Fair questions.

Isn’t this just a front-end to Vast.ai?

Ten feeds, not one — and yes, Vast dominates the cheap interruptible end, because that is genuinely where the market’s floor lives; reporting that honestly is the job, not a flaw in it. The one-row-per-model floor view draws from seven different feeds right now. And the things you come here for — the workload lens that re-prices the book for your task, the price tape recorded since first boot, resting limit orders, accountless tripwires — exist on none of the source marketplaces.

Zero fees — so what’s the business?

The fee lever exists and is public: every API response states fee_bps: 0, and metered usage is priced at the moment it happens, so a future fee can never reach backwards into old usage. Outbound links may carry referral tags; the renter’s price is unchanged and the docs say so. We charge nothing today because an instrument earns trust before it earns basis points — when a fee arrives it will sit on execution services, never on reading the market.

Isn’t idle capacity just unreliable garbage?

It’s reclaimable, and priced like it — that’s not a defect, it’s an asset class. Checkpointed training, batch inference, render queues, CI: restartable work soaks interruptible capacity at roughly half the on-demand price. We quantify exactly that trade — the idle-discount instrument on the front page is nothing else — and on-demand is one toggle away. What we won’t do is pretend the reclaim risk isn’t there: the table’s idle dot says it plainly.

You don’t operate hardware. How is this a market?

We run no metal and sell no SLA — deliberately, and the site says so wherever it matters. What rests here and nowhere else: standing limit orders with signed fill tickets, executed by your own keyholder. The ladder is public — aggregate, index, broker, clearing — and each rung ships when it’s real. Google Flights isn’t an airline; it’s still where flights get found.

Why do some prices look impossibly cheap?

Because interruptible marketplaces are auctions: the listed price of an idle machine is a real, current ask that holds only while you’re the top bid. That’s stated in the table’s fine print, and it’s why a consumer card can print at a cent an hour. It’s not demo data — it’s what the bottom of a bid market actually looks like, labeled instead of hidden.

What will I actually pay beyond the $/hr?

The fine print, read so you don’t have to — each claim quoted from the provider’s own pages with a verified date at /api/costnotes. The traps we’ve verified so far: on Vast, bandwidth is billed both directions at host-set rates and never appears in the $/hr — the classic “credit drained faster than expected”; stopping an instance does not stop storage charges. On RunPod, a stopped pod’s volume disk bills at double the running rate, and a balance at $0 terminates pods with their data. Hyperstack bills per-minute (coarser than the per-second rails) and meters public IPs separately. And where a provider doesn’t publish its granularity or egress policy at all, we say exactly that — undisclosed fine print is itself the warning. Providers we haven’t verified get no claims from us.

Why aren’t AWS, GCP and Azure in the book?

Deliberate. Our buyer is price-driven; the hyperscalers are the price ceiling and the quota wall — useful context, not useful rows. If you need an enterprise SLA and a procurement contract, they are the right answer and we’ll say so to your face. The book stays where the spot economics are.

Is the environmental angle greenwashing?

We publish one measured number — the live idle share of the listed market — and one argument made in the open: the factory footprint of an existing card is already spent, so work routed onto idle silicon adds only its marginal wall-power, while the same demand met with new hardware pays the factory again. What we refuse to publish: carbon calculators, offsets, badges, or any wattage we didn’t meter. /eco tells a Scope 3 reader to their face that they get no certificate here.

Where does the data come from, and how fresh is it?

Ten marketplace feeds, keyed where a key buys depth. The feeds people actually rent from refresh about every five minutes; the sleepy catalog feeds every thirty; the history tape records half-hourly bars — and the split is printed, not hidden. Every feed’s health, age and staleness is public at /api/stats: a dead provider shows as down, never silently reused.

Can an agent really rent through this without a human?

Reading the market is fully keyless — no signup, no token, MCP or plain JSON. Execution is honest about its boundary: rentals run on your provider key passed through a single call and never stored, or on a resting limit order your own keyholder daemon fills — credentials never live here, and neither does custody. Wallet-native keyless rental is the published next rung of the ladder, and it will ship when it’s real, not before.

What could this station still do to me if it were compromised?

Three things, stated here rather than left to be worked out from the design. An external reviewer asked for the plain version and was right to.

It could fill inside your line. The station holds your order secret, so a compromised station can cut a fill ticket at any price up to your maximum — it cannot exceed the line you set, and the price is re-quoted live against the provider before money moves, but “anywhere at or under your limit” is a real range and you should set the limit accordingly.

It cannot spend your key. Your provider key rides one pass-through call and is never stored. A resting order executes on your machine, by your keyholder, so a station that vanishes or misbehaves cannot rent on your behalf.

Your spend cap does not depend on us. The cap lives in your keyholder, not in this station’s memory, and its clock is written to disk — so our restart, our outage, or our compromise cannot lift it. We also run a second guard here, but it is a backstop, not the thing you are relying on. That ordering is deliberate: it was the other way round until 2026-08-25, when the same reviewer pointed out that a station restart dropped the guard while a rental kept billing.