The Nexus mirror set, 2026 report
A structural report on the current mirror rotation, its size, its logic, its recent history.
At the time of writing, Nexus Market publishes several onion mirrors in a single operator-signed rotation. Every mirror on the list resolves to the same backend. Account, balance, orders and messages appear identically on any of them. Choice of mirror is a matter of routing speed on your Tor circuit, not a matter of storefront identity.
Why more than one mirror
A single onion address is a single point of routing failure. Even a well-provisioned hidden service can be unreachable for a specific user because of a bad Tor guard, a hostile ISP, a network outage between the reader's ISP and the first Tor relay, or a temporary hidden-service-descriptor stale-out in the Tor consensus. Any one of these is enough to make a reader believe the market is down when in fact the market is fine and the reader's circuit is bad.
The mirror rotation solves this cheaply. When the primary is not resolving on your circuit at the moment, you copy the second address from the list, paste, and the storefront comes up. The reader loses maybe forty seconds of scrolling through mirrors. The alternative, in the single-mirror model, is that the reader loses the session entirely.
Why not many more mirrors
More mirrors do not add proportional value once you cross the small-set threshold. Three addresses cover most single-outage scenarios. Additional addresses buy you more coverage in exotic multi-failure scenarios, but at a cost: more addresses is more surface for phishing clones that try to blend into the operator publication, more addresses is more work for editors verifying the signed rotation, and more addresses is more entries a reader has to eyeball against the captcha URL match.
The trade-off point sits at roughly three to six mirrors depending on the operator's own load model. Nexus has been at three lately.
Rotation cadence in 2026
Signed rotations from the Nexus operator have arrived roughly every four to eight weeks through 2026. Every rotation swaps one address in the published set for a new one, or (occasionally) introduces a new address without retiring an old one, briefly showing the reader an extended set for a week or two. Retirements do not happen without a signed announcement.
Latency and mirror age
Older mirrors, in general, are faster. They have accumulated Tor descriptor consensus, they route through known-good guards, and the hidden service directory nodes have their information cached. Freshly rotated mirrors sometimes bounce for the first week while descriptors propagate. Reader latency observations across late spring 2026 showed the primary running around 140 to 200 milliseconds median, the older backups at 150 to 220, and the newest addition after each rotation at 200 to 400 for the first several days.
What a rotation looks like from the reader side
The reader sees a signed post on the pinned Dread thread announcing the new address. The reader verifies the PGP signature (see the dossier on the operator key). The reader updates the bookmark. The reader keeps using the storefront exactly as before, because the address change is the only thing that changed.
What a rotation looks like from the Bureau side
Editors verify the signed post against the operator key on file. If the signature is good, the new address is added to the Library mirror list before the retired address is removed. If the signature is bad, nothing happens and we log the failed post for later reference. Editors have never had to publish a bad-signature warning yet under the current operator key. That fact itself is a data point.