The order book, counted

Jensen Huang's headline number for the Blackwell era is a trillion-dollar order book. Air Street Press went and counted how much of it is switched on. Working bottom-up from public filings, operator disclosures, EuroHPC and Top500 records, its State of AI Compute Index tallies about 2.07 million Grace-Blackwell GPUs across the industry — and finds only 100,128 of them deployed and running. Another 15 percent are mid-install. The remaining 80 percent, some 1.66 million chips, exist only as announcements: contracted, or press-released, but not yet built.

A GPU order is not a cluster. A cluster is not always available capacity. And contracted capacity is not the same as a model training run. — Air Street Press

The gap between those buckets is the whole point. A signed order is not a cluster; a cluster is not always available capacity; contracted capacity is not a training run. The two largest 'announced' programs — a 600,000-GPU Saudi build and the 450,000-GPU Stargate site in Abilene — are exactly the sovereign and hyperscale commitments that fill a keynote slide and take years of power, land and permitting to become anything you can rent.

The sharpest line in the data is generational. The previous chip, Hopper, still has 460,904 GPUs deployed — about 4.6 times more silicon actually humming than its celebrated successor, which only crossed 100,000 live units this month. The frontier everyone is racing toward is, for now, mostly running on last year's hardware. That cuts against the reflex reading of the same statistic — that a wall of idle Blackwell is about to flood online. Most of it isn't idle. It doesn't exist yet.