Not available yet Engineering preview

Fleet compute

Later, a job that is too heavy for this machine will be able to rent compute from the Tether fleet and run here anyway. That rental is not offered yet. There is no price, no account, and no way to join a list from this page.

What follows is only the machinery that already exists, and the rule the bill will follow when rental exists. Nothing on this page starts a job or contacts a machine.

What will be billed is time this machine measured

A borrowed interval is the finish time minus the start time, both read from the clock on the machine that asked for the work. A donor cannot hand in a CPU figure, a flop count, or a price, and a reservation a provider claimed is not a measurement. A span the client clock cannot re-check is refused, not rounded into a bill.

Three read-only views, no rental

  1. Borrowed seconds, per donor

    tether usage reads the borrowed-time log on this machine and prints seconds per donor for today, the last 7 days, and all time. It takes only --json. It never writes, and a log it cannot re-read prints an empty summary instead of a guessed total.

    tether usage

  2. The live borrow, per donor

    tether donor-stats reads the store on this machine and prints, per donor, cores, memory, OS, units out, units returned, the rate, the share, the speedup and the efficiency. It always prints one JSON snapshot; its flags only point it at other files, never at a mode. It never dispatches, and a figure it cannot read is written as null rather than as zero.

    tether donor-stats

  3. The donor live view

    The status dashboard has a Donors panel. It shows whether each paired machine is qualified, the measured time per unit, and how many units are out now. A figure the store cannot price reads "not measured". It never reads as zero, and it is never filled in with a stand-in.

    Open the status dashboard

What the fabric measured

Every figure below comes from one committed record, produced by an interleaved paired run: local-only and borrowed, alternating, in the same minute. The record names what it ran on. A record this page cannot read makes the build refuse, rather than printing a number nobody measured.

Speedup, median

—×

spread —–—× over — pairs

Efficiency

—%

of the linear ceiling with — workers

Donors

—

job —

—

The record is fleet-scaling.json, generated by tools/tether_fleet_scaling_run.py and checked by tools/tether_fleet_scaling.py at build time.

You pay for measured saved seconds, not for hours held

Every other compute marketplace bills the time an instance was held, priced by the seller. This one bills the time the fleet actually saved you: the local-only wall of a job minus the wall the same job took with donors, both measured on this machine, in interleaved pairs of the same job. A job the fleet made slower costs nothing, and the bill cannot go below zero.

Rate

$—/s

published list rate, not a measurement

Saved per job, median

—s

measured range —–—s

Cost of that job

$—

derived: rate x saved seconds

A renter's quote from history

A quote for a new job is derived from the measured one below. It scales the measured walls linearly, so it is an ESTIMATE and is labelled as one; the measured figures it scales from are shown unchanged beside it.

— units: estimate would cost $— and save —s (measured: —s)

A donor's earnings from the same job

Each donor is credited the measured saving times the share of the job it ran. The local machine is a worker too and is never paid: a credit to yourself is not a payment. These are derived figures at the published rate, not a payout record.

—

The record is fleet-rent.json, derived by tether_fleet_rent.py from the same committed measurement the section above publishes.

What this page will not do