Every rate SEVEN KG charges is on this page.
Verification, the one-off assessment, evidence retention, certificates and compute — what each costs, and for compute, what the operator is paid and what we keep in between. No rate is held back for a conversation, and none of them moves because you asked.
The verification rates are readable. The compute rates are not published yet. A covered figure is covered in the page source as well as on the screen, so there is nothing to uncover by looking.
These are the numbers each service opens at.
What it costs
Compute, priced per node-hour
Billed per second of execution, rounded up to the next whole second, from the moment the machine reports the work started. There is no minimum charge beyond that one second. Every job returns a signed evidence bundle at no additional charge.
| Line | What you are buying | Price |
|---|---|---|
| Attested node-hour | One machine, running your work, having proved what it is against a pinned manufacturer root. Any isolation level, at the same price — Attested Mobile, Measured Host, Isolated VM and Confidential GPU all bill this one line. | $0.000 per node-hour |
| Per attested GPU, per hour | Added for each GPU the card itself proved, on top of the node-hour. A node that reports eight and proves one is billed for one. | $0.000 per GPU-hour |
| Worked example | One confidential machine with one attested GPU, for one hour. | $0.00 |
The isolation level does not change the price, on purpose. Charging more for stronger isolation would invite you to shop downward, which is the behaviour this whole design exists to prevent. A stronger machine is eligible for more work; it is not sold at a premium. What moves the price is counted hardware, and a GPU is hardware you can decline if you do not need it.
The shape of the workload does not change the rate. A container image and a sealed WebAssembly module cost the same per node-hour, because the rate is for what the evidence proves about the machine and not for how the work was delivered. On Attested Mobile a module is the only shape that runs, since a phone has no container runtime, and the technical limits of that surface are on the API reference rather than here.
Prepaid balance, $00 minimum top up, paid by bank debit or card. Bank debit is offered first, and the minimum exists for the same reason: a charge costs a fixed amount however small it is, so a smaller top up would hand a large share of a fractional-cent product to the processor. $00 is also the figure an operator has to reach before we send a payment, so one number covers both sides of the same problem. Verification and the services below are invoiced monthly instead, also bank debit first.
A balance buys compute from us and is not a wallet. You cannot withdraw it on demand, move it to someone else, or have it paid to an account other than the one it came from. If you close your account, whatever you have not spent comes back pro rata to the card or bank it arrived on. We would rather say what it is not than let the word balance imply something we do not offer.
What operators earn
Your hardware, your share
Operators are paid per node-hour of completed work. Earnings are quoted before tax; operators are independent, not employees.
| Line | Per hour | 25% busy | 50% busy | 100% busy |
|---|---|---|---|---|
| Attested node-hour, any isolation level | $0.000 | $00.00/mo | $00.00/mo | $00.00/mo |
| Per attested GPU, per hour | $0.000 | $000.00/mo | $000.00/mo | $0,000.00/mo |
Monthly figures assume 730 hours in a month and the share of it shown. “Busy” means actually running billable work, not merely enrolled and available.
Earnings accumulate to a $00 balance before a payment is sent. Moving money costs a fixed amount per transfer, so paying out $0 would hand most of it to the processor. At these rates a single device reaches that balance a few times a year rather than monthly.
The per-hour rate is the only figure here we can stand behind.
Every monthly number above is the hourly rate multiplied by a share of 730 hours, and nothing more. It is arithmetic, not a forecast. We cannot promise you any share of any month, and the honest figure for utilisation today is zero: the store is not open, so no operator has yet been paid for a single hour.
Two things decide what a device actually earns, and neither is the rate. It has to be reachable — powered, cooled, on a network, and not in somebody’s pocket. And it has to be able to finish the work it is offered: a job carries the time the customer declared it may take, and a device that cannot finish inside that time never gets the job. Slower hardware therefore sees less work, without us having to rank it.
This is not a reason to go and buy hardware. The rates are built for machines you already own and would otherwise leave idle, and a purchase justified by a number on this page would be justified by our arithmetic rather than by your own measurements. If you are weighing it anyway, weigh it against depreciation and electricity on the hardware in front of you, and against a utilisation figure nobody can give you yet.
Electricity is the running cost that matters at the small end: a tablet drawing six watts costs about $0.00 a month to run. At the large end it is the dominant one — a 600 W GPU plus its host is roughly $0.00 to $0.00 an hour of electricity before anything else.
The share we keep
Both sides of every rate, published
Most marketplaces publish what the buyer pays and what the seller earns, and leave you to work out the difference. Here is the difference.
| Line | Customer pays | Operator receives | SEVEN KG keeps |
|---|---|---|---|
| Attested node-hour, any isolation level | $0.000 | $0.000 | $0.000 |
| Per attested GPU, per hour | $0.000 | $0.000 | $0.000 |
| Worked: one confidential machine, one attested GPU, one hour | $0.00 | $0.00 | $0.00 |
The share is 18.2% of what the customer pays, identical on every line. It does not vary by isolation level, by volume, by operator, or by how the match was made, and there is no separate matching fee.
We publish it because a take rate you cannot see is a take rate that can move. If this number ever changes, it changes on this page first.
For comparison, and it is the reason the single number matters: a comparable attested-compute network charges 30% of the job plus a separate 10% matching fee on top. Ours is one number, it covers the match, and there is nothing added at the end.
A different service entirely
Verifying a fleet you already own
Everything above is about renting compute. This is not that, and the two never meet.
What it is
You have devices — laptops, phones, workstations, terminals. Somebody asks you to prove something about them: an auditor, a customer, a franchisor, a contract. We check the evidence your hardware already produces against the manufacturer’s own root of trust, and issue a signed record of what was true and when.
$3.00 per device, per month is the list rate, and every figure below is quoted against it.
| Under 250 devices | $700 a month, the minimum |
|---|---|
| 250 devices or more | $2.80 per device, so 250 is $700 |
| 2,500 devices or more | $2.55 per device, so 2,500 is $6,375 |
Which is a 6.67% fleet discount at 250 devices and 15% at 2,500. Two things we would rather state than have you work out. No fleet is actually billed at the list rate — below 250 the minimum applies and at 250 or above a fleet rate does, so the $3.00 is the reference the discounts are measured from rather than a figure on anyone’s invoice. And you are never charged more than the next line down. At $3.00 each, 249 devices would come to $747 while 250 devices cost $700, so a fleet just under a line pays the line’s price. Paying more for having fewer devices is not a thing we are willing to put on an invoice.
What it is not
No compute is brokered. No operator is involved. Nothing runs on your devices for anyone else, and nobody joins any network. You are buying evidence about hardware you already own and keep.
If you are here because you have hardware you want to earn from, this is the wrong page — and there is no minimum number of devices for that. Three phones is a perfectly ordinary place to start.
Becoming an operatorOr once, without committing to anything.
A subscription is a strange first thing to ask of a company you have never heard of. So the same work is available as a single engagement: we appraise your fleet once and hand you the result.
The difference is not the depth of the work, it is how often it happens. This is one date; the subscription is every month. If what you need is an annual audit, this is the right product and the subscription would be overcharging you.
| What you get | Price |
|---|---|
| A signed record for every device that presented valid evidence, the list of devices that could not and why, and a written statement of what the evidence proves and what it does not. | $2,100 once, up to 250 devices. Larger fleets quoted. |
If you then subscribe, the fee is credited in full against your first three months — which at the rate above is exactly what it cost, to the dollar. Taking the one-off first should never be the more expensive choice.
What it is not: remediation, consulting, or certification of anything. We do not fix what we find, we do not tell you what policy to adopt, and no assessment here makes you compliant with any standard. It tells you what your hardware can currently prove, which is a question nobody else will answer plainly.
Why there is a monthly minimum at all: a contract this service cannot carry is a promise we would end up breaking. Below that figure a single support conversation costs more than the contract earns in the month it happens in, and the response times we commit to stop being ones we can keep.
For scale, device security tools sell at three to eight dollars per device per month in real negotiated contracts, and full device management runs eight to twenty. Verification is neither of those. It checks evidence against pinned vendor roots and issues a signed record, and it does nothing else to your fleet — it does not manage, patch, configure or lock anything.
Larger fleets
Above 2,500 devices, talk to us first.
The $2.55 rate carries on above 2,500 devices and there is no separate enterprise rate card behind this. The reason to talk first is not the price. Past roughly that size, procurement starts asking for things we do not have yet, and we would rather say so before a purchase order is open than after.
If your procurement requires a SOC 2 report today, we do not have one. Here is what we do have, here is what it proves, and here is what it does not. If you need the report, tell us and we will give you a date rather than a maybe.
This is written as a requirement rather than a headcount on purpose. A five-thousand-device fleet that does not need SOC 2 is not turned away by it, and a four-hundred-device fleet that does need one should raise it with us on day one.
Keeping the proof
Evidence Retention
Every verification produces a signed evidence bundle, and you can download it free for thirty days. This is what happens after that.
What it is
We hold your sealed bundle and keep it retrievable past the free window. Two notices go out before anything expires — fourteen days and three days — and nothing is ever deleted without both of them on record. If our mail is broken, nothing is deleted and the bundles pile up, which is the correct direction for that to fail.
What it is not
We are not your recordkeeper. You retain your own bundles for your own regulatory period; this is a convenience with an expiry and never anyone’s system of record. A regulated firm’s obligation is its own, and we are deliberately not part of that chain.
It is also not storage. The Ed25519 signature is self-contained, so you can check it against our public key for as long as you keep the file, whatever we delete. What a retained copy preserves is the ability to re-audit against the raw certificates inside it.
| Plan | Price | Per month |
|---|---|---|
| Monthly | $34.99 | $34.99 |
| Quarterly | $94.99 | $31.66 |
| Annual | $280 | $23.33 |
Fourteen days of grace on a missed payment, then the ordinary expiry flow with its two logged notices.
Honestly about the margin: a bundle is around twenty kilobytes and holding one for five years costs us a fraction of a cent. This is priced on what it is worth to not own the problem, not on what it costs us, and we would rather say that than imply an infrastructure bill that does not exist.
Retention here is a convenience with an expiry, and it is not the only clock that runs. What we keep publishes every record we hold about you, your device or your machine, and the day each one is deleted.
Something to put on a wall
The certificate
A verification produces a signed evidence bundle, which is a file. Some people need a document instead — something to hand an auditor, frame, or put in a tender.
What it is
Two pages carrying the company seal, the date, what was verified and what the evidence does not establish. A QR code on it resolves to the public verification page for that record, so anyone holding the paper can check it against us without taking our word for anything.
Every printed value also appears in the signed bundle. The seal is a mark of issue, not a second proof — the signature on the bundle is what a verifier actually checks.
What it costs
| One device, on its own | $25 |
|---|---|
| One certificate covering a whole fleet | included with Fleet Attestation |
| A certificate for each device as well | $5 per device |
A fleet certificate is one document about the whole fleet, not one per device. It comes with the verification rather than being sold separately, because a certificate without the work behind it is decoration.
Confidential GPU verification
Seal Check
You rented a confidential GPU instance. Is it actually confidential?
Sealed means the GPU’s own confidential-computing mode is engaged and the accelerator is cryptographically bound to the CPU’s trusted execution environment — so memory moving between them is encrypted and the pair can be attested as one unit. A GPU with that mode switched off is an ordinary GPU in a rented machine, whatever the invoice calls it.
This has nothing to do with the company seal printed on a certificate, further down this page. That seal is a mark of issue on a document. This is a property of the silicon you are paying for.
What it is
You run one collector on the instance. It gathers the CPU’s confidential-computing quote — Intel TDX or AMD SEV-SNP — the attestation report and certificate chain from every GPU, and the confidential-mode state of each. We check all of it against pinned vendor roots and return a signed verdict: whether those GPUs are genuinely sealed, and genuinely bound to that CPU.
As with every verdict we issue, it names what the evidence cannot establish.
Why it exists
Because the reading a tenant is told to check is not the reading that answers the question. On 30 September 2026 we appraised an eight-GPU H100 instance sold as confidential. The host’s own diagnostic printed CC State: OFF while all eight signed GPU reports said the protection was on — and the signed reports were right. The provider’s own instructions told the tenant to check the line that was misleading. The whole appraisal is published.
Today a tenant takes the provider’s word for it, because checking means reading an SPDM report against a vendor root, and almost nobody has that. This is the independent check.
| Monitored, per GPU per month | $3 |
|---|---|
| A single one-off check | Not yet priced |
Priced per GPU to match the per-device verification rate, because a confidential instance is a fleet of accelerators, and each accelerator is what gets checked. The one-off rate opens with the service.
The verifier behind it is live. The coordinator has accepted GPU nodes since 1 October 2026, and the appraisal has run against real signed reports from an H100 deployment and from an RTX PRO 6000 Blackwell deployment — not against anything we made up.
Vendor root monitoring
Root Watch
For teams running their own verifier. A root changes, and your verifier starts refusing genuine devices.
What it is
A monitored feed of the vendor attestation roots — Google, Apple, AMD, Intel, NVIDIA and AWS. We fetch each vendor’s published certificates on a schedule and compare them by SHA-256 against the set we hold. When one is added, replaced, revoked, or approaching expiry, subscribers get a signed notice.
Signed, because a notice telling you to trust a different certificate is exactly the message you would least like to receive unsigned.
Why it exists
A root change is an event you only learn about by going and looking. Google’s move to an ECDSA attestation root left verifiers that had pinned only the older RSA root refusing most new phones — correct behaviour, wrong outcome, and invisible until users complain.
We watch these roots because our own product stops working if we do not. The incremental cost of telling you is close to zero, which is the whole argument for the price being small.
| Subscription, per month | Not yet priced |
|---|
A flat monthly figure, deliberately set below the verification minimum so it is a corporate-card purchase rather than a procurement exercise. One rate, no tiers, no per-root charge.
Both halves already run here on a schedule: the drift check has fired weekly since 26 September 2026 and the expiry check monthly since 4 October 2026. Its first run read 22 certificates across 4 files; the nearest expiry is in 2032.
Hardware capability catalogue
Device Atlas
Will this model attest, and what will it actually prove? Answered before you buy two thousand of them.
What it is
A catalogue of what each device model can and cannot prove under hardware attestation, built from real attestation results rather than from datasheets: which claims the model reports, at what security level, which root its chain terminates at, and its known quirks. Two handsets that both advertise hardware attestation do not necessarily report the same number of claims — one we hold reports eight and another reports six.
What it is never used for
The catalogue never decides whether evidence passes. A model name is self-reported by the device, so a lookup keyed on it would be a control that an attacker fills in. Verification is always against the signed chain. The atlas answers a purchasing question, not a verification one — and keeping those apart is the point.
| Access | Not yet priced |
|---|
Priced on coverage rather than on seats, and the rate opens when the catalogue covers a meaningful share of a real fleet. We will tell you what it covers before you pay for it, because a catalogue you cannot check the extent of is worth nothing.
Characterising a model is paid for once and then answers the question for every customer after, which is why this compounds. A better-funded competitor still has to hold the devices and spend the afternoons.
Confidential GPU capacity
GPU nodes verify. The rate opens with the tier.
Consumer GPUs cannot attest at all.
NVIDIA implements Confidential Computing only on datacenter silicon: H100, H200, B200, HGX B300 and the RTX PRO 6000 Blackwell Server Edition. No GeForce or consumer RTX card has a confidential computing mode, a device identity certificate, or the security processor that signs an attestation report. That is a property of the hardware, not a gap in our software.
We said we would price one when we could verify one.
We can verify one. The coordinator accepts GPU nodes as of 1 October 2026 — an Intel TDX confidential VM together with the NVIDIA Confidential Computing GPUs it vouches for — and the appraisal has been run against real signed reports from an H100 and from an RTX PRO 6000 Blackwell deployment, not against anything we made up. Prices are still pending.
Status
In-Development
The rates above are published so they can be checked before anyone can be charged.
Both open together with the catalogue. Until then the honest form of a next step is a control that names it and does not pretend to do it.
What the price does not buy.
Attestation establishes what was loaded and on what hardware. It does not establish that the code is correct, and a measured host does not protect a workload from the person who administers the machine. Verified is not the same as acceptable: evidence can verify perfectly against a pinned root and still fail your policy.
Read the full list of what we cannot do