The world already built the supercomputer.
We switch it on.

At any moment, an enormous share of the world's computing capacity is powered on and doing nothing. Laptops sit idle on desks. Servers run at a fraction of their memory footprint. Meanwhile the price of fast memory — the scarcest resource in modern AI workloads — keeps climbing.

Exa RAM makes that idle capacity addressable. Devices enroll voluntarily and become nodes. Their pledged memory becomes one pool with a single API. Their processors become a compute fabric that jobs shard across. Nothing is scraped, nothing runs without consent, and a device leaves the instant its owner closes the tab.

The project is deliberately built so that its claims can be checked. Every node on the dashboard is a live socket, every latency figure is measured on the request that produced it, and the verification page lets your own browser prove the system is real — including cryptographic proof-of-possession from the machines holding your bytes.

The Grid — the body. Devices join over a WebSocket from a browser tab or a headless daemon. Each reports its real core count and pledges a slice of RAM or disk. Heartbeats run every five seconds; a node that goes quiet is removed and its data re-replicated automatically.

The Memory — the RAM. The pool is addressed through one API: allocate, write, read, free. Data is sharded into 4 MB chunks and every chunk is stored on two different machines, so a device vanishing loses nothing. MemoryOS moves data between tiers automatically — hot data climbs toward faster memory, idle data sinks to cheaper storage.

The Reserve — the vault. Capacity can be reserved in GB-seconds against what is genuinely online, metered continuously against real usage. Reservations that exceed physical capacity are rejected outright, and an activated reservation guarantees its capacity even under contention.

A single Node.js service hosts both the site and the coordinator, which holds the node registry, the shard map, the job scheduler and the reserve ledger. Nodes speak a compact binary protocol — a JSON header followed by raw chunk payloads — over one WebSocket per device.

Browser nodes allocate their pledged memory up front as ArrayBuffers and run jobs inside sandboxed Web Workers; where WebGPU is available they can additionally offer real GPU buffer memory. The daemon agent is dependency-free Node.js: it holds RAM slabs and writes disk-backed chunks, and executes jobs on worker threads.

Durability comes from replication factor two plus self-healing: when a replica disappears, the coordinator copies the surviving chunk onto a healthy node to restore the safety margin. Tier migrations use copy-then-flip, so data is never at risk mid-move. The public REST API exposes all of it, and the whole system is covered by an automated test suite that kills real nodes and verifies the data survives.

For partnerships, capacity supply, investment, press or anything technical, email james@runexaram.com. We read everything.