Article

The Legacy Address Tax: An Infrastructure Economics Report

This is a report about two servers, one partially defeated absence, and what the pricing of a 45-year-old address format reveals about the internet's relationship with its own history. All figures are live billing values from our provider, not estimates.

The verification plane

EigenSelect's commitment records are DNS TXT records, which is what lets anyone verify a selection with a dig command. They are served by two authoritative nameservers we operate ourselves: one in Amsterdam, one in Warsaw. Each is a small cloud instance with one vCPU, one gigabyte of memory, ten gigabytes of local storage. Authoritative DNS for a zone of text records is one of the lightest workloads in computing; a machine of this size is, frankly, overprovisioned.

Each instance costs €0.43 per month for compute, plus €0.36 per month for the 10 GB local SSD it boots from, which our provider bills separately and which we initially forgot to count. That is €0.79 per nameserver, all in. Both together: €1.58. Add the proof domain (€5.98 per year, flat renewal) and the container image the service ships as (642 MB, mirrored off-box, two cents a month), and the entire apparatus runs at €2.10 per month. That is two servers, two jurisdictions, and one DNSSEC-signed domain whose only purpose is to answer questions about wavefunction collapses.

The disk detail is worth dwelling on, because it is the same disease in miniature. A 10 GB volume costs 84% of the computer it is attached to. We chose local storage over block storage precisely because it was the cheaper option, and then failed to notice that the cheaper option was still an option with a price. Our provider raised that price on 1 June 2026, six weeks before we deployed. We are paying the new rate. We found this out by auditing ourselves for this article, which is an argument for auditing yourself.

For calibration: a single quantum execution on the processor that resolves your selection costs €0.2514. The complete verification plane costs 8.4 executions per month. It is the cheapest component of the entire service, and it is the one you can independently interrogate from any terminal on Earth. We consider both properties features.

The absence, amended

Our founding position was that neither nameserver would have an IPv4 address. The position lasted approximately one hour into production, and the manner of its defeat is worth documenting precisely, not least because we documented it wrongly the first time. On the evening of 13 July the delegation of the proof domain to our nameservers failed, silently. We read the wording of a delegation-testing tool as implying that the registry expected IPv4, and procured one address, for ns1, in Amsterdam. The delegation failed again. We read further, inferred a minimum of two, and procured a second, for ns2, in Warsaw. The delegation failed a third time, and a fourth. Both addresses are held under protest.

Correction, 15 August 2026. An earlier version of this article stated that the .eu registry's delegation acceptance checks refused our nameserver set. They did not. The registry publishes no IPv4 requirement and, as far as its published rules go, runs no acceptance check on a delegation; it scores DNS quality afterwards. The failure was in the registrar's connection to the registry, which fails in the same way with zero, one or two IPv4 addresses, and failed identically on 15 August for a set containing one of the registrar's own nameservers. The address family of our nameservers was never the variable. The addresses were bought on our inference, not the registry's rule. They remain, for the moment, because an IPv4-only resolver can reach them, and the protest stands; it was addressed to the wrong party.

Further note, 15 August 2026, evening. The delegation landed that evening. We moved the domain to a different registrar, submitted the same two nameservers, and the registry accepted them within minutes; the DS record followed, and public resolvers now validate the commitment records. The nameservers did not change. Neither did the address count. The registrar did.

Further note, 16 August 2026. The address count changed. With the delegation obtained and the inference behind it retired, we released ns2's address. ns2 answers on IPv6 only; ns1 remains dual-stack. One address, €3.65 a month, still held under protest. The honest consequence of the release is set out below, where it belongs, rather than in the sentence where it flatters us.

The position itself was not an oversight. At our provider, a routable IPv4 address is a metered add-on costing €3.65 per month. Per address. Set that against the €0.79 nameserver it would be attached to and the arithmetic becomes difficult to write down without editorializing. Each address costs 4.6 times the entire computer it is attached to, disk included. The two together cost €7.30 a month, or, in the unit this company prices things in, 29 quantum executions per month, remitted for the privilege of being reachable by the internet's legacy addressing scheme. The one that survives the release costs 14.5 quantum executions per month, for the same privilege, on half the fleet.

An earlier version of this article said 8.5 times, not 4.6. That figure compared the address to the instance's compute price while ignoring the disk, which is to say it was arrived at by the same failure of attention we are accusing the industry of. It made our case look stronger than it is. We have corrected it downwards, and we note that a complaint you have to inflate is a complaint you should check.

It survives the correction, and it survives the release. The legacy address tax is still the single largest line item in the verification plane: one address at €3.65 against €2.10 for literally everything else, which is 63% of what it costs to run the apparatus. Halving the number of addresses did not dislodge it from first place.

We do not, in the end, have to argue any of this. Our provider's billing API argues it for us. Here are two line items from the same invoice, for the same nameserver, in the same availability zone, over the same 24-hour window, measured in the same unit. The only variable is which format the address is written in:

Zonal Flexible IP · IPv6 · NL-AMS-1     24 ip_minute     €0.00
Zonal Flexible IP · IPv4 · NL-AMS-1     24 ip_minute     €0.12

That is the whole argument, held in a controlled experiment we did not design and cannot be accused of rigging.

The economics here are not our provider's invention. IPv4's address space, 4.3 billion addresses allocated with mid-1980s optimism, was exhausted at the registry level years ago; RIPE NCC allocated its final /22 in November 2019 [1]. Since then, IPv4 has been a scarcity market: addresses trade at tens of dollars each on transfer markets, and cloud providers, holding finite pools, have converted the address from an assumption into a line item [2]. Meanwhile IPv6, standardized in 1998 with 3.4 × 10³⁸ addresses [3], is deployed to roughly half the internet's users, varying by country [4], and costs nothing precisely because nobody needs to ration it.

In other words: the old, cramped format now carries a rent, and the modern, effectively infinite one is free. Infrastructure that clings to IPv4 is paying a tax to the past. We declined to pay it, for these two machines, and the refusal did not survive our own diagnosis of a failing delegation. Half of it has since been reinstated, on the machine where reinstating it costs nothing anyone can query. We are documenting the decision, its defeat and its partial recovery, because infrastructure decisions taken silently are indistinguishable from accidents, and so are infrastructure decisions reversed silently.

The honest consequence

This section has now disclosed three different residual costs, which is itself the story. It began by disclosing partial reachability for IPv4-only resolvers, because we intended to buy no addresses. It then reported that residual as purchased out of existence, by mistake, because we had bought two. Since 16 August 2026 it discloses the version we chose deliberately: a resolver with no IPv6 connectivity reaches ns1 and not ns2.

That is a real reduction and we will not dress it up. Such a resolver has one authoritative server where it previously had two, and its redundancy is now the protocol's retry rather than a second machine in a second country. Every other resolver, and every validating one we have tested, reaches both nodes; the records are identical and identically signed on each; the large public resolvers (Google, Cloudflare, Quad9) are dual-stack, so querying through them (dig @8.8.8.8 …) reaches both nodes from anywhere, and measured IPv6 availability among end users continues to climb [4]. The Nostr commitment channel, operated by third parties, is independent of our DNS entirely.

We priced that reduction at €43.80 a year and decided it was not worth the money. A verification service should be precise about the edges of its own verifiability, and this is where the edge now sits.

That trade-off is stated here, on the verification page, and in the technical appendix of every report we ship. Every report also carries the line item that summarizes the position:

IPv4 allocation    : 1 address (held under protest)
Legacy spend ratio : 4.6x the nameserver it fronts

What this is actually about

A reasonable reader may ask why a selection service maintains opinions about address formats.

The answer is that EigenSelect's product is documented rigor, and rigor is a property of whole systems, not of chosen highlights. The same discipline that publishes a cryptographic commitment before contacting the processor, records every rejected measurement, and enumerates what our verification cannot prove, also requires us to account for why our infrastructure is shaped the way it is. That includes the parts where we optimized for principle over universality, and the parts where the entire apparatus costs less per month than the sandwich you are perhaps eating while reading this.

Two servers. Two jurisdictions. Two legacy addresses bought on a mistake of our own, one of them since released, all of it itemized. €5.75 a month today, €9.40 before the release, fully disclosed, including the €0.74 a month we had been quietly failing to disclose to ourselves. The selection costs €50; knowing exactly what you paid for is included, and so is knowing what we got wrong.

References