Pricing
Priced per vCPU. Never percent-of-savings.
A flat, predictable rate on the capacity Temper actually manages. The measured packing delta alone was worth $4,102/month per 100 nodes at e2-standard-4 list price — you keep all of it.
$0 forever
For proving the kernel-enforcement story on a real cluster — the full engine, not a demo.
- Single cluster
- Full node enforcement engine: QoS tiers, workload profiles, safe mode
- Management dashboard: explorer, logs, perf panels, savings view
- Prometheus
/metrics,/observe, Grafana dashboards - Community support
- Multi-cluster hub
- Multi-user RBAC & audit export
$2 / vCPU / month *
For teams running the platform across production fleets with more than one pair of hands on it.
- Everything in Community
- Unlimited clusters + multi-cluster hub (peer registry, cluster switcher)
- Multi-user RBAC: viewer / operator / admin roles, named tokens, audit export
- L1 cluster intelligence: density-aware scheduler plugin + overcommit webhook
- Workload-profile training pipeline (observe → synthesize → evaluate)
- Priority support
Custom
For regulated, sovereign, and air-gapped environments — the deployments SaaS control planes structurally cannot serve.
- Everything in Pro
- Air-gap install kit: offline images, signed artifacts, SBOM
- Support SLAs with named engineers
- Security review package: whitepaper, threat model, architecture briefing
- Roadmap input (platform targets, kernel matrix)
- On-prem bring-up assistance
* Early access pricing — placeholder rate, subject to change. Billed on vCPUs of nodes where the Temper agent is actively enforcing. Nodes in safe mode or without sched_ext support are not billed.
01 How pricing works
Flat per-vCPU, on purpose.
The unit is the thing Temper manages: CPU capacity. No seats, no per-pod metering, no revenue share on a number we compute ourselves.
Why per-vCPU: Temper’s job is to make each vCPU carry more work safely, so the honest unit of value is the vCPU under management. The count is trivially auditable — it is the sum of CPUs on nodes where the agent is enforcing, the same number your cloud bill already shows — and it scales with the size of the problem, not with how many people log into a dashboard.
Why never percent-of-savings: savings-share pricing pays the vendor more the bigger the “savings” number gets — and the vendor computes the number. That incentive produces inflated baselines, generous counterfactuals, and measurement games. Our savings view deliberately splits realized (measured reclaim) from identified (estimated opportunity) precisely so the number stays honest — tying our revenue to it would corrupt the instrument. A flat rate keeps the measurement neutral: if Temper doesn’t pay for itself, turn it off; the kill switch is one annotation.
Worked example 100 nodes
100× e2-standard-4 (4 vCPU) = 400 vCPU under management.
Pro at the placeholder rate: $800/month.
Against it: the measured 41.9% packing delta on that machine type is worth
$4,102/month per 100 nodes at list price — before the
utilization-reclaim story on latency-critical nodes.
Packing delta measured live (docs/training-artifacts/binpack/SAVINGS-REPORT.md); your workloads will vary.
02 What’s in each tier
The full matrix, no asterisk maze.
| Capability | Community | Pro | Enterprise |
|---|---|---|---|
| Node enforcement engine (QoS tiers, sched_ext)The kernel layer — full-strength in every tier | Yes | Yes | Yes |
| Workload thread profiles (builtin + file-based) | Yes | Yes | Yes |
| Safe-mode kill switch & kernel-native rollback | Yes | Yes | Yes |
| Observability (/metrics, /observe, Grafana, linter, traces) | Yes | Yes | Yes |
| Management dashboard (explorer, logs, actions, savings) | Yes | Yes | Yes |
| Clusters | 1 | Unlimited | Unlimited |
| Multi-cluster hub (peer registry, cluster switcher) | No | Yes | Yes |
| Multi-user RBAC, named tokens, audit export | No | Yes | Yes |
| L1: density-aware scheduler plugin + overcommit webhook | No | Yes | Yes |
| Profile training pipeline (observe → synthesize → evaluate) | No | Yes | Yes |
| Air-gap install kit (offline images, signatures, SBOM) | No | No | Yes |
| Support | Community | Priority | SLA, named engineers |
| Security review package & roadmap input | No | No | Yes |
FAQ Pricing questions
The honest answers, up front.
- What are the real limits of the free tier?
- One cluster. That is the limit. The Community tier ships the full node enforcement engine and the dashboard — the headline latency results on the benchmarks page were measured with capabilities the free tier includes. What it lacks is fleet plumbing: the multi-cluster hub, multi-user RBAC, and the L1 placement layer.
- What counts as a vCPU?
- The CPUs of nodes where the Temper agent is actively enforcing — the same vCPU count your cloud provider bills you for on those nodes. Nodes in safe mode, nodes whose kernels lack sched_ext, and nodes without the agent are not billed. There is no per-pod or per-container metering.
- Is complement mode (under Karpenter / Cast AI) priced differently?
-
No. Complement mode is the same product with the placement layer switched off
(
scheduler.enabled=false) — the node engine, which is what you are paying for, works identically under any placer. Same rate either way. - Is it self-hosted? Do you host anything?
- Everything is self-hosted, in every tier — there is no SaaS control plane to subscribe to. Temper is a helm install into your cluster, makes zero external calls, and requires no account. Billing is a commercial arrangement, not a technical dependency: nothing stops working at a network boundary.
- Are there volume or design-partner discounts?
- Yes on both. Volume pricing kicks in at fleet scale (talk to us with your vCPU count), and design partners — teams who run Temper on real workloads and share benchmark data during early access — get locked-in launch pricing and direct engineering access.
- Why not percent-of-savings? Wouldn’t that be lower risk for us?
- It looks lower risk until you audit the savings number. Savings-share vendors are paid more when the computed “savings” is bigger, and they control the baseline and the counterfactual — a structural conflict of interest. We keep the savings instrument neutral (realized vs. identified, measured from your own telemetry) and charge a flat rate you can cancel by turning the product off. If the numbers ever stop justifying the rate, you have a one-annotation exit.