Nvidia GPUs are everywhere. Here’s how companies access them

Nvidia GPUs are the most sought-after processors in AI, and they’re in such demand that the chipmaker’s stock climbed to yet another record this week, lifting its market cap close to $6 trillion.

Customers can now shop around for access to the chips at the giant clouds from Amazon, Microsoft and Google, as well as at so-called neoclouds like CoreWeave. They can also go to various online marketplaces or even buy the costly hardware directly.

For Nvidia, it all adds up to unrelenting growth, as management anticipates $108 billion in revenue for the October quarter, which would mark an 89% year-over-year jump.

But the paradox of choice can be a headache for companies needing computing power yesterday.

While cloud infrastructure providers have ranked at the top of Nvidia’s customer list for several years, the business is diversifying. Five clients accounted for at least 10% of Nvidia’s accounts receivable in July quarter, up from three in January, according to a filing.

Industry research firm SemiAnalysis counted 323 Nvidia GPU providers as of September, up from 209 less than 11 months earlier.

“You’re going to see a whole new crop of really, really exciting neoclouds with hundreds of billions of dollars backlog together,” Nvidia CEO Jensen Huang said at a Goldman Sachs tech conference in San Francisco last month.

Here’s a rundown of the various options for accessing GPUs, and why each might make sense:

Many big companies spend tens of millions of dollars per year on a smorgasbord of cloud services from Amazon, Google and Microsoft. Since the 2022 launch of ChatGPT, companies have increasingly turned to the hyperscalers for GPUs so they can run generative AI workloads.

The top cloud providers come with a reputation advantage. If a software company relies on Amazon and Microsoft for GPUs and other capabilities, it won’t need to panic about prospective customers questioning its suppliers.

“When you’re talking to enterprises, your subprocessor had better be Azure,” said Bindu Reddy, CEO of AI assistant startup Abacus, referring to Microsoft’s cloud infrastructure.

In the past year, leading AI labs Anthropic and OpenAI have committed to spending over $500 billion between Amazon and Microsoft, which controlled 59% of the cloud infrastructure market in 2025, according to industry researcher Gartner.

“Hyperscalers are in a good position to show trust to the enterprises because of their 10-plus years of full-stack capabilities,” said Gartner analyst Hardeep Singh. But hyperscalers don’t always have as many GPUs as enterprises require, he said.

Andy Jassy, Amazon’s CEO, told analysts in July that the retailer and cloud pioneer won’t be able to serve all the demand it foresees this year.

“I believe this dynamic will also be true in 2027,” he said.

If the hyperscalers were adequate, neoclouds wouldn’t be proliferating.

Modal, a startup operating virtual sandboxes where AI agents work independent of main IT environments, went from running on the hyperscalers to signing up with the major neoclouds, and now it uses 25 of them, said CEO Erik Bernhardsson.

“You can get a few hundred GPUs or maybe a thousand, but at our scale, we needed way more GPUs,” he said.

The hyperscalers themselves are also chasing the neoclouds. Google and Microsoft have started tapping CoreWeave, even as they all compete with one another.

“Some of the hyperscalers have approached us about taking care of customers they’re worried about because they don’t have the ability to service those customers when they need it,” said Marc Boroditsky, chief revenue officer of Nebius, a Netherlands-based neocloud with operations in the U.S.

Video generation startup Reactor uses GPUs through Nebius and hyperscalers, said CEO Alberto Taiuti. The location of data centers matters, he said, because Reactor wants user-created videos to show up right away. Nebius delivers the specific GPUs Reactor needs, solid customer service and sufficient hardware and software, at a good price, Taiuti said.

The most visible neoclouds can require some upfront payment, and chips might not come online for months, because providers raise funding based on contracts and set up the data center equipment, Bernhardsson said.

It would be difficult for CoreWeave to turn over 10,000 GPUs to a new customer with one day’s notice, said Chen Goldberg, an executive vice president. CoreWeave’s near-term capacity remains essentially sold out, CEO Mike Intrator said on the company’s August earnings call.

If companies want GPUs immediately, they might have to go beyond the name brands. Some neoclouds aren’t household names because they target specific countries, which can be viable in certain cases.

“Capacity right now is tight, and your relationships with your suppliers is actually one of the most closely guarded secrets for companies like ours,” said Zhen Lu, CEO of Runpod.

Specialist neoclouds can offer more flexibility than the larger GPU clouds, which often require upfront payments and long-term commitments. Some sell so-called bare-metal GPUs, which give customers more control but also require them to manage more of the technical work themselves.

Companies using these smaller neoclouds share the same concerns: When can they get the GPUs, and at what price? Sunny Smith, co-founder and technology chief at Massed Compute, said customers are often willing to commit to capacity when they expect prices to rise.

Oracle, one of the world’s largest cloud providers, is letting clients bring in their own GPUs. The software maker has more debt than Amazon or Microsoft, and its credit rating is lower so it has less flexibility to go on a GPU spending spree. But it’s happy to operate the technology.

“As we’re generally able to preserve and improve margins in the case of things like bring-your-own-hardware, the ROIC for those types of structures will be even higher,” Oracle CFO Hilary Maxson told analysts on a June earnings call, using the acronym for return on invested capital.

Oracle hasn’t disclosed the names of companies that choose this path. John DiFucci, a Guggenheim Securities analyst who recommends buying Oracle shares, said it would make sense for Advanced Micro Devices and Nvidia, the top two producers of GPUs, to bring their own to Oracle.

Instead of buying thousands of GPUs, early-stage startups with limited capital can borrow them for hours at a time for less through clouds. And for companies with heavy-duty computing needs, Oracle’s new route might be more attractive than building whole data centers. OpenAI committed to spending over $300 billion with Oracle over five years, but it hasn’t mentioned anything about bringing in GPUs.

OpenAI declined to comment.

The method might make sense for companies that have the capital to purchase AI chips but don’t have enough power, data center space or skilled labor. Like its hyperscaler peers, Oracle works hard to secure all three in healthy quantities.

Another emerging option is to cut large deals with firms that have truckloads of GPUs for rent.

SpaceX arranged to turn over excess capacity in separate deals with hyperscaler Google and open-source startup Reflection.

In April, SpaceX agreed to provide Cursor with GPUs and then bought the AI coding startup outright for $60 billion. And in May, SpaceX landed a deal to rent GPUs to Anthropic for $1.25 billion each month through mid-2029. That’s more than what most startups can afford.

For SpaceX, though, the numbers work out nicely.

“The current economics have translated into a less than one-year payback on our new capital deployments for compute,” Bret Johnsen, the company’s finance chief, told analysts in August.

It’s not only SpaceX. In July CNBC reported that Meta was working to form a cloud unit that could sell AI computing power.

Meanwhile, companies continue to install GPU-filled servers in on-premises data centers the old-school way as CEOs work to balance capability with cost control.

Revenue nearly doubled in the enterprise and small and medium business parts of hardware maker Lenovo’s Infrastructure Solutions Group during the June quarter. “We’re seeing more and more enterprises now starting to say, ‘How do I bring AI into my four walls?'” said Vlad Rozanovich, a senior vice president.

The hourly spot price for an Nvidia B200 GPU has more than doubled since March, according to data from Ornn, a startup that maintains indexes.

Collaboration software maker Dropbox relies on GPUs in its data centers, CEO Ashraf Alkarmi said.

“If we want to do a lot more, I think our supply chain connections will still be beneficial and a structural advantage,” he said.

Everpure, which sells data center storage hardware and software, has acquired its own GPUs, to run open-weight AI models for the company’s software engineers, said CEO Charlie Giancarlo.

“In a very dynamic pricing environment, it’s always good to have multiple sources that you can go to,” he said.

WATCH: Hightower’s Stephanie Link: I like that Nvidia is showing growth and confidence