Pay only for what you use, measured to the second. No reserved pods disguised as serverless. No capacity minimums. GPU rates that reflect the isolation tier you actually chose.
Three billing structures. Same hardware, same isolation tiers. Pick the one that fits your risk profile and commit pattern — or mix and match across projects.
Full rate. No commitment.
Provision immediately and run until you stop it. Priority allocation, guaranteed availability, instant start. No minimums, no lock-in. The right choice when you need predictability without risk.
40–60% off on-demand.
Interruptible instances priced aggressively. Spot instances draw from the same hardware pools as on-demand — same GPUs, same isolation tiers, with a pre-emption window for graceful shutdown. Built for workloads you can checkpoint.
30–50% off. Guaranteed capacity.
Pre-commit to GPU capacity for 1-, 3-, or 6-month terms. Locked pricing, zero capacity risk, and SLA-backed uptime. The right structure for production inference servers and predictable training pipelines.
All pricing is per-second. GPU model rates vary by tier and availability. Spot discounts shown as typical range — actual rates are market-driven.
Rates shown are on-demand per-hour equivalents for reference. Actual billing is per-second. Spot rates run 40–60% below these figures. Reserved rates require a volume quote.
— indicates tier unavailable for this GPU model. Rates are indicative and subject to change. Contact us for committed pricing.
Your bill accrues from the moment a GPU is allocated to your workload to the moment it's released. No rounding to the nearest minute or hour. A 90-second serverless invocation costs exactly 90 seconds worth of GPU time.
There's no minimum run duration on any tier. Provision for 10 seconds or 10 days — the meter starts and stops with your workload. No idle charges, no session fees, no account minimums.
GPU time counts when a GPU is allocated and running. Queued time is free. Networking and egress are billed separately at cost. Storage is billed per GB-month. There are no hidden platform fees.
Every billing period ships with per-workload usage records: GPU type, tier, start/stop timestamps, duration in seconds, and the effective rate applied. Export to CSV or pull via API.
The price difference between tiers is real and intentional — it reflects the actual cost of the isolation guarantees provided. Trusted is the most expensive because VM-level hypervisor isolation on hyperscaler infrastructure genuinely costs more. Community is cheapest because it offers less. You choose the trade-off explicitly.
VM-level hypervisor isolation on major cloud infrastructure. You're paying for hardware-enforced boundaries, provider SLAs, and zero shared-kernel risk.
Verified colocation partners with contractual SLAs and physical controls. Aircloud provides platform-level monitoring. Solid isolation at a meaningful cost reduction vs hyperscaler.
Self-registered operators with reputation-based scoring. Docker-hardened containers, not VM isolation. Best prices on the platform. Right for development and non-sensitive batch workloads.
No. Billing tracks GPU allocation, not calendar time. If your serverless function completes in 45 seconds, you pay for 45 seconds. If you shut down a dedicated instance, billing stops immediately. Queued time waiting for GPU allocation is never charged.
When you request a spot instance, you accept that it may be reclaimed if higher-priority demand materializes on the same hardware. You'll receive a configurable pre-emption warning (default: 30 seconds) before the instance is terminated. Spot instances are best suited to workloads that checkpoint state regularly.
Yes. You select a supply tier per workload, not per account. You can run your production inference server on Trusted, your staging pipeline on Secure, and your dev jobs on Community — all under the same billing account with consolidated invoicing.
Spot is available on Secure and Community tiers. Spot pricing on each tier is calculated against that tier's on-demand baseline, so Community spot will be cheaper than Secure spot for equivalent hardware. Trusted tier does not currently offer spot.
Reserved pricing requires a commitment for a fixed term (1, 3, or 6 months). You specify the GPU type and quantity upfront. The discount reflects both the term length and volume. Reserved capacity is guaranteed — you will not be pre-empted or capacity-constrained. Contact sales to get a quote.
The GPU rate covers the compute itself. Networking ingress is free. Egress and inter-region transfer are billed at cost. Persistent storage is billed per GB-month. There are no platform fees, no seat fees, and no undisclosed charges.
Large training runs, production inference fleets, multi-tenant deployments. We'll build a pricing structure around your actual usage — with committed discounts, SLA terms, and dedicated support.