Transforming AI: How Custom Chips are Outpacing Traditional GPUs

In the foundational era of artificial intelligence (AI), graphic processing units (GPUs) emerged as the primary technology for training AI models and enhancing inference capabilities. Originally created to expedite graphics processing in video games, GPUs gained wider applications following Nvidia‘s introduction of a software framework that allowed developers to repurpose these chips.

Recently, however, major tech firms have begun exploring custom AI chips, known as ASICs (application-specific integrated circuits). Unlike GPUs, ASICs are tailored for specific tasks, offering improved performance and efficiency, although they lack the versatility associated with GPUs.

Considering investments? Discover the 10 top stocks recommended right now. View the list »

Let’s explore two leading companies involved in custom AI chip development.

1. Broadcom

In the realm of custom AI chips, Broadcom (NASDAQ: AVGO) stands out as a leader. The company’s first major customer for AI chips was Alphabet, for whom Broadcom aided in the development of its tensor-processing unit (TPU) named Trillium. This chip was specifically designed for integration with Google Cloud’s TensorFlow, a key software library in AI and machine learning.

Alphabet has praised the advantages of these custom chips over GPUs in terms of AI training and inference, highlighting features such as matrix multiply units (MXUs) for efficient computation and SparseCores that expedite data processes critical for recommendation algorithms.

During the last quarter, Alphabet pointed out that its dual use of GPUs and TPUs significantly benefited its cloud computing operations—lowering costs and cutting inference times.

Broadcom has since attracted additional major clients, including Meta Platforms, ByteDance, OpenAI, and, most recently, Apple. The firm reported remarkable growth in its AI chips sector during the fiscal year ending early November, generating AI revenue that surpassed expectations at over $12 billion, well above the anticipated $7.5 billion.

Investors were delighted when Broadcom noted that its initial AI customers could potentially deploy as many as 1 million AI chips by 2027, presenting a revenue opportunity between $60 billion and $90 billion in that year alone. If its newer clients progress similarly, these figures could increase further.

While it’s unlikely Broadcom will dominate the entire AI chip market—some contracts will still go toward GPUs—the company clearly has a significant opportunity ahead. Its stock presently trades at a forward price-to-earnings (P/E) ratio of 35.5, which could be appealing if the company successfully capitalizes on the projected market prospects for 2027.

Image source: Getty Images.

2. Marvell Technology

Marvell (NASDAQ: MRVL), another key player in custom silicon, is collaborating with clients like Amazon on innovative custom chips, such as the Trainium, which aids in the training of large language models (LLMs).

While Broadcom has fully customized chips for Alphabet, Marvell is providing intellectual property related to high-speed SerDes (Serializer-Deserializer) technology, with most of the design work carried out by Amazon, according to analysts from Morgan Stanley. In December, Marvell announced a five-year, multigenerational partnership with Amazon Web Services (AWS) for various custom AI solutions and additional data center components. This agreement is set to boost production volumes considerably.

Beyond Amazon, Marvell is also securing various design deals for AI accelerators and computing, with plans for new initiatives to launch in 2025. The company anticipates welcoming a third major customer soon, with leaders estimating a $40 billion market potential for custom AI chips and aspiring to capture a 20% share, which translates to $8 billion.

Marvell expects its AI revenue to exceed $1.5 billion this fiscal year, with further growth forecasted for fiscal 2026, primarily driven by its custom silicon strategy.

In addition to custom AI chips, Marvell is positioned prominently in the data center sector, where its revenue soared 98% year over year last quarter to $1.1 billion, comprising 73% of total revenue—a marked increase from 39% a year prior. Nevertheless, total revenue growth has been a modest 7%, hampered by declines in other segments.

As data center investments continue to rise, Marvell should see sustained growth in this arena. Currently, its stock trades at 42 times projected earnings, which raises some caution among investors.

Is Broadcom a Wise Investment for $1,000?

Before making a decision on investing in Broadcom, consider the following:

The Motley Fool Stock Advisor team has identified their recommendations for the 10 best stocks to buy currently… and Broadcom was not included in that selection. Stocks on this exclusive list may deliver significant returns in the future.

For context, when Nvidia was featured on this list on April 15, 2005, a $1,000 investment would have grown to an impressive $832,928!*

Stock Advisor equips investors with practical strategies for investment success, including portfolio construction guidance and regular updates from analysts, alongside two new stock recommendations each month. This service has notably outperformed the S&P 500 by over four times since 2002.*

Explore the 10 stocks »

*Stock Advisor returns as of January 6, 2025

Suzanne Frey, an executive at Alphabet, is a member of The Motley Fool’s board of directors. John Mackey, former CEO of Whole Foods Market, an Amazon subsidiary, is a member of The Motley Fool’s board of directors. Randi Zuckerberg, a former director of market development and spokeswoman for Facebook and sister to Meta Platforms CEO Mark Zuckerberg, is a member of The Motley Fool’s board of directors. Geoffrey Seiler has positions in Alphabet. The Motley Fool has positions in and recommends Advanced Micro Devices, Alphabet, Amazon, Apple, Meta Platforms, and Nvidia. The Motley Fool recommends Broadcom and Marvell Technology. The Motley Fool has a disclosure policy.

The views and opinions stated here are those of the author and do not necessarily reflect those of Nasdaq, Inc.