This is not capacity you rent from us.
Almost every infrastructure bill you have is rent on somebody else’s computer, priced per unit consumed. This is the other thing: software that turns machines you already own into one fabric, with the meter running as accounting rather than as a till. We do not sell you compute, and the numbers further down are not a menu to buy from — they are the unit your own work is measured in so that it can be settled.
Anyone. An individual, a team, a lab, a whole institution on its own hardware.
₹0
Available nowNot a trial, and not a free tier with a ceiling — there is no paywalled feature to reach.
An institution pooling its own laptops, workstations and idle GPUs into one sovereign fabric, with us on the hook for it.
Scoped
Scoped per deploymentQuoted against the cloud and VPN spend it replaces, not against our cost.
Workloads that need capacity beyond what you own, drawn from machines other people own.
Not open
Not open yetWhen it opens, this is the unit it will be metered in — and the price will be published before the first receipt, not after.
What it replaces is a portion of what you already spend outside. That is the comparison worth making, and the calculator below makes it against your bill rather than against a rate we picked.
Two kinds of capacity sit behind every infrastructure bill. Committed capacity is the machines you have already bought — their marginal cost is the electricity they draw. Uncommitted capacity is what you rent from somebody else, in cash, every month. Most organisations pay for the second while the first sits idle.
This sizes the overlap. It does not tell you to switch off your cloud account.
Estimated, not measured
Net of the electricity, the overlap is worth ₹89,389 a year. That figure — not a seat count — is what a licence is quoted against.
Checked against the daemon rather than against what would look good in a column.
| Capability | Your machines | Institution | Mesh |
|---|---|---|---|
| All seven resource networks The same node binary runs everywhere. No capability is withheld by tier. | ✓ | ✓ | ✓ |
| Signed self-benchmarks Every node benchmarks itself and signs the result with its own key, whoever owns it. | ✓ | ✓ | ✓ |
| Work stays on hardware you control True by construction on your own devices and inside a gramx you assembled. The mesh is other people’s machines — that is the point of it. | ✓ | ✓ | Not applicable |
| Self-use credited back at settlement Every unit of work produces a signed receipt everywhere, including on your own hardware — the meter is never off. What differs is that work you do for yourself is returned at settlement, so it nets to nothing. | ✓ | ✓ | Not applicable |
| A full account of what your own machine cost The by-product of metering self-use rather than exempting it: a signed, per-resource record of what your own work consumed, which an exempted meter could never produce. | ✓ | ✓ | ✓ |
| Membership you control (add and remove devices) A gramx is an allow-list of device identifiers with Admin and Member roles. One with no members is simply open mesh — enforcement begins with the first member you add. | Not applicable | ✓ | Not applicable |
| Sensitive questions never leave your gramx Model routing honours a privacy tier: questions classified financial or health are answered by a local model, sensitive ones stay within your gramx. Classification is keyword-based and English-only today, so treat it as a floor, not a guarantee. | ✓ | ✓ | Not applicable |
| Pooled memory across machines A model larger than any single machine can draw RAM from the gramx. One device cannot pool with itself. | Not applicable | ✓ | ✓ |
| Setup, patches and an SLA we are on the hook for The commercial obligation. It is what a private-ring licence buys; nobody is on the hook for a machine we never touched. | Not applicable | ✓ | Not applicable |
Every resource is metered wherever it runs — inference, CPU time, memory, watts, disk, egress — and that includes work a gram does for itself, at the same rate a stranger would pay. Nothing is exempt, because an unmetered resource is one nobody accounts for. On hardware you own it is all credited back at settlement, so it nets to zero and what you are left with is a signed record of what your own work actually took.
Standing decides how much comes back where more than one party is involved: the credit is scaled by the weaker party’s, so admitting a stranger to a shared room costs you until they have earned some. Standing is verified identity plus observed time, minus proven fault — and a node’s claims about itself earn nothing.
This is the measure the meter counts in, not a menu to buy from. It is denominated in money because a settlement needs a denomination.
An hour of inference a day
30 GPU-hours × $0.36
$10.80
8 GiB of memory held all month
8 GiB × 730 hours × $0.003
$17.52
A 7 GiB model sitting on disk
7 GiB × $0.02 a GiB-month
$0.14
Twenty grams doing all three
20 × (30 GPU-hours + 8 GiB resident + 10 GiB on disk)
$570.40
On machines you own, each of these is charged and then credited back at settlement — so the cash figure is zero and the number above is a measurement, not an invoice. It is shown so the scale of the meter is legible.
| Resource | On a receipt | µUSD | Per |
|---|---|---|---|
| Compute | vcpu_second | 10 | vCPU-second $0.036 per vCPU-hour |
| Memory | gb_hour | 3,000 | GiB-hour $0.003 per GiB-hour |
| GPU / inference Charged per second of inference, at one rate for every caller — including this gram serving its own turns. | gpu_second | 100 | GPU-second $0.36 per GPU-hour |
| Storage A rate, not a one-time charge. Accrued hourly against a 730-hour month for as long as the bytes are held — by you and by every peer holding a replica. | gb_month | 20,000 | GiB-month $0.02 per GiB-month |
| Egress Charged where it is observable — relay egress, object retrieval, and the daemon’s own research fetches. Per-process network is not sampled, so it is not billed rather than estimated. | egress_gb | 50,000 | GB transferred $0.05 per GB |
| Relay | relay_second | 1 | relay-second $0.0036 per hour |
| Energy Modelled from CPU and GPU class, never read from a meter on your wall. Receipts label it estimated, and an estimate is signed as an estimate. | kwh | 1,50,000 | kWh $0.15 per kWh |
The rate is the rate. Work metered on a Kubernetes pod, on a host process outside the cluster, and on a laptop in lite mode is charged identically — there is no local discount.
Current policy, in the present tense. We are not going to tell you what we will charge in five years.
Your own machines cost you nothing
Work on hardware you own is metered like everything else and credited back at settlement, so it nets to zero. What you pay to run a gram on your own laptop is its electricity.
No egress charge inside your own perimeter
Moving your data between machines you own is not a billable event here. That charge exists elsewhere to make leaving expensive.
No reserved capacity to buy
There is nothing to reserve and no minimum to commit to, because the capacity in question is a machine you already own.
The one thing we do promise
It keeps running if we disappear
Nothing phones home for permission to start. A gram you are running today keeps running whether or not this company exists tomorrow, and that is a constraint on our architecture rather than a policy we could quietly change. Standard containers, standard VMs, content-addressed objects, exportable receipts — the cost of leaving is deliberately low, because that is the only version of trust worth offering.
We already pay AWS and a GPU host. Is this cheaper than that?
It is not the same kind of thing, so cheaper is the wrong axis. Those are rent on somebody else’s computer, billed per unit consumed. This is software that runs on machines you already bought, sold as a licence with support attached — closer to how you buy a self-hosted database than how you buy an instance. It does not replace your cloud account. What it does is let a share of the work currently going out to rent run on hardware you have already paid for, and the calculator on this page sizes that share against your own bill.
Can we buy compute from you?
Not today. Capacity offered by machines other people own is the open mesh, and it is not open — so there is nothing to buy and no rate published for it. Everything you can run now runs on your own hardware.
Why will you not quote a per-hour instance rate?
Because an hour of a machine is not something we sell. A node signs a benchmark and a receipt for a unit of work, and those units — vCPU-seconds, GiB-hours, GPU-seconds — are what the meter counts. An hourly instance price is a rate on a rented machine, which is the product we are not.
What does a private gram actually cost?
It is scoped, and we will tell you on the first call rather than making you fill in a form to learn it. The shape is settled — an annual licence banded by node count, a one-time setup, and support at roughly a fifth of the licence — and what varies is how many machines you have and how much of your existing estate we have to meet.
Is any of this crypto?
No. A private gram is a software licence, invoiced in rupees by Fifty GramX Networks Pvt Ltd, with GST. No token touches you. The internal ledger is mutual-credit and net-zero, and it is not something you interact with, buy, or hold.
If it runs on our hardware, what are we paying for?
The software, the setup, and the obligation. Our marginal cost to run your gram is close to zero, which is the point rather than an embarrassment — the price is anchored to the spend it displaces and the sovereignty it delivers, and we would rather say that plainly than invent a cost base we do not carry.
How many machines can we add?
As many as you have — the daemon enforces no cap, so we do not advertise one. The licence band is set by node count because that is what setup and support actually scale with.
What happens if you go away?
It keeps running. Nothing calls home to be allowed to start, so there is no lever we could pull even if we wanted one. That is a constraint on our architecture rather than a promise in our marketing.
Tell us what you are running and we will scope it on the call, rather than after a form.
Scope your gram →