Meta to Sell AI Computing Resources Amid Stalled Model Development, Shifts Focus to GPU Business

Meta is exploring the launch of a new business initiative called “Meta Compute,” which would allow the company to offer its vast AI computing resources to external customers. Reports from Bloomberg and analysis by SemiAnalysis suggest that Meta could enter the cloud computing business by leveraging its extensive GPU and data center capacity. Amid slower-than-expected progress in developing its AI models, CEO Mark Zuckerberg appears to have shifted the company’s strategy to prioritize monetizing its GPU infrastructure.

The Scale of Computing Resources and

Accelerated Procurement

According to a report by SemiAnalysis, Meta’s procurement of data center and computing resources is accelerating rather than slowing down. In the first half of 2026 alone, the company has signed contracts for over 5GW of capacity in cloud and colocation data centers, excluding the capacity of its in-progress data center construction projects. The two major data center campuses currently being built by Meta are expected to add a combined capacity of 2.5GW.

Since early 2024, Meta’s contracts for data centers and computing resources have already reached nearly 10GW. This scale is extraordinary, even compared to other hyperscalers. Meta plans to allocate these computing resources across multiple use cases.

First, the company will continue investing in its proprietary models. Meta has already released the “Muse Spark” model developed by Meta Super Lab (MSL) and is reportedly training its next-generation model “Watermelon.” Second, the resources will be used to enhance its ad recommendation systems. SemiAnalysis predicts that Meta could increase the complexity of its ad recommendation systems tenfold, using significantly more training and inference resources to boost ad revenue. Third, Meta is considering selling computing resources externally, inspired by “neocloud” deals like those of SpaceX. Finally, the resources could be used to host third-party models.

Revenue Projections from GPU Sales

SemiAnalysis estimates that high-value computing lease agreements, such as those employed by SpaceX, could generate annual revenue of approximately $50 billion per 1GW. If Meta were to offer just 200MW of computing resources to external clients, it could earn $10 billion annually with exceptionally high profit margins. The SpaceX-style agreements, designed to be flexible, allow for termination by either party within 90 days despite being contracted for three years. These agreements essentially function as auto-renewing three-month contracts. By adopting this framework, Meta would be able to lease its resources to external clients while retaining the ability to reallocate them for its own model development as needed.

Wall Street quickly reacted to the news, with Meta’s stock price surging by about 9%. In contrast, shares of neocloud companies like CoreWeave and Nebius were sold off, as the market began to anticipate Meta’s entry into the competitive landscape of AI cloud vendors.

Third-Party Model Hosting Strategy

Further analysis by SemiAnalysis reveals that Meta is in final negotiations with Anthropic, aiming to gain access to private instances of its Claude model. In the future, Meta could create a model service platform similar to Amazon Bedrock, Microsoft Foundry, or Google Vertex AI. Specifically, Meta might deploy third-party models like Claude on its infrastructure and sell them as packaged solutions to corporate clients.

This strategy has three main objectives. First, internal use: reports suggest that Google has restricted Meta’s access to its Gemini model, which might lead Meta to adopt Claude as an alternative. Meta’s AI projects require a massive number of high-quality model tokens, and Claude is currently among the most powerful models available. Second, external sales: customers could access these models via Meta’s platform without needing to directly manage contracts, deployment, or maintenance with companies like Anthropic. Third, vertical applications: Meta could integrate cutting-edge AI agents into its existing ad platform to create sales and marketing SaaS solutions.

SemiAnalysis predicts that Meta is likely to announce similar agreements soon, with Anthropic as the leading candidate, although OpenAI and Google might also become involved.

Strategic Shift Triggered by Challenges in

Model Development

The most direct reason for Meta’s pivot from model development to GPU sales is the enormous cost of model development. Meta’s 2026 capital expenditure guidance has already been revised upward to $125 billion to $145 billion. In Q1 2026 alone, capital expenditures reached $19.84 billion.

However, progress in model development has been underwhelming. While the Llama series has had a significant impact as open-source software, it has been difficult to translate this into direct revenue. Meta’s latest proprietary model, Muse Spark, has also fallen short of bringing the company back to the forefront. Internally, Meta is advancing the training of its next-generation model, Watermelon, which reportedly requires an order of magnitude more computing resources than its predecessor, Avocado. Alexander Wang, who oversees Meta’s AI projects, has stated that Watermelon is already on par with GPT-5.5. Additionally, the current version of Muse Spark is set to receive an update soon, promising significant improvements in programming capabilities and agent performance.

Despite Zuckerberg’s substantial investments in chips, data centers, and talent to catch up with OpenAI, Anthropic, and Google, Meta’s models have yet to convincingly establish themselves as industry leaders. If rapid progress in model development cannot be achieved, Meta’s computing resources may become its most easily monetizable assets in the eyes of Wall Street. GPUs and data centers, at the very least, can be priced and monetized through leasing, model hosting, API sales, services for advertisers, AI agent SaaS, and enhanced internal ad recommendation systems.

Impact on the Competitive Landscape

Should Meta move forward with its computing resource business, its competition will extend beyond just model development companies. It will also face direct competition from major cloud providers like AWS, Azure, and Google Cloud, as well as AI-focused cloud vendors like CoreWeave and Nebius. As one of the largest consumers of GPUs globally, Meta’s entry into this market with its surplus computing capacity could intensify price competition in the AI computing resource sector. Neocloud companies, in particular, may struggle to compete with Meta’s potentially low pricing strategy, as evidenced by the recent drop in their stock prices.

On the other hand, by opening up its vast infrastructure to external clients, Meta could contribute to the democratization of AI development. Startups and research institutions might gain easier access to high-performance GPUs. However, Meta’s dual role—continuing its own model development while hosting competing models like Claude—creates a unique dynamic of “coexistence and competition,” the outcome of which remains uncertain.

Editorial Opinion

In the short term, Meta’s GPU sales strategy is expected to diversify its revenue streams and alleviate pressure from investors. While its model development is not advancing as anticipated, rebranding its GPU infrastructure as a revenue-generating asset is a commendable move. However, the AI computing resource market is already highly competitive, and Meta’s ability to differentiate itself will depend on its pricing and service quality. Competition with neocloud companies like CoreWeave and Nebius seems inevitable.

In the long term, Meta will need to balance its proprietary model development with its infrastructure sales. While external GPU sales may provide short-term cash flow, cutting back too much on internal resources could further weaken the competitiveness of its AI models. Meanwhile, the strategy of hosting third-party models like Anthropic’s Claude to establish itself as an ecosystem hub mirrors the successful approaches of AWS and Azure.

References

• Meta Joins the Gold Rush: Selling GPUs Amid Model Delays - QbitAI, Published July 5, 2026
• Meta Compute: Everyone Wants to Be… - SemiAnalysis
• Meta Is Building a Cloud Business to Sell Excess AI Compute - Bloomberg, Published July 1, 2026