Block's Dessa: One Rack, Zero Text, and the 1000x Token Gap

Somewhere inside Block's infrastructure, a single GB200 NVL72 rack ran for roughly a month. Seventy-two Blackwell GPUs, about 120 kilowatts, a room that pulls more power than a small office building. The output of that month is a model that never read a word. That is the entire claim. No text corpus. No pretraining on the open web. No instruction tuning on the pile of filtered human writing that every frontier model is built on. Just human conversation audio, tokenized, and an autoregressive loop trained from zero until it could listen and speak at the same time, faster than real time, on a MacBook Pro. Strip the framing away and you have two hard numbers and one hard absence. The numbers: one rack for training, one laptop for inference. The absence: every architectural detail that would let anyone outside the building verify the claim. Dessa is being described, by people close to it, as a GPT-3 moment for voice. I want to test that sentence against arithmetic. Because arithmetic does not care about press releases.