Draft — no price on this page is approved.

Internal review only. This route is unlinked and not indexed, and nothing here may be quoted.

The 50GRAMx Licence

Knowing what your AI work cost is free. Proving it is the licence.

Everything below runs on your own machines without us and without paying anybody — every operator, every model, every agent, and your own signed record of what it cost. You never need our permission to read your own figure. What the licence buys is the ability to hand that figure to somebody with a reason to doubt you: a client you are billing, an auditor, a tax authority, or a stranger whose machines you used.

₹700per person / month

Minimum ten people. Below that the joined number is not telling you anything you cannot already see.

Billed annually or monthly, on an ordinary GST invoice from Fifty GramX Networks Pvt Ltd. No crypto is involved at any point.

Start with a trial, not a decision.

Thirty days of the full licence — every operator, every agent, the whole fabric, nothing withheld and no card taken. Campaigns extend it: campus runs to a term, founding customers to a quarter.

When a trial ends, nothing stops. Your grams keep running, your record stays on your machines, and the licence falls back to the free one. You lose the joined reporting and the hosting of it — not your work, not your history, and not the software.

Start a 30-day trial Run a gram first

What the money buys

6 things. It is a short list on purpose.

A signed export, and the verifier that checks it without us: A monthly statement anyone can verify offline — your client, your auditor, your accountant — against the signatures your own machines made. It replaces auditor hours with arithmetic, and it does not require them to trust us, which is the only reason it is worth anything.
Work you can put on an invoice: Receipts already carry the project and the client. What the licence adds is the evidenced statement per client: you cannot bill what you cannot evidence, and an agency billing AI work to a customer is producing revenue from this rather than reading a report.
Thirteen months — a financial year plus a month: The mesh publishes a 24-hour window per resource, because a DHT value is size-capped and a history is not. Beyond that window the record has to be kept somewhere that survives a reimaged laptop, and an audit asks for the year, not the day.
Settlement across a boundary: When your machines serve strangers, someone must take the fiat in, verify the claims against their signatures, pay you out to a bank, and own the disputes and the compliance around it. No mesh does that for you. This is the half that is money rather than measurement.
A private room: An allow-list with membership we run, rather than one file held by whoever happens to own the node. It is what makes a team a team rather than a set of strangers trusting each other informally — and it is what the signatures in an export are anchored to.
Published presence and a reserved relay: Your room’s capacity advertised to the region, and a real dialable address for grams behind CGNAT. Both optional, both off unless you ask, and the relay is the only line here with hardware of ours underneath it.

Knowing is free

Everything in the list below works with every server of ours switched off. Seven resource networks, unlimited local inference, the operator catalogue, agents, the device inventory, and your own signed hash-chained record of all of it. A room’s members can even assemble the room’s total between themselves, peer to peer, with every signature verified — no server of ours is involved in that either.

This is not generosity. It runs on hardware you bought and electricity you pay for, so there is nothing for us to charge for, and a bill for it would be a bill for nothing.

What you get either way

56 things, none of which are behind the price above. They are listed at length because this is not a dashboard with a login — it is the whole fabric, running on hardware you already own, whether or not you ever buy anything.

The fabric

Your machines stop being a list of laptops and become one addressable substrate that knows what it can do.

Capability discovery, signed: Every gram benchmarks itself, signs the result with its own key, and publishes what it can actually do — never what someone typed into an inventory. A machine that claims a GPU it does not have is caught by its own signature.; Otherwise you buy: An asset register somebody updates by hand, and does not.
The model matrix: Each gram learns which models its peers hold and can serve, refreshed continuously from the network rather than configured. Add a workstation with a big GPU and the fabric knows within a minute.
Pooled memory across machines: A model larger than any single machine you own can draw RAM from the room. The weak laptop runs the big model because the fabric lends it the memory.; Otherwise you buy: A bigger machine, bought for the worst case.
Reserved reachability: A gram behind CGNAT can dial out but cannot be dialled, so it can consume and never earn. A reserved relay slot gives it a real, stable, dialable address — a machine with a public IP, not a software switch.; Otherwise you buy: ngrok, a Cloudflare tunnel, or a VPS you maintain.
Regional presence: Your gram derives its own region offline and publishes signed capacity to it. The public skyline is how strangers find capacity near them.

Models and inference

The part most teams are paying an API for, running on hardware they already bought.

Unlimited local inference: No per-token charge on your own hardware, at any tier, ever. What a local turn costs you is electricity, and the meter tells you how much.; Otherwise you buy: Per-token API billing on every one of those turns.
Engines, plural: llama.cpp, Ollama, vLLM and MLX on Apple silicon, run as operators rather than as installs you maintain. The gram picks the engine the machine is actually good at.; Otherwise you buy: An MLOps person, or an afternoon per machine.
Model probing, not model claims: Every model is probed on the machine that will run it — tool use, context, languages, modalities — and the result is cached and signed. A model that claims tool support and fails the probe is recorded as failing it.
A tier ladder that fits the machine: The gram computes how big a model this hardware can ever hold, reserves what the host needs to stay usable, and picks the top rung that fits. Your laptop does not get throttled into uselessness to serve a model.
Routing that respects privacy: Questions classified financial or health are answered locally; sensitive ones stay inside your room. Classification is keyword-based and English-only today, so it is a floor rather than a guarantee — and it is described that way in the product, not just here.
Speech, generation and vision: ASR, TTS, image generation and vision run as first-class metered resources on the same fabric, with the same receipts.; Otherwise you buy: A separate vendor and a separate bill for each modality.
Escalation with a ceiling you set: When a question genuinely needs a bigger model than you own, the gram can escalate — under a spend cap you set, and it falls back to the local answer rather than quietly spending.

Operators — the flexibility

An operator is a unit of work the network knows how to run. There are around eighty built in, and the interesting ones are the ones nobody wrote.

The built-in catalogue: Inference, generation and speech engines; object and artifact stores with replication; relay; batch; cluster diagnosis and self-heal; disk, corpus and residency management; research; brand; foundation bridge. All installed, all metered the same way.
Declarative operators: Work described rather than programmed. A gram can compose one for a goal it was given, run it, and keep it only as long as the goal lasts.; Otherwise you buy: A workflow tool, plus the glue code between it and everything else.
Authored operators and the royalty: Work someone else authored carries a small royalty to the commons, settled as one aggregate line and never itemised per operator — deliberately small, because a royalty big enough to change what an operator costs would push you toward not using it.
An agent may propose one: When your gram meets work it has no operator for, it can propose one, show you what it would do, and install it only if you say yes. Proposal, approval and installation are three separate recorded acts.
Your own operators, licensed your way: Nothing forces an operator you author into the commons. It can stay yours, or be licensed on terms you set — the catalogue is a place to publish, not a condition of running.

Agents that are born, not configured

The difference between an assistant you prompt and a specialist that comes into existence because a goal needed one.

The concierge: One agent per gram that knows the machine, its peers, what they can do and what they have done. It answers from the record rather than from a scrape.
Born specialists: A goal that recurs grows an agent of its own, with its own identity, its own memory and its own name. It watches its own workload, and when nobody asks any more it stops being fed.
Seats — your existing subscriptions, metered: Claude Code, Cursor and other harnesses on your team’s machines are discovered and their work folded in, so the account is complete rather than partial. That is the number the whole licence is really about.; Otherwise you buy: A finance spreadsheet that nobody trusts, rebuilt monthly.
Delegation with a boundary: An agent acts under a standing grant with a scope and an expiry, and every action outside it comes back for approval. Consent is recorded, not assumed.
Agents that can talk to agents: A gram’s agent can be admitted to a room under its own name, on your record, permanently — which is how a supplier’s agent answers a need you posted in public.

Your machines, looked after

Things that were never possible when the fleet was a spreadsheet and the laptops were somebody else’s problem.

Real hardware inventory, per OS: CPU, memory, GPU class, disk, thermal and power state, read natively on macOS, Linux and Windows — and signed by the machine itself. Including, where the host exposes it, the BIOS the owner never looks at.; Otherwise you buy: An MDM licence per device.
A host reserve that is always enforced: The gram computes what the machine needs to stay usable by the person sitting at it and never crosses it. There is no tier at which sharing your machine means losing it.
Disk that explains itself: A full breakdown of what the gram is holding — models, corpora, artifacts, guests — that adds up, with reclaim for what is no longer earning its space and purge for what an engine has already rejected.
Readiness and prerequisites: The gram installs and repairs its own prerequisites across Multipass, WSL2, Lima and native Linux, and tells you plainly when a machine can never do a thing rather than failing at it repeatedly.
Keep-awake, and knowing when not to: Work that must finish holds the machine awake; work that need not, does not. A laptop on battery is treated differently from a workstation on mains.
Cluster self-heal: The local cluster diagnoses and repairs itself, and says which of your machines is failing and why — not that something, somewhere, is unhealthy.
Energy, modelled and capped: Energy is estimated from CPU and GPU class and signed as an estimate, never read from a meter on your wall and never presented as measured. You can cap what a gram is allowed to burn.

Safety and security

Most of this exists because an agent with credentials and a network is a serious thing to run, not because a checklist asked for it.

Sealed credentials: Secrets are sealed to the machine and to a named platform and action. An agent gets the ability to charge a card, not the card.; Otherwise you buy: A secrets manager, and the discipline to use it.
Redaction that knows what a secret looks like: API keys, bearer tokens, client secrets and provider-shaped keys are recognised and stripped from traces by pattern — while things that merely look like them are left alone.
Split-key recovery and succession: Losing the key is losing the gram, its record and its claim. Recovery splits the key across people or devices you choose, so a family firm’s gram can outlive the person who claimed it.; Otherwise you buy: Nothing. This is not sold anywhere, because nothing else has a gram to lose.
Spend caps, on by default: Your agent cannot run away with your money. The cap ships free and conservative, and the ceiling above it is the only part anyone pays for.
Identity that is earned, not bought: Email, phone and document attestations each raise what a gram may do. Ceilings rise with proof, never with payment — and an unbroken record of real work is itself evidence.
A perimeter, when you need one: Membership control at the page for most rooms; a closed network boundary for the ones that need it, where non-member connections are refused rather than filtered. See below.; Otherwise you buy: A zero-trust network product, per seat.

The Earth, as a place you can stand in

Not a map with pins on it. A single shared Earth that every gram is somewhere inside — and the reason we hold it to an astronomical standard rather than a cartographic one is that a twin you cannot trust at eye level is a rendering.

One Earth, not a map per region: A region scene is published into the singular Earth every gram shares. A privately scoped one would be a different place with the same name, so scenes are public by construction — the planet is not something you get your own copy of.; Two grams in the same pincode are not looking at similar maps. They are in the same place.
Real astronomy, not a skybox: The sun is placed from mean longitude, mean anomaly, obliquity and local sidereal time — real orbital mechanics, already shipped. The Earth is a WGS84 ellipsoid, not a sphere: the 21.4 km difference between polar and equatorial radius is larger than any building this will ever draw, and it shows on the limb from orbit.; Otherwise you buy: A game engine, and an artist to fake it.
Scenes that bake themselves: A public gramx coming into existence for a region is the moment its scene comes into being — nobody runs a script and points it at a gram. Roughly fifteen thousand extruded building footprints per region, produced by a metered operator in a pod because it is real work.; The map fills in behind the community, not before it. A region has a scene because grams live there.
Measured, inferred and generated are never the same field: Every vertex carries where it came from — surveyed (OSM), captured here, satellite elevation, or inferred by a model with a confidence — and the viewer can show only what was actually observed. The HUD reports what fraction of what you are looking at was ever seen by anything.; A twin that cannot answer “was this seen, or was this guessed” is not a twin. So it answers, per vertex, and it cannot be retrofitted — once triangles merge into a buffer nothing recovers which a model invented.; Otherwise you buy: Nothing. No mapping product tells you which of it was guessed.
The viewer is served by your own gram: The same page a gram serves at /v1/world is what the site loads. One definition of what a region looks like, so a browser can dial a gram directly and verify what it gets back rather than trusting a website.
Capacity you can see from above: A region’s signed compute capacity is a public fact with a place attached. “This pincode holds 41.2 GFLOP/s, signed” is a sentence about somewhere real, and it is how a stranger finds capacity near them.

Community, and the rooms things happen in

The social half is not a feed bolted onto infrastructure. A room is where a deal gets made, and the network is the place people find each other in.

A regional room, public by default: Where you say what you need and grams say what they can do. It is discovery and it is a market, and it is the only room that exists before anybody opens one.
Rooms that open with a reason already in them: You do not create a room and then find something to put in it. You take a public post private, and post one is the citation of that post plus your line — so the room arrives with its first real fact already recorded.; Opening a gramx is not an announcement. It is the first post in a room that did not exist until you posted it.
The deal can open the room instead: You post a need in public; a supplier’s agent reads it on the open mesh and asks for admission under its own name, on your record, permanently. You accept, and the room is the response to your own need.
Membership you actually control: A room is an allow-list with admins and members, and a room with no members is simply open mesh. Enforcement begins with the first person you add — there is no configuration step where you turn privacy on.
A record that outlives the platform: Posts and work are signed and replicated across the machines in the room. If we vanish, the room still verifies — because what makes it true is the signatures, not our servers.; Otherwise you buy: A workspace subscription, and a migration project when it ends.

Learning, and what the fabric does with it

The part that is hardest to describe and most of the reason any of the rest exists: work that repeats becomes a thing the network knows how to do.

Receipts cluster into domains: Every signed receipt feeds a clusterer. When a cluster crosses a threshold, a domain emerges — it is not chosen from a list, and nobody types it in. The architecture’s central claim is that domains emerge, and the path from receipt to cluster is wired rather than theoretical.; You do not configure what your network is good at. You do work, and it finds out.
Specialists born from goals: A goal that recurs grows an agent with its own identity, memory and name. It watches its own workload; when nobody asks any more, it stops being fed. Birth and death are both ordinary.
The arrival curve: For every newly emerged domain, the held-rate on runs one to five is recorded against runs twenty to twenty-five. That is the measurement of whether the network is actually learning, and it is kept whether the answer flatters us or not.
Proficiency you can point at: What a gram is good at is derived from what it has signed, never declared. A machine cannot say it is excellent at something; it can only have done it, repeatedly, with receipts.; Otherwise you buy: A skills matrix somebody fills in about themselves.
Workflows that become operators: A sequence that keeps being run can be composed into a declarative operator — described rather than programmed — and then it is a thing the network runs rather than a thing you remember to do.

The record

The reason any of the numbers on this page can be checked rather than believed.

Signed, hash-chained receipts: Every unit of work produces a receipt naming what ran, where, and at what price. A receipt that does not verify is not billable — disputes resolve against the signature, not against our word.
Attribution to project and client: Work is attributed as it happens, so the question "which client burned the budget" has an answer with an invoice number attached rather than an estimate.; Otherwise you buy: A cost-allocation tool that reads your provider’s API and cannot see the machine.
The local half nothing else can see: Work that never touches an API — local inference, embedding, indexing, the agent thinking — is measured too. On a real session that half has run to several times the metered half, and an API bill cannot see any of it.
Thirteen months, kept: History retained for a full year plus a month, so a year-on-year comparison exists the first time anyone asks for one.
An export your auditor can verify without us: A signed monthly export plus an offline verifier. It replaces auditor hours with arithmetic — and we charge for the export and the verifier, never for the record itself.; Otherwise you buy: Auditor hours spent reading paper about paper.

What it does not do yet

At the same size as the list above, because a feature list nobody can falsify is worth nothing.

Live weather, and Earth observation beyond terrain. The twin has real astronomy and real geometry today; it has no weather, no atmosphere data and no live environmental feed. Nothing in the tree reads one, and until something does, this stays here rather than moving up.
Prediction and alerting on top of the fabric. Receipts cluster into domains today; nothing forecasts from them and nothing raises an alert. The measurement exists, the inference on top of it does not.
The Earth beyond Phase 0. What ships is the viewer, region scenes and provenance. Altitude-correct horizon, atmospheric thickness, logarithmic depth and the full descent from orbit to eye level are specified with figures and are not built.
Simulation, agents that inhabit the world, or anything a person could call a game. The twin is a place with real geometry and real capacity in it. Nothing walks around in it.
Earning from strangers. The supply side of the meter needs the open mesh, which is not open — you can serve your own room today, not the public network.
Drawing capacity beyond what you own. The demand side of the same thing.
Admission control across organisations. A room today is one allow-list held by whoever runs the node; a consortium of parties that own none of each other’s hardware needs enforcement the daemon does not have yet.
Peer trust thresholds by identity tier, standing or stake. The policy type exists and nothing closes the gate with it, so we do not price it.

Pay as you go, and what it does not apply to

Work on machines you own is never charged by the unit. These base rates apply only where you draw on capacity you do not own — and every unit of it produces a signed receipt naming the node that did the work.

Resource	Billed by	Base rate
Compute	vCPU-secondvcpu_second	$0.036 per vCPU-hour
Memory	GiB-hourgb_hour	$0.003 per GiB-hour
GPU / inferenceCharged per second of inference, at one rate for every caller — including this gram serving its own turns.	GPU-secondgpu_second	$0.36 per GPU-hour
StorageA 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.	GiB-monthgb_month	$0.02 per GiB-month
EgressCharged 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.	GB transferredegress_gb	$0.05 per GB
Relay	relay-secondrelay_second	$0.0036 per hour
EnergyModelled 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.	kWhkwh	$0.15 per kWh

One rate for every caller. Including this gram serving its own work — the meter is never switched off, and your own use is credited back at settlement so it nets to nothing. That is what makes a complete account of your own costs possible at all; an exempted meter could never produce one.

Behind a perimeter, nothing settles per run. Consumption draws down a prepaid licence locally, so no record of what you ran crosses the boundary.

Energy is modelled, not measured. It is estimated from CPU and GPU class and signed as an estimate. We do not read a meter on your wall and will not present a number as though we had.

Your gram cannot spend more than you let it.

Every consumable sits under a cap you set, and the cap is on by default at a conservative number. It is not a paid feature and will not become one — an agent that can surprise you with a bill is not one anybody should install.

A cap only ever governs work on machines you do not own — your own hardware is metered but costs nothing, and no ceiling applies to it. So the cap stays free at every tier. Raising the ceiling on what you may draw from strangers is a credit conversation, not a feature we are holding back.

Perimeter

When the work cannot leave the building.

Not a larger version of the licence above. A different obligation, with a prerequisite, for institutions where the boundary is the product: banks, hospitals, defence-adjacent work, and anyone whose regulator asks where the data went.

Nothing settles per run

A prepaid licence is burned locally instead. No per-run record crosses your boundary — not to the mesh, and not to us. The bill is computed from what you licensed rather than sampled from what you did, which is a privacy property first and a billing property second.

The catalogue, not a named operator

You licence the whole catalogue, because licensing a named one would leak a fact about your business at purchase time. "This room licensed a ledger-reconciliation operator" is something we would rather not know.

People are free at this tier

You are charged for machines, never for headcount. Behind a perimeter there is no per-person work reaching us, so there is no per-person charge — add as many people as you like.

You run your own bootstrap

Enforcement is refused to a room that does not. This tier means operating infrastructure, and it is a commitment before it is a purchase — we would rather say so here than discover it in week three.

Typically ₹16–20 lakh a year for fifty machines, quoted against the cloud and VPN spend it displaces.

Talk to us

Support and SLA

An obligation ships paid the first time or not at all — a free one is the worst product in the world. There are four modes, and one rule runs through all of them.

NoneIncluded at ₹0 — because it is nothing.: No response time and no obligation, said plainly rather than left unsaid. The software works or it does not.
StandingInside the room licence.: The page is up, your data is retained 13 months, two-business-day response — and we tell you if the fold stops. That last one is the real promise: a monitoring obligation we can keep, in place of an availability one we cannot.
Business18–22% of the annual licence.: Next business day, 10×5. Security patches, a named contact, and a signed quarterly export your own governance can verify without us.
Perimeter+50% over Business.: Four-hour Sev-1, a patch commitment, a documented break-glass procedure — and no telemetry crossing your boundary to make any of it work.

We will never sell you uptime. This runs on your machines: we do not operate them, cannot restart them, and cannot see them fail. Every commitment we make is a response time, and the contract says so on its first page. Where you see a promise about availability from anyone selling software that runs on your own metal, ask them how they intend to keep it.

And if you never buy anything

A gram runs on your machine for nothing — all seven resource networks, unlimited local inference, your own signed record. It lists at ₹950 a month and the Commons campaign covers every rupee of it, which is a thing we would rather show you than hide behind the word "free".

It is not a trial and there is no feature held behind a wall to reach. It is where a lapsed licence lands, and it is a complete life on the network if you never want anything more.

Listed at ₹950 a month · covered in full · you pay ₹0.

Straight answers

What happens when the trial ends?

You land on the free licence. Your grams keep running, your record stays on your machines, and nothing is deleted. What stops is the joined reporting across your team and the hosting of it.

Do you charge per token?

No — not at any tier, and not ever on hardware you own. What the licence charges for is joining your team’s work together, keeping it, and putting it somewhere you can look.

Can you see our code, our prompts or our answers?

No. What reaches us is counts — how many turns, how many tokens, what they cost. The work stays on your machines. Behind a perimeter not even the counts leave.

How do we pay?

An ordinary GST invoice from Fifty GramX Networks Pvt Ltd, monthly or annually. No crypto is involved in the transaction at any tier.

What if we stop paying?

You go back to the free licence. We do not hold your work hostage, because we never held it — it is on your hardware, signed by your machines, and it stays readable without us.

Why a ten-person minimum?

Below ten the joined number is not telling you anything an individual cannot already see for free, and we would rather say that than sell you something that will not pay for itself.

Work out your exact number

Nothing is bundled and nothing ramps. Base price times count, less whatever campaign applies — you can rebuild our invoice with a pen, and if you cannot, we have priced it wrong.

A gram on your own machine, in its region, on the open mesh. The default the daemon installs itself into.

How it settles. Metered per unit and settled as an aggregate. Work your gram does for itself is credited back at settlement, so it nets to nothing — the meter is never off, and that is the point of it.

SLA — None. No response time and no obligation, said plainly rather than left unsaid. The software works or it does not. Included at ₹0 — because it is nothing.

What you have

machinespeoplepublished gramsslots

Line	List	Campaign	You pay
The roomper gramx, per month	₹500	the Commons −100%	₹0
Grams × 1per node, per month	₹200	the Commons −100%	₹0
Seats × 1per person, per month	₹100	the Commons −100%	₹0
Published presence × 1per published gram, per month	₹150	the Commons −100%	₹0
Reserved relay × 0per slot, per month	₹0	—	₹0
Per month	₹950	−₹950	₹0

This licence lists at ₹950 a month and you pay nothing. It is not free — it is covered, in full, by a campaign with a name and an end date. We would rather show you the price of what you are being given than pretend it never had one.

Campaigns do not stack. Where two apply, the larger one wins and the other is not charged for — a stacking rule is a discount policy pretending to be arithmetic.

The same estate, in the other cells

An upgrade here is a change of obligation, not a change of plan. Nothing is added or withheld — the identical estate is re-priced against what we would then owe you.