Meta, SpaceX turning excess AI compute into cash

Key Takeaways

• Bloomberg reported on July 1, 2026, that Meta is developing plans for a cloud infrastructure business that would sell AI computing power and, potentially, access to its own models to outside customers — a move that would put it in direct competition with AWS, Microsoft Azure, and Google Cloud. Meta has not confirmed the plans.
• The reported effort is centered on an internal group called Meta Compute, overseen by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs’ Daniel Gross, and Meta President Dina Powell McCormick.
• Meta raised its full-year 2026 capital expenditure guidance to $125–145 billion, up from $115–135 billion, and up from $72.2 billion spent in all of 2025 — nearly double the prior year, mostly for AI infrastructure. (An earlier draft of this article cited a $182.9 billion figure; that number doesn’t match any confirmed Meta guidance and has been corrected here.)
• SpaceX, which merged with Elon Musk’s xAI in early 2026, already tested this model: in May 2026, Anthropic agreed to pay roughly $1.25 billion per month — potentially over $40 billion through May 2029 — for the entire output of xAI’s Colossus 1 data center in Memphis, more than 220,000 Nvidia GPUs and over 300 megawatts of power.
• SpaceX’s confidential S-1 filing for an IPO reportedly targeting a $1.75–2 trillion valuation disclosed that xAI’s compute segment posted a $2.47 billion operating loss on $818 million of revenue in the first quarter of 2026 — a reminder that reselling compute doesn’t automatically mean profiting from it.
• Financial Times analysis cited by multiple outlets estimates Meta’s AI infrastructure investment currently generates roughly a negative 29% return, versus a positive 7.2% for Amazon — the only major hyperscaler reportedly showing clearly positive AI infrastructure returns so far.
• Meta cut about 8,000 jobs (roughly 10% of its workforce) in May 2026, with Zuckerberg directly linking the cuts to the scale of the company’s AI infrastructure spending.

What’s Actually Happening: Meta’s Reported Cloud Compute Plans

Meta’s stock jumped more than 6% in premarket trading — and more than 10% at points during the session — on July 1, 2026, after Bloomberg reported that the company is developing plans for a cloud infrastructure business. According to people familiar with the matter, cited by Bloomberg and later corroborated by CNBC, Meta is exploring whether it can sell “excess” AI computing capacity to outside customers, and is separately considering letting outside developers pay to run queries against Meta’s own AI models — including Muse Spark, part of its Superintelligence Labs work — on infrastructure Meta owns and operates.

It’s worth being precise about what’s confirmed here and what isn’t, because the coverage has occasionally blurred the two. Meta has not made an official announcement. Everything currently public comes from anonymous sourcing in Bloomberg’s initial report, followed by CNBC’s additional detail. TechCrunch and other outlets reached out to Meta for comment and, as of this writing, had not received a substantive response. That doesn’t mean the reporting is wrong — Bloomberg and CNBC both cited multiple people familiar with internal planning, and Zuckerberg himself had already flagged the possibility publicly — but it does mean some of the specifics (which models, what pricing, what launch timeline) remain unconfirmed.

What gives the story weight is that Zuckerberg effectively previewed it himself. At Meta’s late-May 2026 shareholder meeting, he said selling spare compute capacity was on the table, a notable shift from his stance in Q3 2025, when he’d suggested Meta would only become a compute supplier if it had genuinely overbuilt its infrastructure. Seven months and a further capex increase later, that scenario appears to be exactly where the company finds itself.

Why Meta Has “Excess” Compute in the First Place

Capex Reality Check: The Real Numbers Behind the Headlines

Meta’s AI infrastructure spending has escalated in stages. The company spent $72.2 billion on capital expenditures in all of 2025 — itself up roughly $30 billion from 2024. On its Q1 2026 earnings call, CFO Susan Li told investors the company now expects full-year 2026 capex of $125–145 billion, raised from a prior guidance range of $115–135 billion, citing higher component prices and additional data center costs tied to future-year capacity. Q1 2026 capex alone reached $19.84 billion. That upper-bound figure is nearly double what Meta spent in all of 2025, and more than its combined 2024 and 2025 spending.

For context on how that fits into the wider industry: Google, Microsoft, Meta, and Amazon combined are on pace to spend roughly $725 billion on capex in 2026, up 77% from the prior year, and the “Magnificent Seven” as a group are projected to spend more than $700 billion on AI infrastructure this year alone, up from roughly $400 billion in 2025. Meta’s spending is a large piece of a genuinely unprecedented industry-wide buildout, not an outlier.

That scale of investment is also why the company can plausibly claim to have compute it doesn’t immediately need. Building data center capacity for “future-year” demand — Li’s own phrase — means some fraction of that capacity is, by design, ahead of current internal usage. Reselling it converts a multi-year cost center into near-term revenue, at least on paper.

Meta Compute: Who’s Running This Effort

Bloomberg’s reporting named the internal organization at the center of these plans as Meta Compute, created to oversee the buildout and operation of the company’s AI infrastructure. It’s led by three people with distinct mandates: Santosh Janardhan, Meta’s head of infrastructure, who oversees the physical data center buildout; Daniel Gross, a leader inside Meta Superintelligence Labs, the AI research unit Zuckerberg has aggressively staffed over the past year; and Meta President Dina Powell McCormick, whose involvement signals this is being treated as a strategic business initiative, not purely an engineering side project.

Notably, Meta doesn’t break out Meta AI or Llama-family revenue separately in its financial reporting, and executives have generally emphasized internal AI applications — improving ad targeting and content recommendations, primarily — over public-facing AI products as the company’s near-term priority. That makes a compute-resale business a logical way to generate a visible, near-term revenue line from AI spending, independent of whether consumer AI products like Meta AI itself become major moneymakers.

The timing adds a harder edge to the story. The same week Meta raised its capex guidance to $125–145 billion, Zuckerberg told employees at a company town hall that roughly 8,000 planned layoffs — about 10% of the workforce, beginning in late May — were a direct consequence of the AI infrastructure budget. “We basically have two major cost centers in the company: compute infrastructure and people-oriented things,” he reportedly told staff, according to Reuters’ account of the meeting, declining to rule out further headcount reductions later in the year. A cloud compute business, in that light, isn’t just an opportunistic pivot — it’s one of the more direct ways to offset costs that are already reshaping the rest of the company.

SpaceX and xAI: The Blueprint Meta Is Reportedly Following

The Anthropic-Colossus 1 Deal, Explained

If Meta’s reported plans sound familiar, that’s because SpaceX effectively ran this experiment first — and in an unusually public way, thanks to its own IPO paperwork. Elon Musk merged SpaceX with his AI venture xAI in early 2026, in a deal that reportedly valued the combined entity at $1.25 trillion. On May 6, 2026, Anthropic announced it had signed an agreement with SpaceX to use the entire compute capacity of xAI’s Colossus 1 data center in Memphis, Tennessee — more than 220,000 Nvidia GPUs (a mix of H100, H200, and GB200 accelerators) and over 300 megawatts of power, available within a month of signing.

The deal was, on its face, an unusual pairing: Musk had spent the prior months publicly criticizing Anthropic, at one point writing on X that the company “hates Western Civilization.” Days before the deal was announced, Musk said he’d spent time with senior Anthropic staff and came away “impressed,” softening his public stance considerably.

According to reporting that later cited SpaceX’s own IPO filing, Anthropic is paying roughly $1.25 billion per month through May 2029 — a contract that could bring xAI more than $40 billion in total revenue, or roughly $15 billion annually. Simon Willison’s widely cited write-up of the deal noted that xAI had already shifted its own Grok model training to a newer, larger cluster called Colossus 2, which made leasing out the older Colossus 1 to a direct competitor more palatable — reported internal utilization on Colossus 1 had fallen to around 11% before the deal, well below the 35–45% range considered normal for production-grade model training.

Other Buyers: Google, Cursor, and a Widening Customer List

Anthropic wasn’t the only company buying into xAI’s spare capacity. Reporting that emerged in the weeks following the initial deal indicated Google also purchased compute capacity from xAI, and that Anthropic itself later booked additional capacity in Colossus 2 as well.

Separately, in April 2026, xAI announced that Colossus data center capacity would support Cursor, the AI code-editing company — around the same time SpaceX extended a conditional $60 billion takeover offer to Cursor (with a $10 billion break fee if the acquisition didn’t proceed). Whatever the underlying corporate logic, the pattern is consistent: an AI lab that overbuilt its own infrastructure quietly repositioning as a compute landlord to companies that, in some cases, directly compete with its own AI products.

What the SpaceX IPO Filing Revealed

SpaceX confidentially filed an S-1 with the SEC on April 1, 2026, reportedly targeting an IPO valuation between $1.75 trillion and $2 trillion, with a roadshow said to be planned for the week of June 8. That filing is also where some of the more sobering numbers surfaced: xAI’s compute segment posted a $2.47 billion operating loss in Q1 2026 against just $818 million in segment revenue.

SpaceX’s own language in the filing described the Anthropic arrangement as a way to “monetize unused compute capacity,” while acknowledging the companies could each terminate the deal with 90 days’ notice. In other words, even the highest-profile compute resale deal in the industry so far is, by the seller’s own disclosed numbers, not yet turning a profit at the segment level — it’s converting a sunk cost into partial revenue recovery, not into a standalone profitable business, at least not yet.

Comparison: The Emerging AI Compute Resale Market

The table is worth sitting with for a second, because it shows two genuinely different business models being described with the same language. CoreWeave, AWS, Google Cloud, and Azure are compute providers by design — selling capacity is their core business. What Meta and SpaceX/xAI are reportedly doing is different: taking infrastructure originally built for internal AI ambitions and repositioning the surplus as a revenue stream after the fact. That distinction matters for evaluating how durable or strategic this trend actually is, versus how much of it is a financially convenient response to having built more than was needed.

Why Sell Compute Instead of Just Using It?

The Overbuild Question

The infrastructure analysts and journalists keep circling back to is whether the AI industry, collectively, is building more compute capacity than current or near-term demand justifies. Some analysts have explicitly compared the current AI infrastructure buildout to the fiber-optic overbuild that preceded the dot-com crash — enormous capital investment made on the assumption that demand would eventually catch up to supply.

Others point out that unlike that earlier cycle, today’s AI compute buildout is backed by companies with far larger existing cash flows and profitable core businesses, which changes the risk calculus considerably. Both views can be defensible at once: the infrastructure spending is real and enormous, and whether it was strictly necessary at this scale is a genuinely open question that won’t be settled until well after the fact.

Related: US Iran talks resume amid gas fears

The ROI Problem Hyperscalers Don’t Like Discussing

When Meta’s CFO was pressed on short-term return on investment from AI infrastructure spending at the same shareholder meeting where compute resale first came up, Zuckerberg reportedly deflected, calling it “a very technical question.” Financial Times analysis cited across multiple outlets estimated Meta’s AI infrastructure investment currently generates a negative 29% return — a striking figure, especially set against Amazon’s estimated positive 7.2% return, reportedly the only major hyperscaler currently showing clearly positive returns on AI infrastructure spend.

Investors have drawn an explicit comparison to Meta’s Reality Labs division, which has lost an estimated $70–90 billion since its creation with minimal revenue to show for it — a comparison Meta’s own leadership would presumably prefer not to invite, but one that’s become difficult to avoid given the scale of the new spending.

Selling excess compute doesn’t fix a negative-ROI infrastructure investment on its own, but it does something psychologically and financially useful for a public company: it converts an asset sitting on the balance sheet, generating no return, into one generating measurable revenue — even if, as the SpaceX/xAI numbers suggest, that revenue doesn’t yet cover the underlying cost of running it.

Precedent: CoreWeave’s Business Model

Bloomberg’s reporting specifically framed Meta’s potential approach as similar to CoreWeave — a company that built its entire business model around leasing raw GPU compute capacity to AI labs and enterprises, rather than building consumer-facing AI products of its own. CoreWeave already has a direct commercial relationship with Meta: in September 2025, the two companies signed a $14.2 billion agreement for CoreWeave to supply Meta with Nvidia GB300-based compute capacity through December 2031, with an option to extend into 2032.

If Meta does move into direct compute resale, it would be entering a market it has, until now, primarily participated in as a customer — a shift that raises an interesting question about whether Meta is preparing to compete with, rather than continue relying on, some of its own infrastructure vendors.

What This Means for the Cloud Market

A Meta-run cloud compute business, even a partial one focused only on raw compute rather than full-service cloud offerings, would add a new and unusually well-capitalized competitor to a market long dominated by three players: AWS, Microsoft Azure, and Google Cloud. Amazon’s stock dipped following Bloomberg’s initial report — a signal that markets, at minimum, read the news as a credible near-term competitive threat rather than background noise. Google’s position is worth watching closely too, given its parallel push to expand AI into healthcare and other verticals that depend on the same underlying compute infrastructure Meta would be competing for.

The more direct competitive overlap, though, is probably with neocloud providers like CoreWeave and Nebius — companies built specifically around reselling GPU capacity rather than offering the full menu of traditional cloud services (storage, databases, managed software) that AWS, Azure, and Google Cloud provide. Meta entering that specific niche, backed by a balance sheet several orders of magnitude larger than CoreWeave’s, would be a meaningfully different competitive dynamic than another hyperscaler entering the general cloud market. It’s the kind of infrastructure-as-a-service positioning that smaller providers offering cloud infrastructure as a service have built entire businesses around, except at a scale few companies besides Meta could realistically match.

There’s also a marketing and advertising angle worth noting, since it’s Meta’s original core business. The company’s own AI-driven marketing analytics already depend on internal compute at massive scale, and any move to formally productize spare capacity would sit alongside — not replace — that internal use case. Businesses evaluating cloud-based marketing tools more broadly may eventually see Meta positioned on both sides of that market: as an advertising platform and, potentially, as infrastructure provider to other AI-driven marketing tools built on its compute.

Risks and Open Questions

Unconfirmed Reporting

It bears repeating: Meta has not confirmed these plans. Bloomberg’s report relies on unnamed sources, and while CNBC’s follow-up reporting corroborated the core claim, key details — pricing, which models would be accessible, launch timing, whether this becomes a formal product line or stays a smaller pilot — remain unknown. Readers should treat the specifics as developing, not settled.

Compute Demand Risk

The entire premise of a compute resale business depends on sustained external demand for AI compute at prices that justify the investment. If AI adoption growth slows, or if enough new capacity comes online industry-wide to ease the current shortage, the pricing power that makes deals like SpaceX’s $1.25-billion-a-month Anthropic contract attractive today could erode quickly. The xAI compute segment’s disclosed Q1 2026 operating loss is a concrete illustration of how easily a resale strategy can generate revenue without generating profit.

Power and Infrastructure Constraints

Compute resale businesses are ultimately bottlenecked by electricity, not just chips. Meta’s major AI data center projects in Louisiana and Ohio depend on securing power at a scale that increasingly intersects with regional energy politics and, at times, broader geopolitical dynamics — the kind of pressure already visible in geopolitical energy tensions playing out elsewhere in the energy market this year. Data center power demand has become one of the least glamorous but most consequential constraints on how fast any of these companies — Meta included — can actually scale a compute resale business, regardless of how much capital they’re willing to spend.

Reputational and Competitive Risk

Related: Micron poised to follow Nvidia’s success

The SpaceX-Anthropic deal was unusual precisely because it involved leasing infrastructure to a direct competitor, following a period of public hostility between Musk and Anthropic’s leadership. That dynamic — and the broader questions it raises about Musk’s legal dispute with OpenAI and his approach to AI industry rivalries generally — is a reminder that compute resale deals can carry reputational complexity beyond the straightforward commercial logic. If Meta moves forward with hosting outside developers’ workloads, including potentially rival AI labs, it will face similar questions about where commercial pragmatism ends and competitive strategy begins, especially amid ongoing scrutiny of how major AI labs are governed across the industry.

The Broader AI Infrastructure Race

Meta and SpaceX/xAI are the most visible examples of this pattern right now, but they’re operating inside a much larger industry-wide buildout. The “Magnificent Seven” collectively are on pace for more than $700 billion in AI capital expenditures in 2026, and even well-funded challengers face very different resource constraints than the largest players — a dynamic visible in how smaller, capital-constrained AI labs like Mistral have had to build fundamentally different strategies around compute scarcity rather than compute surplus.

The contrast is instructive: while Meta debates what to do with too much compute, plenty of AI companies elsewhere in the world are still structuring their entire business models around not having enough of it, including labs pursuing rapid AI expansion in markets like India, where infrastructure access remains a genuine constraint on growth.

The layoffs tied to Meta’s infrastructure spending also fit into a wider conversation about how AI investment is reshaping career paths inside the companies building it, not just the industries AI is expected to disrupt. And as more of this compute gets resold into enterprise use cases — including sectors like fintech, where AI chatbots and automated services are already spreading — the infrastructure decisions being made by Meta and SpaceX today will shape pricing and availability well beyond the companies making them.

What to Watch Next

• Whether Meta formally confirms or launches a compute resale product, and what it’s initially priced or scoped to include.
• Whether early compute-resale deals like SpaceX/xAI’s Anthropic contract become profitable at the segment level, or continue running at a loss as reported in Q1 2026 filings.
• How AWS, Google Cloud, and Microsoft Azure respond competitively, particularly given Amazon’s stock reaction to the initial Meta report.
• Whether Meta’s next capex guidance update (typically alongside quarterly earnings) moves further from the current $125–145 billion range, and whether any of that spending is explicitly earmarked for external, revenue-generating capacity rather than internal use.
• Power availability, which may prove a harder ceiling on this trend than capital or chip supply, given how tightly AI data center buildouts are already bumping up against regional electricity constraints.
• How to evaluate the real return on this spending, a question every company in this space will eventually have to answer with numbers, not just guidance ranges — a discipline worth applying using the same basics behind how to calculate software ROI, scaled up to infrastructure investment.

Related: Pichai booed at Stanford over Google Israel ties

Frequently Asked Questions

1. Has Meta officially confirmed it’s building a cloud compute business? No. As of this writing, Meta has not made an official statement. The reporting comes from Bloomberg, citing unnamed sources familiar with internal planning, later corroborated by additional detail from CNBC. TechCrunch and other outlets reached out to Meta for comment without receiving a substantive response.

2. How much is Meta actually spending on AI infrastructure in 2026? Meta’s official guidance, given during its Q1 2026 earnings call, is $125–145 billion in capital expenditures for 2026, up from a prior range of $115–135 billion, and nearly double the $72.2 billion it spent in all of 2025.

3. What is the SpaceX-Anthropic compute deal, and is it related to Meta’s plans? In May 2026, Anthropic agreed to pay SpaceX (which owns xAI) roughly $1.25 billion per month for the entire compute capacity of xAI’s Colossus 1 data center — more than 220,000 Nvidia GPUs. It’s widely cited as a precedent for exactly the kind of compute-resale business Meta is reportedly now considering.

4. Is SpaceX’s xAI compute business actually profitable? Based on figures disclosed in SpaceX’s confidential IPO filing, xAI’s compute segment posted a $2.47 billion operating loss on $818 million in revenue in Q1 2026. The Anthropic deal generates substantial revenue but hasn’t been shown to make the segment profitable so far.

5. Why would Meta sell computing power to outside companies, including potential competitors? Meta has built AI infrastructure capacity partly in anticipation of future internal needs, which may exceed what it currently uses. Reselling that surplus converts otherwise idle, cost-generating infrastructure into a near-term revenue source, similar to the model CoreWeave already uses.

6. Does this mean the AI infrastructure boom is a bubble? That’s genuinely disputed. Some analysts see the scale of spending, mixed return data (like Meta’s reported negative 29% ROI on AI infrastructure), and now, compute resale, as signs of overbuilding. Others note the spending is backed by companies with substantial existing profitable businesses, which differs meaningfully from past infrastructure bubbles. The evidence supports both caution and the case that this cycle is different — it isn’t settled.

7. How does this affect AWS, Google Cloud, and Microsoft Azure? A Meta-run compute resale business would introduce a large, well-funded new competitor, though likely focused on raw GPU compute rather than full-service cloud offerings. Amazon’s stock fell after the initial Bloomberg report, suggesting markets see it as a credible competitive threat.

8. What is Meta Compute? Meta Compute is the internal organization reportedly overseeing Meta’s AI infrastructure buildout and the exploration of a compute resale business, led by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs’ Daniel Gross, and Meta President Dina Powell McCormick.

9. Why did Meta lay off 8,000 employees around the same time it increased AI spending? Zuckerberg told employees the cuts were a direct consequence of the company’s growing AI infrastructure budget, describing compute infrastructure and headcount as the company’s two major cost centers, with more capital going toward the former.

10. What should readers watch for next in this story? Whether Meta formally confirms a compute business, whether early compute resale deals like SpaceX/xAI’s become profitable, how competitors respond, and whether power availability — not just capital or chip supply — becomes the real constraint on how far this trend can scale.

Conclusion

The core fact behind this story is straightforward: Meta, SpaceX, and a growing list of AI infrastructure builders spent enormous sums securing compute capacity for a future that hasn’t fully arrived yet, and reselling the surplus is an increasingly obvious way to generate revenue from that spending in the meantime. What’s less straightforward is whether this represents a genuinely new, durable line of business or a financially convenient response to overbuilding — and the disclosed numbers so far, particularly xAI’s reported quarterly operating loss on its compute segment, suggest the answer isn’t fully settled even for the company that pioneered this specific playbook.

For Meta, a confirmed move into compute resale would mark a real strategic shift: from a company that has historically kept its infrastructure entirely in-house to one competing directly with cloud providers it has, until now, mostly done business with as a customer. Whether that shift pays off will depend on questions still unanswered as of this writing — how much external demand actually exists at sustainable prices, how power constraints evolve, and whether the return-on-investment math that currently worries some analysts improves once idle capacity starts generating revenue instead of just costs.

Editorial note: This edit corrects one unverified figure from the original draft (a $182.9 billion Meta capex claim, which does not match any confirmed Meta guidance) and replaces it with Meta’s actual disclosed 2026 capex range of $125–145 billion, sourced from the company’s Q1 2026 earnings call. All financial figures regarding SpaceX, xAI, and Anthropic are drawn from public reporting on SpaceX’s confidential S-1 filing and company statements as of July 2026; some details of Meta’s reported cloud compute plans remain unconfirmed by Meta itself and should be treated as developing.