By Adam Clark
Tech earnings later this month are likely to reveal more massive outlays for artificial intelligence. Amazon.com, Google, Meta Platforms, and Microsoft remain locked in a costly race to build data centers -- and buy Nvidia chips to fill them.
There's one way out of the cost spiral: sourcing chips not made by Nvidia. None of the hyperscalers are aiming to replace pricey Nvidia chips entirely -- at least not yet. But, they are hoping that using their own chips can slow the pace of spending. Enter chip designers Broadcom and Marvell Technology.
Both companies develop a type of chip called an application-specific integrated circuit, or ASIC. While such processors lack the all-round capabilities of Nvidia's graphics-processing units, or GPUs, they can be designed for predictable, high-volume workloads. They also cost an average of several thousand dollars compared with more than $30,000 for Nvidia's latest GPUs.
How much of the market they can take is a matter of debate. Switching to different chips can involve a frustrating transition for customers to new software, meaning ASICs are often limited to internal workloads. Morgan Stanley analysts estimate that the custom AI chip share of the market stood at 11% in 2024, and will rise to 15% in 2030.
That might not sound that impressive, but it's a larger piece of a growing pie. By 2030, the AI accelerator market is set to grow to $390 billion from $124 billion in 2024, according to Susquehanna analysts. That means that while Broadcom and Marvell might not be Nvidia killers, they offer valuable diversification in the AI trend.
Custom chips aren't new. Alphabet's Google is a relative veteran, with the search company on the seventh generation of its Tensor Processing Units, or TPUs.
Google's latest custom chips are powerful enough to resemble Nvidia's hardware. They are packing in more high-bandwidth memory, a crucial component for performing the computations behind AI applications.
The latest Google TPU, named Ironwood, has 192 gigabytes of high-bandwidth memory, in line with Nvidia's latest Blackwell chips. Measured by TFLOPS -- the number of trillions of floating-point operations a chip can perform per second -- Ironwood was consistently among the top performers, alongside Nvidia's Blackwell Ultra hardware, according to Susquehanna.
"TPUs excel at the tensor operations fundamental to deep learning and generative AI applications, while GPUs remain highly versatile options for traditional machine learning and tasks involving a broader range of computation," a Google spokesperson told Barron's. "Ironwood shines when serving large and complex generative AI models."
Google is launching a new TPU essentially every year and runs 100% of the training and serving of its Gemini AI model on its own hardware. That's good news for Broadcom, which is the primary partner for the search company's chip project. Google's program likely accounts for more than 80% of Broadcom's AI compute sales, according to BofA Securities.
Broadcom is now entrenched as the No. 2 pick for AI chips behind Nvidia. That is recognized in its valuation, with its stock up 58% in the past 12 months to a recent price of $275.50. It now trades at a forward price/earnings multiple of around 35 times according to FactSet, more expensive than Nvidia's 32 times.
Still, Susquehanna analyst Christopher Rolland has a Positive rating and $300 target price on Broadcom, arguing it is set to take around 11% of the overall AI chip market by 2030, compared with 67% for Nvidia.
Broadcom puts its AI revenue opportunity at between $60 billion and $90 billion by 2027 just from three existing hyperscale customers, generally believed to be Google, Meta, and TikTok parent ByteDance. Broadcom has said it is working with seven customers on custom AI chips in total.
Marvell is the other main custom chip player. Marvell said in December that it had signed a five-year deal with Amazon's cloud-computing business which includes supplying custom AI products, while Microsoft executives have also acknowledged working with the company on custom chips.
Marvell is still a secondary player in AI chips, with data-center revenue of $4.16 billion in its most recent fiscal year, compared with $115.2 billion for Nvidia. However, it's growing quickly -- Marvell's data-center revenue rose 76% in the first quarter of its current fiscal year from the same period a year earlier.
The stock is down 3.9% over the past 12 months and it trades at a forward P/E multiple of 23 times, a significant discount to both Nvidia and Broadcom.
That suggests opportunity for investors, but first Marvell has to convince Wall Street it can win long-term deals.
Marvell recently said it had won two new AI chip projects with unnamed customers. Wall Street analysts have speculated the companies could be OpenAI or Elon Musk's xAI.
"It fundamentally comes down to winning those contracts across multiple generations. That's where having a lot of differentiated IP [intellectual property] is fundamental," William Blair analyst Sebastien Naji says. "The more IP you have the more you can contribute to the overall chip design, and the better you can monetize that relationship."
Naji has an Outperform rating on Marvell stock, arguing the risk-to-reward ratio is in its favor. Marvell recently raised the forecast for its total addressable market in custom AI chips to $55 billion in 2028 from $43 billion.
Nvidia remains the king of the AI chip sector and its annual pace of improvement makes it hard to match. But Amazon, Google, Meta, and Microsoft aren't used to being dependent on a single supplier and that alone means there should be a growing market for custom AI chips.
Write to Adam Clark at adam.clark@barrons.com
To subscribe to Barron's, visit http://www.barrons.com/subscribe
(END) Dow Jones Newswires
July 11, 2025 21:31 ET (01:31 GMT)
Copyright (c) 2025 Dow Jones & Company, Inc.
Comments