As the demand for AI large models surges,companies specializing in AI custom chips are experiencing rapid growth,igniting interest from shareholders and capital markets alike.This trend is not limited to major GPU manufacturers; it extends to firms dedicated to providing tailored solutions for AI applications.The intertwining of AI technology with various industries is driving an unprecedented shift in the chip market.

Recently,Broadcom released its fiscal fourth-quarter earnings report,revealing a staggering increase of over 200% in revenue from AI chip-related sales year over year.Such impressive performance has thrilled investors,leading to a notable surge in Broadcom’s stock price,allowing the company to surpass a market valuation of one trillion dollars for the first time.

Marvell Technology,another entity with a similar operational model to Broadcom,showed remarkable performance in the market as well.Following Broadcom's positive news,Marvell’s stock also reflected this optimism,experiencing a significant increase.This interconnected growth among firms in the AI semiconductor space indicates a wider trend that capitalizes on the booming AI sector.

However,it is crucial to differentiate between customized AI chips and general-purpose GPU chips.The former are designed with specific scenarios in mind,where differentiation in performance metrics is paramount.In contrast,general-purpose GPUs are built for broader applications.As NVIDIA advocates for its GPU products with slogans emphasizing bulk purchasing,cloud giants like Google and Microsoft are increasingly exploring their own AI chip designs tailored to their unique operational needs.These custom-designed chips do not require the extensive computational power boasted by general-purpose GPUs but are optimally aligned with specific business requirements and come at a lower cost.

Sensing the lack of deep chip design capabilities among cloud service providers,companies like Broadcom and Marvell have stepped in,playing the crucial role of helping these businesses develop chips based on their outlined needs.Beyond basic partnerships,the competitive landscape emphasizes NVIDIA's stronghold,which stems from a combination of robust GPU hardware,a substantial CUDA software ecosystem,and advanced NV Link connectivity.Given that both Broadcom and Marvell are involved in networking switch technologies,they have essentially secured positions at two vertices of this “capability triangle.” It's evident why these industry leaders are drawing investor attention.

Previously,NVIDIA had acknowledged ambitions to develop custom ASIC chips,though recent actions have not made any further headlines.With the new directives emerging from market demands,the battle for supremacy in the AI chip industry is entering a new phase.

Demand for customization has reached new heights.Broadcom reported a record-setting revenue of $51.6 billion for the fourth quarter,representing a 44% increase from the previous year.Their semiconductor revenue hit $30.1 billion,driven notably by the integration of VMware’s operations.The company's CEO,Hock Tan,elaborated that the surge in demands for AI XPU chips and Ethernet products enabled Broadcom’s AI revenue to climb by 220% to $12.2 billion.

During an earnings call,Tan painted a picture of tremendous growth potential in the AI chip arena over the next three years,revealing that they have three major customers embarking on extensive multi-generation AI XPU road maps.Each of these clients is projected to deploy massive clusters in a single network structure by 2027,with an estimated market demand of $60 to $90 billion.Broadcom aims to capture a significant slice of this burgeoning sector.

The company is also working on next-generation AI XPU products for two additional large-scale customers,expected to start generating revenue prior to 2027.The much-anticipated 3nm XPU chips are set to begin large-scale shipments in the latter half of 2025.

Notably,major customers of Broadcom’s ASIC chips include industry giants such as Google and Meta.Emerging reports suggest that Apple is also orchestrating its AI server chip initiatives,with Broadcom likely being a cooperative partner.

This influx of positive news has propelled Broadcom’s stock price up by 24.43%,closing at $224.80 per share,while the company's market value reached an impressive $1.05 trillion for the first time.Following suit,Marvell also garnered favorable market reactions,with its stock climbing by 10.79% to $120.77 per share.

Earlier this month,Marvell had posted its earnings for the third quarter of fiscal 2025,showing a revenue of $1.516 billion,marking a year-over-year increase of 7% and a quarter-over-quarter rise of 19%.The data center segment witnessed a particular boom with a year-over-year growth of 98%,underscoring the increasing demand for AI custom chips.

Marvell’s president and CEO,Matt Murphy,emphasized that this growth primarily stems from the demands for customized AI chips,alongside persistent needs from cloud service clients for interconnected products.The company predicts this trend will persist into fiscal 2026.

Despite challenges in non-data center sectors such as automotive,consumer,and communication infrastructure—which saw declines of 22%,43%,and 73% respectively—there are signs of recovery,as revenues in these areas experienced slight quarter-over-quarter increases of 9%,9%,and 12%.

In December,Marvell announced an expanded strategic partnership with Amazon Web Services (AWS),launching a five-year plan for generational product collaboration,which includes a range of AI custom chips and networking solutions tailored to enhance AWS’s capabilities in data center computing,networking,and storage.Shortly afterwards,they introduced what are touted to be the industry’s first 3nm high-speed interconnect platforms.

Both Broadcom and Marvell share a strategic focus,prioritizing aiding cloud service providers with their chip customization needs rather than focusing solely on designing general-purpose GPUs.This specialization is key to their rapid growth in the ASIC chip segment.

Moreover,both companies are actively expanding their networks for AI chip transmissions,such as Ethernet switches and silicon photonic technologies,ensuring the ability for comprehensive capabilities in AI computing.In their quest to challenge NVIDIA’s NV Link through an IB switch,both Broadcom and Marvell are prominent members of the “Ultra Ethernet Consortium” (UEC).

This focus on ASIC custom chips,combined with advanced transmission technologies,diminishes the necessity of NVIDIA’s CUDA ecosystem when empowering single manufacturers with dedicated chip business.These developments have ushered in a new battleground for AI chips beyond NVIDIA’s traditional domain.

The landscape of AI inference presents a further avenue of growth.Industry insights indicate that the ASIC chip is adept not only at training but also in executing AI inference tasks tailored for specific applications.This capacity aligns with a significant and expanding market need as AI models evolve commercially.

During one earnings call,NVIDIA's founder and CEO Jensen Huang acknowledged a rapid escalation in demand for AI inference capabilities.

According to a veteran in the cloud services sector,research indicates that 90% of expenditures related to AI in the future will be allocated to inference operations.“Particularly as we observe many Chinese firms optimizing training with open-source models,it becomes essential to acknowledge the significance of AI inference in terms of costs and technical complexity,” he remarked.

TrendForce’s senior vice president,Guo Zuorong,expressed that the market potential for AI inference is vast,projecting it to ultimately eclipse the size of the AI training market.While numerous vendors are currently focused on catering to AI inference with chip offerings,Guo foresees that an oversupply is unlikely to happen within the next two to three years.

He discussed the complexities of AI training,which involves substantial parameters and varies significantly across different contexts,making it a domain where GPUs dominate,acknowledging that NVIDIA thrives in this space.Yet,when it comes to AI inference markets,ASIC chips exhibit superiority due to their fixed designs that allow for quick analysis after existing training data,coupled with reduced power consumption.

NVIDIA has not established a definitive barrier in the AI inference chip realm,with major cloud providers all developing specialized ASICs tailored to their unique requirements,such as Meta focusing on community algorithm optimization and Google enhancing search engine capabilities.

Given these challenges,U.S.cloud service giants will still need to maintain strong partnerships with NVIDIA,making it evident that while there are opportunities for ASIC chips,collaboration remains essential.

While recent speculations suggested that NVIDIA might venture into the ASIC market,Guo believes these rumors may have been fleeting,as the company has not provided extensive follow-up details.Therefore,such a shift should remain under scrutiny moving forward.

As the demand for AI computing chips escalates,the question of when terminal applications will genuinely ignite the next AI demand cycle in the chip market continues to be a topic of interest.During the first half of this year,the storage chip sector,often viewed as an industry bellwether,initially experienced an upward price movement.However,this momentum did not sustain itself,as killer applications for AI large models have yet to surface in the terminal sectors.

Despite a slight resurgence in global mobile and PC markets more recently,sentiment remains relatively tempered,with several smartphone companies admitting that AI applications are not yet persuasive catalysts for upgrading devices.With demand still lacking vigor,storage chip prices began to retreat.The second half of the year saw a marked decline in storage market prices,with indications of sellers aggressively reducing prices to clear inventory.

According to statistics from third-party agencies,the global storage market continues to show growth trends in the third quarter,reaching $44.871 billion,albeit with reducing growth rates.Sequential growth plummeted from 22.1% in the second quarter to 8.3% in the third,as challenges in the smartphone and PC markets precipitate substantial uncertainty for fourth-quarter pricing,particularly in NAND Flash and DRAM segments.

Du Xiting,a senior vice president at Innodisk’s CAS department,analyzed the bifurcated situation within the NAND flash market.This year,the winds of AI have spread to data centers for intensive training of large language models,which serves consumers better informed by nuanced AI capabilities.

This surge in enterprise-grade storage demand persists,highlighted by growth in both AI server and general server markets; thus,requiring a sharp increase in demand for large-capacity enterprise SSDs.

“We are latecomers in the enterprise data center space; it took us extensive time to adjust products.We’ve finished sending samples for testing,with expectations of mass production by the first half of 2025.We anticipate a surge in business activity,” he explained.

He projected that NAND demand linked to data centers will maintain robust levels until at least mid-2025.However,in the consumer-grade market,demand is weaker,compelling certain storage module manufacturers to engage in price competition to alleviate inventory pressures,leading to a softening of prices in this niche market.

“Many AI firms frequently engage us in discussions about progress in edge applications,yet I believe the focus should start from the 'top' and work its way down,” he added.Currently,the cloud remains keen on storage capacity,with demand for AI servers outstripping standard data storage configurations by 2 to 4 times; yet when edge applications become more prevalent,the growth in storage requirements at the edge may surpass that of the cloud.