In the rapidly evolving arena of artificial intelligence (AI), the increasing demand for customized AI hardware has captured the attention of investors and industry analysts alikeBroadcom, a significant player in this field, recently reported an impressive performance for its fourth fiscal quarter, outperforming expectationsThe news of its collaboration with the top three cloud computing giants in the U.Sto develop bespoke AI chips has propelled its market capitalization past the $1 trillion markThis accomplishment underscores the formidable growth potential of the custom AI chip industry, and various Chinese companies are positioning themselves to capitalize on this trend.

On December 13, 2024, Broadcom's market valuation soared, driven largely by its financial results and optimistic market forecast

The company reported revenues of $14.054 billion for the fourth quarter, marking a year-on-year growth of 51.2%, alongside a net profit increase of 22.7% to $4.324 billionSuch robust performance not only reflects its operational strength but also plays into the broader optimism surrounding AI ventures.

Broadcom’s CEO pointed out the strategic partnership with leading cloud service providers is crucial for the development of tailored AI chipsThis announcement elicited enthusiastic responses from Wall Street, prompting various investment banks to revise their price targets for Broadcom shares upwardFor instance, Goldman Sachs raised its target from $190 to $240, while Barclays increased its target from $200 to $205, highlighting the growing confidence in Broadcom's future potential.

Notably, the custom AI chip scene is not limited to Broadcom or American firms

• The Challenges Behind Hikvision's Peak Amid Layoff Turmoil
• A-share Market Shifts into Rapid Rotation Phase
• Semiconductor Equipment Enters a Period of High Prosperity
• Institutional Demand Fuels Dividend ETF Growth
• Robust U.S. Economic Data

Chinese technology companies such as Cambricon (688256.SH), Baidu (09888.HK), and Tencent Holdings (00700.HK) have been proactive in this burgeoning sector, making early inroads and achieving substantive milestonesThe demand from domestic cloud service providers—including Alibaba Cloud, Tianyi Cloud, China Mobile Cloud, and Tencent Cloud—position ensures a steady need for customized AI chips, further encouraging innovation and investment in this field.

AI chips have emerged as specialized processors designed explicitly for high-performance execution of artificial intelligence algorithmsAt the core of these chips are artificial neural network models that mimic biological neurons, utilizing a multitude of processing units to perform mathematical calculations and data processing efficientlyWhile Nvidia's GPUs dominate the AI market and are commonly equated with AI chips, the main categories of AI chips in the current landscape include General Purpose GPUs (GPGPU), Application-Specific Integrated Circuits (ASICs), and Field Programmable Gate Arrays (FPGAs), with GPUs and ASICs being the primary technological pathways.

According to data from Huibo Research, GPUs are predominantly suitable for AI training due to their capability to manage vast amounts of data and complex computations efficiently

This makes them invaluable in applications that require extensive model training involving frequent parameter adjustments and iterative techniquesHowever, GPUs are not without limitations; they tend to have a relatively high power consumption and may waste energy on tasks that could be more effectively executed by specialized hardware.

In light of these limitations, ASICs have become increasingly favored for inference tasks due to their cost-effectiveness and low power consumptionASICs are customized chips designed for specific applications or algorithms, with subtypes including Tensor Processing Units (TPUs), Data Processing Units (DPUs), and Neural Processing Units (NPUs). Google’s TPU emphasizes tensor computation, DPUs accelerate internal data center tasks, and NPUs are associated with Convolutional Neural Networks (CNNs) extensively utilized in edge computing solutions.

ASICs offer several advantages over GPUs, particularly regarding cost and energy efficiency

Research from Guotai Junan indicates that ASICs' simpler hardware design allows for significant cost reductions by eliminating unnecessary components typically required for general computing prowessDespite their lower costs, the single-card computational power of ASICs, such as Google’s TPU v6 and Microsoft’s Maia 100, demonstrates they can still achieve around 80-90% of the performance of leading GPUs like the Nvidia H100, making them a more appealing option for specific use cases.

As the demand for AI computing surges, the market for customized data center chips is expected to expand dramaticallyCurrently, custom chips only account for about 16% of the total market for accelerated computing chips in data centers, estimated at $6.6 billionHowever, projections suggest this number could grow to $42.9 billion by 2028, with a compound annual growth rate (CAGR) of 45% between 2023 and 2028.

Broadcom's CEO has expressed even more optimistic projections regarding the ASIC market

In a recent earnings conference, he stated that the collaboration with three major cloud service providers to develop million-unit XPU clusters could potentially yield market opportunities for Broadcom’s AI ASIC products between $60 billion and $90 billion by 2027.

The global cloud computing landscape remains a key battleground for AI chip manufacturers, with China witnessing substantial growthAccording to the China Academy of Information and Communications Technology, the Chinese cloud computing market reached 616.5 billion yuan in 2023, representing a growth rate of 35.5%, surpassing global averagesWith the AI wave driving technological innovation and larger-scale applications making headway, the Chinese cloud computing industry is set to embrace new growth, likely surpassing 2.1 trillion yuan by 2027.

Chinese cloud service providers like Alibaba Cloud and China Telecom's Tianyi Cloud show significant revenue increases, signaling robust market health

In the first three quarters of 2024, Alibaba Cloud reported revenues of 81.754 billion yuan, a 5.55% year-on-year increaseSimilarly, China Telecom's Tianyi Cloud experienced a revenue growth of 20.4%, while China Unicom and China Mobile also reported impressive gains, aligning with trends observed in North America.

Cambricon, known as China's unicorn in the AI chip realm, has made substantial strides, with over 100 million terminal devices integrated with its intelligent processorsWhile the company’s stock has seen considerable fluctuations, peaking close to 300 billion yuan, it has faced financial struggles, recording losses over recent periods as a consequence of extensive R&D investmentsNotably, the heavy investment in R&D underscores its ambition to carve out a competitive edge in the rapidly evolving sector.

Despite the financial losses, the company has developed numerous products that show promising technological advancements

In the inference chip market, Cambricon has released several series, showcasing innovations such as the MLU370, which adopts 7nm technology and is recognized for supporting LPDDR5 memory, positioning itself favorably within high-density cloud inference domains.

Other notable companies like Huawei and Baidu have also entered the ASIC landscape with their iterations of AI chipsHuawei's Ascend series, which debuted in late 2018, focuses on energy efficiency and performance, even launching the Ascend 910B in 2023 that approaches Nvidia's A100 in single-precision computational performanceConversely, Baidu’s Kunlun chips have witnessed two iterations and have seen application across various sectors beyond the cloud, including autonomous driving and intelligent transportation.

Tencent has made similar significant strides, developing its three proprietary chips designed for different applications