- Meta’s attempt to build frontier AI has faltered badly: its Muse Spark model — built by Superintelligence Labs, staffed with $300 million pay packages and anchored by Scale AI founder Alexandr Wang (acquired for $14.3 billion) — hasn’t matched OpenAI or Anthropic’s benchmark performance, several star researchers have departed to rival labs, and Meta itself acknowledged gaps in coding and agents; its parallel open-source Llama effort has also struggled to keep pace with the frontier.
- Zuckerberg is now exploring a pivot to selling computing power to rival AI labs — a strategic retreat from model-building toward infrastructure monetization, echoing the logic of Amazon Web Services, which launched as a way to monetize spare server capacity and now generates roughly $45.6 billion in annual operating income at ~35% margins, contributing the majority of Amazon’s operating profit despite representing only 17% of revenue.
- The parallel to Elon Musk is direct: SpaceX has been leasing compute from its Colossus 1 and 2 data centers to Anthropic (reportedly $15 billion in annual revenue) and Google, after Musk’s Grok AI model fell behind on capability benchmarks and failed to gain meaningful enterprise market share — both Zuckerberg and Musk built massive compute infrastructure for AI model ambitions they couldn’t fully realize, and are now monetizing that infrastructure by renting it to the labs that beat them.
- Analyst Richard Windsor at Radio Free Mobile estimates that Meta could sell a thin slice of spare data center capacity in 2026 and scale to 5 gigawatts by 2030 — priced at roughly construction cost — potentially adding ~$70 to Meta’s share price (currently ~$583); Meta’s existing 4.8 GW North American footprint across 30+ data centers makes it the fourth-largest compute operator behind Amazon (10.6 GW), Microsoft (5.5 GW), and Google (5.2 GW), with $115-135 billion in 2026 capex targeting 10+ GW by year-end.
What Happened?
Bloomberg Opinion columnist Parmy Olson argues that Meta and xAI have both effectively failed to build competitive frontier AI models — and that both Zuckerberg and Musk are now pivoting to the more reliable business of selling compute. Meta’s Muse Spark system, built by the lavishly funded Superintelligence Labs team (including Scale AI’s Alexandr Wang, acquired for $14.3 billion at a 49% stake), hasn’t matched the benchmarks of OpenAI or Anthropic, with acknowledged gaps in coding and agentic capabilities. Multiple star researchers have departed. In parallel, Meta’s open-source Llama models have struggled to stay at the frontier. Zuckerberg is now exploring turning Meta’s vast data center network into a cloud business that sells compute to other AI labs — including potentially the same labs that have outcompeted Meta on model quality.
Why It Matters?
The pivot carries real logic. Meta has committed $115-135 billion in 2026 capital expenditure largely to build AI infrastructure — a massive fixed-cost investment that generates returns only if the underlying compute is utilized. If Meta’s own models can’t fill that capacity, selling the spare to external AI labs converts a stranded cost into recurring high-margin revenue. The AWS analogy is instructive: Amazon didn’t set out to build a cloud business — it built infrastructure for its own needs and discovered that selling spare capacity was enormously profitable. Meta’s situation is even more acute: it derives ~98% of revenue from online advertising, making diversification into any stable, high-margin business strategically urgent. A compute-leasing revenue stream would reduce Meta’s dependence on the cyclical ad market and provide a hedge against the structural threat of AI chatbots diverting attention from the social media feeds that drive advertising. Whether the math pencils out depends heavily on how much capacity Meta actually frees up and what customers pay — two variables Windsor’s $70 share price uplift estimate explicitly treats as uncertain.
What’s Next?
Watch for any formal announcement from Meta of a cloud compute product or partnerships with external AI labs — the shift from “exploring” to signing contracts would be the moment this pivot becomes real. Olson frames Musk’s SpaceX compute leasing to Anthropic as the clearest precedent, and the reported $15 billion in annual revenue from that arrangement suggests the demand is real and the pricing is attractive. The risks are material: the AI infrastructure boom could overshoot, leaving the industry with excess compute capacity that drives prices down — the same dynamic that periodically crushes cloud pricing in commodity computing. Meta is also building toward 10+ GW by end of 2026, which assumes sustained AI demand growth that isn’t guaranteed. The broader question is whether compute leasing becomes a structural business for Meta or merely a transitional hedge while it figures out its AI model strategy.
Source: Bloomberg Opinion
