This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now.

Startups selling alternatives to Nvidia’s famed graphics processing units are unlikely to truly challenge Nvidia for a long time. But they defiantly claim superior performance, speed, energy efficiency, cost, or all of the above.

Gusto is mandatory for teams wishing to compete against its 10-year headstart, virtually unlimited resources, and more than 70% market share.

The answer to who’s buying Nvidia alternatives is coming into focus alongside the financial value of AI tools and the total volume of AI use.

Related video

“Competitors have a window to really carve out their niche,” Karl Mozurkewich, senior principal AI architect for cloud firm Valdi, told Business Insider.

For efficiency’s sake

Nvidia chips are considered at the cutting edge of accelerated computing. But they’re also expensive and require immense power at scale, and in some cases, companies seek more targeted, energy or space-efficient solutions.

When companies want to diversify away from Nvidia, they often turn to AMD, which also makes a GPU. Mark Papermaster, AMD’s CTO, told BI that moving workloads from one GPU to another is always going to be the easiest move, from a technical perspective. And AMD is directing resources toward its software to make this as easy as possible.

But some industries are particularly keen to look outside those two contenders.

If inference volume is high enough and uniform enough, it makes financial sense to invest in making an alternative chip architecture work for certain workloads, Robert Wachen, cofounder of Etched, one of the newer entrants to the chips space, told BI.

“When you’re inferencing, it really matters to find the right processor that matches your use case,” David Driggers, CEO of Cirrascale Cloud Services, which offers chips from startups Cerebras, SambaNova Systems, as well as Qualcomm, Nvidia, and AMD, told BI.

Here are the industries turning to Nvidia alternatives.

High-frequency trading

High-frequency stock trading, or HFT, is an area where the speed and accuracy of computation is “mission-critical,” Rodrigo Liang, CEO of SambaNova Systems, told BI. Firms like Citadel Securities, Susquehanna, and Jane Street can make millions by getting a fraction of a second ahead of market movements.

“The entry point for the new chips is going to come in these higher value use cases first,” Liang said.

HFT firms hire top machine learning talent, often competing directly with frontier companies like OpenAI and Anthropic.

Jane Street has thousands of H100 and H200 chips from Nvidia, according to the firm’s website. However, the company also participated in Etched’s $120 million Series A in 2024. Etched is betting on transformer models, the kind used for chatbots — models specifically suited to its Sohu chip. The company has also raised funds from PayPal founder Peter Thiel and GitHub CEO Thomas Dohmke, among others.

In addition to a need for speed, HFT firms often require at least some of their computations to be completely private.

“For a long time, hedge funds with truly proprietary code, that were truly trade secret strategies, were the last holdouts to cloud,” Sourcegraph CEO Quinn Slack told BI. On-premises data centers are more likely to have space and energy-related constraints that Nvidia alternatives are looking to seize upon.

Targeting and recommendations

Some of the earliest returns on the billions of dollars invested in AI were from within tech giants’ social media and e-commerce businesses.

“AI has already made us better at targeting and finding the audiences that will be interested in their products than many businesses are themselves, and that keeps improving,” Mark Zuckerberg said on Meta’s April earnings call. Using AI to improve recommendations across the platform has led users of Instagram, Threads, and Facebook to spend more time on the apps.

“There are a lot of opportunities also for us to improve our core business by putting more compute against our ads and recommendation work,” Zuckerberg said.

Wachen told BI recommendations workloads are primed for alternatives to Nvidia, especially as ad-targeting becomes custom-generated ads.

Ad targeting and content recommendations are also major use cases for the chips that cloud firms have developed themselves. Google’s TPU, a chip originally designed in partnership with Broadcom, for instance, is particularly suited for these tasks.

Sovereign AI

Sovereign clouds — state-developed data centers built for national security and other purposes, often share some constraints with the financial services sector.

Related stories Business Insider tells the innovative stories you want to know Business Insider tells the innovative stories you want to know

Saudi entities have been particularly committed to diversifying their chips. Saudi Aramco, for example, has deals with Cerebras, Groq, SambaNova Systems (as well as AMD and Nvidia).

G42, a major Saudi AI data center project, has partnered with Cerebras, AMD, and Nvidia. Humain, the latest entity to emerge from Saudi Arabia’s voracious appetite for AI, which was first announced during President Donald Trump’s recent visit to the Kingdom, counts Groq, Nvidia, and AMD as partners.

Canadian telecom firm Bell Canada just announced Groq as its inference provider. And SambaNova chips are installed in one of SoftBank’s data centers in Japan.

Up the stack

When Meta introduced its Llama 4 model, it also launched a first-of-its-kind API — a direct way for developers to access the model from Meta itself, with computing power coming from Cerebras and Groq. Both companies have their own novel chip architectures; they also offer their own inference service alongside selling chips.

The inference market is crowded. But setting up their own data centers to provide it as a service, allows chip startups to bring in revenue faster than the much longer process of selling the chips themselves. Still, there are tradeoffs.

Hardware choices present both opportunity and risk for companies, said Driggers. Going with an alternative to Nvidia can save on time and money in some cases, but most alternatives are less flexible than a GPU. That the risk of commitment in these early days of AI is what’s keeping many companies from buying non-Nvidia chips, he said. In the meantime, startups are contending with a very finicky inference market.

“If our offering is better, customers will stay,” said Liang. “If somebody else has a better offering, the market will move.”