs, which is also an important market opportunity for both Broadcom and Marvell. Therefore, I think there will be different corresponding situations, and GPUs and ASICs have growth potential."
The same applies to applications in the intelligent driving industry. Yao Jiayang analyzed to reporters that intelligent driving, currently known as "advanced assisted driving" or ADAS, is still in a relatively early stage of development. Whether in China, the United States, or Europe, there are currently different regulatory restrictions, insurance claims, and other requirements. "From this perspective, in the short term, in the next 3-5 years, I believe that the demand for intelligent driving ASIC chips will not grow so quickly; GPUs may have better development in response to different countries' regulatory requirements due to their versatility. However, in the future after entering the L3 era, as regulations and other aspects become more mature, the demand for ASIC chips will gradually emerge."