Meta is gearing up to venture into the cloud computing business, with plans to roll out two distinct service offerings. The first involves opening up its AI infrastructure's model capabilities to external clients, while the second centers around leasing out underlying "bare computing power." However, the true picture is more nuanced. Rather than abandoning its quest for high-end computing power, Meta is likely recycling cash flow generated from older computing resources.

In mid-to-late June, reports surfaced that Meta had inked a deal with Crusoe, securing approximately 1.6 GW of AI computing capacity from two data centers located in Texas and Missouri. Simultaneously, Meta revised its full-year capital expenditure guidance for the first quarter of 2026, raising it to a staggering $125-145 billion. Taken together, these moves suggest that Meta is strategically reallocating resources across different generations and use cases.

The company continues to invest in purchasing new cards for training cutting-edge models. At the same time, it leverages older cards (such as the H series) for inference tasks in high-traffic products, hosting external models, and other applications. Additionally, Meta is open to leasing out a portion of its computing power. Nevertheless, this strategy does not imply that Meta will ease up on acquiring the most sought-after high-end cards, which remain a critical component of its technological ambitions.