When Meta starts selling computing power

July 1st, Bloomberg broke the news: Meta is internally advancing a project codenamed “Meta Compute,” planning to sell excess AI computing power to external clients. Two paths are being pursued simultaneously: - Managed model access — opening up models like Llama and Muse Spark to enterprise clients, charging by token usage, benchmarking AWS Bedrock; - Direct sale of raw compute — renting GPU clusters by the hour, benchmarking CoreWeave. As soon as the news came out, the market responded decisively: META surged about 9% that day (CNBC). Investors interpreted this as Zuckerberg’s direct answer to doubts about whether the tens of billions sunk into AI can be recovered. Meanwhile, CoreWeave and Nebius both dropped about 15% (247 Wall St) — these two “Neocloud” companies (emerging AI cloud vendors that make a living by “selling GPU time”) suddenly have a rival dozens of times their size. This is not a coincidence. Zuckerberg had already said in the May shareholder call: “Absolutely, this is one of our options. Almost every week there are external companies reaching out, hoping we launch API services or asking to purchase our computing power, even offering to pay more than our procurement costs.” His full statement back then was: “We haven’t done this yet because we believe there is still internal utility for these resources. But if in the future we see excess capacity, this becomes an option we can take.” Now, those words are turning into action. ## $145 Billion Bet To understand why Meta is selling compute, you first need to know how much it’s built. In April 2026, Meta raised its annual capital expenditure guidance to $125 billion–$145 billion (confirmed by SEC 10-Q filings), nearly doubling the actual capex of $72.2 billion from 2025. This number made investors panic briefly during the earnings call—the stock dropped 10% that day. But Zuckerberg did not retreat. His logic: “The biggest bottleneck facing the industry is still compute supply, so we should stock up as much as possible now, and decide later how to use it.” How much was actually built? A few numbers: ![image] It's not just Meta’s issue. Microsoft, Google, and Amazon are also pouring money in at the same time. In 2026, the combined capex of the four tech giants is nearing $700 billion. This money isn’t buying software. It’s buying rebar, electricity, NVIDIA GPUs, and vast data centers rising across the landscape. ## Why There's Excess: Utilization Gap Between Training and Inference Compute has a physical problem: it’s not consumed evenly. A large language model training job can saturate tens of thousands of GPUs for months. But after training, the cluster utilization drops sharply to 30%-50%—only inference workloads are running, which require far less compute than training. Meta’s training cadence is public: Llama 4 is trained, Llama 5 is on the way. In between, the clusters sit there, burning electricity and generating no return. Zuckerberg’s strategy is called “Hoard now, decide later.” Build the infrastructure for peak training demand now; as for how to use it after construction—decide when the time comes. The premise: there will always be a reason to use this compute in the future. And “sale to outsiders” is one such reason. From a physics perspective, this is not Meta’s mismanagement causing waste—it’s the inherent cycle of compute infrastructure. Anyone doing frontier AI research will face this problem sooner or later. Meta differs in that it was first to admit it—and now is acting on it. ## First Movers: SpaceX/xAI’s Compute Business Meta isn’t the first to do this. In May 2026, Musk’s SpaceX/xAI closed two market-shaking compute leasing deals: First: Anthropic rents Colossus 1 - Monthly rent: $1.25 billion - Contract term: until 2029 - Total contract value: about $45 billion - Subject: all available compute at the Colossus 1 data center in Memphis, Tennessee (200,000+ NVIDIA GPUs) Second: Google rents Colossus 2 - Monthly rent: $920 million - Subject: compute cluster at Colossus 2 data center Combined, SpaceX/xAI can earn over $26 billion a year just by “renting GPUs.” More importantly: it validated the “build own data center → sell compute” model. Meta is following that path—and its scale dwarfs xAI. It has pledged more than $182.9 billion to infrastructure. ## Who’s Hurt Most: Double Squeeze on Neoclouds With Meta’s entry, pressure lands directly on Neocloud companies. Their business model is simple: bulk purchase/rent GPUs from upstream (NVIDIA or super-scale buyers like Meta), then subdivide for hourly rental to downstream AI startups, research institutes, and enterprise users. CoreWeave is the best known—it IPO’d in March 2025, at one point reaching a $50 billion market cap. Its core assets: a pile of GPUs and a batch of long-term client contracts. But with Meta entering, two cracks appear in CoreWeave’s business model: Crack one: downstream clients are diverted. If AI startups can rent directly from Meta—with newer GPUs, larger scale, possibly lower prices—why stick with CoreWeave? Crack two: largest client becomes competitor. CoreWeave and Meta already have deep cooperation. In April 2026, CoreWeave signed a $35 billion compute supply agreement with Meta (to 2032), including $21 billion in new quantity for 2027–2032. Nebius has a similar $27 billion agreement with Meta. These deals are CoreWeave/Nebius supplying compute to Meta—Meta is the buyer. If Meta builds its own compute and sells externally, it will likely reduce procurement from CoreWeave and Nebius. Contract renewals and new quantities will be discounted. The market is not just pricing in “Meta becomes a rival”—but also “Meta is no longer a reliable big customer.” There’s a deeper risk: valuation of collateral for financing. Neocloud’s expansion relies heavily on debt financing, and the collateral is their GPU clusters. In March 2026, CoreWeave closed an $8.5 billion GPU-secured term loan—the industry's first investment-grade GPU-backed debt (Nasdaq announcement). If massive players like Meta enter compute leasing, GPU hourly rental rates will fall—asset values shrink—debt refinancing gets harder. This isn’t theoretical. On news day, CoreWeave dropped 15%. The market is repricing. ## Bigger Picture: $700 Billion Infrastructure Gamble—Who Moves First? Zoom out. Meta selling compute isn’t just a Meta story. In 2026, the four tech giants’ combined capex nears $700 billion. Most of it flows into the same thing: AI infrastructure. The question: After all this is built, what utilization will be reached? Bears’ logic: - GPU compute prices keep falling: B300 cloud instances as low as $7.4/hour on-demand, $4.3/hour for spot (GPUFinder, July 2026). More broadly, LLM inference costs have dropped about 1,000-fold in three years (GPU Nexus). - Inference efficiency leaps: DeepSeek R1, Anthropic’s latest models do more with less compute. - Some analysts compare this to the late-90s fiber overbuild—telcos laid fiber madly, resulting in oversupply, price collapse, mass bankruptcies. Bulls’ logic: - Jevons Paradox: As compute gets cheaper, demand rises—not linearly, but exponentially. - Inference workloads are exploding. In 2026, inference accounts for about two-thirds of all AI compute, up from one-third a year ago. - The current AI penetration rate is about where the internet was in 1995—you may think we’re overbuilding, but retrospectively, it wasn’t enough. Both sides have a point. But one fact is indisputable: **Meta didn’t wait for “confirmed surplus” to start selling compute. It prepared an outlet for “potential surplus.”** This is the signal worth watching. If you’re 100% confident in your own compute needs, you don’t need a backup plan to sell compute—you just build. Meta made a backup plan. What about others? Microsoft, Google, Amazon—their business is cloud, so there’s no “should we sell compute” question; they always have. The real question: they are also expanding madly, and at a speed not slower than Meta. If even the biggest buyers are preparing their own exit strategies—the market may not be as deep as once thought. ## The First Crack in Infrastructure Investment For the last two years, the logic behind AI infrastructure investment was: “Demand is infinite, compute is never enough.” Now, that logic has shown its first crack—not because demand has disappeared, but because supply construction may have already outpaced demand. Meta’s selling compute marks AI infrastructure transitioning from “build with no regard for cost” to “start accounting.” **Several things worth watching next:** 1. Will Microsoft follow? With deep OpenAI collaboration and tons of Azure compute, it doesn’t need to build new—it’s already selling. But will the expansion pace slow? 2. CoreWeave’s next quarterly report. Can it prove with contract data that it’s not impacted? 3. Trends in GPU hourly rates. If Meta officially enters the market, will a price war start? Risk Warning and Disclaimer Markets are risky, investments should be cautious. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial situation, or needs of any particular user. Readers should evaluate whether any opinions, viewpoints, or conclusions herein fit their own circumstances. Acting on this content is your responsibility.