Broadcom vs. Nvidia: The Battle for AI Chip Supremacy

Advertisements

Last Friday, Broadcom's shares skyrocketed by an impressive 24.43%, propelling the company's market valuation to over $1 trillionFollowing this surge, on Monday, Broadcom's stock continued its upward trajectory, gaining an additional 11.21%, ultimately reaching a market cap of $1.17 trillionThis upward momentum was fueled by the release of a quarterly earnings report that exceeded market expectations, sparking continued enthusiasm surrounding artificial intelligence (AI) customized chipsNonetheless, despite a dip of 3.91% on Tuesday amidst a broader downturn in the semiconductor sector, Broadcom's market capitalization remained above $1.1 trillion.

Within the realm of AI, Broadcom is actively engaged in designing application-specific integrated circuits (ASICs) and Ethernet networking components, collaborating with three major cloud providers to develop customized AI chipsASICs offer a specialized alternative to the more generic graphics processing units (GPUs), which are predominantly developed by NVIDIA and AMD

The landscape is divided, with tech giants like Google, Meta, and Amazon taking the lead in ASIC development while NVIDIA continues to dominate GPU production.

The recent leap in Broadcom's stock performance serves as a significant indicator of the ongoing battle between ASICs and GPUsIn addition to major cloud providers opting to replace NVIDIA GPUs with proprietary ASICs, the growing trend of startups in the ASIC field is noteworthyThese startups are expanding their reach globally, forging relationships and securing clientsIndustry insiders believe that the competition between GPUs and ASICs resembles a larger conflict between general-purpose and specialized processing, suggesting that neither technology will completely displace the other until AI processes stabilizeThis ongoing rivalry does not necessarily lead to a clear-cut victory for one side.

So, who is driving Broadcom's impressive performance?

For quite some time, NVIDIA has been the standout player in the GPU market

With its extended prominence, it is easy to overlook the massive efforts from various cloud providers to design their own chipsThe penetration rate of privately developed ASICs in the cloud sector is likely deeper than many realize.

ASICs encompass a range of specialized processors, including tensor processing units (TPUs), language processing units (LPUs), and neural processing units (NPUs). Major players like Google began their foray into ASIC technology years ago, launching TPUs that are now in their sixth generationThis month, Google officially made its TPU Trillium available for clientsLikewise, Meta introduced its custom chip for AI training and inference, labeled MTIA v2, while Amazon is preparing to roll out its Trainium3, building on previously developed Trainium2. Microsoft has also joined the fray with its Azure Maia AI chip.

Despite their non-commercial approach, these cloud providers have increasingly deployed ASIC chips in their data centers, focusing on expanding their use

Google, for instance, has quietly ascended to the third-largest data center processor design company globally, trailing only CPU titan Intel and GPU leader NVIDIA, according to TechInsightsGoogle's TPU operates as an internal solution without being available for external sale.

Amazon has made significant investments in Anthropic, a competitor of OpenAI, further binding the two companiesAnthropic uses Amazon's Trainium, and recently, Amazon announced the rapid progress of its Rainier supercomputer project tailored for AnthropicAdditional capacity is being built to satisfy Trainium demand from other customers.

Notably, Broadcom and Marvell have found themselves on the receiving end of substantial orders from these cloud providersSpecifically, Google's and Meta’s ASICs are co-developed with BroadcomAnalysts from JPMorgan predict that Meta could soon emerge as Broadcom’s next $1 billion ASIC client

alefox

Meanwhile, Amazon is collaborating with Marvell; just earlier this month, Amazon AWS entered a five-year agreement with Marvell aimed at expanding AI and data center connectivity product collaboration, facilitating Amazon's deployment of a semiconductor product portfolio and dedicated network hardware.

The results are telling: For fiscal year 2024, Broadcom's revenue surged by 44% year-over-year, hitting a record high of $51.6 billionNotably, its AI-related revenue skyrocketed by an astonishing 220% compared to the previous year, totaling $12.2 billion and propelling the company’s overall semiconductor revenue to a record-breaking $30.1 billionLooking ahead, Broadcom anticipates another 22% revenue growth in the first quarter of fiscal year 2025.

Similarly, Marvell's recent fiscal report for the third quarter of 2025 recorded revenues of $1.516 billion, reflecting a 7% year-over-year increase and a 19% quarter-on-quarter increase

Marvell attributes this performance primarily to custom AI chip projects entering mass production, expecting sustained strong demand through fiscal year 2026.

In addition to Google, Meta, and Amazon, companies like OpenAI and Apple have also indicated partnerships with ASIC custom chip manufacturersApple reportedly is in the process of developing AI server chips in collaboration with Broadcom, while OpenAI has been constructing an AI inference chip with Broadcom for several months.

Meanwhile, ASIC startups are fiercely pursuing clientsThese chip entrepreneurs, in contrast to cloud providers that seek to design chips for their proprietary models, explore various chip foundries and actively search for clientsFor example, Cerebras Systems has been developing wafer-level chips produced by TSMC, while the Etched's Sohu chip makes use of TSMC's 4nm processThe Groq LPU chips, adopting a near-memory computing architecture, have less stringent production process requirements, utilizing GlobalFoundries’ 14nm process.

These ASIC startups are expanding their client base worldwide, notably focusing on AI development in the Middle East

According to data from Cerebras Systems, its net sales nearly reached $79 million in 2023, with $136.4 million in revenue for the first half of the yearA staggering 83% of revenue stemmed from G42 in Abu Dhabi, which has pledged to purchase $1.43 billion worth of Cerebras Systems products and services in the coming year.

In September, during an AI summit in Saudi Arabia, emerging companies like Cerebras Systems, Groq, and SambaNova Systems showcased their capabilitiesCerebras Systems signed a memorandum of understanding with Saudi Aramco, which plans to utilize Cerebras Systems' products for training and deploying large models.

On the other hand, Groq has partnered with Saudi Aramco’s digital and technology subsidiary to develop what is set to be the world's largest inference data center, projected to go live by year's end with an initial capacity of 19,000 Groq LPUs, with potential expansion to 200,000 LPUs

Similarly, SambaNova Systems collaborates with Solidus AI Tech in Dubai, aiming to provide SymbaNova Cloud for high-performance computing data centers across Europe, as well as working with Canvass AI to offer AI solutions to enterprises operating in the Middle East, South Asia, Europe, and Africa.

Furthermore, SymbaNova Systems is working with the Argonne National Laboratory in the United States, while Groq is partnering with Carahsoft to provide IT solutions to U.Sand Canadian government agencies and is also collaborating with Earth Wind & Power to establish an AI computing center in Norway.

The ongoing tussle between specialized ASICs and general-purpose GPUs presents a distinctive contrast, each trailing a set of advantages and disadvantagesWhile GPUs excel in versatility and possess a mature ecosystem through NVIDIA's CUDA, they often face challenges related to power efficiency—general-purpose GPUs may consume more power than necessary

On the flip side, ASICs offer tailored capabilities, potentially achieving better performance and efficiency for specific algorithmsFor instance, Groq claims that its LPU operates ten times faster than NVIDIA's GPUs while consuming only one-tenth of the powerHowever, the highly specialized nature of ASICs also means they may struggle with multiple algorithms, making the transition from GPU-run models to ASIC compatibility not necessarily straightforward.

As ASICs gain substantial traction, there remains an open question regarding whether one will reign supreme over the otherThe capital markets' favorable view of Broadcom raises inquiries about whether this confidence is leading to a decline in market expectations for NVIDIAIndeed, as Broadcom's valuation crossed the trillion-dollar mark, NVIDIA saw its share price decline consecutively for three daysAs Keith Lerner, co-chief investment officer of Truist Investment, remarked, “You need NVIDIA, but I think the market is saying there are other beneficiaries as well.” Nevertheless, some semiconductor insiders argue that the GPU-ASIC rivalry should be viewed more broadly as a struggle between general-purpose and specialized chips

From this perspective, both types of chips can coexist for some time, rather than a straightforward replacement scenario.

From the standpoint of application, an industry expert explained that while GPUs will continue to be utilized for various parallelized general-purpose tasks, ASICs could be deployed for cost-effective alternatives in distinct settings, particularly in low-power inference scenariosAdditionally, a study by McKinsey highlights that the primary focus of AI workloads is likely to shift toward inference, with ASIC-equipped AI accelerators predicted to handle the majority of AI workloads by 2030.

Yet, the extent to which ASICs will capitalize on the AI chip market share remains uncertain, as there is potential for GPUs to adapt by integrating ASIC advantagesAccording to the product director of ARM Technologies, Bo Minqi, it is essential to keep in mind that GPUs will not necessarily be replaced by alternative chips

GPUs are primarily employed in AI cloud applications and can more easily engage within openCL cuda or SYCL coding ecosystems, which enhances usabilityHowever, the efficiency trade-offs associated with GPUs—notably in maintaining multiple threads—cannot be overlookedTherefore, it becomes apparent that in edge computing scenarios, a blending of GPU and other chip architectures may emerge rather than a straightforward partiality.

Board chairman of Thousand Chip Technology, Chen Wei, also posits that GPUs’ high energy consumption can be improved as they tap into ASIC advantagesHe notes that both constraints and opportunities will continue to shape the evolution of technology“In the contest between the architectures of GPU and other AI chips, there is a constant shift,” Chen commented, pointing to how companies like Microsoft, Tesla, and Google are investing heavily in specialized chip design, while NVIDIA, despite producing GPUs, is increasingly channeling resources into more specialized computational frameworks.

Currently, there is a growing roster of ASIC chips meticulously crafted for large model applications, elevating efficiency through hyper-specialization

For example, Etched has anchored mainstream models based on the Transformer architecture directly onto the Sohu chip, asserting that an eight-unit Sohu server can rival the performance of 160 NVIDIA H100 GPUsChen speculates that there is a high likelihood that specialized GPUs may also arise in the future, resulting in GPU firms refining their Tensor Core structures, potentially sacrificing some capacity for memory support in the process.

Yet, the very specialization that drives efficiency can also serve as a double-edged swordAs expressed by another industry insider, the current dominant architecture for AI is the Transformer, and its evolution may usher in new architecturesShould such a transition occur, especially toward alternative uses, overly specialized ASICs may struggle to adapt, whereas more general GPUs will continue to be effective.

From this viewpoint, ASIC manufacturers must also consider the downside of relinquishing generality