Meta is launching a GPU compute product offering ('selling shovels'), betting that GPUs must be profitable before model performance matters.
Meta has faced a series of setbacks: Gemini models have been restricted from use, Zuckerberg admitted that internal AI agent technology development is slower than expected, and employee morale has hit a 20-year low. According to Bloomberg, Meta is considering launching Meta Compute, opening its massive AI infrastructure to external customers.
According to SemiAnalysis, Meta's data center and compute purchases won't slow down; instead, they will continue to accelerate. In just the first six months of this year alone, Meta has already signed for over 5GW of capacity in cloud and managed data centers—not including the proprietary data centers it is rapidly developing.
Meta has multiple uses for this compute capacity. First, it will continue powering its own models, including the already-launched Muse Spark and the next-generation model Watermelon currently in training. Second, it will enhance its advertising recommendation system. SemiAnalysis believes Meta may want to increase the complexity of its advertising recommendation system tenfold, using more training and inference compute to boost ad revenue. Third, it enables vertical applications—Meta can leverage its advertising platform to build sales and marketing SaaS integrated with cutting-edge AI agents.
The economics are compelling. With just 200MW of compute allocated to external customers, Meta could generate $10 billion in annual revenue at extremely high margins. SpaceX has pioneered a contracting model that Meta could adapt: nominally three-year contracts where both parties can cancel within 90 days—effectively rolling three-month terms with automatic renewal.
SemiAnalysis suggests that Meta is in final negotiations with Anthropic to gain private instance access to Claude. In the future, Meta will build a model services platform similar to Amazon's Bedrock, Microsoft's Foundry, and Google's Vertex. Meta can deploy third-party models like Claude on its own infrastructure and resell them as a package to enterprise customers. This creates multiple revenue streams: selling Claude-as-a-service like Amazon's Bedrock, where customers don't need to negotiate with Anthropic or handle deployment and operations themselves—they simply call the model through Meta's platform. This move is particularly strategic given that Google just restricted Meta's use of Gemini; Meta may turn around and offer Claude as a replacement.
SemiAnalysis predicts that Meta may soon announce similar agreements, with Anthropic as the top candidate, though OpenAI or Google could also become partners.
If Meta's compute business materializes, its competitors won't just be model companies like OpenAI, Anthropic, and Google. It will also compete directly with cloud providers like AWS, Azure, and Google Cloud, as well as AI cloud companies like CoreWeave and Nebius. Meta's stock surged nearly 9% on this news, while neocloud companies like CoreWeave and Nebius faced sell-offs.
However, Meta faces an ongoing challenge: its latest in-house model, Muse Spark, hasn't yet truly returned Meta to the top tier of AI capabilities. Meta is internally training the next-generation model Watermelon, reportedly with compute investment an order of magnitude higher than its previous Avocado model. The current version of Muse Spark is also slated for a major update with significant improvements in coding capabilities and agent functionality. Yet despite the spending, Meta still needs to convince developers and customers that its models stand at the forefront of the industry.
Ultimately, the strategy is flexible: these compute resources can be rented to others, can host third-party models, can sell APIs, can serve advertisers directly, can provide AI agent SaaS, or can continue to enhance Meta's own advertising recommendation system internally.