Nvidia GPUs are everywhere. Here’s how businesses can access it

Nvidia GPUs are the most popular processors for AI, and they’re in such high demand that chipmakers’ stocks hit another record this week, pushing their market caps to nearly $6 trillion.

Customers can now shop for access to chips in the giant clouds of Amazon, Microsoft, and Google, as well as in so-called neoclouds such as CoreWeave. You can also access various online marketplaces and purchase expensive hardware directly.

All of this adds up to relentless growth for Nvidia, as management expects revenue of $108 billion for the October quarter, an 89% year-over-year increase.

But the paradox of choice can cause headaches for companies in need of computing power.

Cloud infrastructure providers have been at the top of Nvidia’s customer list for several years, but the business is diversifying. Five customers accounted for at least 10% of NVIDIA’s accounts receivable in the July quarter, up from three in January, according to the filing.

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

“We’re going to see a whole new, really exciting neo-cloud with a combined backlog of hundreds of billions of dollars,” Nvidia CEO Jensen Huang said at the Goldman Sachs Technology Conference in San Francisco last month.

Here’s an overview of the different options for accessing GPUs and why each one makes sense.

Many large companies spend tens of millions of dollars annually on various cloud services from Amazon, Google, and Microsoft. Since the announcement of ChatGPT in 2022, enterprises have increasingly turned to GPU hyperscalers to run generative AI workloads.

Top cloud providers have reputational advantages. If your software company relies on Amazon or Microsoft for GPUs and other features, there’s no need to panic about potential customers questioning your suppliers.

Bindu Reddy, CEO of AI assistant startup Abacus, mentioned Microsoft’s cloud infrastructure, saying, “When we talk to enterprises, their subprocessor is better off in Azure.”

Leading AI labs Anthropic and OpenAI have committed more than $500 billion in spending between Amazon and Microsoft in the past year, with Microsoft set to control 59% of the cloud infrastructure market by 2025, according to industry researcher Gartner.

“Hyperscalers are well-positioned to demonstrate confidence to enterprises thanks to more than a decade of full-stack capabilities,” said Gartner analyst Hardeep Singh. But hyperscalers don’t necessarily have as many GPUs as companies need, he said.

Amazon CEO Andy Jassy told analysts in July that the retailer and cloud pioneer would not be able to meet all the demand expected this year.

“We believe this dynamic will hold true in 2027,” he said.

If hyperscalers are in place, neoclouds will not proliferate.

Modal, a startup that operates a virtual sandbox in which AI agents operate independently of the main IT environment, has moved from running on hyperscalers to contracting with major neoclouds and now uses 25 neoclouds, CEO Eric Bernhardson said.

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

Hyperscalers themselves are also chasing neocloud. Google and Microsoft have started using CoreWeave, even in competition with each other.

“Some of the hyperscalers have approached us about how to deal with their concerned customers because they don’t have the ability to serve them when they need it,” said Mark Boroditsky, chief revenue officer at Nevius, a neocloud based in the Netherlands with operations in the United States.

Video generation startup Reactor uses GPUs through Nebius and hyperscalers, CEO Alberto Taiuti said. He said the location of the data center is important because Reactor wants user-generated videos to be viewed immediately. Nebius provides the specific GPUs Reactor requires, solid customer service, and sufficient hardware and software at an affordable price, Taiuti said.

Bernhardson said the most visible neoclouds may require upfront payments, and chips may not come online for months, as providers raise capital and install data center equipment under contracts.

Executive Vice President Chen Goldberg said it would be difficult for CoreWeave to deliver more than 10,000 GPUs to new customers with one day’s notice. CoreWeave’s near-term production capacity remains nearly sold out, CEO Mike Intrator said during an August earnings call.

If your business needs GPUs right away, you may need to go beyond a big name brand. Some neoclouds target specific countries and are therefore not household names, but can be viable in some cases.

“Capacity is tight right now, and supplier relationships are actually one of the best-kept secrets for a company like ours,” said Zhen Lu, CEO of Runpod.

Specialist neoclouds can offer more flexibility than larger GPU clouds, which often require upfront payments or long-term contracts. Some places sell so-called bare metal GPUs. This gives customers more control, but also requires them to manage more technical tasks themselves.

Companies using these smaller neoclouds share the same concerns. When will GPUs be available and at what price? Sunny Smith, co-founder and head of technology at Massed Compute, said customers are often more willing to reserve capacity if they expect prices to rise.

oracleone of the world’s largest cloud providers, allows clients to deploy their own GPUs. The software maker has more debt than Amazon or Microsoft and a lower credit rating, giving it less flexibility to spend heavily on GPUs. But it’s nice to be able to operate the technology.

“With things like bring-your-own hardware, you’re generally able to maintain and improve margins, so the ROIC for these types of structures will be even higher,” Oracle CFO Hilary Maxon told analysts on an earnings call in June, using the acronym for return on invested capital.

Oracle has not named the companies that will choose this path. Guggenheim Securities analyst John DiFucci, who recommends buying Oracle stock, said it makes sense for the top two GPU makers, Advanced Micro Devices and Nvidia, to put their stock into Oracle.

Instead of purchasing thousands of GPUs, early-stage startups with limited funds can rent GPUs for hours at a time at low cost through the cloud. And for companies with heavy computing needs, Oracle’s new route may be more attractive than building an entire data center. OpenAI has committed to spending more than $300 billion over five years with Oracle, but has made no mention of GPU implementation.

OpenAI declined to comment.

This approach may make sense for companies that have the capital to buy AI chips but don’t have enough power, data center space, or skilled labor. Like competing hyperscalers, Oracle is working hard to secure healthy amounts of all three.

Another emerging option is to sign large contracts with companies that rent truckloads of GPUs.

SpaceX has arranged separate deals with hyperscaler Google and open source startup Reflection to hand over excess capacity.

In April, SpaceX agreed to provide GPUs to Cursor and later fully acquired the AI ​​coding startup for $60 billion. And in May, SpaceX signed a deal to rent GPUs to Anthropic for $1.25 billion a month until mid-2029. This is more than most startups can afford.

For SpaceX, however, the numbers work.

“Given the current economic climate, this new capital commitment for computing will pay for itself in less than a year,” Brett Johnsen, the company’s finance director, told analysts in August.

It’s not just SpaceX. CNBC reported in July. Meta was working to create a cloud division that could sell AI computing power.

Meanwhile, as CEOs strive to balance functionality and cost control, enterprises continue to deploy GPU-equipped servers the old-fashioned way in on-premises data centers.

Hardware maker’s enterprise and small business segment revenue nearly doubles lenovo’s Infrastructure Solutions Group for June Quarter. “More and more companies are starting to say, ‘How can we bring AI to our four walls?'” said Vlad Rozanovic, senior vice president.

The hourly spot price of Nvidia B200 GPUs has more than doubled since March, according to data from Ornn, the startup that manages the index.

collaboration software maker drop box CEO Ashraf Alkarmi said the company relies on GPUs in data centers.

“If we want to do more, I think the supply chain linkage will still be useful and a structural advantage,” he said.

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

“In a highly dynamic pricing environment, it’s always good to have multiple sources of information available,” he said.

Spotted: Hightower’s Stephanie Link: I like how Nvidia is showing growth and confidence