The chain.

Interruptible machines are cheap because they die; on‑demand is dear because it doesn't. That spread is the reliability premium — and reliability is a portfolio property, not a hardware property. Chain deep-discount machines behind a standing order and the portfolio survives what any one machine can't. Priced here off the live book, this minute.

gpu

Run the hot pair today.

Two standing orders, one checkpoint bucket, two sidecars — whichever machine dies refills itself from the book while the other keeps serving:

# order A and order B — same spec, same line
curl -X POST https://compute.wick.pics/api/orders -H 'Content-Type: application/json' \
  -d '{"gpu":"RTX 4090","max_price_per_gpu_hr":0.05,"offer_class":"interruptible",
       "standing":true,"auto_destroy_budget_usd":5}'
# one keyholder per order, same CHECKPOINT_URL in both containers
ORDER_ID=A ORDER_SECRET=... PROVIDER_KEY=... CONFIRM=1 python3 keyholder.py &
ORDER_ID=B ORDER_SECRET=... PROVIDER_KEY=... CONFIRM=1 python3 keyholder.py &

The limits, plainly.

The chain re-provisions machines, not memory. It fits checkpointable training and stateless inference; sticky in-memory state dies with its host, and no chain fixes that. Gaps are minutes, not zero for a solo machine; the hot pair covers the gap by already running. Floors move — the numbers above are this minute's book, not a contract. And none of this is an SLA: as the trust tape and heartbeat ledger accrue, this page will quote measured preemption rates and delivered uptime per class — measured history, labeled young until it isn't, never a promise.