Murmur runs on the devices you already own. By default, your fleet stays Local — your devices, your network, a flat monthly fee, no traffic with the outside world. When you want more capacity than your fleet can provide, or you would like to monetize your idle assets, opt in to Network — and tap the spare capacity of every other Murmur customer who's done the same. The opt-in goes both ways: you use, and your devices contribute, together.
Each dot is a contributor node — a fleet of laptops, workstations, or GPUs whose owners chose to plug in. The amber pin is the seed cluster in Cherry Hill, NJ. We're showing simulated data while we onboard the first wave of real contributors; the shape of what you're seeing is what the network looks like at modest scale.
Your devices, your network, your work. A flat monthly fee, predictable as a software license. Nothing leaves your premises and nobody else's traffic touches your fleet.
Local plus a two-way connection to every other Murmur customer's spare capacity. Use what you need; your idle devices help with theirs. The bill nets out at the end of the month.
Local is the default. You enroll your devices, you run your workloads, and the bill is the same $5 every month. Nothing leaves your network and nobody else's work touches your devices — it's a private compute platform that happens to run on the hardware you already own.
Push the Murmur agent through your existing IT tools, or send enrollment links to specific people. Devices show up in your console as they come online.
Run risk simulations, model training, batch jobs — whatever your team builds for. The platform splits each workload across the devices in your fleet.
No metering, no per-workload billing, no overage. If a workload exceeds what your fleet can absorb, it queues until devices free up — or you flip the Network toggle for overflow. Either way the base fee doesn't change.
Privacy posture: on Local, no traffic ever leaves your network. Your devices don't see anyone else's work, and nobody else sees yours. Same architecture you read about on the home page — only smaller.
Network is a single switch on top of your Local plan. Flip it on and your fleet joins a much larger pool — every other Murmur customer who's also opted in. From that moment, two things happen at once: when your fleet is overloaded, the network lends you spare capacity; when your devices are idle, they help finish someone else's work. You don't pick — both directions go on together. That reciprocity is what lets the math work.
Network is a single switch in your console, off by default. Turning it on enrolls your fleet in both directions: you can use the broader pool, and your idle devices contribute to it.
When your own fleet is busy or you're running a workload bigger than it can handle, the platform borrows capacity from other contributors automatically. You don't schedule it; it just happens.
Idle devices in your fleet pick up small jobs from the rest of the network. They pause the moment your team needs them back — the same elasticity rules as any other Murmur workload.
Reciprocity sounds risky until you see what actually crosses the wire. Whether work flows to your devices or away from them, all that ships is a math operation and the raw numbers it operates on — stripped of labels, field names, schemas, and any other metadata. The receiver sees something like “multiply this 128×256 tensor by that 256×64 tensor” and a stream of floats. They have no way to know whether the numbers are loan defaults, model weights, or sensor readings. Your code, your models, and the meaning of your data stay on your server.
Read the full architectureWhen Network is on, your bill is your $5 base, plus what you used from other companies' spare capacity, minus what you contributed back from your own. Net out and you pay the base. Use more than you contributed and the difference is added. Contribute more than you used and the difference rolls forward as credit. We borrow the idea from residential solar, where your meter runs both directions.
Most teams start in one and switch later. There's no penalty for changing your mind — both run on the same network, the same way.
Yes. The Network toggle is in your console, off by default. Flip it on whenever you'd like — it takes effect at the start of the next billing cycle, so you'll always know exactly which mode any given workload ran under.
Yes. Flip the toggle off and you go back to the flat $5 base. Your devices stop helping the broader pool, you stop drawing from it, and any credits you've accumulated either roll into your next Local bill or get cashed out at the end of the cycle.
Yes. Network is a single reciprocal opt-in — you can't draw from the pool without also adding to it. That's how the economics stay fair. If you'd rather only ever use your own devices, stay on Local.
Numbers do cross — that's how compute works — but they cross stripped of anything that says what they are. No field names, no model context, no schemas, no labels. The receiver sees something like “multiply this 128×256 tensor by that 256×64 tensor” and a stream of floats. They have no way to know whether the numbers are loan defaults, model weights, or sensor readings. Your code, your models, and the meaning of your data stay on your server.
This works the same way in both directions, so the math is symmetric: when other companies' work runs on your devices, your devices receive the same kind of context-free numbers, and they can't tell what those numbers mean either.
No. Your devices receive a math operation (e.g. multiply two matrices) and the bare numbers to operate on — the same shape as work coming from your own server. There's no executable code from another company touching your machines, no scripts, no runtime payloads. Just primitives and unlabeled tensors.
A credit is an internal unit of compute time, like the build minutes you'd see on a CI service. Whatever rate you're billed at when you use the network is the rate you're paid at when you contribute. We don't skim the spread. Current rates are on our talk-to-us conversation rather than the public site while we tune them.
We're publishing $5 today because the network is still small and we want to make joining frictionless. As capacity and reliability grow, the rate will rise — but anyone already on a plan keeps their existing rate for a meaningful grandfather window, and every change is communicated at least one billing cycle in advance.
Not yet. Today the toggle is account-wide. Per-device modes are on the roadmap; tell us if that's blocking you and we'll prioritize.
Yes. Larger fleets, custom SLAs, dedicated support, and stricter data-residency commitments are handled on a per-customer basis. The fundamentals don't change — same architecture, same data-stays-on-premises guarantee — just at a scale and contract shape that fits an enterprise procurement process.
Tell us a bit about your fleet and the work you'd like to run, and we'll point you at the model that makes sense — and the one to switch to if your usage shifts later.
Talk to us