Last Friday, Broadcom's stock soared by an impressive 24.43%, propelling its market value to surpass the one trillion dollar markThen on Monday, the shares continued their upward trajectory, climbing an additional 11.21% and reaching a market valuation of $1.17 trillionThis surge followed the release of a quarterly report that significantly outperformed market expectations, reflecting a growing interest in AI custom chipsEven after a dip of 3.91% on Tuesday — when many semiconductor stocks were struggling — Broadcom managed to maintain a robust market cap exceeding $1.1 trillion.

In the AI landscape, Broadcom is positioning itself in the realm of custom or application-specific integrated circuits (ASICs) and Ethernet network componentsThe company is collaborating with three major cloud service providers to develop tailored AI chipsThese specialized ASICs stand in contrast to the more general-purpose graphics processing units (GPUs), representing a divide between companies like Google, Meta, and Amazon versus competitors such as Nvidia and AMD

This rivalry encapsulates a broader industry shift toward tailored hardware solutions capable of handling specific computational tasks more efficiently.

The rise of Broadcom’s stock could be seen as the beginning of a counterattack by ASICs against the dominance of GPUsAlongside cloud providers developing their proprietary ASICs to replace Nvidia's GPUs, a wave of startups in the ASIC space is eagerly searching for clientsIndustry insiders suggest that the GPU versus ASIC battle is not merely a contest of superiority but a dynamic interplay of general-purpose versus specialized technologyUntil AI solutions fully stabilize, both chip types will likely coexist, and the outcome of this competition isn't strictly one-sided.

What factors are contributing to Broadcom's burgeoning success?

GPU titan Nvidia has enjoyed the spotlight for a prolonged period, leading many to overlook the endeavors of cloud service providers in semiconductor design

In reality, their ASIC penetration may be more profound than initially perceivedASICs include various chip types, such as Tensor Processing Units (TPUs), Language Processing Units (LPUs), and Neural Processing Units (NPUs). Google, for instance, laid the groundwork for TPUs several years ago, with its sixth-generation TPU, named Trillium, having been recently made available to clientsMeanwhile, Meta has introduced custom chips designed specifically for AI training and inference with its MTIA v2, and Amazon is rolling out Trainium2 with plans for Trainium3 next yearMicrosoft has also developed its AI chip, Azure Maia.

Perhaps due to their lack of external chip sales, these cloud firms' AI chips have garnered limited attention in the marketplaceNevertheless, they have strategically deployed ASIC chips within their data centers and are working to expand their use.

Google, in particular, has quietly ascended to become the world’s third-largest data center processor design company, trailing only behind Intel and Nvidia

The internal workloads at Google leverage TPUs, which are not sold externallySimilarly, Amazon has made significant investments in competitors to OpenAI, such as Anthropic, which utilizes Amazon's TrainiumRecently, Amazon announced that a supercomputer cluster project, Rainier, dedicated to Anthropic, is nearing completion, alongside plans to boost capacity for other customers seeking to use Trainium.

Within this environment, semiconductor manufacturers like Broadcom and Marvell are securing orders from these cloud playersGoogle and Meta's collaborations with Broadcom in ASIC customizations are notable, with analysts predicting Meta could soon emerge as a significant client for Broadcom, potentially bringing in a billion dollars in revenueAmazon, too, has struck a five-year agreement with Marvell to expand cooperation on AI and data center connectivity solutions, allowing Amazon to roll out a full semiconductor product line and dedicated network hardware.

The results are evident in financial performance: For the 2024 fiscal year, Broadcom's revenue surged by 44%, hitting a record $51.6 billion

alefox

AI-related revenue skyrocketed by 220% to $12.2 billion, significantly contributing to the semiconductor revenue reaching a record $30.1 billionBroadcom anticipates a 22% year-on-year increase in revenue for the first quarter of the 2025 fiscal year.

Marvell, in a recent earnings report, showcased a 7% year-on-year revenue growth for the third quarter of the 2025 fiscal year, amounting to $1.516 billion, with expectations for an additional 19% growth in the forthcoming quarter, propelled by custom AI chip projects that have commenced mass production and are expected to see robust demand through the 2026 fiscal year.

Besides major players like Google, Meta, and Amazon, OpenAI and Apple have also been linked to potential collaborations with ASIC custom chip manufacturersRecently, reports surfaced indicating that Apple is developing AI server chips and is working with Broadcom on networking technology related to these chips

OpenAI has also been rumored to have partnered with Broadcom for several months to create AI inference chips.

Startups in the ASIC sector are actively seeking clients globally, particularly engaging with Middle Eastern countries that are ramping up their AI capabilitiesTake Cerebras Systems, which has reported nearly $79 million in net sales in 2023, with $136.4 million in the first half of the yearNotably, revenue from the UAE’s G42 accounts for 83% of total income, with G42 committing to purchases topping $1.43 billion in Cerebras products and services next year.

During an AI summit in Saudi Arabia in September, representatives from Cerebras Systems, Groq, and another AI chip startup, SambaNova Systems, were presentCerebras signed a memorandum of understanding with Saudi Aramco, which plans to utilize Cerebras’s products for training and deploying large-scale models.

Groq is collaborating with Saudi Aramco's digital and technology subsidiary to establish the largest inference data center globally, set to be operational by the end of this year, starting with 19,000 Groq LPUs and potentially scaling up to 200,000 in the future

SambaNova Systems is also engaging with Solidus AI Tech in Dubai to deliver high-performance computing solutions for Europe and is collaborating with Canvass AI, which operates across the Middle East, South Asia, Europe, and Africa, to provide AI solutions for businesses.

While the competition between GPUs and ASICs is defining the landscape, each has its strengths and weaknessesGPUs excel in versatility, capable of running a myriad of algorithms, supplemented by the well-established Nvidia CUDA ecosystem, which facilitates ease of useHowever, general-purpose GPUs often waste computational power and energy due to their broad applicabilityIn contrast, ASICs are designed with specific tasks in mind, offering potentially superior computational performance and energy efficiency tailored for particular algorithmsFor example, Groq’s LPU boasts speeds ten times faster than Nvidia’s GPUs while consuming only one-tenth of the power and cost.

Yet, the very specificity that gives ASIC chips an advantage may also be their vulnerability, as adapting large models originally designed for GPUs may pose challenges

Meanwhile, the sustained usability of GPUs remains critical for a multitude of parallel processing applications, while ASICs can be more cost-effective for specialized needs, like inference with low power consumptionMcKinsey's research also indicates a transition in AI workloads towards inference, with ASICs expected to handle a majority of AI tasks by 2030.

Yet, the future of ASICs capturing a significant market share in AI chips remains uncertain, particularly as GPUs evolve by incorporating ASIC-like advantagesBo Minqi, a product director at Arm, shared insights indicating that GPUs may not be usurped in their entiretyTheir fit within AI cloud applications offers specific ease of integration via programming frameworks like OpenCL, CUDA, or SYCLWhile GPUs do incur some overhead related to multi-threading context switching, their flexibility allows them to adapt to various workloads, which ASICs, with their dedicated designs, may struggle to accommodate effectively.

Chairman Chen Wei of Qianxin Technology also acknowledged the potential for GPUs to improve on efficiency deficits, as they assimilate strengths from dedicated chips

As companies like Microsoft, Tesla, and Google lean towards researching increasingly specialized chips, Nvidia is also pivotingTheir focus has shifted from traditional GPUs to more specialized computing architectures, particularly enhancing their Tensor Core capabilities beyond their original functionality.

We are witnessing an emergence of ASICs tailored explicitly for large models that boost chip efficiency due to their high specializationEtched, for instance, has integrated the mainstream Transformer architecture directly onto its Sohu chip, claiming that a server equipped with eight Sohu chips can match the output of 160 Nvidia H100 GPUs"We could potentially see dedicated GPUs specifically for large model applications, with GPU manufacturers likely refining their Tensor Core setups to sacrifice some capacity for memory support," noted Chen.

However, this high level of specificity carries inherent risks