Market Analysis: Meta's Compute Power Leasing Resembles "Using Old to Support New" Rather Than Halting Pursuit of High-End Compute Power

According to a report by Jinse Finance on July 2, market analysis suggests that Meta is planning to launch a cloud computing business for leasing compute power. This includes two types: one involves opening up the model capabilities deployed on its own AI infrastructure to external clients, and the other involves leasing more fundamental "bare compute power." However, the current situation seems to lean more toward Meta using its existing older compute power for cash flow recovery, rather than halting its pursuit of high-end compute power. In mid-to-late June, Meta was reported to have signed an agreement with Crusoe to acquire approximately 1.6GW of AI computing capacity from two data centers located in Texas and Missouri. Meanwhile, in Q1 2026, Meta raised its full-year capital expenditure guidance to $125–145 billion. Viewed together, these two developments appear more like a reallocation of resources across different generations and use cases: continuing to purchase new GPUs to train cutting-edge models, while older GPUs (such as the H series) are used for inference in high-traffic products, hosting external models, and other application scenarios. Leasing out a portion of these older GPUs is acceptable, but this does not indicate a slowdown in the acquisition of highly sought-after high-end GPUs.