← Machinery Machinery · 05

A listing is a claim.
Billing starts anyway.

A rented GPU costs money from the moment it boots and long before it can do anything useful, so five measurements decide — inside a hard time budget — whether it ever gets work, and the ones that fail are destroyed rather than parked.

The funnel

The cheap filters read a description. The expensive ones read the hardware.

Renting compute from a marketplace is a two-stage problem, and the stages have very different prices. The first stage is cheap: filter a catalogue by an allow-list of acceptable GPU models, a ceiling on hourly price, a floor on memory, a floor on the host's own reliability history, then take the cheapest offer that still qualifies. Nothing has been spent at this point — and nothing has been verified either, because a catalogue entry is a description written by the seller.

The second stage is expensive, because it happens on hardware you are already paying for. The billing clock starts at "rented", not at "ready to work." Boot, first answer, and every measurement taken afterwards are all spend against a machine that has demonstrated nothing. That window is the reason the audit is bounded by wall-clock time rather than retried until it succeeds.

SEARCH CATALOGUE FILTERS OFFER PICKED BILLING STARTS MACHINE RENTED NOTHING PROVEN YET IT BOOTS THE AUDIT FIVE MEASUREMENTS IN A FIXED ORDER ALL PASS REGISTERED MAY TAKE WORK ANY FAIL DESTROYED NOT PARKED BILLED · NOT YET USEFUL THIS WINDOW IS WHY THE AUDIT HAS A DEADLINE
The clock starts at rented, not at registered. Everything inside the bracket is money spent on a machine that has not yet shown it can do anything — which is why the audit is given a deadline rather than patience.

The home page shows the short version of this: ask a pool for five machines, watch more than five get booted, watch each one get measured, and watch the ones that fail get killed. This page is what measured means, and why the window in which you are paying for an unproven machine is deliberately short.

The checks

Five measurements, in order, fail-fast.

Each check runs on the machine itself, in a fixed order, cheapest first. The first failure ends the audit — there is no reason to time the disk of a box whose card is already wrong.

The sequence starts with authentication, and that is not a formality. A marketplace can mark a machine ready as soon as its login port accepts a connection; that is a statement about the port, not about the machine. Completing an actual authenticated session is the first thing that distinguishes a running host from an open socket, and a whole class of broken rentals lives in the gap between those two facts.

AUTHENTICATE A REAL LOGIN, NOT AN OPEN PORT THE GPU CARD PRESENT, MEMORY MEASURED THE SPECS CORES, MEMORY, DISK CAPACITY DISK SPEED A TIMED WRITE OF 512 MB NETWORK A TIMED PULL OF 100 MB FIRST FAILURE ENDS THE AUDIT — DESTROYED THE MEASUREMENTS ARE KEPT, THE MACHINE IS NOT
Authentication runs first because a machine that answers on a port has not yet proved it is a machine. Any step can end the audit, and the step that ends it is the step that gets recorded.
CheckWhat it actually measuresWhat it catches
Authenticate A login session completes, not merely a port answering A host that is listening but not usable
The GPU The card responds and its memory is read from the driver A different card, fewer cards, or less memory than listed
The specs Core count, system memory, disk capacity An oversubscribed host handing you a slice of what was advertised
Disk speed A timed 512 MB write to the working volume Shared or throttled storage — a model load that takes an hour
Network speed A timed 100 MB download A link too slow to pull model weights in usable time
Why the audit has a deadline The whole sequence runs under a hard wall-clock budget of roughly two and a half minutes. That is containment, not a performance target: an audit that hangs is a machine billing indefinitely with nobody watching. The audit is allowed to be wrong about an unusually slow machine. It is not allowed to be patient.
The trap

The check that rejects the hardware you asked for.

This is the check anyone would write, us included, and it is the reason this page exists. A card does not report the size it is sold as. The driver holds back a slice for its own bookkeeping, so the figure you can measure sits a little under the figure on the listing. Write the comparison the obvious way — measured memory must be at least the requested memory — and it rejects exactly the hardware you asked for, every single time, with nothing that looks like a bug anywhere in the code.

The consequence is not a rejected machine. It is a loop. The rental is destroyed for failing a check it can never pass. The next search returns the same shape, because the same shape is still the cheapest offer matching the request. It is rented, it reports the same number, it is rejected again. Each pass around that circle is a real rental with a real bill and no output — and from the outside it does not look like arithmetic. It looks like scarcity: capacity that can never quite be found.

RENT THE SAME SHAPE AGAIN — BILLED ON EVERY PASS THE LISTING A ROUND NUMBER ON A MARKETPLACE CLAIMED THE HARDWARE REPORTS SLIGHTLY LESS THE DRIVER KEEPS SOME MEASURED THE CHECK MEASURED < LISTED ? WRITTEN THE OBVIOUS WAY SHORTFALL REJECTED GOOD HARDWARE THROWN AWAY WITH A 5% ALLOWANCE ACCEPTED THE CARD YOU ASKED FOR
One comparison closes the circle. The tolerance is the only thing standing between a working fleet and a treadmill that bills on every turn and produces nothing.

The fix is small and unglamorous: measured capacity is compared against the requested figure less five percent. That absorbs what any driver holds back and is nowhere near enough to admit a card of a genuinely smaller class. A threshold set to exactly the right number is a threshold that is sometimes wrong, because hardware does not report round numbers.

The part that is easy to miss The same allowance has to be applied by every later decision that asks the same question. If admission at rental time and placement at run time disagree about whether a card is big enough, a machine gets admitted and then permanently refused work — or rejected by one path and immediately re-rented by the other. One number, one rule, applied everywhere it is asked. Two components quietly disagreeing about a threshold is the same failure as two components disagreeing about a fact.
Evidence

The measurements outlive the machine.

A machine that fails is torn down within seconds of the verdict. If the measurements lived on that machine, every failure would be permanently unexplainable: you would know that something failed, never what was measured or by how much it missed. So each check writes its result away from the machine as it completes — which step, what value, and the reason for the verdict, not merely pass or fail.

Passes are recorded as well as failures, and that is the part people skip. A single box failing a disk timing is a bad box. Every box of one shape failing the same timing this week is a supply problem worth acting on, and the second pattern is only visible if you kept the numbers from the machines that passed. Records that only exist when something goes wrong cannot tell you that things are getting worse.

The verdict

What the platform does with the answer.

Four outcomes, two of which are refusals. The refusals are the ones that cost nothing and save the most.

  1. Act — admit it to the fleet All five measurements passed. The machine is registered with its measured specification rather than its advertised one, becomes visible in the fleet, and may be given work. Every later scheduling decision uses the number that was measured, not the number that was promised.
  2. Refuse — destroy it, do not park it Any single failure ends the audit and the rental is torn down immediately. A parked machine bills at exactly the same rate as a working one, so "leave it up, we will look at it later" is the most expensive sentence in this system. There is no quarantine tier, because a quarantine tier is a subscription.
  3. Act — keep the numbers, not the box The verdict, the value behind it and the step the audit reached are written before the teardown starts. A rejection can be answered a week later without anyone having to reproduce it, and the same record is what makes a provider's drift measurable rather than anecdotal.
  4. Refuse — stop renting the shape soon When the same combination of image, card and host keeps failing the same way, the combination itself comes out of the search rather than being retried. Retrying an unclassified permanent fault is not resilience; it is the same money burned on a slower schedule.
The difference this makes Without an audit, a bad rental is discovered when a real job dies on it hours later, "the card was smaller than advertised" is an argument rather than a measurement, and the same shape is rented again tomorrow. With one, the discovery happens in minutes, the claim is a number taken before the machine ever saw work, and a shape that keeps failing stops being offered to you at all.

Run cheap GPUs without the bill shock.

Request early access →