When Meta starts selling computing power.

On July 1, Bloomberg broke a story: Meta is internally advancing a project code-named "Meta Compute," planning to sell surplus AI computing power to external customers.

Two paths are being pursued simultaneously:

• Hosted model access — making models like Llama and Muse Spark available to enterprise clients, charging based on token usage, comparable to AWS Bedrock;
• Direct bare-metal computing sales — renting out GPU clusters by the hour, comparable to CoreWeave.

As soon as the news broke, the market voted with its feet: META rose about 9% on the day (CNBC), with investors interpreting it as Zuckerberg’s direct response to concerns about "whether the $200k AI investment can be recouped."

On the other hand, both CoreWeave and Nebius fell about 15% (247 Wall St) — these two Neoclouds (emerging AI clouds) that rely on "selling GPU time" suddenly faced a competitor dozens of times their size.

This is no coincidence. Zuckerberg had already said during the May shareholder call:

"This is definitely one of the options we are considering. Almost every week, external companies contact us, hoping we will launch API services, or asking if they can purchase our computing power, even willing to pay higher than our procurement costs."

His full statement at the time was: "We are not doing this yet because we believe these computing resources still have their own uses. However, if we find in the future that there is an overcapacity in computing power construction, then this would be an option we could take."

That statement is now turning into action.

The $145 Billion Bet

To understand why Meta wants to sell computing power, you first need to know how much it has built.

In April 2026, Meta raised its full-year capital expenditure guidance to $125 billion–$145 billion (confirmed in SEC 10-Q filing), nearly doubling its actual 2025 capex of $72.2 billion. This figure briefly spooked investors during the earnings call — the stock fell 10% that day.

But Zuckerberg did not back down. His logic:

"The biggest bottleneck facing the entire industry remains the supply of computing power, so we should stockpile as much computing resources as possible now and decide later how to use them."

How much exactly has been built? A few numbers:

This is not just Meta's problem. Microsoft, Google, and Amazon are also pouring money simultaneously. In 2026, the combined capex of the four tech giants is approaching $700 billion.

This money is not buying software — it’s buying steel, electricity, Nvidia GPUs, and data centers rising one after another.

Why the Surplus: The Training-Inference Utilization Gap

Computing power has a physics problem: it is not consumed evenly.

A large language model training task can fully saturate tens of thousands of GPUs at 100% for months. But once training is complete, the cluster utilization plummets to 30%–50% — only inference requests remain running, and inference demands far less computing power than training.

Meta’s training schedule is publicly known: Llama 4 is trained, Llama 5 is on the way. In the gap between, clusters sit there, consuming electricity, generating no returns.

Zuckerberg’s strategy is called "Hoard now, decide later." Build infrastructure based on peak training demand; figure out how to use it afterward.

The premise of this strategy is: there will always be a reason to use these computing resources in the future. And "selling them" is one of those reasons.

From a physical law perspective, this is not wasteful mismanagement by Meta — it is the inherent cycle of computing infrastructure. As long as you are doing frontier AI R&D, you will eventually face this problem.

Meta’s difference: it was the first to acknowledge it and started acting.

First Movers Emerge: SpaceX/xAI’s Computing Business

Meta is not the first to do this.

In May 2026, Musk’s SpaceX/xAI completed two shocking computing lease deals:

First deal: Anthropic leases Colossus 1

• Monthly rent: $1.25 billion
• Contract term: until 2029
• Total contract value: approximately $45 billion
• Asset: All available computing power at the Colossus 1 data center in Memphis, Tennessee (200k+ Nvidia GPUs)

Second deal: Google leases Colossus 2

• Monthly rent: $920 million
• Asset: Computing cluster at the Colossus 2 data center

Together, these two contracts bring in over $26 billion annually for SpaceX/xAI simply by "renting GPUs."

More importantly: it validated the feasibility of the "build your own data center → sell computing power" model.

Meta is walking the same path. And Meta’s scale is far larger than xAI — it has committed $182.9 billion to building infrastructure.

Who Gets Hurt Most: The Double Squeeze on Neoclouds

Meta’s entry puts the most immediate pressure on Neoclouds.

These companies’ business model is simple: bulk purchase/lease GPUs upstream (from Nvidia or hyperscalers like Meta), then break them down into hourly computing rentals for downstream AI startups, research institutions, and enterprise customers.

CoreWeave is the most well-known — IPO in March 2025, with a market cap once exceeding $50 billion. Its core assets are stacks of GPUs and a batch of long-term customer contracts.

But with Meta’s entry, CoreWeave’s business model shows two cracks:

Crack one: Downstream customers diverted. If AI startups can rent computing power directly from Meta — and Meta’s GPUs are newer, larger scale, and possibly cheaper — why go to CoreWeave?

Crack two: Biggest customer turns rival. CoreWeave and Meta already have deep ties. In April 2026, CoreWeave signed a total $35 billion computing supply agreement with Meta (until 2032), with $21 billion in new additions from 2027–2032. Nebius also signed a $27 billion similar deal with Meta. These contracts have CoreWeave/Nebius supplying computing to Meta — Meta is the buyer.

If Meta decides to build its own computing and sell externally, it will likely reduce external procurement from CoreWeave and Nebius. The renewal rates and new volume of these contracts will be discounted. The market is pricing not just "Meta becomes a competitor" but also "Meta is no longer a reliable large customer."

There is a deeper risk: the valuation of financing collateral.

Neocloud expansion relies heavily on debt financing, with the collateral being their GPU clusters. In March 2026, CoreWeave closed an $8.5 billion GPU-backed term loan — touted as the industry’s first investment-grade GPU-backed debt (NASDAQ announcement). If a hyperscaler like Meta enters the computing rental market at scale, hourly GPU rental rates will fall — collateral value will shrink — debt refinancing will become harder.

This is not theoretical. On the day the news broke, CoreWeave fell 15%. The market is repricing.

The Bigger Picture: The $700 Billion Infrastructure Bet — Who Runs First?

Zoom out. Meta selling computing power is not just about Meta.

In 2026, the total capex of four tech giants is approaching $700 billion. The vast majority goes in one direction: AI infrastructure.

The question: After all this is built, what will utilization be?

The bear case:

• GPU computing prices keep falling. B300 cloud instances on-demand minimum is about $7.4/hour, spot instances as low as $4.3/hour (GPUFinder, July 2026). More broadly, LLM inference costs have dropped about 1,000x in three years (GPU Nexus).
• Inference efficiency is jumping. DeepSeek R1, Anthropic’s latest models are doing more with less computing.
• Some analysts compare this to the 1990s fiber optic overbuild — telecom companies laid fiber frantically, resulting in oversupply, price collapse, and a wave of bankruptcies.

The bull case:

• Jevons paradox: The cheaper computing becomes, the more people use it. Not linear growth — exponential growth.
• Inference demand is exploding. In 2026, inference workloads account for about two-thirds of all AI computing power, up from one-third a year ago.
• Current AI penetration is roughly equivalent to the internet in 1995 — you might think it’s overbuilt now, but in hindsight, it won’t be enough.

Both sides have merit. But one fact is undeniable:

Meta did not wait until "overcapacity was confirmed" to start selling computing power. It reserved an exit for "possible overcapacity."

That is the truly notable signal. If you were 100% confident in your own computing demand, you wouldn’t need to plan a computing sales strategy in advance. Just build.

Meta has a backup plan. What about others?

Microsoft, Google, Amazon — cloud selling is their core business; they don’t have a "whether to sell computing" question — they have always been selling. The real question is: they are also expanding frantically, and at a pace not slower than Meta.

If the biggest buyers are all leaving themselves an escape route — then the market might not be as deep as it seems.

The First Crack in Infrastructure Investment

For the past two years, the logic of AI infrastructure investment has been: "Demand is unlimited; computing will never be enough." Now, the first crack appears — not because demand vanished, but because the pace of supply construction may have raced ahead of demand.

Meta selling computing power marks the shift from "build at any cost" to "start counting the costs."

Things to watch next:

2. Will Microsoft follow? It has deep ties with OpenAI and plenty of Azure computing — it doesn’t need to build new; it’s already selling. But will it slow its expansion?
3. CoreWeave’s next quarterly report. Can it prove with contract data that it is unaffected?
4. The trend of hourly GPU rates. If Meta formally enters the market, will a price war erupt?

Risk Warning and Disclaimer

        Market risk exists; investment requires caution. This article does not constitute personal investment advice, nor does it account for individual users’ specific investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investment based on this content is at your own risk.