Pricing

Priced on workloads, not on promises

Turbokern is sold as a platform subscription sized by workloads and GPU targets, plus the GPU compute a search uses. Estimate the GPU-hours first; we put a price on it after the baseline benchmark, when both sides know what is at stake.

What you pay forModel
Platform
Workloads, GPU targets, seats
Search compute
Benchmark GPU-hours
Your GPUs
No compute charge
Edge
Per device class

Plans

Four ways to start

Design partner

Pilot

Pilot terms

One workload, one GPU target, with our engineers alongside.

  • Profyle and the full loop
  • Engine, recipe and report
  • Weekly reviews
Request a benchmark

Teams

Team

Quoted per scope

For a team with a handful of production models.

  • Up to 10 workloads
  • Cloud or VPC
  • Re-runs on model changes
Estimate GPU savings

Organizations

Platform

Quoted per scope

Every model, every GPU, as a standard step before release.

  • Unlimited workloads
  • CI integration and policies
  • Pooling opt-in
Talk to an engineer

Regulated and edge

Enterprise

Quoted per scope

On-prem, air-gapped and device-class work.

  • On-prem via AI Enterprise (planned)
  • Edgyn device classes
  • Custom guards and approvals
Talk to an engineer

No public prices yet. Pilot terms are agreed per design partner.

How pricing works

Three parts, all visible on the run report

01

Platform subscription

Annual, sized by the number of workloads and GPU targets you optimize and the seats on your team.

02

Search compute

GPU-hours spent benchmarking candidates on our fleet. Runs on your own GPUs carry no compute charge.

03

Value-based option

Once savings are proven in production, an option to price on a share of measured savings. Under consideration.

Under consideration

Estimator

Estimate GPU savings, and the search compute it takes

Everything here is in GPU-hours, not dollars. The speedup is the target you set, not a Turbokern result; a baseline benchmark is what turns it into a measured number.

Estimated GPU savings: serving GPU-hours freed at your target

0GPU-hours / month

GPUs needed at your target
0
Benchmark compute for the search
0 GPU-hours
Without Simyx pruning
0 GPU-hours
Net GPU-hours freed
0
Runs per month
0
Candidates benchmarked
0
Plan that usually fits
Team

Planning assumptions: GPUs needed = GPUs today ÷ throughput gain, rounded up; serving GPU-hours freed = (GPUs today − GPUs needed) × 730 hours. Runs = workloads × GPU targets × runs per month; search GPU-hours = runs × candidates × share benchmarked × minutes per benchmark ÷ 60. Real runs vary with model size and GPU. Illustrative only, not a quote or a measured result.

Compare plans

What each plan includes

FeaturePilotTeamPlatformEnterprise
Profyle, Kernl, Quantra, ServexYesYesYesYes
Edgyn device classesOneAdd-onAdd-onYes
Simyx what-ifNoNoPlannedPlanned
DeploymentCloud or VPCCloud or VPCCloud or VPCPlus on-prem (planned)
CI integrationNoYesYesYes
Engineer supportWeekly reviewsBusiness hoursNamed engineerNamed engineer

FAQ

Questions engineers ask

Why no prices on the site?

Turbokern is pre-launch, and the value depends on the size of your inference bill. We price a plan after a baseline benchmark shows what is achievable.

Do we pay for GPU time on our own hardware?

No. Search compute is charged only for runs on our fleet.

What does the pilot cost?

Pilot terms are agreed per design partner and depend on scope. Ask us when you request the benchmark.

Is there a free tier?

Not at launch. We may offer free profiling with Profyle later.

Next step

Measure first, then talk price

Request a benchmark with your workload and GPU target. Plan options follow once the baseline run shows what is achievable.