Meta Weighs Selling Its Spare AI Cloud Capacity

Meta Wants to Sell AI Infrastructure to Outside Customers

Bloomberg reported that Meta is considering a cloud business that would offer its surplus AI computing capacity to external customers. On the day the story broke, Meta’s stock briefly jumped 8.6% in pre-market trading. This isn’t an official launch yet — just a report that the company is exploring the idea.

The news brought back something from a few months ago. There were reports back then that Meta had asked employees to cut back on their AI token usage. But limits on using external models and spare capacity in your own data centers are two different issues. The fact that a particular service is hard to use internally and the fact that you can rent out other computing resources aren’t mutually exclusive — both can be true at once.

Renting out computing capacity is already a business that several cloud companies run. What caught my attention in this story is that Meta is trying to turn infrastructure built for its own services into external revenue. What actually matters now is what kind of service it ends up offering, and how much profit it can wring out of it.

Two Business Models Under Consideration

Bloomberg’s report outlines two broad directions.

The first resembles Amazon Web Services’ Bedrock: offering multiple AI models for a usage fee. Meta is reportedly considering including its own model, Muse Spark, in the lineup. Customers could use the models without building their own GPU servers. Whether they choose to depends on model performance, pricing, stability, and data-handling terms.

The second is renting out compute capacity so customers can run their own models and software—the business neoclouds¹ like CoreWeave have been running. This approach reduces the need to develop model services in-house, but it’s capital-intensive, since it requires securing GPUs and data centers. According to the report, Meta’s compute organization along with infrastructure and AI executives are driving this effort. If it materializes, Meta would gain a new revenue stream beyond advertising.

At the shareholder meeting last May, Zuckerberg himself said that if Meta secured more data center capacity than it needed, it could sell off the surplus compute. He also mentioned that other companies were already inquiring about using Meta’s infrastructure. Still, whether those inquiries turn into actual contracts or profit remains to be seen.

What’s needed and what’s sellable can differ

Compute supply and demand shift depending on model type, timing, and data center location.

Earlier this year, there were reports that Google had restricted Meta’s access to Gemini, disrupting internal AI work. But limits on this kind of external service usage don’t necessarily reflect the utilization rate of all the GPUs Meta owns outright. Nor is it easy to judge an individual company’s spare capacity just from industry-wide investment growth or projected data center expansion.

Efficiency gains in training, or securing infrastructure ahead of expected demand, can both free up capacity. But this reporting alone doesn’t tell us how much of Meta’s spare capacity came from which cause. Meta has also signed deals to secure compute from CoreWeave, Google, and Oracle. If the region, timing, or hardware needed differ, a company could rent external capacity while simultaneously offering its own idle resources to other customers.

The reporting also includes remarks from an internal Meta town hall. Zuckerberg reportedly said that AI agent development over the past four months had fallen short of expectations, while AI chief Alexandr Wang explained that the next-generation model, Watermelon, is being trained with roughly 10 times more compute than existing models. There were also claims that it performs at a GPT-5.5 level on major benchmarks, though this should be distinguished from any published, independently verified results. Increasing the compute allocated to a next-generation model while also considering external sales of other equipment isn’t inherently contradictory.

The report also covered SpaceX’s move—following its acquisition of xAI—to make data center capacity available to companies like Anthropic. Bloomberg Intelligence estimated that this strategy could grow xAI’s revenue to $50 billion by 2028 and $100 billion by 2030. These are projections, not realized revenue. What’s notable is that selling compute to outside customers is being floated as a way to recoup massive infrastructure investment.

Renting Compute vs. Selling Model Services: Where the Margin Lives

Semiconductor analyst Dylan Patel believes that tokens—the output of AI computation—will become the critical commodity. If that’s true, companies can’t stop at accumulating GPUs; they have to figure out at what cost they can deliver the results customers actually need.

Renting out GPU capacity and offering AI models as a service have fundamentally different cost structures. Model services carry added costs for development, training, and operations, but in exchange they can charge for functionality customers can use immediately. Which side ends up more profitable depends on utilization rates, equipment depreciation, model costs, and selling price. Simply offering a model doesn’t guarantee higher margins.

The response speed customers demand also affects pricing. Coding assistants and real-time chat need fast responses, while bulk document processing can trade waiting time for lower cost. Patel thinks this divergence will split inference pricing accordingly. We shouldn’t read the pricing gap of one particular high-speed service as a margin that applies to all workloads.

Another approach Patel emphasizes is co-design² across chips, software, and models. The idea is to tune all three together so the same hardware handles more requests or lowers the cost per response. How much improvement you get depends on the task and the baseline you’re comparing against. From a business standpoint, what matters is value capture³—how much of that efficiency gain actually translates into pricing power or profit.

As more compute providers enter the market, Nvidia’s customer base could diversify further. That would reduce dependence on any single large customer—a real benefit. But there isn’t enough evidence to treat this as the direct cause behind Meta’s reported review of its cloud business.

In Korea, the government announced its “Three Mega-Projects”—semiconductors, physical AI, and AI data centers—on June 29th. This includes a ₩800 trillion (~$580B) plan to build a semiconductor production hub in the southwestern region, plus ₩550 trillion (~$400B) for Phase 1 data center investment from SK, GS, and Naver. This shouldn’t be read as ₩1,350 trillion (~$980B) in government spending. Data center capacity targets are 8.4GW for Phase 1, expanding to 18.4GW.

China Plays Efficiency, Japan Plays Command, Korea Plays… ₩800 TrillionChina = efficiency, Japan = command, Korea = the physical layerinlevel9.com

Korea intends to build on its semiconductor supply capabilities—HBM⁴ chief among them—to expand into data centers and physical AI. There’s real substance here: the country can draw on experience accumulated on the manufacturing floor. But physical AI is also a field where U.S. Big Tech is investing heavily. Korea’s edge won’t come from facing less competition—it will come from whether domestic manufacturing capability can actually be converted into product and service performance.

What I want to look at more closely is the execution plan for operating software alongside the capital investment. The government’s announcement did include support for cloud technology, a domestic NPU ecosystem, and physical AI model development—so it’s not fair to say the software plan is simply missing. What I’m watching is what products and customer contracts this support actually leads to. I think the investment will only pay off broadly if the models and software needed to run domestically made components and equipment become competitive in their own right.

Oswarld’s Lens

I see this news less as a story about surplus resources and more as a process of figuring out how to turn infrastructure investment into revenue.

I’ve had experience examining both the explanations companies give and the actual revenue targets behind their businesses when building market-entry strategies. Meta may genuinely be able to sell its spare capacity, but I think there’s likely also a motive to demonstrate a revenue outlook that justifies its massive capital spending. This is my interpretation, not a confirmed fact about the company’s internal intentions.

External revenue gives Meta more ways to recoup its infrastructure investment, and that could plausibly have influenced the stock’s rise. But pre-market price movement alone can’t confirm whether training efficiency actually improved, nor can it cleanly tell us what exactly the market was reacting to.

Going forward, I’ll be watching whether Meta actually launches this business, what kind of customers it lands, and how it allocates resources between developing its own AI and selling capacity externally.

From a GTM standpoint, too, GPU stockpiles alone don’t tell you much about business viability. I pay closer attention to what services domestic companies actually build with that hardware, and who’s paying how much for them. Revenue only converts into profit when the ability to lower costs is paired with a real reason for customers to keep coming back.

Closing

Meta’s plan is still at the review stage. We need to see whether it actually launches and whether it’s genuinely profitable.

Selling compute and offering model services are both ways to recoup infrastructure investment. Which one pays off better depends on customer demand and costs. And Korea’s large-scale investment plans, too, can only be judged by looking beyond facility size — at operational know-how, customer acquisition, and utilization rates.

Rather than the size of the investment, what interests me is what the invested facilities actually produce, and how much money they make.

If you’ve dealt with cloud/GPU costs before, I’d like to ask: if Meta started selling compute, would you consider switching over from AWS, Azure, or Google Cloud? Tell me in the comments what would be decisive for you — price, stability, or quality.

---

📨 Share this piece with a colleague who’s interested in AI infrastructure.

---

References & Further Reading

Primary sources

• Bloomberg, “Meta Is Building a Cloud Business to Sell Excess AI Compute”, 2026. 7. 1. : This is the original source for today’s news. It lays out the two business models and the organization driving them.
• Reuters, “Meta building cloud business to sell excess AI capacity, Bloomberg News reports”, 2026. 7. 1. : A good companion piece for Zuckerberg’s remarks at the May shareholder meeting and the stock’s reaction.
• Sequoia Capital, Dylan Patel (SemiAnalysis) interview, 2026. : An interview covering Patel’s view on AI compute demand and chip-model co-design.

Background

• SemiAnalysis, “AI Value Capture: The Shift to Model Labs”, 2026. : An analysis of the profit structures of compute providers versus model developers.
• Ministry of Trade, Industry and Energy, Korea’s Great Leap Forward: Three Megaprojects Underway, 2026. 6. 29. : The official announcement of investment and technical-support plans for semiconductors, physical AI, and data centers.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. NeoCloud: a new breed of cloud provider that specializes in renting out GPU compute for AI workloads. CoreWeave is the archetypal example. Depending on the vendor, some also offer development and operations tooling beyond raw GPU rental. ↩

2. Co-design: an approach where the chip (hardware), software, and AI model are designed together from the outset rather than separately, so they mesh with one another. The goal is to jointly reduce data-movement overhead and wasted computation across these layers. ↩

3. Value Capture: the share of value created in an industry that actually flows back to you as profit. It’s similar to how selling raw materials versus a finished product yields very different margins even for “the same” underlying goods. ↩

4. HBM (High Bandwidth Memory): specialized memory designed to move data extremely fast. It boosts the “data-movement speed” that bottlenecks AI computation, making it nearly indispensable in high-performance AI chips. Samsung Electronics and SK Hynix supply the vast majority of the world’s HBM. ↩