What this infrastructure costs

EigenSelect publishes its infrastructure economics in full. The figures below are live billing values, read from the provider's billing API and checked against its published price list. They are not estimates. The single exception, the proof domain's line, is flagged under Accounting notes.

Earlier versions of this page were wrong twice: once in the figures, once in the account of why the nameservers carry IPv4 addresses. Both are corrected below, at length, including the parts that are embarrassing to us. A service whose product is documented evidence does not get to be selective about which of its own records it corrects.

The commitment chain is served by two authoritative nameservers we operate in separate jurisdictions, one dedicated proof domain, and the image storage that keeps the service reproducible. This is every line item, in full:

This page previously reported the verification plane at €8.66 a month. The true figure was €9.40. We were under-reporting our own costs by €0.74 a month, on a page whose stated standard is the cent. Two line items were missing. (The plane costs €5.75 today, because we released one of the addresses on 16 August 2026. That is a change, not a correction, and it is recorded under The IPv4 position, below. The audit that follows is left exactly as it was written.)

The nameservers' disks. We reported each nameserver at €0.43 a month, which is its compute price. Storage on a Scaleway instance is billed separately from the instance, a fact stated plainly in the provider's documentation and overlooked entirely by us. Each nameserver carries a 10 GB local SSD at €0.000049 per GB per hour, which is €0.36 a month. The disk costs 84% of the machine it is bolted to. We chose local storage over block storage specifically because it was cheaper, and then failed to notice that cheaper is not the same as free. The rate also went up on 1 June 2026, six weeks before we deployed, so we are paying the new one.

The image storage. The service ships as a 642 MB container image, mirrored off the server so that losing the machine cannot lose the image. That storage costs €0.02 a month. We had excluded it on the grounds that it rounds to zero. It does not round to zero. It rounds to two cents, and two cents is not zero on a page that claims to account to the cent.

The correction cuts against us in one more place, and this is the part worth reading. Our sharpest complaint about IPv4 was that an address costs 8.5 times the server it is attached to. That compared €3.65 against a €0.43 machine. But the machine is not a €0.43 machine. With its disk it costs €0.79. The honest multiple is 4.6 times, not 8.5. Our best line was inflated by our own sloppy accounting, and we would rather be accurate than aggrieved.

It remains 4.6 times.

This page previously stated that the .eu registry's delegation acceptance checks refused our IPv6-only nameserver set, then refused a set with one IPv4 address, and that the two IPv4 addresses were procured to satisfy the registry. That account was wrong in its central attribution, and we would rather correct it than let it stand.

There is no evidence the registry refused anything. The .eu registry publishes no minimum number of IPv4 addresses and, as far as its published rules go, runs no acceptance check on a delegation at all; it scores DNS quality afterwards, weekly. The "minimum of two" we acted on was our own reading of one sentence in a third-party delegation-testing tool. It was an inference. We bought infrastructure on it.

What failed was the registrar's connection to the registry. Every request to delegate the domain to our nameservers failed there, in the same way, from 13 July to 15 August: with zero IPv4 addresses, with one, and with two. It failed in under a minute and gave no reason. On 15 August we requested a set consisting of one of the registrar's own nameservers beside one of ours; it failed identically. The address family of our nameservers was never the variable.

On the evening of 15 August we moved the domain to a different registrar, using the transfer code the registry issues to the registrant, and submitted the same two nameservers. The registry accepted them within minutes. eigenselect.eu has been delegated to our nameservers since 21:05 UTC that day, the DS record for our key was published a minute later, and public resolvers answer the commitment queries with the validation flag set. Same nameservers, same key, same zone. The variable was the registrar.

We said here that whether both addresses stayed was under review, and that this page would say so when it changed. It changed on 16 August 2026. One address remains, on ns1. The protest stands. It was addressed to the wrong party.

One of our nameservers has an IPv4 address. We would like it on record that we did not want either of them.

A public IPv4 address costs €3.65 per month, which is 4.6 times the all-in cost of the nameserver it fronts (€0.79), and the equivalent of 14.5 quantum executions on cryogenic hardware. Per month. For a 45-year-old address format. Our founding position was to allocate that budget to physics instead, on both nameservers, and the verification plane was built IPv6-only accordingly.

In July the delegation of our proof domain failed, and we concluded that the registry would not accept nameservers without IPv4. Delegation is the one function a proof domain cannot decline to perform, so both nameservers acquired IPv4 addresses, procured in stages and held under protest. The conclusion was ours and it was wrong; the second correction above has the details.

On 16 August 2026 we released ns2's address. The reason we bought it did not survive the audit, and the delegation it was supposed to buy had already been obtained by moving registrar. ns2 now answers on IPv6 only. We are stating the consequence rather than letting it be discovered: a resolver with no IPv6 connectivity can now reach one of our two nameservers instead of both. DNS is designed for that, and ns1 answers every query ns2 would have; the redundancy that remains for such a resolver is the protocol's retry, not a second machine. We think that is the correct trade at €43.80 a year, and we would rather write it down than have it inferred from a table.

The address on ns1 stays, because delegation is not a thing a proof domain gets to be principled about twice.

The remaining address costs €3.65 a month. Everything else in the verification plane, combined, costs €2.10. The legacy address tax is still the largest single line item here, at 63% of what it costs to run this apparatus: one address, still more expensive than the servers, their storage, the proof domain and the image registry added together.

We do not need to argue this point. Our provider's billing API argues it for us. Below are two line items from the same invoice, for the same nameserver, in the same availability zone, over the same 24-hour period, measured in the same unit. The only variable is the address format:

The modern format, with 3.4 × 10³⁸ addresses, is free because nobody needs to ration it. The exhausted format, with 4.3 billion, carries a rent because everybody does. We pay the rent on the second one so that a resolver with no IPv6 can reach our evidence about quantum measurements. This is the internet we have.

Everything below is quoted per resolved pick, because that is also how the service is priced. A coin flip is one pick. A pick‑7 lottery draw is seven picks: seven separate commitments, seven separate measurements, seven separate entries in the report, and seven times €50. The costs scale with it, and so does our exposure.

Why a pick can cost us twice what it should. A quantum measurement returns one of 2^q basis states, but your outcome set is rarely a power of two. Nine raffle entrants need four qubits, which produce sixteen states, so seven of them mean nothing. When the processor lands on one of those, the honest options are to fold the result back into the domain, which would bias the draw, or to throw it away and pay to run the circuit again. We throw it away and pay again. That is the entire reason this line exists in our accounts.

Not a projection. This is the execution log, to date:

The five-number draw is the instructive one. Five resolved picks took seven physical executions, because two measurements landed outside the domain and were rerun. The customer paid for five picks. We paid for seven. That is the arrangement, and it is stated here rather than discovered later.

The direct quantum cost of a €50 pick is 0.5% to 1.0% of its price. The remainder funds the report, the verification plane, the engineering, and the margin. We consider the ratio itself part of the product. You are not paying for electrons through a refrigerator. You are paying for the documentation that they went through it for you.

Note the total. Every quantum computation this company has ever performed cost €2.77, which is less than a sandwich. The IPv4 addresses overtook it on 25 July 2026, and the one we kept will keep going forever.

What is excluded, and why. The application runs on a shared server that hosts unrelated services. Its marginal cost to EigenSelect is not zero, but it is not separable either, and we would rather exclude it openly than invent an allocation key and present the invention as a measurement. Everything that is separable is in the table above, including the two cents of image storage.

Generated reports accumulate on disk and are backed up nightly. At the present volume this is four PDFs weighing 280 KB in total, which costs a genuine and verifiable €0.000004 a month. That one really does round to zero, and we mention it only so that you know we checked.

The proof domain's line is the audited figure, not the current one. €0.50 a month is €5.98 a year, prepaid to the previous registrar to July 2027. The domain moved to Porkbun on 15 August 2026 at $5.88 a year, and that is what the next audit will use, against an invoice rather than a price page. We would rather carry a number we have checked than a currency conversion we have not.

The commercial domain (eigenselect.com, to 2029) and the proof domain (eigenselect.eu, auto-renewing, to 2028) are held at different registrars with no shared failure mode. Registrar fees are stated at their flat renewal rates, with no promotional first-year pricing. Figures on this page are updated when billing values change, and corrected in public when they turn out to have been wrong.