Out of the distributed platform: one host, and the microVMs it runs.
Lower-cost, flexible compute for recurring data jobs and model training, without long-term commitments,
run to a deadline and measured by what each completed job actually costs.
Out of the distributed platform: one host, and the microVMs it runs.
Most of the compute a product depends on is the recurring work behind it: ingesting and transforming data, refreshing models, scoring batches, building reports. Jobs arrive as containers with a schedule, a hardware requirement and a deadline.
Each is placed on the capacity that fits it, across Reserved Instances, capacity reservations, Spot and On-Demand, and moved when that capacity disappears. Retries, spend and waiting time are bounded, and the measure is what a successful run cost and whether it finished on time.
Training runs are large, bound to accelerators and expensive to interrupt, so each is planned as a window with the hardware, duration and checkpoints fixed in advance. In early scans, short GPU reservations for some types were listed at roughly 62 to 87 percent of the On-Demand rate; they cannot be cancelled, so a window is booked only when the run is ready.
Interruptible GPU capacity costs less but can be taken back without notice, so it is used only with frequent checkpoints. Each run is measured by the cost of a finished, validated model.
Reserved Instances discount baseline compute; capacity reservations hold physical CPU for a set time and are never promised to two clients at once; Spot and On-Demand cover flexible work and overflow. A discount is not a reservation, so we say which one stands behind your job.
Flexible jobs shift between locations as supply and prices change. Interruptible capacity can be taken back without notice, so we never promise uninterrupted execution on it, and you see the capacity available for your job before you commit.
Compute is placed in distinct regions, each an independent concentration of capacity.
Hardware is priced differently from one region to the next, and the gap moves. A spread is only worth taking after everything it costs: moving data, the time a transfer adds, and any rule about where the data may live.
Jobs with small or replicated inputs can run where the hardware is cheapest; jobs tied to large datasets or residency rules stay where they are. We show which region a job ran in, and why.
One workload. One measured result.