NVIDIA GPUs are the most sought-after processors in the field of artificial intelligence, with such strong demand that they have pushed the chipmaker’s stock to a new record this week, bringing its market value close to $6 trillion.
Today, customers can purchase access to computing power on major cloud platforms such as Amazon, Microsoft, and Google, as well as on so-called neoclouds like CoreWeave; they can also buy it on various online marketplaces or directly purchase expensive hardware.
For NVIDIA, all of this means sustained growth. Management expects revenue for the October quarter to reach $108 billion, representing an 89% increase year-over-year.
But for companies that needed computing power yesterday, too many choices may become a problem.
Although cloud infrastructure providers have long topped NVIDIA’s customer list, this business is becoming more diversified. According to a document, five customers accounted for at least 10% of NVIDIA’s accounts receivable in the July quarter, up from three in January.
Industry research firm SemiAnalysis reported that, as of September, the number of NVIDIA GPU suppliers had increased significantly to 323, up from 209 less than 11 months earlier.
NVIDIA CEO Jensen Huang said last month at the Goldman Sachs Technology Conference in San Francisco: "You'll see a wave of entirely new, extremely exciting neoclouds, collectively with hundreds of billions of dollars in order backlogs."
Here are several ways to obtain a GPU, along with potential reasons why each might be suitable:
Hyper-scale cloud
Many large companies spend tens of millions of dollars annually on various cloud services from Amazon, Google, and Microsoft. Since the launch of ChatGPT in 2022, enterprises have increasingly turned to these hyperscale clouds to acquire GPUs for running generative AI workloads.
Leading cloud service providers come with inherent credibility advantages. If a software company relies on Amazon and Microsoft for GPU and other capabilities, it doesn’t need to worry about potential customers questioning its suppliers.
Bindu Reddy, CEO of AI assistant startup Abacus, said, “When you’re talking to enterprise customers, your sub-processor should ideally be Azure.” She is referring to Microsoft’s cloud infrastructure.
Over the past year, leading AI labs Anthropic and OpenAI have committed over $500 billion in spending between Amazon and Microsoft. Industry research firm Gartner states that these two companies controlled 59% of the cloud infrastructure market in 2025.
Gartner analyst Hardeep Singh said, “Hyperscale clouds are well-positioned to demonstrate trustworthiness to enterprises because they have over a decade of full-stack capabilities.” However, he also noted that hyperscale clouds do not always have as many GPUs as enterprises require.
In July, Amazon CEO Andy Jassy told analysts that the retail and cloud pioneer would be unable to meet all of the demand it anticipates.
"I think this situation will be the same in 2027," he said.
Flagship Neocloud
If large-scale clouds were sufficient, neoclouds wouldn't keep emerging.
Modal is a startup that operates a virtual sandbox where AI agents can work independently of the main IT environment. CEO Erik Bernhardsson says the company initially ran on hyperscale clouds, later shifted to partnering with major neoclouds, and now uses 25 of them.
He said, "You might be able to get hundreds of GPUs, or even a thousand, but at our scale, we need more."
Large-scale cloud providers themselves are also pursuing neocloud. Google and Microsoft have begun using CoreWeave, even as they compete with each other.
Marc Boroditsky, Chief Revenue Officer at Nebius, a company headquartered in the Netherlands and operating in the United States, said: “Some hyperscale cloud providers have reached out to us, hoping we could take over customers they’re concerned about because they’re unable to serve those clients when needed.”
Alberto Taiuti, CEO of video generation startup Reactor, said the company uses GPUs through Nebius and hyperscale cloud providers. He emphasized that the location of data centers is critical because Reactor wants user-generated videos to appear instantly. Taiuti noted that Nebius provides the specific GPUs Reactor needs, excellent customer service, and sufficient hardware and software at a competitive price.
Bernhardsson said that the most well-known neoclouds may require some upfront payments, and since providers finance and set up data center equipment under contract, the chips may not be online for several months.
Chen Goldberg, Executive Vice President of CoreWeave, said it would be difficult to transfer 10,000 GPUs to new customers within a single day. Mike Intrator, CEO of CoreWeave, stated on the company’s August earnings call that its short-term capacity is essentially sold out.
Small neocloud
If a company wants to obtain GPUs immediately, it may need to go beyond well-known brands. Some neoclouds are not household names because they target specific countries, and in some cases, this can work.
Runpod CEO Zhen Lu said, “Current capacity is tight, and relationships with suppliers are among the most closely guarded secrets for companies like ours.”
Neocloud offers greater flexibility than large GPU clouds, which often require upfront payments and long-term commitments. Some also offer so-called bare-metal GPUs, giving customers more control but requiring them to handle more technical work themselves.
Companies using these small neoclouds face the same issues: when will they get access to GPUs, and at what price? Sunny Smith, co-founder and head of technology at Massed Compute, says customers are often willing to commit to capacity when they anticipate price increases.
Bring your own hardware to the cloud
One of the world’s largest cloud providers, Oracle, allows customers to bring their own GPUs. Because this software company carries more debt than Amazon or Microsoft and has a lower credit rating, it has less flexibility to make large-scale GPU purchases. However, it is eager to operate these technologies.
Oracle's CFO, Hilary Maxson, told analysts during the June earnings call: "Because we typically maintain and even improve margins in arrangements such as bring-your-own-hardware, these structures have higher ROIC." ROIC stands for return on invested capital.
Oracle has not disclosed the names of the companies that chose this path. Guggenheim Securities analyst John DiFucci recommended buying Oracle stock, noting that it would make sense for the two major GPU manufacturers, AMD and NVIDIA, to bring their GPUs to the cloud.
For early-stage startups with limited capital, renting GPUs by the hour via the cloud is more cost-effective than purchasing thousands of GPUs. For companies with heavy computational demands, Oracle’s new approach may be more attractive than building an entire data center. OpenAI has committed to investing over $300 billion in Oracle over five years, but has not mentioned bringing its own GPUs.
OpenAI declined to comment.
This approach may be ideal for companies that have the funds to purchase AI chips but lack sufficient power, data center space, or skilled labor. Like hyperscale cloud providers, Oracle is also working to ensure adequate availability of all three resources.
Tactical trading
Another emerging option is to enter into large contracts with companies that have substantial GPU resources available for rent.
SpaceX arranged to offload excess capacity through separate deals with hyperscale cloud provider Google and open-source startup Reflection.
In April, SpaceX agreed to provide GPUs to Cursor, followed by acquiring the AI programming startup outright for $60 billion. In May, SpaceX also reached an agreement to lease GPUs to Anthropic until mid-2029 at a monthly cost of $12.5 billion—a expense far beyond what most startups can afford.
However, for SpaceX, these numbers are very favorable.
The company's CFO, Bret Johnsen, told analysts in August: "The current economics have reduced our new computing capital investment payback period to less than one year."
Not just SpaceX. In July, CNBC reported that Meta is building a cloud division that may sell AI computing power.
Return to traditional methods
Meanwhile, companies continue to deploy GPU-equipped servers locally in traditional ways, as executives strive to balance capability with cost control.
Lenovo Infrastructure Solutions Group's enterprise and SMB revenue nearly doubled in the June quarter. Senior Vice President Vlad Rozanovich said, "We're seeing more and more enterprises asking, 'How can I bring AI inside my four walls?'"
According to data from Ornn, the startup that maintains the relevant index, the hourly spot price of NVIDIA’s B200 GPU has more than doubled since March.
Ashraf Alkarmi, CEO of collaboration software provider Dropbox, said the company relies on GPUs in its data centers.
He said, "If we want to do more, I believe our supply chain relationships will still bring benefits and constitute a structural advantage."
Everpure, which sells data center storage hardware and software, has purchased its own GPUs to enable company software engineers to run open-weight AI models. CEO Charlie Giancarlo said, “In a highly dynamic pricing environment, having multiple available sources is always beneficial.”
Watch: Stephanie Link of Hightower says, "I like NVIDIA's demonstrated growth and confidence."