Cedana is the compute multiplier for frontier inference.
We increase tokens per GPU-hour at fleet-level scale. Transparent checkpoint, migrate, and resume turns GPU-hours lost to cold starts, failures, idle windows, and stranded placement into delivered tokens. Real-time GPU virtualization optimizes runtime execution to increase token throughput. These proprietary primitives deliver new capabilities for inference.
Expand your existing capacity without new GPUs
Fleet output is useful GPU-hours times tokens per hour. Cedana raises both, from a single GPU to the fleet, with no changes to your models, engines or orchestrator.
Frontier models resume 10 to 40x faster.
An 8-GPU, trillion-parameter model restores from checkpoint in about 60 seconds instead of a 10 to 40 minute cold start.
Idle GPU windows convert to finished work.
A spot tier runs stateful inference through idle windows across the fleet. When the GPU is reclaimed, the job moves with its state.
Failures, patches, and rebalancing run as policy.
Jobs resume on healthy hardware after a failure, pause and resume through a patch, and follow changing priorities, with no loss of state.
Each GPU delivers more tokens per hour.
Runtime execution is optimized beneath the engine for 10 to 20% more throughput, on top of your speculative decoding and kernel work.
Built for frontier inference at fleet scale.
Single GPU to multi-node. Prefill and decode pools autoscale independently and rebalance in seconds.
Scale agentic workloads with less VRAM.
Tasks route to the right model and escalate only when needed. Cedana swaps models in and out of VRAM in seconds, for over 2x the tasks at lower cost.
Full-stack inference out of the box.
Bring your own stack, or start with NVIDIA Dynamo, vLLM and SGLang support and Cedana Storage Express.
Compute multiplier in action
Same hardware, same models, native versus Cedana. Every number links to the run it came from.
- BenchmarkCold startsFour frontier models on 8× B200 · 57–70 s restore vs 9.4–34.2 min native29×faster restoreView benchmark
- BenchmarkAgentic workloadsFour coding models on 8× RTX PRO 6000 · 14 GPUs of models served on 8 · 52% lower cost2.1×tasks per hourView benchmark
- Interactive replayIdle spot tier on Slurm24-hour run, 8 GPUs · 588 preemptions · 8 of 8 jobs completed87%of idle time convertedOpen the replay
- Interactive replayIdle spot tier on Kubernetes28.9-hour run on Amazon EKS · 630 preemptions · 28 s median handover to the priority job68%of idle time convertedOpen the replay
- Write-upModel portfolio on one nodeNVIDIA Dynamo and NeMo Switchyard · each swap costs 3.4% of the cold start it replaces1,891 GiBon a 1,341 GiB nodeRead the write-up
Multiply the fleet you already have.
Same hardware, same models, same serving stack. Bring one workload and we will show you the GPU-hours it gets back and the tokens it delivers.