Meta Launches Meta Compute to Sell Computing Power: Model Progress Falls Short, GPU Leasing Becomes New Growth Driver

Meta announced in July 2026 the launch of an AI computing power leasing business, internally codenamed "Meta Compute," planning to open its massive AI infrastructure to external customers, including direct GPU leasing and hosting third-party models. This move comes as Meta's in-house model progress has fallen short of expectations, while its computing power investments continue to escalate—with 2026 capital expenditure guidance reaching $125-145 billion. Following the announcement, Meta's stock surged nearly 9% before retreating; the U.S. semiconductor sector plunged on concerns of oversupply, with Micron, SanDisk, and others dropping over 10%.

Background: Model Setbacks, Computing Power as a New Path

Meta has recently faced repeated setbacks in AI models: the Gemini model was restricted, internal AI agent development has been slow, and employee morale hit a 20-year low. Meanwhile, Meta has invested heavily in computing infrastructure—signing over 5 GW of data center capacity in the first half of 2026 alone, with two major campuses totaling 2.5 GW under construction, and nearly 10 GW in related transactions since early 2024. Faced with massive investments and difficulty monetizing models, Meta chose to lease out some computing power to improve asset utilization.

Key Details: Two Models of Meta Compute

• Bare-metal GPU leasing: Similar to the CoreWeave model, directly leasing GPU computing power. Referencing SpaceX's neocloud contract (three-year term, either party can cancel within 90 days), each GW can generate approximately $50 billion in annual revenue. If Meta allocates 200 MW of computing power, it could bring in $10 billion in annual revenue with very high margins.
• Model service platform: Similar to Amazon Bedrock, deploying third-party models on the infrastructure and selling them as a package. According to SemiAnalysis, Meta is in final negotiations with Anthropic to gain access to private instances of Claude, and may integrate models from Claude, OpenAI, or Google into the platform in the future.

Reactions and Market Impact

• Capital markets: On the day of the announcement, Meta's stock surged 8.81%, but fell 4.9% the next day. The U.S. semiconductor sector suffered heavy losses, with the semiconductor index dropping 5.44%, Micron and SanDisk falling over 10%, and Nvidia also weakening. A-share AI hardware chains also plummeted, with the ChiNext index closing down 5.71% and the STAR 50 index falling 7.70%.
• Analyst views: Some believe that Meta's leasing of computing power may signal a shift in AI infrastructure from supply shortage to periodic oversupply, undermining the scarcity logic that previously supported the hardware bull market. However, other analysts point out that Meta is still ramping up next-generation model training—its superintelligence lead Alexander Wang revealed that the next-generation model "Watermelon" requires an order of magnitude more training compute than its predecessor, and its benchmarks have already caught up with GPT-5.5. Additionally, Cailianshe reported that Meta signed a contract with Samsung worth over 10 trillion Korean won (approximately RMB 50 billion) for 2nm AI chip foundry, planning to mass-produce hundreds of thousands of chips, indicating that Meta has not stopped expanding its computing power.

Impact and Outlook

Once Meta's computing power business takes shape, it will directly compete with AWS, Azure, Google Cloud, and AI cloud providers like CoreWeave and Nebius. At the same time, Meta may leverage its advertising platform to build sales and marketing SaaS, integrating cutting-edge AI agents. Despite market concerns about oversupply, Meta's move is more likely about optimizing existing asset utilization rather than cutting investment. In the short term, Meta's stock volatility reflects the market's reassessment of AI investment returns; in the long term, whether computing power leasing can become a stable revenue source for Meta depends on its ability to find a differentiated advantage in the model competition.