the hardware is already inside the cities

50km

the radius real-time work has to live in

Light in glass covers two hundred kilometres per millisecond and a round trip pays that twice, so no data centre five hundred kilometres away wins by trying harder. Millions of capable machines already sit inside every city on earth, bought by somebody else for their own reasons. Right now, 50 to 75 percent of them are switched off.

what the network sells
somebody is playing on a stranger’s graphics card, fifty kilometres away sold by walter.gg, the consumer storefront that runs on this network start a session yourself

nobody built it this way because nobody could

the first ceiling is a price

in a data centre

on this network

the hour there, the session here (a session runs up to one hour)

~$2.00

$0.08

hardware

bought years before it earns

already bought, to play games on

cooling and power

30 to 40% of operating cost

it is a room, it is already cool

bandwidth

every frame travels the distance

the distance is 50 kilometres

idle time

the rack is paid for at 3am

bought per session served

real estate

built before the first customer

somebody lives there

We know it from the inside. In 2020 we streamed live 3D for TEDx and GSK, and this is the floor we kept hitting. The product was technically correct and economically dead.

you cannot buy your way past the grid

the second ceiling is electricity

the power here is already committed through 2028
the grid adds one to two percent a year

228 electricity data centres will demand

111 grid capacity to supply it

2023, indexed to 100 2030

18 to 36 months to connect one new site

Every month in that queue is a month a competitor cannot get back. Capital is abundant. Substations are not.

IEA, Energy and AI LBNL 2024 data centre energy report Goldman Sachs

the costs are already paid by somebody else

Not reduced. Paid years ago by somebody else, for reasons that had nothing to do with us. That is the difference between a discount and a cost structure, and it holds at roughly sixty percent gross margin.

what a kilowatt hour actually costs, at the GPU

  • hyperscaler, industrial agreement

    $0.04 to $0.07

  • hyperscaler, retail tariff

    $0.10 to $0.15

  • a carrier tower

    $0.04 to $0.07

    pre-permitted, pre-powered, no queue, and one runs a live node today

  • a house

    ~$0.00

    the operator pays their own bill

The bottom two rows are ours: houses where density already is, towers where coverage has to be. A competitor can match the chip price eventually, and the latency with capital. Neither buys a place in a line that is already full.

three kinds of work, one price logic

gaming was the hard part, and it is running

Ten milliseconds, both directions, held for an hour. State that survives every frame. Millions of sessions at once, each one somebody’s evening. Batch inference is easier than that. So is a configurator. So is a model answering a question.

Gaming is not the market. Gaming is the proof: the workload that forces you to build orchestration, session state, security and routing properly, because nothing about it forgives you. Clear it and everything looser runs on the same substrate.

the network today

  • 30 live regions
  • 27,686 unique rigs scanned for compatibility, as of 19 Sept, 11:00 UTC

+1,316%

daily active users

  • 7gb became 2mb

    the download stopped being a decision

  • +54%

    retention at thirty days

Roughneck Rumble, by Skyhook Games, over its first six months on the network. Nothing about the game changed. The studios and partners already on the network.

what would have to be true

  1. a session cannot survive on hardware that comes and goes

    Stateless networks retry a job elsewhere when a card vanishes. Retrying a session is not a recovery, it is the end of it. This network is both stateful and stateless on the same hardware. One of the largest in the category left interactive work entirely in 2025.

  2. my build would sit on a stranger’s machine

    Titles run in secure-boot containers. The operator has no access to what is inside, and no copy stays behind.

  3. what if there is no node near my user

    Then it runs on ordinary commercial cloud instead of failing, at roughly 6x what an operator node costs. We carry that, because a network has to work everywhere on day one. That share falls as operators join.

  4. why would supply keep growing

    An operator earns more where sessions are being paid for. Supply follows demand into a region instead of being built ahead of it, and that is what carries the cost curve down.

  5. where does the session fee go

    A publisher pays per session. The fee is split four ways when it settles: ~60% to the operator who served it, 30% to the network, 5% to the staking pool and 5% burned, in $YOM, settled on-chain. Operator payouts track session volume, not token price.

how the orchestration actually works
No existing system combines decentralised consumer-grade GPU infrastructure, real-time interactive game streaming, browser-native delivery via embed SDK, and an orchestration layer with QoS-based routing.
Dr. Slinger Jansen, Associate Professor, Utrecht University. Independent expert opinion, March 2026.

the layer keeps the value

Nobody knows which application layer wins, and it has stopped being the interesting question. What does not get abstracted away is latency, locality, energy and hardware. A frame still has to be rendered somewhere, near a person.

Every session pays a fee to the protocol regardless of what the session is. A game tonight, a configurator next quarter, an inference call the year after.

  • $255B

    AI inference a year by 2030, growing 19% a year

    MarketsandMarkets; Deloitte

  • $130B

    GPU compute sold as a service

    Analysys Mason

  • $57B

    the latency-bound slice, growing 37% a year

    BCC Research

we are building the GPU layer of the post-AI economy

what runs next runs here

Same nodes, a different workload. Three things and we can price it: what runs, where your users are, and how long a session lasts.

Every figure here comes from the model we run the company on. Ask for a source and you get one, on the source of truth page. Here with a machine instead of a workload? See what it earns.