Enterprise Support packages
Subscriptions are priced per GPU running the supported software, per year, on hardware you own or lease. Three packages; the engine is the same in each. What changes is the security posture, the tooling, the lifecycle commitment and the support hours. We publish what each package includes and how it is metered. Quotes follow a short conversation about your fleet, because the price depends on GPU count, term and coverage hours, and a number on a page that does not fit you is worse than no number.
Base is available from LTS-1 launch. Enhanced follows the hardened LTS-1 build and the baseline Tuner. Premium follows the advanced Tuner and Traction Inference Reliability. The Runtime Assessment is available now.
| Package detail | BaseEnterprise support for the community edition In developmentavailable from LTS-1 launch Per GPU, per year | Recommended starting point EnhancedA hardened build you can stay on for a yearIn development Per GPU, per year | PremiumGuarantees when upstream does not deliver, and the full picture Roadmap Per GPU, per year or on measured savings after sixty days | |||
|---|---|---|---|---|---|---|
| Comparison | Comparison | Comparison | Comparison | |||
| Who it is for | Teams on upstream who want someone to call | Teams that need a hardened build they can stay on | Fleets that need guarantees beyond upstream and round-the-clock cover | |||
| Engine track | Latest two upstream releases, fix-forward | LTS branch, twelve months of backports, published EOL | LTS branch with extended lifecycle and branch overlap | |||
| Security fixes | Next upstream build | Upstream fixes backported to your branch | Plus our own patch when upstream is silent; day-zero advisories | |||
| Hardened profile | —Not included | Included, with checklist | Included, with checklist | |||
| Signed builds, SBOM, scan | —Not included | Included | Included | |||
| Traction Tuner | —Not included | Basic: rule-based, recommend-only or verify-or-revertEarly access | Advanced: regime detection, warm-started optimization, multi-replicaIn development | |||
| Visibility | Diagnostics bundle | Diagnostics bundle, open metrics emitter, Prometheus export | Plus Traction Inference ReliabilityIn development | |||
| Coverage | Business hours | Business or extended hours | Business, extended or round-the-clock; named engineer | |||
| Air-gapped bundle | —Not included | Included | Included | |||
| Source escrow | —Not included | On request | Included | |||
| Response targets | In the data sheet and agreement | In the data sheet and agreement | In the data sheet and agreement | |||
| Status | In developmentavailable from LTS-1 launch | In development | Roadmap | |||
| The three packages | The three packages | The three packages | The three packages | |||
| What each package includes | For platform teams already running upstream SGLang who need a vendor on the hook.
| For regulated enterprises and operators who need secure-by-default SGLang on a pinned branch, with measured tuning and without outcome pricing.
| For sovereign and regulated buyers, large GPU fleets, and air-gapped sites that need lifecycle continuity beyond what upstream provides.
| |||
How metering works
The unit.
A GPU running the supported software in the trailing quarter. In plain words: at each quarterly true-up we count the GPUs that ran Traction Runtime during the previous three months, from the engine's own inventory, and that is what the year's subscription covers. Elastic pools, scale-to-zero deployments and mixed-engine fleets pay for what actually ran.
Per node.
A per-node option covers an 8-GPU server as one unit, for teams that buy and account for whole nodes.
Fleets.
Volume and fleet terms apply above 1,000 GPUs.
Channels.
Available by direct agreement, through cloud and GPU-operator partners, and through cloud marketplaces as listings go live.
Start with the Runtime Assessment
How most subscriptions begin. Available now.
A fixed-fee engagement on your own cluster, typically two to three weeks. We run the benchmark harness on your workload, size it with our predictor, apply the hardened profile, and leave you a measured baseline, a hardening report and a written recommendation for package and configuration. The fee is credited against a subscription signed within ninety days.
Premium follows the same path: Enhanced first, the tuner in recommend-only mode for sixty days, and savings measured on your own cluster before anything is priced on them.
Questions buyers ask
Why are there no prices on this page?
Because the price depends on GPU count, term, coverage hours and whether you need the LTS branch, and a published figure that does not fit your fleet wastes everyone's time. We publish exactly what each package includes and how it is metered, and we quote after one conversation. Premium can be priced per GPU per year or on measured GPU savings after sixty days on your own cluster.
Is Traction Runtime endorsed by the SGLang project?
No. Traction Runtime for SGLang is an independent downstream distribution built on the open-source SGLang engine, which is released under the Apache License 2.0 and hosted by LMSYS. Traction Layer AI is not affiliated with, sponsored by or endorsed by the SGLang project. We send our scheduler handlers, backports and security fixes upstream and keep our patch set public.
Is the engine modified?
Upstream SGLang pinned to a tested release, plus a small published patch set: backported fixes, hardening and the scheduler handlers we submit upstream. The API is unchanged. A workload that runs on upstream runs here.
Who answers the ticket?
The engineers who build the branches, carry the backports and contribute to upstream SGLang. There is no first-line script. Severity definitions are on the support policy page; response targets per package are in the data sheet and agreement.
What is the LTS cadence?
One long-term-support branch per quarter, cut from an upstream release that has passed our hardened-profile, upgrade-path and performance-regression tests, with twelve months of backported fixes and an end-of-life date published at cut and never moved. LTS-1 is cut from v0.5.20. Base tracks the latest two upstream releases instead.
Can we run air-gapped?
Yes. Offline bundles carry the images, charts and signatures, and verify without a network. The license verifies offline with a grace period. No component makes an outbound connection. Air-gapped delivery is included in Enhanced and Premium.
Do you support vLLM?
Not yet. Traction Runtime for vLLM, applying the same lifecycle, hardening and tuner model to vLLM, is on the roadmap. Teams standardized on vLLM should talk to us about timing.
Do we still need NVIDIA AI Enterprise?
They are different things and they coexist. NVIDIA's stack packages the engine inside Dynamo; Traction Runtime gives you a pinned branch with backports, the hardened profile and vulnerability response on that branch. The tuner operates beneath Dynamo Planner and consumes its metrics.
Can we run it on GPUs we rent?
Yes. Metering counts GPUs that ran the software in the trailing quarter, so dedicated clusters from a GPU operator, elastic pools and scale-to-zero deployments pay for what ran.
What if upstream governance changes?
Your LTS branch keeps receiving fixes for its published lifetime regardless. Builds are reproducible from source, and source escrow is available on request.
Can we try it first?
The public mirror carries evaluation builds with no registration at launch. For production candidates, the Runtime Assessment is the trial, and its fee is credited.