density
homes
Gaming hardware sits bought, plugged in and cooled before we arrive. The operator owns the machine, so cooling, bandwidth and real estate are never ours.
two kinds of host, and we own neither
certainty
towers
Carrier sites come pre-powered and cooled at four to seven cents a kilowatt hour. 168 live game sessions ran over 20 days on UK tower sites.
ours
the layer
Three components, built by us, and the only part of this picture we own.
built for stateful sessions
the same open connection carries a live conversationThe clearest evidence comes from the market itself: in 2026 the largest cloud-gaming services capped play hours or cut features. That is not a pricing choice. It is the cost of GPUs that sit waiting for the evening peak.
YOM
sessions
stateful: one player on one live connection, input up and frames back for the whole session
middleware
hardware encoding (NVENC) and WebRTC streaming built in, so the buyer ships no streaming stack
latency
each session placed on the nearest qualifying node; sub-10 ms one-way is the target
content security
a container per session, destroyed when it ends, on an encrypted node disk; unreleased builds only on capacity YOM controls
overflow
rented cloud GPUs take the overflow where operator supply is thin
defensibility
patent-pending technology across the secure boot medium, pixel streaming and orchestration
stateless GPU networks
sessions
GPU-hours, mostly batch and stateless inference
middleware
raw GPU rental
latency
latency tolerant by design
content security
the buyer secures its own workload
overflow
the buyer's problem
defensibility
commodity aggregation
three components, one network
the orchestrator
HyperOrch
The control plane. It checks every node near the player for GPU capability and free capacity, and places the session on the nearest one that qualifies, so the player gets the shortest path the network can offer.
Five steps from a click to a paid operator, the same route the world above takes.
-
request
Press play in the browser. Nothing to install.
-
match
The nearest qualified node takes it.
HyperOrch -
boot
A fresh container for this player, destroyed when the session ends.
NodeOS -
stream
Frames encoded on the GPU and sent over WebRTC; input comes back up.
Universal Streamer -
settle
The fee is split and the operator is paid.
the runtime
NodeOS
Each player gets a Docker container with GPU access (NVIDIA Container Toolkit), destroyed when the session ends, so nothing carries over. Windows titles run through Proton, and the host agent pulls its work, so a node opens no inbound ports.
the last meter
Universal Streamer
Each session runs on the GPU's hardware encoder (NVENC) and streams over WebRTC to the browser, so players install nothing and the title needs no change.
live network
the network, live
Every green point is a host machine, sized by the sessions it served. Every line is real sessions between a player area and a host. Player positions are rounded to about 110 km.
measured
host machines offline players sessions52 host machines · 16,100 sessions · last 30 days · measured as of 10 Oct, 00:00 UTC
every figure comes with its method
15ms
median one-way latency, players within 50 km of a node
Median one-way latency (half the WebRTC round trip), last 30 days, players within 50 km of the node, measured as of 10 Oct, 00:00 UTC. The figure the rest of the site quotes.
40.5ms
median, all sessions
The same measurement over every session in the window, including players far from any node.
Full benchmark detail on the source of truth page.
read the docs or bring a workload
The docs go one level deeper on every component above. If you would rather test it than read it, tell us what you want to run.