Your AI spend has a control plane now.
Roundhouse sits over the cloud accounts and hardware you already own. You bring your provider keys — we never resell a token, and your provider bills stay yours, at your own rates. For thirty days it runs in full, free, and shows you exactly what you're overpaying. Then it's a flat monthly fee to keep routing: every job to the cheapest place that can finish on time — your own machines at $0, your own cloud accounts only when a deadline demands it — with budgets enforced and savings you can reconcile to the cent.
Works with Ollama, vLLM, LM Studio, and llama.cpp on your hardware — AWS Bedrock and the Anthropic API on your cloud accounts.
Deadline-aware routing across everywhere a job could run
Three steps between a job and the cheapest place it can run on time.
Submit with a deadline
Every job declares a priority, a deadline, and the capability tier it needs — basic, standard, or frontier. That is the whole contract.
It runs in the cheapest capable place
The router estimates completion time and cost for every executor that could take the job — each of your own machines and each qualifying cloud model — and picks the cheapest one that meets the deadline. Hardware you own wins whenever it can finish in time: $0 marginal cost, no per-token bill.
It escalates only when it must
Queued work is re-checked every 30 seconds. The moment local capacity falls behind — a backlog grows, a machine goes offline — at-risk jobs move to the cheapest capable cloud model.
Already on AWS Bedrock? Roundhouse exposes a Converse-compatible endpoint — keep your request/response code, change only the endpoint URL and the auth header, and every call becomes a deadline-aware, fully metered job. See the Bedrock migration path →
Thirty days to see the number. Then decide.
The trial — 30 days, $0
Connect your own provider accounts and your own machines and run the whole product. Every job is metered — cloud calls at the prices on your bill, local runs at the cloud price they would have cost.
You finish the month with a number computed on your own traffic. If it doesn't clear the fee, don't subscribe.
route and up — a flat monthly fee
Keep the router on. Every job runs in the cheapest place that meets its deadline, budget caps hold spend by policy, and the savings report becomes a bill that's actually shrinking.
No per-request fees, no percentage of your savings — the fee is flat, published, and on the pricing page.
Pay cloud prices only for work that actually needs the cloud
Own hardware? A team generating 200M output tokens a month on a standard-tier API pays about $2,000 a month for output alone. If machines you already own absorb 70% of that at $0 marginal cost, roughly $1,400 a month never gets billed.
| Monthly workload — 200M output tokens | All-cloud | Routed with Roundhouse |
|---|---|---|
| Runs on your own hardware | 0 | 140M tokens (70%) · $0 |
| Billed by the cloud | 200M tokens | 60M tokens |
| Output bill at ~$10 / M tokens | ~$2,000 | ~$600 |
| Never billed | — | ~$1,400 / month |
Illustrative math based on public standard-tier API pricing (~$2 per million input tokens, ~$10 per million output). Actual savings depend on your workload, hardware, and deadlines. Roundhouse meters every execution — including on-prem runs at $0 — so your dashboard shows real spend against a real all-cloud baseline, not an estimate.
No hardware? Optimization still pays.
Capability tiers keep bulk work off frontier-priced models, the router picks the cheapest cloud model that qualifies for each job, and daily budget caps hold spend instead of letting it drift. Every decision is recorded and every execution metered — so you can see, job by job, what each choice cost and why.
And capacity you don’t own is coming.
Next: capacity from providers who have GPUs to spare, used whenever their price beats the cloud baseline — over the same pull-based, credentialed, fully audited worker protocol your own fleet uses, opt-in per job. The router already treats any executor uniformly, so the marketplace is a routing strategy, not a rewrite. Join the waitlist to hear when it opens.
Different applications, different model classes
Each application declares the model class its jobs require, and every job is routed independently. Bulk work stays cheap; hard problems get frontier models.
Bulk classification & extraction
Tagging, routing, entity extraction — high-volume work a small model on your own machine handles all day at $0 marginal cost.
Summarization & drafting
Reports, structured generation, everyday language work — solid mid-class models, local when they can finish on time, cloud when they can’t.
Hard reasoning
Complex analysis and high-stakes output — jobs that require a frontier-class model and escalate to the cloud when the tier demands it.
Built to never surprise you
Deadlines are contracts
The queue is swept every 30 seconds. A job escalates to the cloud before its deadline slips — not after. A machine going offline triggers the same automatic failover.
Daily cloud budget caps
Set a daily cloud budget. When the cap is reached, jobs hold rather than silently overspending — and individual jobs can be flagged to override when they truly must run.
Every decision explained
Each routing decision records which executors were considered, their ETAs and costs, and why the winner won. No black-box placement.
Metered to the token
Every execution is metered — on-prem runs at $0 included — so the dashboard shows exactly what you spent versus what an all-cloud setup would have cost.
A two-sided market for AI capacity
Next, the router’s reach extends beyond hardware you own. Orgs with idle GPUs will earn per-token compensation completing jobs; orgs without hardware will use that capacity whenever a provider’s price beats the cloud baseline. Not live yet — the early-access waitlist is open on both sides.
Invitation only, for now
We’re onboarding a small number of organizations that run real AI workloads. Tell us about yours and we’ll reach out as capacity opens up.
Already invited? Sign in.