NVIDIA GPU is the most sought-after processor in the field of artificial intelligence, and the high demand has even driven the stock price of this chip manufacturer to set a new record this week, with a market value approaching 6 trillion dollars.

Today, customers can purchase access to computing power on major cloud platforms such as Amazon, Microsoft, and Google, as well as on so-called neocloud like CoreWeave; they can also go to various online markets or even directly buy expensive hardware.

For NVIDIA, all of this means continued growth. Management expects revenue for the October quarter to reach $108 billion, which represents an increase of 89% compared to the same period last year.

But for those companies that need computing power yesterday, having too many options can actually become a problem.

Although cloud infrastructure providers have long been among the top of NVIDIA's customer list, this business is becoming more diversified. According to a document, in the July quarter, there were 5 customers accounting for at least 10% of NVIDIA's accounts receivable, compared to 3 in January.

Industry research firm SemiAnalysis stated that as of September, the number of NVIDIA GPU suppliers had reached 323, a significant increase from 209 just over 11 months prior.

NVIDIA CEO Jensen Huang said at the Goldman Sachs Technology Conference held in San Francisco last month, "You will see a large number of brand-new, extremely exciting neocloud, which together have orders backlog worth hundreds of billions of dollars."

The following are several ways to obtain GPU, as well as the possible reasons why each method might be applicable:

Ultra-large-scale cloud

Many large companies spend tens of millions of dollars each year on various cloud services provided by Amazon, Google, and Microsoft. Since the launch of ChatGPT in 2022, more and more enterprises have turned to these hyperscale clouds to obtain GPU for running generative AI workloads.

Leading cloud service providers come with an inherent advantage in terms of credibility. If a software company relies on Amazon and Microsoft to provide GPU and other capabilities, it doesn't have to worry about potential customers questioning its suppliers.

AI Assistant, the CEO of Abacus startup, Bindu Reddy said, "When communicating with corporate clients, it's best if your sub-processors are Azure." She was referring to Microsoft's cloud infrastructure.

Over the past year, leading AI laboratories Anthropic and OpenAI have committed to investing over $500 billion between Amazon and Microsoft. According to industry research firm Gartner, these two companies will control 59% of the cloud infrastructure market by 2025.

Gartner Analyst Hardeep Singh said, "Ultra-large-scale clouds are in a favorable position when it comes to demonstrating credibility to enterprises, as they possess full-stack capabilities of over 10 years." However, he also noted that ultra-large-scale clouds do not always have as much GPU as enterprises require.

Amazon CEO Andy Jassy stated to analysts in July this year that the retail and cloud pioneer would not be able to meet all of the anticipated demands.

“I think this will be the case in 2027 as well,” he said.

Flagship neocloud

If ultra-large-scale clouds are already sufficient, then neocloud would not continue to emerge.

Modal is a startup company that operates virtual sandboxes, where AI agents can work independently from the main IT environment. The CEO, Erik Bernhardsson, stated that the company initially ran on hyperscale clouds, later switched to collaborating with major neocloud, and now utilizes 25 of them.

He said, "You might be able to get a few hundred GPU, or maybe a thousand, but with our scale, we need more GPU."

Ultra-large-scale clouds themselves are also pursuing neocloud. Google and Microsoft have already started using CoreWeave, although they are also competing with each other.

neocloud, a company headquartered in the Netherlands and operating in the United States, Nebius, Chief Revenue Officer of neocloud, said: "Some very large-scale clouds have already approached us, hoping that we can take over the customers they are concerned about, as they do not have the capability to provide services to those customers when needed."

The CEO of the video generation startup Reactor, Alberto Taiuti, stated that the company utilizes Nebius and ultra-large-scale cloud computing for its operations. He emphasized that the location of the data center is crucial, as Reactor aims for users to see the generated videos immediately. Taiuti mentioned that Nebius can provide the specific Reactor required, excellent customer service, as well as sufficient hardware and software at a reasonable price.

Bernhardsson said that the most well-known neocloud may require some upfront payment, and since the provider will raise funds according to the contract and set up the data center equipment, the chips may not be available for several months.

CoreWeave Executive Vice President Chen Goldberg said that it would be very difficult to transfer 10,000 GPU to new customers within a single day. CoreWeave Chief Executive Officer Mike Intrator stated during the company's August financial report conference call that its short-term production capacity has basically been sold out.

Small neocloud

If a company wants to obtain GPU immediately, it may have to surpass those well-known brands. Some neocloud are not household names, as they target specific countries, and in some cases, this approach can also be viable.

Runpod CEO Zhen Lu said, "Currently, production capacity is very tight, and the relationship with suppliers is actually one of the best-kept secrets for companies like ours."

Professional neocloud can offer more flexibility than large-scale GPU clouds, which often require upfront payments and long-term commitments. Some also sell so-called bare metal GPU, allowing customers to have more control, but they also need to take on more technical work themselves.

Companies that use these small neocloud face the same issue: when will they be able to obtain GPU, and what will the price be? Massed Compute, co-founder and technical director Sunny Smith, said that customers are usually willing to commit to production capacity when they expect prices to rise.

Bring your own hardware to the cloud

One of the world's largest cloud service providers, Oracle, allows its customers to bring in their own GPU. Since this software company carries more debt than Amazon or Microsoft and has a lower credit rating, it does not have as much flexibility to make large-scale purchases of GPU. However, it is willing to operate these technologies.

Oracle's Chief Financial Officer Hilary Maxson said to analysts during the June earnings conference call, "Since we are usually able to maintain and improve profit margins in arrangements that involve our own hardware, the ROIC of such structures will be higher." ROIC refers to return on invested capital.

Oracle has not disclosed the names of the companies that have chosen this path. Analyst John DiFucci recommends buying Oracle stocks, stating that it would be reasonable for the two major manufacturers AMD and NVIDIA to move to the cloud on their own.

For early-stage startups with limited capital, it is more cost-effective to rent by the hour through the cloud rather than purchasing thousands of GPU. For companies with high computational demands, Oracle's new approach may be more attractive than building their own data centers. OpenAI has committed to investing over $300 billion in Oracle within five years, but has not mentioned that they will bring their own GPU.

OpenAI refuses to comment.

For those companies that have the funds to purchase AI chips but lack power, data center space, or skilled labor, this approach may be suitable. Similar to its large-scale cloud competitors, Oracle is also working hard to ensure an adequate supply of these three resources.

Tactical trading

Another emerging option is to sign large contracts with companies that have a large amount of GPU available for rent.

SpaceX arranged to transfer excess production capacity through separate transactions with the super-large-scale cloud provider Google and the open-source startup Reflection.

In April this year, SpaceX also offered GPU to Cursor, and subsequently acquired this AI programming startup for a total of 60 billion US dollars. In May, SpaceX also reached an agreement to lease GPU to Anthropic for a term until mid-2029, at a cost of 12.5 billion US dollars per month. This cost far exceeds what most startups can afford.

However, for SpaceX, these numbers are quite advantageous.

The company's financial director, Bret Johnsen, stated to analysts in August: "The current economic conditions have reduced the payback period for our new capital investments to less than one year."

Not only SpaceX. In July, CNBC reported that Meta was establishing a cloud department and might sell AI computing power.

Return to traditional methods

Meanwhile, companies are still deploying servers containing GPU in their local data centers using traditional methods, while CEOs are striving to find a balance between capacity and cost control.

Lenovo's Infrastructure Solutions Group, a hardware manufacturer, saw its business revenue from enterprises and small and medium-sized businesses almost double in the June quarter. Vlad Rozanovich, the senior vice president of the group, said, "We are seeing more and more enterprises starting to ask, 'How can I bring AI into my four walls?'"

According to data from Ornn, a startup company that maintains relevant indices, the hourly spot price of NVIDIA B200 GPU has more than doubled since March.

The CEO of the collaboration software manufacturer Dropbox, Ashraf Alkarmi, stated that the company relies on GPU within its data center.

He said, "If we want to do more things, 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 GPU for the company's software engineers to run the open-weight AI model. CEO Charlie Giancarlo said, "In an environment where prices are highly dynamic, it's always a good thing to have multiple sources to turn to."

Watch: Hightower's Stephanie Link says, I like the growth and confidence demonstrated by NVIDIA.