HOW WE MEASURE · LAST UPDATED Q3 2026
Every number on this site has a source.
We state what we measured, how we measured it, and where the data came from. If a stat needs a hedge, it has one. If it needs attribution, it has that too.
GPU UTILIZATION
Measured per deployment
Source
Measured on your own cluster during a Proof of Value. We report average GPU utilization inside active time-sliced sessions, a production average and not a peak.
How to read it
We don’t publish one number for every deployment. We measure it on your cluster during a Proof of Value. The industry baseline for enterprise on-prem GPU utilization is 10 to 15% (Goldman Sachs and Cast AI, 2025).
Cite as Juno Innovations Measurement Methodology, Q3 2026. juno-innovations.com/methodology#gpu-utilization
COMPUTE COST REDUCTION
up to ~40%
Source
R3D Studios cloud billing while they ran Orion, April to September 2025 against October 2024 to March 2025. Compute-only line items, no credits applied.
How to read it
GPU time-slicing replaced 1:1 GPU assignment, and suspend-on-idle ended forgotten-GPU billing. Cited only with R3D attribution. Measures compute spend, not total infrastructure.
Cite as Juno Innovations Measurement Methodology, Q3 2026. juno-innovations.com/methodology#compute-cost
PROVISIONING TIME
60s
Source
The Orion product spec. Request submitted to container ready, with a pre-pulled image. Kindwell Solutions cut deployment time by about 60% with Orion provisioning during a 9-day agent build.
How to read it
Cold start with an image pull takes 3 to 8 minutes depending on image size.
Cite as Juno Innovations Measurement Methodology, Q3 2026. juno-innovations.com/methodology#provisioning-time
WORKLOAD CAPACITY
2:1 GPU density
Source
R3D Studios, during their production run on Orion, ran 10 concurrent artist workstations on 5 g5.2xlarge instances through time-slicing, a 2:1 ratio.
How to read it
Two artist workstations per GPU instance through time-slicing. Density depends on the workload mix, and GPU-heavy render jobs still get a whole GPU.
Cite as Juno Innovations Measurement Methodology, Q3 2026. juno-innovations.com/methodology#workload-capacity
VMWARE PRICING
150% to 10x+
Source
CloudBolt survey, January 2026. Forrester analysis. Public enterprise reporting.
How to read it
Smaller deployments see 150 to 300% increases; large agreements have seen 5 to 10x and more. vSphere 8 reaches end of life on October 11, 2027.
Cite as Juno Innovations Measurement Methodology, Q3 2026. juno-innovations.com/methodology#vmware-pricing
DEFINITIONS
Terms we use.
Unified compute plane
One layer of software across the machines you already own. Your team launches workstations, VMs, notebooks, GPU jobs and apps on it from a browser, and admins run it from Orion Admin.
Managed Core Hour
One worker CPU core, managed by Orion, for one hour. Management nodes are free, cores count as physical cores (never threads), and GPUs are not licensed.
Workload
Anything Orion launches for someone, such as a workstation, a VM, a notebook, a GPU job or an app.
Customer-hosted
Orion runs in your environment, on your machines or in your cloud account. Juno never hosts it.
Built to run air-gapped
Orion needs no connection back to Juno. There is no phone-home and no log export, so it runs the same with or without the internet.
GPU time-slicing
Several container workloads take turns on one whole GPU and share its memory. An admin sets how many slices each card gets, four by default, and can change it later. Windows VMs get a full GPU through passthrough.
MIG
Multi-Instance GPU splits a supported data-center GPU into hardware partitions, each with its own fixed memory. Admins can set each GPU node to MIG, time-slicing, or VM passthrough.
vGPU
NVIDIA vGPU shares a GPU between virtual machines and needs NVIDIA vGPU software licenses. Orion shares GPUs between containers through time-slicing, and VMs get a full GPU through passthrough.
