We've made some improvements that improve usage on the long tail for power users of Astra when logged in with your ChatGPT account.
No change in quality and a pure win that on the long tail can result in up to 3-4X less usage being drawn from the subscription.
A first-party product update says recent changes can reduce the subscription usage drawn by long-tail Astra power-user workloads by up to 3–4x, with no quality change. That is a material cost update to the September 3 benchmark picture, although the affected workload distribution is still unspecified.
On power, Abilene’s original full-campus plan was 1.2 GW, with an eventual deployment reported at 450K+ GB200 GPUs. The training cluster represents only part of that campus.
If Jensen’s “400K next” refers to Vera Rubin GPUs, the power requirement changes materially. Using…
On power, Abilene’s original full-campus plan was 1.2 GW, with an eventual deployment reported at 450K+ GB200 GPUs. The training cluster represents only part of that campus.
If Jensen’s “400K next” refers to Vera Rubin GPUs, the power requirement changes materially. Using Supermicro’s 227-kW-per-rack sizing implies ~1.26 GW at the racks alone (400K ÷ 72 × 227 kW), or ~1.5+ GW at the facility using similar overhead assumptions.
So, that exceeds the original Abilene power budget before retaining any existing Blackwell capacity? Accommodating it would require additional sites or power, or a lower-power operating configuration. Tbf, the GPU count alone doesn’t establish where that capacity will run.
The worked estimate turns a GPU headline into a siting constraint. At 72 Rubin GPUs and 227 kW per rack, 400,000 GPUs imply about 1.26 GW at the racks and more than 1.5 GW at facility level under similar overhead assumptions, above Abilene's original 1.2 GW plan. Read the full post for the crucial caveat: the GPU count does not establish where that capacity will run.
Sentiment
AI infrastructure demand meets power constraints+0.12