Meta Compute Deals: Where Meta Rents AI Capacity and Why

Training and running a large AI model takes rooms full of special chips. It also takes the power to feed them and the staff to keep them cool and online. A company that wants that much computing power has two plain choices: build its own data centers, which takes years, or rent time on someone else's machines and start within months.

Renting looks simple until you read the contract. The biggest deals lock the customer into paying for a block of servers for five or six years, whether or not it ends up using all of them. Those payments rarely show up in the headline spending figure, so a reader who checks only that number misses a large part of the bill.

Investors track rented computing in two places. The supplier often makes the contract and its value public, while the customer lists what it owes in a note to its quarterly accounts. Reading both side by side shows how much has been promised, to whom and for how long.

Meta Platforms owns Facebook, Instagram and WhatsApp. It rents AI computing power from at least six outside companies, and the payments it cannot cancel, for rented computing, servers and data centers together, now run into the hundreds of billions of dollars. BloFin gives you two ways to follow those bills through the share price: own Meta as the METAX token through METAX/USDT Spot trading, or trade its rise or fall with leverage through the METAUSDT Perpetual.

What are Meta's compute deals?

Meta's compute deals are contracts to rent AI computing power from cloud firms such as CoreWeave, Nebius and Google Cloud. On June 30, 2026, Meta had $349.3 billion of contractual commitments it cannot cancel. It says they mostly relate to third-party cloud capacity arrangements and to investments in servers, network gear, data centers and Reality Labs hardware (source: Meta Form 10-Q for the June 2026 quarter).

Meta does not split that total by supplier or by type, so no public figure shows how much of it is rented cloud computing (source: Meta Form 10-Q for the June 2026 quarter). Cloud capacity is the first item Meta names, and it is the only one Meta pays for as a service; the others are investments in things Meta will own.

Renting is one of three ways Meta pays for its AI build, and each one lands in a different place in the accounts. The first is buying servers and buildings with its own cash, which shows up as capital spending, or capex. The second is a long lease on a data center that a partner owns, so Meta pays rent on the building but runs its own servers inside it. The third is renting finished computing power from a cloud company that owns and runs the machines.

Say you rent a car for a week. You never buy the car, you never service it, and you pay the rental company for the days you drive it. A cloud contract works the same way for servers, but the rental runs for years and the price is set in advance. None of the rent paid to cloud companies is counted in Meta's AI capex, the money Meta spends on its own machines and buildings.

Meta's cloud and neocloud contracts

Meta has signed or is reported to have signed at least eight large compute deals with six outside suppliers since September 2025. Only CoreWeave and Nebius have published what their contracts are worth. Google Cloud, Amazon Web Services, Oracle and Crusoe have given no figure, so the values for those deals come from news reports or are missing.

The deals with a named value come from the supplier, because a contract this large matters to a smaller company's results and it has to tell its investors. Meta's commitments note does not name the cloud companies. So for each deal, the supplier's own word is the best source, and a news report is the next best when no company has spoken.

CoreWeave's two contracts with Meta

CoreWeave's first disclosed order from Meta was signed on September 25, 2025. Under it, Meta initially committed to pay CoreWeave up to about $14.2 billion through December 14, 2031, as long as CoreWeave delivers the service (source: CoreWeave Form 8-K). The same filing gave Meta the option to expand its commitment through 2032, and it says either side may end the agreement for cause.

Meta widened the relationship on April 9, 2026, when the two companies announced a second deal worth about $21 billion. It runs through December 2032, spreads the servers across several sites, and includes some of the first systems built on Nvidia's Vera Rubin platform (source: CoreWeave's April 2026 announcement). The two contracts add up to about $35.2 billion.

Nebius and the $15 billion backstop

Nebius, a cloud company based in Amsterdam, announced a contract with Meta in November 2025, worth about $3 billion over five years (source: Nebius third-quarter 2025 results). A larger one, signed on March 16, 2026, is worth up to about $27 billion and has two parts. The first is $12 billion of servers set aside for Meta, delivered from early 2027 on Nvidia's Vera Rubin platform (source: Nebius announcement).

The second part is a backstop, a promise to buy whatever other customers do not. Meta committed to buy spare time on some new Nebius clusters, up to $15 billion over five years. Nebius plans to sell that time to other customers first, and Meta takes what is left (source: Nebius announcement).

Meta's own filing describes a commitment of the same shape, without naming the supplier. It lists contingent obligations, promises that apply only in some cases, to buy up to $14.7 billion of cloud computing over five years. The amount can fall if the cloud provider sells it to other customers (source: Meta Form 10-Q for the June 2026 quarter).

Google Cloud, Amazon, Oracle and Crusoe

In February 2026, The Information reported a multibillion-dollar deal for Meta to rent Google's own AI chips through Google Cloud. The chips are called tensor processing units, or TPUs. The same report said the two companies were also discussing a purchase of TPUs for Meta's own data centers, with no agreement reached (source: SiliconANGLE on The Information's report).

The Amazon deal is the one Meta announced itself. On April 24, 2026, Meta said it would bring tens of millions of Amazon's Graviton processor cores into its computing setup, with room to grow (source: Meta's AWS announcement). Graviton chips are general processors that Amazon designs for its own cloud, alongside the AI chips Amazon builds itself. Meta says AI agents, systems that reason, plan and carry out complex tasks, need more processor power, and it wants Graviton chips for that work (source: Meta's AWS announcement).

Oracle's deal has a reported value and a confirmed customer but no confirmed figure. On September 19, 2025, Reuters reported that the two were negotiating a contract of about $20 billion, and Bloomberg that its value could exceed $20 billion (source: SiliconANGLE on the Reuters and Bloomberg reports). In its December 10, 2025 results, Oracle said its future contracted revenue was "highlighted by new commitments from Meta, NVIDIA, and others" and gave no amount for any of them (source: Oracle second-quarter fiscal 2026 results).

Crusoe is a company that builds and runs AI data centers. On June 18, 2026, Bloomberg reported that Meta had contracts for roughly 1.6 gigawatts of computing power at two Crusoe sites. They are in Childress, Texas and Warrenton, Missouri, and the report cited people familiar with the deals. Neither Meta nor Crusoe has given a value for them (source: Bloomberg on the Crusoe deals).

Why Meta rents computing power it could build

Meta rents because it needs more computing power than its own data centers can deliver in time. A rented cluster can start within months, while a new campus takes years and a large power supply. Meta's own reason is that "diversifying our compute sources is a strategic imperative", in the words of its head of infrastructure, Santosh Janardhan (source: Meta's AWS announcement).

The same post sets out how Meta divides the work. "We invest in our own data centers and custom hardware," it says, and "we partner with cloud providers who bring differentiated capabilities" (source: Meta's AWS announcement). In plain terms, Meta builds what it needs every day and rents what someone else can supply faster or better.

Speed is the first reason Meta rents. Nebius delivers Meta's dedicated servers from early 2027, a little under a year after signing, and CoreWeave's newer contract includes some of the first Vera Rubin systems.

The second reason is choice of chip. Meta rents Nvidia graphics chips from CoreWeave and Nebius, Google's TPUs through Google Cloud, and Amazon's Graviton processors, alongside the chips it designs itself. Some AI tasks run better on one chip and some on another, so spreading the work keeps Meta from leaning on one supplier's prices and delivery dates.

The third reason is flexibility at the end of the contract. When a five-year cloud deal ends, Meta can renew it, move the work to its own data centers or walk away, and the cloud company keeps the old chips. Owning those chips would leave Meta holding servers that lose value as newer models arrive.

The cost is that renting gives up some control. Meta needs the supplier to deliver on time, and a cloud company has its own debts to pay. Meta counts contracts like CoreWeave's among the commitments it cannot cancel, so it keeps paying even if it later needs less computing power.

How take-or-pay cloud contracts work

A take-or-pay contract makes the buyer pay for the servers it booked, whether or not it uses all of them. CoreWeave, Meta's largest disclosed supplier, says its committed contracts are take-or-pay, while its on-demand customers pay only for what they use (source: CoreWeave Form 10-K for 2025).

The supplier asks for this because it has to buy the chips and pay for power before the customer uses a single hour. For Meta, the trade is a guaranteed supply in return for a fixed bill.

Cost per hour actually used equals the agreed hourly rate divided by the share of reserved hours you use.

Say you are Meta and you reserve servers for $1 billion a year, an illustrative round figure. If your engineers keep them busy all year, each hour costs exactly what you agreed. If demand drops and you use only 80% of the hours, you still pay $1 billion, so each hour you do use costs 25% more.

That is the risk a take-or-pay contract carries for the buyer. The bill stays fixed while the use can fall. Meta has chosen to avoid the opposite risk: running short of computing power while rivals train bigger models.

The Nebius backstop works differently, because the amount Meta pays depends on how well Nebius sells. Suppose Nebius sells $9 billion of the $15 billion to other customers. Meta would then owe up to the remaining $6 billion. If Nebius sells all of it, Meta owes nothing under that part of the contract.

Meta's cloud commitments in its quarterly filings

On September 30, 2025, Meta's commitments it cannot cancel stood at $81.2 billion (source: Meta Form 10-Q for the September 2025 quarter). By June 30, 2026 they had reached $349.3 billion, 4.3 times as much in nine months. Each filing says they mostly relate to two things: third-party cloud capacity arrangements, and Meta's own investments in servers, network gear, data centers and Reality Labs hardware (source: Meta Form 10-Q for the June 2026 quarter).

The December figure comes from Meta's annual report (source: Meta Form 10-K for 2025), and the June figure from its quarterly filing for that quarter (source: Meta Form 10-Q for the June 2026 quarter). The total grew by $218.3 billion in the first half of 2026, the months in which Meta's deals with Google, Nebius, CoreWeave, Amazon and Crusoe were announced or reported.

The timing of the payments matters as much as the total. Of the $349.3 billion, about $53.5 billion falls due in 2026 and about $81.7 billion in 2027 (source: Meta Form 10-Q for the June 2026 quarter). That leaves about $214.1 billion due in 2028 and later, which fits contracts that run to 2031 and 2032.

Set the contracts with published values next to that total. CoreWeave's two contracts at $35.2 billion, Nebius's $3 billion contract from November 2025 and its $12 billion of dedicated servers come to $50.2 billion, about 14.4% of $349.3 billion. Meta does not say which deals sit inside its total, so the comparison shows scale and nothing more.

Why the commitments stay off the balance sheet

Meta lists its cloud capacity arrangements among its contractual commitments, apart from its leases. So the promise sits in a note, a written part of the accounts that comes after the main tables (source: Meta Form 10-Q for the June 2026 quarter). It is left off the balance sheet, the statement of what a company owns and what it owes. A data center lease is different. Once a lease starts, Meta adds the rent it still owes to its balance sheet as a lease liability. Only the leases that have not yet begun go in that note, a route explained in how Meta is financing the AI build.

When Meta uses rented servers, the rent shows up among its costs. In its June 2026 quarter, Meta named "third-party cloud services" among the reasons its costs rose, alongside its own data centers (source: Meta Form 10-Q for the June 2026 quarter). Capital spending, by contrast, reaches profit slowly, as depreciation spread over a server's life.

So rented capacity changes how Meta's spending reads. A dollar spent on a server shows up in capex at once and in profit over several years. A dollar of cloud rent never appears in capex and hits profit in the quarter it is used. Read only the capex figure, and you miss a part of the bill that never shows up there.

Meta Compute and selling spare capacity

Meta Compute is the unit Meta set up on January 12, 2026 to plan and build the computing power behind its AI. Mark Zuckerberg called it a "top-level" initiative, led by Santosh Janardhan and Daniel Gross. He said Meta plans to build tens of gigawatts this decade, "and hundreds of gigawatts or more over time" (source: Axios on the Meta Compute launch).

In January the unit was about getting enough computing power for Meta's own apps and models. A later report described a second purpose for it. On July 1, 2026, Bloomberg reported that Meta was forming a business to sell its spare computing power to outside customers. That would put it up against Amazon Web Services, Microsoft Azure and Google Cloud (source: Bloomberg on Meta's cloud plans). Ynet, citing Bloomberg, reported that the effort is part of Meta Compute (source: Ynet on Meta's cloud plans).

The idea flips Meta's usual role in these deals. Meta rents from CoreWeave and Nebius because it needs more, and a Meta cloud would sell because it has more than it needs at some times. Both can be true at once, because a company that reserves servers years ahead will sometimes hold more than its teams use, and selling those hours softens the take-or-pay risk.

Two weeks later came the first named possible customer. On July 17, 2026, CNBC reported that Anthropic, the maker of the Claude AI models, was in very early talks to lease computing power from Meta. The New York Times put a possible deal at up to about $10 billion (source: CNBC on the Anthropic talks).

No sale has been confirmed since then. On Meta's July 29, 2026 results call, Mark Zuckerberg listed "potentially selling compute directly" among the things Meta could offer businesses (source: TechCrunch on the call). On September 28, 2026 Meta launched Meta Enterprise Platform to sell its AI models and agents to businesses. It named its large-scale infrastructure as a strength but announced no computing-power product (source: Meta's Enterprise Platform announcement). Meta has not announced a cloud customer, a price list or a separate reporting line, so selling compute is an option Meta has named, not a business it reports.

Anthropic's request and the October 2026 report

On October 9, 2026, The Wall Street Journal reported that Meta had turned down a request for computing power from Anthropic. Its chief executive, Dario Amodei, had asked Meta's chief AI officer, Alexandr Wang, earlier in the year. The report cited people familiar with the talks (source: The Wall Street Journal via Techmeme).

The report placed the request in a wider race for computing power, in which rivals team up and top executives get involved themselves (source: The Wall Street Journal via Techmeme). The report gives no sign of a sale.

Two facts help explain a refusal like this one. Anthropic builds AI models that compete with Meta's own, which makes it one of Meta's competitors in AI. Renting it servers would earn Meta money while helping a rival train and serve better models.

The other is Meta's own demand. Every new model Meta releases needs more computing power, first to train it and then to answer users. Meta's Muse AI runs inside apps used by billions of people, so spare hours can disappear quickly.

What to watch in Meta's third-quarter report

Wednesday, October 21, 2026 is the date Meta set for its third-quarter results, which it releases once Nasdaq has shut for the day (source: Meta Investor Relations). Its quarterly filing will update the commitments figure, last given as $349.3 billion on June 30, and show whether the contingent cloud obligation of $14.7 billion has changed.

Beside the revenue and profit figures that lead Meta's earnings report, four things in it speak to rented capacity. The first is the new commitments total and how much of it falls due in 2027. The second is whether Meta still lists "third-party cloud services" among the reasons its costs rose. The third is any comment on selling compute through Meta Compute or Meta Enterprise Platform, such as a first customer or a decision to keep every server for itself. The fourth is any new supplier named on the call.

Each of the four tells you something different. A higher total means more promises signed in the summer. A larger 2027 figure means the bill arrives sooner. A first Meta Compute customer would mean some of that bill could be offset by outside revenue.

The figures land after the close, and METAX trades on a 24-hour weekday schedule, so it can react while Nasdaq is shut. You can watch the token's first reaction to the commitments figure on the METAX/USDT Spot page. For a leveraged view, the METAUSDT Perpetual page puts the contract's price next to its funding rate, a fee that traders on one side of the contract pay to the other at set times. Outside the Nasdaq session, BloFin may accept only orders that reduce a position on the METAUSDT Perpetual (source: BloFin stock futures notice). On the night of the report, you may be able to close or shrink a position but not open one until the next session.

What rented capacity means for METAX and METAUSDT holders

METAX follows the price of Meta's shares, so the compute deals reach a token holder through what they do to Meta's profit and cash, two of the things that move Meta's stock price. The commitments are fixed costs that Meta pays whether its ad business grows fast or slowly.

That fixed bill cuts both ways. In a strong year, rented capacity lets Meta launch new models and features quickly, and the cost spread over more revenue looks small. In a weak year, the payments stay the same while revenue slows, so profit falls faster than sales. A token that tracks Meta, as tokenized Meta (METAX) does, carries that same trade-off.

Leverage makes any move bigger for a trader. If a report shows commitments growing faster than expected, the shares can fall sharply after the close, and a leveraged METAUSDT Perpetual position gains or loses several times that move on its margin. Size any position so that a gap after the report cannot wipe out your margin when trading Meta around earnings.

Looking to trade Meta around its next figures on rented computing power? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the METAX/USDT Spot trading page or the METAUSDT Perpetual trading page.

Frequently asked questions

What is a neocloud?

A neocloud is a cloud company built mainly to rent out graphics chips for AI work, a narrower business than a general cloud such as Amazon Web Services. CoreWeave, Nebius and Crusoe are the best-known examples. CoreWeave sells most of its computing time under committed take-or-pay contracts lasting years, and sells the rest on demand, with customers paying as they go (source: CoreWeave Form 10-K for 2025).

Who were CoreWeave's biggest customers in 2025?

Microsoft was CoreWeave's biggest customer in 2025. It brought in about 67% of CoreWeave's revenue that year, and no other customer reached 10% (source: CoreWeave Form 10-K for 2025). The same report names OpenAI and Meta as other major customers. OpenAI agreed in March 2025 to pay up to about $11.9 billion through October 2030. With so few large buyers, a change at one of them can move CoreWeave's results.

Can Meta cancel a cloud computing contract?

Meta's contracts with CoreWeave can be ended by either side for cause. Meta's orders with CoreWeave sit under one main contract, a master services agreement the two firms signed on December 10, 2023. It stays in place until every order under it has run out or been ended (source: CoreWeave Form 8-K).

Does Meta say how much it spends on cloud services?

Meta gives no separate figure for cloud rent. It lists third-party cloud services as one reason its costs rose, alongside its own data centers, staff and legal costs. In the June 2026 quarter it also named third-party AI token costs, the fees for using outside AI models (source: Meta Form 10-Q for the June 2026 quarter). The commitments note is the closest public measure of how much rented capacity Meta has agreed to buy.

What is the difference between renting cloud capacity and leasing a data center?

With cloud capacity, the supplier owns the building and the chips, and Meta pays for computing time. With a data center lease, Meta rents the building and runs its own servers inside it, so the cost of those servers reaches its accounts as depreciation. Meta's filing lists cloud costs in two places: the cost of running its apps, called cost of revenue, and the cost of building new products, called research and development (source: Meta Form 10-Q for the June 2026 quarter).

Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include Meta's 2025 annual report and its quarterly filings for September 2025 and June 2026, Meta's AWS announcement and investor-relations notice, CoreWeave's 2025 annual report, Form 8-K and April 2026 announcement, Nebius's third-quarter 2025 results and March 2026 announcement, Meta's Enterprise Platform announcement, Oracle's second-quarter fiscal 2026 results and BloFin's stock futures notice. News sources include Bloomberg, Reuters, The Wall Street Journal, CNBC, TechCrunch, Axios, Ynet and The Information via SiliconANGLE, current as of October 21, 2026.

Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold Meta in any form. Meta's cloud contracts are fixed payments that continue if its revenue slows, and news about them often lands with its results after Nasdaq closes. METAX follows any move in Meta's shares, and a leveraged METAUSDT Perpetual position can lose its whole margin if the price gaps past its liquidation level. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.