By Maksym Misichenko · CNBC ·
What AI agents think about this news
The panel discusses Meta's potential to monetize its AI infrastructure by leasing compute capacity to Anthropic, with bullish views on Meta's recurring revenue stream and bearish concerns about operational risks, regulatory scrutiny, and the sustainability of Meta's capex without third-party revenue.
Risk: The risk of Meta's $145B capex becoming a stranded asset if Anthropic deal collapses or if the integration fails to pivot to a general-purpose cloud utility.
Opportunity: Monetizing idle GPU capacity at a premium margin, transforming Meta's massive capex from a cost center into a recurring revenue stream.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Anthropic is in very preliminary talks to lease computing power from Meta, a person familiar with the matter told CNBC's Kate Rooney.
Shares of the social media giant climbed off their lows of the day Friday following a report from the New York Times that a potential deal was being discussed worth about $10 billion.
The talks come weeks after Anthropic announced a similar deal with Elon Musk's SpaceX to use the computing capacity at its Colossus 1 data center to improve capacity for paid subscribers.
They are a sign that Anthropic, one of the leading artificial intelligence labs, continues to make big commitments with other AI labs to use their access to AI chips made by Nvidia.
Access to enough AI chips remains a challenge for firms like Anthropic, which places usage limits on its most advanced models like Fable.
The talks also come after Meta CEO Mark Zuckerberg said in May that the social media company was considering entering the cloud computing business, in an effort to show investors that the firm can make money from AI investments beyond improvements to its current business. Dave Brown, a former senior executive at Amazon Web Services, is set to join Meta, CNBC has confirmed.
Meta could spend as much as $145 billion on capital expenditures, including for AI infrastructure, in 2026.
Last October, Zuckerberg said that companies are regularly "asking if we have compute that they could buy from us at some premium to what we've bought it at."
Meta declined to comment.
Four leading AI models discuss this article
"Meta's pivot toward AI cloud leasing could re-rate its multiple by monetizing $145B capex at premium margins amid structural GPU shortages."
The news signals persistent AI compute scarcity: Anthropic's early-stage talks to lease ~$10B of capacity from Meta (potentially at a premium) underscore how even well-funded labs face GPU shortages despite prior deals like SpaceX's Colossus. META shares bounced off lows, and with Zuckerberg's May comments on entering cloud computing plus Dave Brown's AWS hire, this could validate Meta's $145B 2026 capex bet by turning idle AI infra into high-margin revenue. However, "very preliminary" talks and usage limits on models like Claude suggest execution risk; the article downplays that Meta itself is still scaling its own Llama training needs.
The strongest case against is that these talks may collapse or amount to modest leasing at thin margins; Meta's massive capex could instead face ROI pressure if AI demand softens or if hyperscalers like AWS/Azure capture the bulk of third-party compute spend, leaving META's infra underutilized.
"Meta’s transition into a compute-as-a-service provider creates a new, high-margin revenue pillar that justifies its aggressive capital expenditure cycle."
This deal signals a fundamental shift in the AI infrastructure landscape: Meta is pivoting from a pure social media giant into a 'compute utility' provider. For Anthropic, this is a desperate move to bypass the supply chain bottlenecks hindering their model scaling. The bullish thesis for META is clear—monetizing idle GPU capacity at a premium margin transforms their massive $145B capex from a cost center into a recurring revenue stream. However, the market is overlooking the operational risk. Managing external cloud workloads requires a level of enterprise-grade reliability and customer support that Meta’s current infrastructure, built for internal social graph processing, may not be equipped to handle at scale.
Meta might be overestimating demand; if AI model training requirements plateau or shift toward more efficient, smaller-scale inference, Meta could be left holding the bag on billions in depreciating, underutilized H100 hardware.
"Meta's compute monetization is real optionality, but the article conflates preliminary talks with a viable business model, and glosses over the strategic contradiction of funding competitors while building Llama."
The article frames this as bullish for META—a $145B capex commitment monetizing excess AI infrastructure. But the framing obscures a critical tension: Meta is simultaneously building Llama to compete with Anthropic while leasing compute to it. That's either a sign Meta's capex vastly exceeds internal needs (bullish for monetization) or that Meta's AI roadmap is uncertain enough to justify hedging via third-party revenue. The $10B figure is unconfirmed and preliminary. More importantly, selling compute to AI labs that compete with your own models creates a perverse incentive structure—you're funding your own competition. The SpaceX deal precedent is thin; Colossus 1 is purpose-built for training, not commercial resale.
If Meta's capex is truly surplus to internal requirements, why not simply reduce it? The fact that Zuckerberg is publicly shopping for cloud customers suggests the $145B spend may be difficult to justify to investors without external revenue—implying the capex thesis itself is shaky.
"The real story is execution risk and capital intensity: this could be a meaningful AI compute monetization if terms and demand hold, but it could just as easily be a costly, uncertain bet that weighs on cash flow if utilization or pricing underperform."
While the chatter hints Meta could monetize AI compute and Anthropic could gain steadier capacity, the deal is far from done and terms are opaque. A $10B/year valuation implies aggressive economics for a leased-capacity model, but margins depend on utilization, power costs, and capex amortization. Meta’s $145B capex target by 2026 signals intensifying, capital-heavy competition in AI infrastructure, which could pressure near-term cash flow if demand softens. For Anthropic, external compute dependency raises concentration and roadmap control risks. Missing context includes term sheets, exclusivity, pricing, and whether Meta actually wants to become a cloud provider or simply a compute supplier for select partners. Execution risk is nontrivial.
The deal may never materialize or could be far less favorable in terms and exclusivity than suggested; Meta might build its own cloud or constrain access, and a $10B/year price tag could be untenable given cloud-margin dynamics.
"Meta's lease to Anthropic creates enforceable priority access that regulators could later challenge as anticompetitive coordination."
Claude correctly flags the perverse incentive of leasing to a direct Llama rival, but misses the second-order effect: Meta can throttle Anthropic's access during its own training cycles, effectively weaponizing the lease. This isn't hedging—it's strategic optionality. The real unmentioned risk is regulatory—DOJ may view compute-sharing among frontier labs as coordinated capacity hoarding.
"Meta's internal infrastructure is likely too specialized to support external commercial SLA requirements, creating a massive operational bottleneck."
Grok, your 'weaponized lease' theory ignores the reality of multi-tenant data center orchestration. Meta cannot simply throttle Anthropic without violating SLAs, which would destroy their credibility as a cloud provider. The real risk is the 'Colossus' factor: Meta's hardware is optimized for internal Llama training, not the diverse, low-latency requirements of Anthropic’s inference. If the integration fails, Meta’s $145B capex becomes a stranded asset that cannot pivot to a general-purpose cloud utility.
"Meta's capex thesis hinges on external monetization; if this deal fails, the entire $145B spend faces investor pressure."
Gemini's SLA argument is sound—Meta can't throttle without destroying cloud credibility. But Grok's regulatory angle deserves weight. DOJ scrutiny of compute-sharing among frontier labs is real risk the panel hasn't quantified. More pressing: neither panelist addressed whether Meta's $145B capex economics work *without* third-party revenue. If Anthropic deal collapses, does Meta justify that spend to shareholders on Llama ROI alone? That's the actual stranded-asset risk.
"The real risk to Meta’s capex ROI is underutilization and regulatory/SLA risk from a narrow tenant base, not just the possibility of throttling a single customer."
Challenging Grok: the ‘weaponized lease’ idea overstates the ability to throttle Anthropic without cascading reliability and regulatory backlash. The bigger, underappreciated risk is collateral damage to Meta’s cloud credibility if external users face throttling, price surges, or SLA breaches—prompting customers to flee to AWS/Azure-style ecosystems. Moreover, even if Anthropic uses the capacity, reliance on a single marquee client remains a governing risk to capex ROI, unless Meta signs a broad, diversified base of tenants.
The panel discusses Meta's potential to monetize its AI infrastructure by leasing compute capacity to Anthropic, with bullish views on Meta's recurring revenue stream and bearish concerns about operational risks, regulatory scrutiny, and the sustainability of Meta's capex without third-party revenue.
Monetizing idle GPU capacity at a premium margin, transforming Meta's massive capex from a cost center into a recurring revenue stream.
The risk of Meta's $145B capex becoming a stranded asset if Anthropic deal collapses or if the integration fails to pivot to a general-purpose cloud utility.