Compute Is the New Oil: Why Meta Might Rent Its Data Centers to a Rival

The reported Anthropic–Meta talks reveal a new reality in AI: the scarce resource is no longer just talent, chips or models. It is access to industrial-scale compute — and the companies that own it may become AI’s landlords.

Meta and Anthropic are not natural partners.

One is the owner of Facebook, Instagram and WhatsApp, building its own frontier AI systems and open-weight Llama models. The other is the fast-growing maker of Claude, backed by Amazon and Google and competing directly for enterprise AI users, researchers and talent.

Yet Meta is reportedly in preliminary discussions to lease computing capacity to Anthropic in a deal that could be worth up to $10 billion over two years. Neither company has confirmed a final agreement, and the talks could still go nowhere. But even the possibility matters.

It suggests that in the AI era, a company may be willing to rent out the most strategic part of its operation — data-centre capacity — to a rival, provided the price is high enough.

That is a profound shift. It is also why “compute is the new oil” is more than a catchy metaphor.

The real AI bottleneck is becoming physical

For years, AI was discussed as software: smarter models, better algorithms and more useful chatbots. But the race has rapidly become an industrial one.

Frontier-model development now requires huge clusters of specialised chips, extraordinary quantities of electricity, high-speed networking, cooling systems, land, construction capacity and years of planning. A lab cannot simply decide to train a larger model on Monday and have the necessary infrastructure by Friday.

Even companies with billions in funding can find themselves compute-constrained. Anthropic has already pursued large-scale capacity arrangements across the industry because access to enough reliable compute can determine how quickly it can train new models, serve enterprise customers and keep pace with OpenAI, Google, Meta and others.

To understand why this matters, it helps to remember that large language models are not simply apps. They are vast systems trained and run across expensive, highly specialised hardware. The smarter and more widely used they become, the more the physical infrastructure beneath them starts to matter.

This is the uncomfortable truth beneath the AI boom: the best model is not always the model with the cleverest research team. It may be the model backed by the most megawatts.

Meta’s potential pivot: from AI builder to AI landlord

Meta has historically built enormous infrastructure for itself. Its data centres supported social feeds, video, messaging, ads and, more recently, the company’s AI ambitions.

Now, after a colossal investment cycle, it appears to be considering whether parts of that infrastructure could become a product in their own right.

The reported Anthropic proposal would make Meta something closer to a specialist AI cloud provider — competing, at least in part, with Amazon Web Services, Microsoft Azure and Google Cloud. Rather than only using its chips and data centres to improve Facebook’s recommendations or train Llama, Meta could sell access to raw computing power.

The logic is hard to ignore.

Data-centre capacity is expensive to build and often arrives in large blocks. Internal demand is not always perfectly aligned with construction schedules. If Meta has capacity coming online before every internal workload is ready to consume it, a major external tenant could turn a cost centre into a revenue stream.

For Anthropic, the logic is equally clear. It does not need Meta to become a long-term ally. It needs dependable capacity now.

In oil terms, the companies may compete to sell petrol, but they can still buy crude from the same producer.

The $700 billion infrastructure binge

The backdrop is an infrastructure spending race unlike anything the technology sector has seen before.

Amazon, Alphabet, Microsoft, Meta and Oracle are collectively expected to spend well above $700 billion on capital expenditure in 2026, much of it directed towards AI-related data centres, chips, networking and energy infrastructure.

That spending is not merely a bet on demand for AI products. It is an attempt to secure a strategic position before the physical bottlenecks become even tighter.

Owning models matters. Owning customers matters. But owning the land, power agreements, fibre connections, chip supply and operating expertise may prove just as important.

This is why data centres are increasingly being treated less like back-office technology assets and more like ports, pipelines or power stations. Once a site has secured several hundred megawatts of electricity, planning approval, transmission access and installed GPU capacity, it becomes extremely difficult for a competitor to replicate quickly.

The result is a market where compute itself can be packaged, contracted, sublet and traded.

That has consequences for the wider AI race too. The debate over open-source versus closed AI models often focuses on access to the software. But access to the underlying compute could become just as important. An open model is far less open in practice if only a small circle of companies can afford to run it at serious scale.

Nuclear deals are the clearest sign of the shortage

The move towards nuclear power shows how seriously Big Tech now views the energy constraint.

Microsoft’s agreement connected to the planned restart of Three Mile Island’s Unit 1, Google’s plans with Kairos Power, and Amazon’s nuclear investments and agreements are all part of the same message: intermittent renewable power alone will not be enough for every AI workload, especially where companies want round-the-clock, predictable electricity.

Meta, too, has been associated with efforts to secure long-term nuclear capacity.

Nuclear is not a quick fix. New reactor projects take years, are expensive and face regulatory and construction risk. Small modular reactors remain promising rather than proven at the scale the industry wants. But that is precisely the point: the biggest AI players are making energy commitments for the early 2030s because they expect today’s compute shortage to remain strategically important for years.

The race is no longer just for Nvidia chips. It is for the electrons that keep those chips running. The International Energy Agency’s analysis of AI and energy demand makes clear how closely the future of AI deployment is tied to the availability of electricity.

Communities are pushing back

The industry is also discovering that money does not automatically buy permission to build.

Data centres bring construction spending, local tax revenue and jobs. But they also create concerns over electricity prices, water consumption, noise, diesel backup generators, land use and whether residents are effectively subsidising private AI infrastructure through grid upgrades.

New York has introduced a one-year pause on new hyperscale data-centre development while it develops stronger rules around local benefits, energy and environmental impacts. Its official executive order points to nearly 12 gigawatts of data-centre load requests in the state’s interconnection queue as of May 2026.

Ireland offers another warning about what happens when data-centre development becomes concentrated. Data centres consumed roughly 23% of the country’s electricity in 2025, nearly as much as all households combined. The country has already had to tighten rules around new grid connections.

These tensions will matter commercially. A company with an already-operational, powered site may have an advantage that cannot be recreated simply by announcing another multi-billion-dollar data-centre campus.

Scarcity raises prices. Regulation can make scarcity even more valuable.

Will hyperscalers become AI landlords?

The traditional cloud giants already rent computing infrastructure to customers. What is changing is who might join them and what exactly is being rented.

Meta has not historically been viewed as a cloud provider in the same category as Amazon, Microsoft and Google. But if it begins leasing AI capacity to companies such as Anthropic, the lines blur quickly.

The future could look less like a handful of software firms selling AI subscriptions and more like a layered economy:

• Frontier labs build models.
• Hyperscalers finance and operate giant AI campuses.
• Energy companies supply dedicated power.
• Chip makers sell the scarce equipment.
• Landlords of compute rent capacity to everyone else.

Some companies will operate across all of those layers. That may be the most powerful position of all.

As AI agents become more capable and widely used, demand will not come only from occasional model-training runs. It will also come from millions of ongoing AI tasks: research, coding, customer support, analysis and automated workflows. That turns compute into a recurring operating requirement, not a one-off research expense.

The danger, of course, is that this makes AI even more concentrated. If only a small number of firms can finance gigawatt-scale infrastructure, then startups and governments may become dependent on a handful of compute landlords.

But it also creates an opportunity. Renting capacity from rivals could make the market more flexible than a world where every AI lab is trapped inside one cloud ecosystem.

The Anthropic–Meta talks are bigger than one deal

A $10 billion agreement would be striking in its own right. But the more important story is what it represents.

Meta potentially leasing compute to Anthropic is a sign that infrastructure is becoming valuable enough to transcend ordinary competitive boundaries. The company with spare capacity is not simply sitting on unused servers; it may be holding a scarce industrial asset that another AI contender urgently needs.

Oil transformed economies because it powered everything else. Compute is beginning to play a similar role in artificial intelligence.

The winners may not just be those who create the smartest models. They may be the companies that control the wells.