Mark Zuckerberg is trying to compete with Jensen Huang's "darling" for a share of the market.

Is Meta Set to Become the Fourth Cloud in North America?

On May 28th, according to CNBC, Meta CEO Mark Zuckerberg stated that if the company over - invests in data centers and ends up with surplus computing power, Meta might enter the cloud computing market.

Zuckerberg reiterated what he said in last year's earnings call. He said, "Almost every week, different companies come to us. Some hope we can build API services for them, while others ask if we have computing power to sell. They're even willing to buy it at a price higher than our procurement cost."

As the last company among the four major North American internet giants to have hyperscaler - level infrastructure but not yet productize its infrastructure into a public cloud, Meta doesn't lack cloud capabilities; it just hasn't sold its cloud capabilities to others in the past.

So, is Meta really going to create a cloud service to compete head - on with AWS, GCP, and Azure? And who will be the biggest 'victim' when Meta enters the cloud computing market?

01 A New Player in the AI Computing Power Market

Meta is not an old - hand in the public cloud market, but it has been involved in cloud infrastructure for many years.

As early as 2011, Facebook launched the Open Compute Project, making public the servers, power supplies, cabinets, backup power systems, and building designs in its first self - built data center in Prineville, Oregon.

According to Facebook at that time, this self - developed infrastructure increased data center energy efficiency by 38%, reduced construction costs by 24%, and achieved a PUE close to 1.07.

This isn't something an ordinary internet company would do. However, due to the businesses of Facebook, Instagram, WhatsApp, advertising systems, recommendation systems, content review, and short - video distribution, which serve billions of people, they all require a large number of servers, networks, and data centers to support them. So Meta had to invest heavily in data infrastructure.

In 2012, Facebook's capital expenditure was only $1.24 billion; it increased to $2.52 billion by 2015, jumped to $13.92 billion in 2018, and reached $19.24 billion in 2021 before the explosion of generative AI.

These funds have been repeatedly verified in earnings reports and calls to be invested in servers, data centers, and network infrastructure.

So, Meta has always had good cloud capabilities. It just didn't productize these capabilities into a public cloud business that charges externally. But with the arrival of AI, the times have changed.

In the past few years, the four major North American internet companies have almost all been driven by FOMO to do the same thing: pour profits into data centers. Amazon needs to expand AWS and develop the Trainium chip; Microsoft needs to provide computing power for Azure and OpenAI; Google needs to strengthen the foundation for Google Cloud, Gemini, and its search business.

It's easy to understand why they invest in capital expenditure because these three companies are already cloud providers, and data centers are their means of production. But Meta's uniqueness lies in that it's not a public cloud provider, yet it has started to spend money on a scale comparable to cloud providers.

Since 2023, the capital expenditure of the four major North American internet companies has been rising rapidly. Amazon's capital expenditure has soared from $52.7 billion to $200 billion; Google's has increased from $32.3 billion to $185 billion; and Microsoft's has risen from $28.1 billion to $190 billion.

Meta hasn't been left out either.

In 2023, Meta's capital expenditure was only $27.3 billion; it reached $69.7 billion in 2025 and is expected to further reach $135 billion in 2026. Although the absolute value is still slightly lower than the other three, this is far beyond the normal level of a social advertising company's'server - buying' and is a typical hyperscaler - level investment.

From an investment perspective, it's not surprising that Meta wants to enter the cloud market.

It already has data centers, server clusters, a global network, a large number of business scenarios, an internal scheduling system, and years of engineering experience serving billions of users. And in the recent AI wave, it has accumulated a large number of AI chips and huge computing power.

But the question is, will the cloud service Meta might offer be similar to AWS, Azure, or GCP?

Our judgment is that it's unlikely. The reason is simple: the traditional public cloud market is no longer a place where you can enter just by having money to buy servers.

The real barriers for the three major North American CSPs are not just the number of their data centers. Over the past decade, they have captured the IT budgets of enterprise customers, the habits of developers, the partner ecosystem, and the distribution channels of software vendors.

Even if a new public cloud provider has a large number of GPUs, it's difficult to persuade a large enterprise to migrate its core systems, databases, application architectures, and operation and maintenance systems entirely. For customers, switching clouds is not just about changing suppliers; it's about changing the operating system. This is why no new global public cloud giant has emerged in the past decade.

Moreover, Meta doesn't need to do so. In the AI era, what's truly scarce is no longer database and Kubernetes hosting platforms, but NVIDIA's B300, electricity, data center capacity, and the engineering ability to stably schedule these resources.

This is exactly what Meta is most likely to have in surplus.

So, even if Zuckerberg really enters the cloud market, he won't create a general - purpose public cloud for all developers, enterprises, and workloads. Instead, he'll create a more niche, more resource - intensive, and upstream AI infrastructure cloud for large customers.

This is not the narrative of a typical public cloud but the narrative of large - customer computing power procurement.

Under the narrative of 'Meta entering the cloud market', the ones truly affected may not be AWS, Azure, and Google Cloud, but another type of Neocloud that stands between NVIDIA and AI customers and has grown rapidly by supplying GPUs, AI clusters, and reselling computing power.

Among them, the most typical and subtle one is CoreWeave, NVIDIA's 'darling'.

02 Are Meta and CoreWeave Friends or Foes?

CoreWeave's origin is closely related to AI.

It wasn't originally a proper cloud provider but a company that emerged from cryptocurrency mining. In 2017, three Wall Street investors founded the company under the name Atlantic Crypto to engage in mining, and the core asset for mining is GPUs.

However, the good times didn't last long. As the cryptocurrency cycle declined and Ethereum completed its mechanism switch in 2022, mining was no longer a profitable business. So CoreWeave had to seek transformation and started to shift its accumulated GPU infrastructure to a wider range of high - performance computing scenarios.

This transformation wasn't easy. Transforming from a mining company to a cloud company isn't just about changing the official website; it requires improvements in customers, products, scheduling systems, data centers, electricity, networks, and financing capabilities.

But the arrival of the AI wave turned CoreWeave into a rising star.

In 2023, large - model companies, AI application companies, and traditional technology giants were all scrambling for GPUs, and the computing power quotas of AWS, Azure, and Google Cloud were far from enough. At this time, CoreWeave, a Neocloud specializing in GPU clouds, suddenly became a scarce supply in the market.

Its business model is quite straightforward. It first locks in long - term computing power contracts with large customers, then uses these contracts to raise funds to buy GPUs, rent data centers, connect to electricity, and set up servers. Finally, it rents out the GPU clusters and collects payments in installments. More radically, it uses customer contracts and existing GPU assets to support financing, then uses the funds to buy more GPUs and rents them to customers.

Surprisingly, this self - sustaining business model has become a sought - after option for giants in the AI wave.

CoreWeave disclosed in its prospectus that Microsoft was its largest customer in 2023 and 2024, contributing 35% and 62% of the company's revenue respectively.

Then there's OpenAI. In March last year, CoreWeave signed a computing power agreement with OpenAI, and OpenAI promised to pay up to approximately $11.9 billion.

In addition to Microsoft and OpenAI, the customer list disclosed by CoreWeave also includes companies such as Meta, IBM, Mistral, Cohere, and NVIDIA. Especially model companies like Mistral can well illustrate CoreWeave's value. Many enterprises don't have the ability to build their own computing power infrastructure; they need large - scale GPU clusters that can run quickly.

This is CoreWeave's most attractive aspect. It transforms NVIDIA's chips into cloud - based computing power for AI companies, turns the scarcity of GPUs into contract revenue, and turns the anxiety of large - model companies into a growth story that the capital market is willing to invest in.

More importantly, this approach has received strong support from NVIDIA.

Before its IPO, NVIDIA was already an important shareholder of CoreWeave. When CoreWeave went public, NVIDIA also participated in the IPO investment. Later, the two parties signed a deeper capacity agreement. When CoreWeave's data center capacity isn't fully utilized by other customers, NVIDIA will purchase part of the remaining capacity.

The reason why NVIDIA is willing to support CoreWeave with real money is simple: it has become the largest 'pool' for NVIDIA to sell GPUs.

When the market is good, CoreWeave can continue to purchase GPUs, expand data centers, and sign large customers, further amplifying the demand for AI computing power. When the market is not so good, it can also help NVIDIA absorb part of the chip supply and convert hardware sales into cloud - based computing power consumption.

This makes CoreWeave a very special entity: It is not only a major buyer of NVIDIA's GPUs, a distribution channel in NVIDIA's AI computing power ecosystem, but also an important means for NVIDIA to extend its chip demand to the cloud service market.

Reflected in its performance, there has been a soaring revenue and capital expenditure.

In 2022, CoreWeave's revenue was only $16 million; it increased to $229 million in 2023, reached $1.9 billion directly in 2024, and by 2025, its revenue had reached $5.131 billion, a more than 320 - fold increase in three years. In the AI infrastructure industry, this is a textbook - level explosive growth.

But there are also controversies. CoreWeave is not a light - asset software company but a heavy - asset computing power leasing enterprise. While its revenue is rising, its capital expenditure is also expanding rapidly. In 2022, CoreWeave's expenditure on purchasing property and equipment was only $72.4 million; it jumped to $2.943 billion in 2023, further reached $8.702 billion in 2024, and $10.309 billion in 2025.

This means that its growth is built on a large number of GPUs, data centers, debts, and long - term contracts.

Supporters see that CoreWeave seized the window of the most acute shortage of AI computing power and achieved large customers and revenue scale through aggressive expansion. Skeptics, on the other hand, see that the company is highly dependent on NVIDIA GPUs, a small number of large customers, financing ability, and a premise: the shortage of AI computing power must continue.

If GPUs remain in short supply, CoreWeave will be a 'cloud lord' holding scarce resources. But if the demand for AI fades, CoreWeave's story will change from that of a new cloud giant in the AI era to a short - lived phenomenon during the period of computing power shortage.

This is also the most subtle aspect of Meta's potential entry.

Meta may not really create an AWS - style public cloud, nor will it immediately compete with CoreWeave for customers. But as long as it starts selling its surplus AI computing power to large external customers, the market will re - evaluate the scarcity of Neoclouds like CoreWeave.

Because the most valuable part of CoreWeave isn't just the number of GPUs it has today, but the market's belief that in the era of the most acute shortage of AI computing power, it's one of the few entrances that can obtain GPUs preferentially through its 'father - son' relationship with NVIDIA, deliver clusters quickly, and serve large customers.

Ultimately, CoreWeave benefits from the shortage of AI computing power. Once Meta enters the market, the market will start to question how long this shortage will last.

This also leads to the next question: if Meta really opens its data centers and rents out a large amount of computing power, what will the AI cloud market look like?

This article is from the WeChat official account "Super Focus foci", author: Sean. Republished by 36Kr with permission.