For inference providers

Maximize tokens per GPU.

Cedana restores frontier models in about a minute: a 1.6T model in 70 seconds. Collapse the warm pool, hold SLO through failures and upgrades, with zero changes to your engines or schedulers.

Schedule a demo

29.3×

faster cold starts

57–70s

restore, 229B–1.6T params

code changes

In production trials with leading AI infrastructure companies.

Benchmark / 8× B200

Native cold start grows with the model. Cedana restore doesn't.

Scale deployments for frontier inference.

8× NVIDIA B200 · 1.7TB memory · CUDA 12.9 · official sglang cookbook recipes · models ordered by parameter count

Cold start / native vs. Cedana restore

Key frontier models, measured.

Model	Params	Checkpoint (GiB)	Native	Cedana	Speedup
MiniMax-M2.7MoE	229B	244	564s	57s	9.9×
GLM-5.2-FP8MoE · Multimodal	753B	734	1322s	61s	21.7×
Kimi-K2.6MoE · Multimodal	1100B	670	1217s	63s	19.3×
DeepSeek-V4-ProMoE	1600B	873	2051s	70s	29.3×

Restore time scales with checkpoint size, not parameter count: MoE and quantization keep it sublinear as models grow.

Full benchmark: methodology and per-model results →

What it protects

Three taxes on your margin, removed.

Cold starts

More models on the same GPUs. Higher utilization.

Start multi-GPU frontier models in seconds to reduce overprovisioning.

Failures

Failures become fast, stateful resumes.

Automatically resume multi-GPU inference workloads through failures, without losing customer work.

Immobility

Drain, upgrade, rebalance without killing workloads.

Migrate live inference workloads off nodes for maintenance and engine upgrades, and rebalance fragmented capacity across the fleet.

Build vs. buy

The primitive your scheduler is missing.

Your stack including scheduler, router, cache tiering stays the same. Cedana sits transparently beneath them, capturing and restoring live execution state. Inference and engine agnostic.

No code, stack, or config changes.

Inference enginesvLLM · SGLang · TensorRT-LLM

OrchestratorsKubernetes · SLURM

Cedanacheckpoint · restore · migrate

CUDA driver & GPUsB200 · H100 · prior generations

Read the architecture →

The economics

Widen the spread between cost per GPU-hour and revenue per GPU-hour.

Cold starts: Reduce overprovisioning of warm replicas. The full catalog and every dedicated endpoint stay seconds away instead of always loaded.

Failures: Stop paying for failures twice. Stateful resume keeps multi-GPU workloads serving through faults, so an incident costs seconds, not SLA credits and lost customer work.

Immobility: Run the fleet hotter. Consolidate and rebalance live workloads, and drain nodes for upgrades and maintenance without taking capacity offline.

Why now

Inference is becoming stateful. State is becoming the bottleneck.

Agents run longer.

Long-horizon agents accumulate session state worth real money. Every eviction, failure, or cold start throws it away.

Models keep growing.

Frontier open-weight models push cold starts from seconds to half an hour, and warm pools stop being affordable.

More models, longer contexts.

Multi-model routing means constant swapping; growing context windows mean gigabytes of KV and session state per workload. More rebuilds, each one more expensive.

FAQ · What platform teams ask

Questions we hear first.

01We optimize cold start already. Why Cedana?

Because Cedana is complementary: it executes at the kernel level and leverages your existing model and kernel optimizations.

02Why should we try this now?

Inference is becoming increasingly stateful, complex, and expensive to serve, as we explain here →

03Does this work with NVIDIA Dynamo?

Cedana seamlessly integrates with Dynamo and provides additional capabilities: balancing prefill/decode with faster latency and higher granularity, and increasing reliability for higher SLA compliance.

04How long will this take to integrate and try out?

Integration happens within minutes. We work with you on a testing process to give you high confidence it works and you can scale.

05Will it work with our autoscaler and orchestrator?

Cedana works with Kubernetes and Kueue with a Helm chart. We have seamless integrations for SkyPilot and Slurm.

06How does it handle our serving framework (vLLM, SGLang, Triton)?

Cedana captures at the OS level, so the serving framework is irrelevant to the checkpoint layer. We show resume across vLLM and SGLang on the call.

07We've built our own warm-pool optimization. Why Cedana?

Most platforms have. Most result in overprovisioning as they scale across the model catalog, and this gets worse as model sizes grow. By accelerating cold starts that take 10 minutes or more to seconds, the warm-pool equation completely changes.

08What about compliance and data handling at our scale?

Cedana operates on GPU memory state inside your existing security boundary. It is SOC 2 compliant, ships on-premise, and supports fully air-gapped deployments.

09Does it handle distributed multi-GPU inference?

Yes. Distributed multi-GPU coordinated checkpoint is the differentiating capability. We show an 8-GPU model-parallel job restored as one unit on the call.

What Cedana is[primitive]

Cedana is the missing primitive for stateful inference. It checkpoints, restores, and migrates live GPU workloads across the fleet, without rebuilding state from scratch. By turning execution state into a schedulable object, Cedana lets fleets produce more tokens with higher reliability and lower cost.

save ▸ migrate ▸ resumezero code changes

More tokens per GPU.
More margin per token.

Schedule a demo Why inference is becoming stateful