In recent discussions within the tech industry, the growing focus on alternative models for training AI systems has sparked a significant debate regarding the necessity of traditional training methods, particularly those reliant on NVIDIA's GPUsA recent report by Morgan Stanley has shed light on this evolving scenario, revealing that despite the initial excitement surrounding alternative GPU models, most of these substitutes remain unrecognized and are often abandoned in favor of the established ecosystem that NVIDIA provides.

This particular shift in focus has brought application-specific integrated circuits, or ASICs, into the spotlightRecent trends indicate that there has been a noticeable shift in investment within the AI industry from GPUs to ASICs, partly due to the waning momentum of NVIDIA's market presence and relative underperformance of AMDFor context, NVIDIA currently boasts a staggering market valuation of $3 trillion, driven by quarterly AI revenues exceeding $32 billion, while Broadcom's valuation stands at $1.1 trillion with $3.2 billion in revenuesThe market appears to anticipate that the growth potential of ASICs may outstrip that of GPUs by several magnitudes.

The question arises: could ASICs truly outpace general-purpose chips over time? According to the report, the prevailing perspective is that, unless new developments occur, NVIDIA will likely continue to hold a dominant market shareAn illuminating case in point is Google's TPU, which exemplifies the success of ASICsThe triumph of the TPU was rooted in Google’s innovative development of the Transformer technology for large models, guiding Broadcom in creating a chip specifically optimized for this application.

As of now, the success of Google’s TPU stands as a testament to the tangible benefits of customized solutions for cloud customers, contributing significantly to over $8 billion in TPU revenue for Broadcom

Advertisements

However, it is equally important to note that part of Google's expenditure directed towards NVIDIA in 2025 will be related to cloud investments, as general-purpose chips often outperform ASICs in cloud computing realmsIndeed, NVIDIA has not been idle, focusing on optimizing its offerings around Transformer models aimed at training multi-modal general AI systemsDespite potentially appearing excessive for some legacy applications, NVIDIA remains an unparalleled player in high-end training capabilities.

The report forecasts that Google's procurement of NVIDIA products may double in 2025, with some purchases stemming from its investments in enterprise cloud pursuitsHowever, much of this procurement can be attributed to NVIDIA's unmatched performance in large model TransformersWhile one might argue that certain ASIC chips perform below that of NVIDIA's offerings like the H100, these ASICs usually carry a much lower price tag—approximately $3,000 versus the H100's $20,000—leading to a more attractive total cost of ownership.

Yet, can merely manufacturing a low-cost AI chip allow one to compete with NVIDIA effectively? The report highlights Intel's struggles over the past decade, where despite numerous acquisitions of companies that had begun shipping lower-priced products, Sony failed to make a significant impact in the fieldSimilarly, AMD struggled with its earlier generations until seeing breakthroughs with its MI300 in 2024. If there were a pressing necessity, one might expect NVIDIA's competitors to focus solely on producing these lower-cost $3,000 chipsEven NVIDIA itself has rolled out several lower-priced chips aimed at traditional inference applications, only to find a compelling draw toward their premium, high-performance graphics cards.

Why is there such a pull towards premium offerings despite the availability of cheaper models? While manufacturing costs for processors can indeed be lowered, the overall system cost may become more expensive

Advertisements

For instance, creating clusters with ASIC chips could incur greater costs than using NVIDIA productsNVIDIA's unique advantage, represented by its copper cable technology capable of interconnecting 72 GPUs, contrasts sharply with ASICs that frequently utilize more costly optical communication technologies.

Moreover, critical cost elements such as high bandwidth memory are fundamentally similar between ASIC and NVIDIA chips, but due to NVIDIA’s dominance in the market, they have greater leverage over pricing in HBM procurement.

While software compatibility remains a crucial consideration for chip clients, it often represents a daunting and time-consuming challengeEase of use is paramount when adapting to software changes and managing varied workloadsThe report notes that American big data company Databricks, after acquiring Amazon Web Services' Trainium chips, anticipated it would require 'weeks or months' to configure and deploy the chipsSuch deployment delays could leave chip clients lagging significantly behind their counterparts still leveraging NVIDIA's products, which are supported by the broadly accepted CUDA software development toolkit.

Despite the achievements of Google's TPUs and AMD's MI300, the report indicates that enthusiasm for NVIDIA's ecosystem is still robust, predicting a market share increase for NVIDIA in 2025. This does not negate the value of lower-cost processors, but their widespread adoption has not transpired as initially anticipated.

A crucial question remains: did DeepSeek truly pose a threat to NVIDIA? Within the scope of the report, ASICs are not inherently superior or inferior to commercial GPUs—both represent distinct pathways to achieving similar outcomes

Advertisements

The development budgets for ASICs generally hover below the $1 billion mark, quite contrasted with NVIDIA's anticipated R&D investment of around $16 billion in 2025 alone, sustaining a protracted development cycle.

Moreover, NVIDIA plans to pour billions into connectivity technologies to bolster the performance of system-rack scale and cluster operationsThis scale of investment allows NVIDIA to solidify its dominance in the software ecosystem, and given their global presence across various cloud platforms (provided that U.SCommerce Department regulations permit), any enhancement in NVIDIA's ecosystem will have extensive ripple effects.

Although DeepSeek's methodologies might invoke critical evaluations of the previous "hardware stacking" approach, the report does not foresee a seismic shift from training to inference nor endorse the view that ASICs are unequivocal 'winners.' The positive returns projected across large cluster scales, as echoed by CEOs from tech giants and firms like OpenAI and xAI, suggest that the most competent solutions will continue to emerge victorious.

As large model workloads become increasingly prevalent, it is projected that GPUs will be repurposed for inference tasks without extensive cluster arrangementsThus, the report suggests that ASICs are unlikely to emerge as definitive victorsNVIDIA retains an edge regarding raw processing power per watt, and while ASIC chips are cheaper, the overall cost of constructing clusters may outweigh this advantage.

A cloud computing executive has openly stated that every two years their ASIC team delivers technology that trails NVIDIA’s by 2-3 years, resulting in limited benefitsThis sentiment isn't uncommon among cloud providers that have widely adopted ASIC models, as many initially viewed them as future investments, yet found that most goals remained unattained.

AMD had previously anticipated significant revenue potential from its MI400 product, yet investors raised concerns about confidence in AMD's 2026 offerings compared to NVIDIA’s forthcoming innovations

Advertisements

Advertisements