The ASIC Supply Chain: Key Beneficiaries

Advertisements

The landscape of artificial intelligence (AI) is evolving at a breakneck pace, driven primarily by generative AI applications that have sparked an unprecedented demand for computing powerAs we increasingly look towards the future, a pressing topic of discussion has emerged: Can AI Application-Specific Integrated Circuits (ASICs) establish themselves as viable alternatives to NVIDIA's GPUs, which have long dominated the market? This discussion has gained momentum, especially after Morgan Stanley published a recent report titled "AI ASIC 2.0: Potential Winners," forecasting a promising trajectory for ASICs in this rapidly expanding field.

The analytics from the report suggest that the AI ASIC market could swell from $12 billion in 2024 to an astounding $30 billion by 2027, reflecting a compound annual growth rate of 34 percentSuch growth highlights the increasing relevance of ASICs designed specifically for AI tasks, a domain where performance and cost optimization are paramount.

Despite the rapid advancements in ASIC technology and potential market capture, NVIDIA retains a robust position due to its strategic advantages in training large language models

Prominent players such as Broadcom, Alchip, and Socionext are poised to benefit, alongside companies like Cadence and TSMC, who are well-positioned within the ASIC design and manufacturing ecosystem.

However, the emergence of ASICs does not spell the demise of GPUsOn the contrary, experts suggest that these two technologies will likely coexist, catering to diverse requirements across the AI landscapeAs organizations maneuver through varied operational needs, both ASICs and GPUs could serve as complementary solutions in this transformative era.

The escalating demand for AI computation is undeniableAccording to Morgan Stanley's findings, by 2027, the cloud AI semiconductor market is expected to reach approximately $238 billion under the most conservative scenarios, with optimistic projections pushing the figure even higher to around $405 billionWithin this surge of growth, ASICs are anticipated to claw back market share from NVIDIA's GPUs by leveraging their tailored optimizations and favorable cost structures.

While NVIDIA's GPUs are recognized for their superior performance in handling diverse workloads, major cloud service providers (CSPs) such as Google, Amazon, and Microsoft are actively pursuing ASIC designs to enhance their operational efficiencies

This is driven by two primary motivations: first, the need to optimize internal workloads through bespoke chip development that meets specific AI inference and training demands, and second, the pursuit of better cost-performance ratiosAlthough NVIDIA's GPUs excel in computational prowess, their high costs—especially during AI training phases—are pushing organizations toward the more cost-efficient ASIC alternativesMoreover, for instance, Amazon's proprietary Trainium chip is reported to offer a cost advantage of about 30% to 40% over NVIDIA's H100 GPUs for inference tasks, while Google's latest TPU v6 has achieved a 67% increase in energy efficiency over previous iterations.

Looking forward, while NVIDIA remains the preferred option for a vast majority of CSPs, the maturity of ASIC designs over the next few years may empower these tech giants to negotiate from a stronger position during procurement discussions.

As we assess the future market dynamics, Morgan Stanley's report identifies six potential beneficiaries within the global ASIC supply chain

These include:

  • NVIDIA, which is expected to retain its dominance in the AI GPU arena, particularly in large-scale language model training.
  • ASIC suppliers like Broadcom, Alchip, and Socionext are highlighted as significant players due to their ongoing innovations and evolving partnerships, such as Alchip's deep collaboration with AWS to broaden their market footprint.
  • Electronic Design Automation (EDA) tools, with firms like Cadence anticipated to see structural growth amidst the ASIC surge.
  • Foundries like TSMC and its partners, including ASE and KYEC, are expected to thrive from the rapid expansion in ASIC design and fabrication.
  • Testing service providers, with Advantest leading the charge in AI chip testing, focusing on amplifying its market share through enhanced testing solutions.
  • Lastly, Samsung Electronics is poised to benefit from the rising demand for High Bandwidth Memory (HBM) solutions in the non-NVIDIA segment.

In contrast, traditional semiconductor manufacturers may find themselves in challenging positions

alefox

Companies such as AMD could face a shrinking market share due to their inability to catch up with NVIDIA in the AI GPU spaceMeanwhile, foundries lacking cutting-edge process nodes, such as UMC, could struggle to compete in the high-end AI chip arena.

When drilling down into total cost of ownership (TCO) analyses, comparisons delineate the cost-effectiveness of ASIC versus GPUs in AI training and inference tasksFindings indicate that although NVIDIA's GPUs boast high-performance metrics, ASICs often present lower initial costs, particularly appealing to cloud service providers with budget constraintsAn example is AWS's Trainium 2, which reportedly accomplishes inference tasks more quickly than NVIDIA's H100 GPU while enhancing cost-effectiveness by 30% to 40%. Additionally, the upcoming Trainium 3 is projected to debut in the latter half of 2025, promising double the computational performance and a 40% increase in energy efficiency compared to its predecessors.

However, it’s essential to acknowledge that NVIDIA’s well-established system integration and superior software ecosystems continue to provide competitive leverage, especially in scenarios requiring flexibility across various AI applications

Leave a Comment

Stay Tuned for Updates

Get exclusive weekly deals with our newsletter subscription