Meta starting to sell excess compute is essentially a re-pricing of the valuation logic for AI semiconductors. Moreover, the market has begun to reward the first “surrender” in AI capital expenditures, which itself suggests that the market recognizes this logic. Bulls may have their own opinions: they say demand is still tight, but the issue is that the market doesn’t agree.
For Meta and SpaceX, which lack model and cloud service resources, selling compute is their optimal solution—because they have compute but no services to monetize.
1. Compute isn’t truly scarce; OpenAI and Anthropic’s compute capacity that can actually monetize users is severely insufficient, while Meta and SpaceX have excess compute. Their compute utilization rates are too high (i.e., too much idle capacity), and depreciation alone can swallow profits.
2. Meta and SpaceX, the first to start selling compute, don’t themselves have sufficiently strong models, paid users, or cloud servers for business operations.
3. Morgan Stanley estimates Meta has 3GW of compute, while Anthropic and OpenAI have 1.4GW and 1.9GW, respectively. The sense of compute scarcity that comes with the supply shortage you feel is largely driven by Anthropic and OpenAI, so the market perception is the most obvious. Meta doesn’t have much in terms of services and doesn’t need that much compute. Also, the grok model used for SpaceX is well-known to be subpar, so the market obviously can’t sense how abundant their compute is—only they know.
4. In 2026, cloud providers will increase capital expenditures by 70%; more than half of that is absorbed by hardware price hikes.
For Meta and SpaceX, which lack model and cloud service resources, selling compute is their optimal solution—because they have compute but no services to monetize.
1. Compute isn’t truly scarce; OpenAI and Anthropic’s compute capacity that can actually monetize users is severely insufficient, while Meta and SpaceX have excess compute. Their compute utilization rates are too high (i.e., too much idle capacity), and depreciation alone can swallow profits.
2. Meta and SpaceX, the first to start selling compute, don’t themselves have sufficiently strong models, paid users, or cloud servers for business operations.
3. Morgan Stanley estimates Meta has 3GW of compute, while Anthropic and OpenAI have 1.4GW and 1.9GW, respectively. The sense of compute scarcity that comes with the supply shortage you feel is largely driven by Anthropic and OpenAI, so the market perception is the most obvious. Meta doesn’t have much in terms of services and doesn’t need that much compute. Also, the grok model used for SpaceX is well-known to be subpar, so the market obviously can’t sense how abundant their compute is—only they know.
4. In 2026, cloud providers will increase capital expenditures by 70%; more than half of that is absorbed by hardware price hikes.