ExeleritGPU cloud without the surprise bill
H100 from $1.19/hr, L40S from $0.32/hr — roughly 20% under the lowest marketplace rate. Metered by the second, with hard spend caps so the invoice never outruns the plan.
No commitments · No quote calls · Cancel by destroying the box
Cheaper than the marketplaces, without the marketplace roulette
Marketplaces get cheap by auctioning someone's spare desktop — variable reliability, mystery hosts, surprise interruptions. We price under them anyway, on hardware we actually run.
| GPU | Exelerit | Vast.ai | RunPod | AWS |
|---|---|---|---|---|
| H100 80GB · frontier training | $1.19 | $1.49 | $1.99 | $3.90+ |
| L40S 48GB · serving & training | $0.32 | $0.40 | — | — |
| RTX 4090 24GB · inference & fine-tuning | $0.23 | $0.29 | $0.34 | not offered |
| Spot interruptible, same hardware | from $0.08 | varies by host | ||
Per GPU, per hour, on-demand. Competitor figures are the lowest advertised public on-demand rates observed in July 2026; marketplace rates fluctuate by host and availability, and AWS reflects H100 instance pricing normalised per GPU. Storage, egress and per-token endpoint usage are billed separately and shown live in the console.
A control plane, not just a box with a GPU in it
Everything between "I have a model" and "it is serving traffic" — already built, already metered.
Instant instances
Pick a GPU and a template and launch in under a minute. Whole cards or fractional slices, on-demand or spot, with SSH, shells and Jupyter built in.
One-click serving
Deploy any model to a vLLM-backed endpoint with an OpenAI-compatible API. Point your existing client at a new base URL and it just works.
Scale-to-zero endpoints
Serverless endpoints sleep when idle and wake on the next request. Bill per token or per GPU-second — no paying for an idle replica overnight.
Code sandboxes
Ephemeral isolated compute for agents and untrusted code. Create, exec, move files, destroy — over a clean SDK, with fractional GPUs from $0.04/hr.
Batch & fine-tuning
Queue batch jobs onto spot capacity at a discount, or run LoRA fine-tunes and serve the adapter straight from the same platform.
Real storage
Persistent volumes, snapshots, shared filesystems, a model-weight cache and an S3-compatible gateway — encrypted, and priced by the GB-hour.
Built for people who would rather type than click
A versioned REST API, typed SDKs, a single-binary CLI, and an MCP server so your agents can drive it directly.
# install the CLI curl -fsSL https://app.exelerit.ai/install/cli.sh | sh # launch a GPU box from a template, then drop into a shell exelerit instances create --template pytorch-2.3-cuda12.1 --gpu-count 1 --name trainer exelerit shell trainer # forward a local port straight into the box exelerit port-forward trainer 8888:8888
from exelerit.sandbox import Sandbox # an ephemeral, GPU-backed sandbox — destroyed on exit with Sandbox.create(api_key=KEY, base_url=URL, gpu="l40s/8", ttl_sec=600) as sbx: print(sbx.commands.run("nvidia-smi").stdout) sbx.files.write("/work/data.txt", "hello sandbox") print(sbx.files.read("/work/data.txt").decode())
from openai import OpenAI # your endpoint speaks the OpenAI API — just change the base URL client = OpenAI( base_url="https://app.exelerit.ai/api/inference/<endpoint>/v1", api_key="exl_live_...", ) resp = client.chat.completions.create( model="my-model", messages=[{"role": "user", "content": "Why is my GPU bill so weird?"}], )
The bill can't surprise you if it can't exceed the cap
Most GPU bill shock is structural: hourly rounding, idle replicas, forgotten boxes, and a dashboard that only tells you last month. We fixed the structure.
- Per-second metering. Stop at 04:12 and you pay for 4 minutes and 12 seconds — not the hour.
- Hard spend caps. Set a ceiling per account. Cross it and new workloads stop, instead of quietly billing on.
- Scale-to-zero by default. Idle endpoints suspend themselves and stop charging for GPU time.
- Live cost, not a monthly reveal. Every instance, endpoint and volume shows its running rate and accrued spend as it happens.
- Spot when it fits. Interruptible capacity at a steep discount, with a notice window before reclaim.
The things people actually ask
How can you be cheaper than a marketplace?
Marketplaces are cheap because they resell idle consumer hardware from anonymous hosts — you take on the variance in reliability, bandwidth and interruption. We run our own fleet and price under the lowest marketplace rate we can find, so the cheap option is also the predictable one.
What happens when I stop an instance?
Billing stops with it, to the second. You can suspend a box to keep its state and stop paying for GPU time, or destroy it outright. Persistent volumes keep billing at the storage rate until you delete them — and the console shows that rate the whole time.
Do I have to rewrite anything to serve a model?
No. Endpoints expose the OpenAI-compatible API, so an existing OpenAI client works by changing the base URL and key. Anthropic-style /v1/messages is supported too, which means coding agents can point straight at your own served model.
Is my data isolated from other tenants?
Yes. Every tenant-scoped query runs through an application-level isolation layer with Postgres row-level security behind it, workloads are sandboxed at runtime, per-tenant networks are isolated, and secrets and volumes are envelope-encrypted. Administrative actions are recorded in a tamper-evident audit log.
How do I pay?
Accounts run on prepaid credit today, and billing is invoiced manually while we onboard early users — so there is no card on file to run away with. Usage is metered continuously and visible in the console the whole time.
Can I get in right now?
We are onboarding early users by request while capacity scales. Email us and we will set up an account, seed it with credit, and get you to a running GPU the same day.
Get your first GPU running today
Tell us what you're building. We'll open an account, add starting credit, and you can be training or serving within the hour.