Meta renting out computing power is more like "using the old to support the new" rather than stopping the pursuit of high-end computing power.

But currently, the real situation seems to be: Meta is cashing out its stock of old computing power, rather than stopping its pursuit of high-end computing power. In mid-to-late June, it was reported that Meta had signed an agreement with Crusoe, planning to obtain a total of about 1.6GW of AI computing capacity from two data centers in Texas and Missouri. At the same time, Meta's Q1 2026 still raised its full-year capital expenditure guidance to $125-145 billion. Looking at these two things together, it appears more like a reallocation of resources of different generations and purposes: continuing to buy new cards to train cutting-edge models, while old cards are used for inference in high-traffic products, hosting external models, and other application scenarios, making it acceptable to rent out a portion. However, this does not mean that purchases of the most in-demand high-end cards will slow down.