How Big A Threat Are Custom AI Chips To Nvidia Stock?

Nvidia stock (NASDAQ:NVDA) had a stellar 2024, rising by almost 3x to about $135 per share. Business has boomed led by the surging demand for graphics processing units (GPUs) which have emerged as the backbone of the generative artificial intelligence era. That said, over the last month or so investors appear to have their eyes on another sector of the semiconductor market, namely ASICs (application-specific integrated circuits) that could play a bigger role in AI computing.

This comes after two major ASIC players, Broadcom and Marvell Technology, reported a surge in demand for their ASICs from large cloud customers in the most recent quarters. Do Custom AI Chips Make Marvell Stock A Buy? For instance, Broadcom’s sales from its custom AI chips and networking processors surged by 220% to $12.2 billion in 2024, up from $3.8 billion in revenue that the company generated from AI silicon in FY’23. To be sure, Nvidia’s sales are head and shoulders above, estimated to come in at about $129 billion this fiscal year, but its growth rates are slowing.  So could ASICs threaten Nvidia’s AI dominance as the market matures? Separately, if you want upside with a smoother ride than an individual stock, consider the High Quality portfolio, which has outperformed the S&P, and clocked >91% returns since inception.

ASICS vs GPUs

ASICs have been around for more than five decades. However, they are seeing renewed interest in the AI era. While GPUs from companies such as Nvidia are pretty versatile and can be programmed for AI as well as other tasks, ASICs are custom-designed semiconductors built to perform specific tasks giving them certain advantages versus general processors. By focusing on targeted functionalities these chips offer several advantages versus GPUs for AI. For instance, these specialized chips could be more cost-effective than GPUs, which are designed for a wider range of applications.

ASICs also consume less electricity and this makes them ideal for data centers aiming to reduce electricity costs – a key cost of operating large AI systems. ASICs can also achieve higher performance for dedicated tasks than general-purpose GPUs from Nvidia or AMD as they are purpose-built. These chips could be well suited for large cloud computing providers given that they operate at a scale that can justify the design and development costs of ASICs. For instance, Broadcom, a company that is viewed as the biggest beneficiary of a potential pivot toward ASICs, recently said that three of its hyperscaler customers intend to build clusters of 1 million custom chips across a single network.

Change In The AI Space Benefits ASICs

Companies have devoted immense resources to building AI models over the last two years or so. Now training these massive models is more of a one-time affair that requires considerable computing power and Nvidia has been the biggest beneficiary of this, as its GPUs are regarded as the fastest and most efficient for these tasks.  However, the AI landscape may be shifting. Incremental performance gains are expected to diminish as models grow larger in terms of several parameters. Separately, the availability of high-quality data for training models is likely to become a bottleneck as much of the Internet’s high-quality data is already run through large language models. Considering this, the significantly front-loaded AI training phase could wind down. The underlying economics of the end market for GPU chips and the broader AI ecosystem are weak, and most of Nvidia’s customers likely aren’t generating meaningful returns on their investments just yet. See How Nvidia Stock Could Fall To $65.

As shareholders eventually seek better returns, we could see more customers looking toward ASICs to reduce upfront costs as well as operating costs. The longer-term focus of AI will be on inference, where trained models are used in real-world applications. This phase is less computationally intensive and could open the door for alternative AI processors that are less powerful. ASIC chips customized for inference could also be a top choice for these tasks. There is some historical precedent to this as well. For example, in the the crypto industry, Bitcoin miners initially used  GPUs for mining, but some large players have since switched to ASICs as they scaled up given their better hash rates (effectively how many guesses the system makes per second to solve the crypto puzzle) and overall cost-effectiveness.

What This Means For Nvidia

While NVDA stock has seen strong growth over recent years, the Trefis High Quality (HQ) Portfolio, with a collection of 30 stocks, has provided better returns with less risk versus the benchmark S&P 500 index over the last four years; less of a roller-coaster ride as evident in HQ Portfolio performance metrics. So what lies ahead for Nvidia stock?

To be sure, we don’t expect Nvidia’s business to be supplanted by these new waves of processors, given the company’s head start in the AI market and its deeply entrenched CUDA software stack and development tools which results in high switching costs for customers. Moreover, if the market sees a strong shift, Nvidia too could potentially expand its presence in the space. However, Nvidia’s premium valuation may not fully account for the potential risks and slowing growth. We value Nvidia stock at about $93 per share, about 32% below the current market price. See our analysis of Nvidia valuation: Expensive or Cheap.

[1] Returns as of 1/1/2025
[2] Cumulative total returns since the end of 2016

Invest with Trefis Market-Beating Portfolios
See all Trefis Price Estimates