reply I estimate ~8-18 GW AI-operational today. Total US data center capacity is ~50 GW across all workloads. Frontier lab run-rate revenue is ~$110B annualized (mid-2026) but that doesn't capture internal hyperscaler AI use, which may be larger than all frontier lab inference combined. As for efficiency, Blackwell delivers ~30x more inference tokens per watt vs H100, so yes capability is scaling faster than power draw. But cheaper tokens create more demand than they displace (Jevons.) Reminder: DeepSeek cut prices 75% in May, got overwhelmed by demand, raised peak-hour prices in July, and is now warning of a broader hike.

reply I estimate ~8-18 GW AI-operational today. Total US data center capacity is ~50 GW across all workloads. Frontier lab run-rate revenue is ~$110B annualized (mid-2026) but that doesn't capture internal hyperscaler AI use, which may be larger than all frontier lab inference combined. As for efficiency, Blackwell delivers ~30x more inference tokens per watt vs H100, so yes capability is scaling faster than power draw. But cheaper tokens create more demand than they displace (Jevons.) Reminder: DeepSeek cut prices 75% in May, got overwhelmed by demand, raised peak-hour prices in July, and is now warning of a broader hike.

Source:https://x.com/matthew_sigel/status/2086276613721755841