Meta Compute: The Neocloud Revolution
The recent buzz around Meta's potential transformation into a Neocloud has sparked intense market reactions, with a notable sell-off in Neocloud stocks like Coreweave and Nebius. However, I argue that these reactions are misguided, and Meta's datacenter and compute procurement strategy is set to accelerate, not slow down. In 2027, we can expect shockingly high capital expenditures, with Meta already contracting over 5GW of capacity in the first half of the year, not accounting for their self-build initiatives.
Meta's datacenter expansion is relentless. The two largest campuses, as depicted, represent 2.5GW of capacity under construction. This rapid growth challenges the notion that US datacenters are delayed, as some headlines suggest. My previous article, 'Stop Saying Half of US Datacenters Are Delayed', delves into this misconception.
The question arises: what will Meta do with all this compute? I outline four high-value use-cases that differentiate Meta's approach from traditional Neoclouds:
- Frontier AI Models: Meta's commitment to training frontier models remains strong, with significant capacity allocated to Meta Superintelligence Labs. A follow-up report will explore their data strategy and chances of catching up with industry leaders.
- RecSys: Meta aims to scale up Ads recommendation systems by over 10x, requiring substantial inference and training compute. This will enable more generative targeted ads, driving revenue growth.
- (SemiAnalysis Exclusive) Meta is in advanced talks with Anthropic to access private instances of Claude, similar to Bedrock, Foundry, and Vertex from other hyperscalers. This opens up various use-cases, including internal usage and the development of Sales & Marketing SaaS powered by Frontier AI Agents.
- SpaceX-type Deals: Meta is likely to pursue large-scale on-demand compute deals, similar to SpaceX's approach. These deals offer high revenue potential, with a few hundred MWs driving over $10B in yearly revenue.
The diversity of these use-cases provides Meta with high optionality. Meta Superintelligence Labs remains a key focus, but other high-margin alternatives, such as RecSys, Bedrock-type token services, and SpaceX-like deals, offer flexibility. This strategy makes it easy for Meta to invest heavily in compute, even if MSL doesn't pan out.
Meta's datacenter model, which includes self-build, leasing, and cloud renting, has seen significant growth since early 2024. The bulk of capacity additions are now through third parties, positioning Meta as a significant source of RPO growth for Neoclouds like Coreweave and Nebius. The model also categorizes capacity into MSL, Other AI, and Non-AI, highlighting the potential for various monetization avenues.
The AI Cloud TCO model team tracks GPU cloud deals, providing valuable insights into pricing and contract terms. The recent SpaceX-Anthropic deal, with its short-term, high-revenue potential, is a unique phenomenon. This deal's structure, with a 3-year term and a 90-day cancellation option, is rarely seen due to the high financing burden for Neoclouds.
Meta's focus on Superintelligence and its recent datacenter designs, like the 'tent' ultra-fast datacenter, position it well to capitalize on this new market. I predict that Meta will announce a SpaceX-like deal soon, potentially with Anthropic, OpenAI, or Google. This deal will kick-start the flywheel, generating substantial revenue.
Additionally, Meta's partnership with Anthropic to sell Claude models presents several monetization opportunities. These include internal usage, selling Claude as a service, and building applications. The distribution potential and network effect make this a strategic move for both Meta and industry leaders like OpenAI and Anthropic.
The key to Meta's success lies in RecSys, which has significantly contributed to revenue acceleration. The company has dramatically re-accelerated revenue growth, driven by GPU investments, improved yields for advertisers, and increased time on platform. The question of sustainability remains, but Meta's non-SuperIntelligence AI chip fleet is delivering outstanding ROI.
In conclusion, Meta's compute strategy is a strategic move towards becoming a Neocloud, with a focus on high-value use-cases and flexible monetization options. The company's rapid datacenter expansion and innovative designs position it well to capitalize on the AI infrastructure market, potentially reshaping the industry.
