Microsoft says its topological qubits now hold their state for 20 seconds and that a scalable quantum computer is due by 2029. If that alert reached you, it probably came with the usual subtext: encryption is about to break.
I think the claim is contested and unproven, and I think that even if it holds, it does not change the date on which you should act. That date comes from how long your data has to stay secret and from standards that are already final. Below: what was claimed, how to read it, what the real yardstick for breaking RSA looks like, and a short audit for your own servers.
What Microsoft actually claims
On 2 June 2026 Microsoft announced Majorana 2. Its post claims a mean qubit lifetime of 20 seconds, with some instances reaching a minute. The headline multiple is "1,000 times" the first generation, as Microsoft's German-language EMEA post puts it. The company now expects a scalable machine by 2029, cutting its original timeline in half.
Some vocabulary, because the numbers mean little without it. A topological qubit is meant to store information in a way that resists noise by design. It relies on Majorana zero modes, exotic quasi-particles in specially built superconducting nanowires. And "lifetime" here means parity lifetime: how long a charge-parity state persists. That is not a gate error rate, and it is not the figure you would use to judge a working computer.
Microsoft's pages say what the demo was, a four-qubit array, but they give no total qubit count, no error rates and no parity-measurement fidelity. The technical paper, "20 Second Parity Lifetime in an InAs-Pb Tetron Device", has 165 authors and sits on arXiv in two versions from 2 and 3 June. The page lists no journal reference. It is a preprint and has not been peer reviewed.
Read a quantum press release like a vendor benchmark
Four questions work for any vendor benchmark: what was measured, by whom, what is missing, and has anyone else reproduced it. They work just as well here.
What was measured is one parity lifetime on one device type. One secondary analysis reads the data as covering a single nanowire of one tetron and says it "does not demonstrate a qubit". That is one critic's reading, but it shows how narrow the measurement is.
Who measured it is Microsoft. heise's English report calls the figures self-reported and notes that the DARPA evaluation does not replace independent scientific review. What is missing is everything a cryptographer cares about: gate errors, logical qubits, scale. And as for replication, public reporting describes none for Majorana 1 or Majorana 2. The critiques so far are reanalyses of Microsoft's own data, not rebuilt devices.
| Claim | Evidence present | Evidence missing |
|---|---|---|
| 20 s parity lifetime | Microsoft-authored preprint | Peer review, independent replication |
| A qubit | Four-qubit array demo | Qubit count, error rates, measurement fidelity |
| Scalable machine by 2029 | A roadmap statement | Any quantified target |
Why physicists are still arguing
The history explains the caution. The same line of research produced a 2018 Nature paper that was retracted in 2021, and the central dispute has never gone away: whether the measured signals are unambiguous evidence for Majorana zero modes.
Henry Legg of St Andrews argues in an arXiv comment that Microsoft's topological gap protocol, the test used to claim a topological phase, "lacks a consistent definition of 'gap' or 'topological'". A St Andrews release dated 24 June 2026 describes a peer-reviewed comment in Nature on the 2025 paper. On Majorana 2, Legg told The Register it "is not proven to even be a single qubit". Microsoft replied: "We stand by our results and our roadmap." Sergey Frolov told Scientific American that the preprint "is not based on a research track record that can be considered a solid foundation".
Not everyone is hostile. In 2025 Sankar Das Sarma said "I'm persuaded" about one measurement, and Kartiek Agarwal thinks the gap protocol is sound, according to Science News. So the fair stance is "unproven, not debunked". Even heise's German commentary only goes as far as saying it is unclear whether Majorana 2 is any good.
The yardstick is a resource estimate, not a press release
The number that matters for your keys is how many qubits it takes to break them. Two terms: a physical qubit is the noisy hardware element, a logical qubit is an error-corrected qubit built from many physical ones.
Craig Gidney's 2025 estimate puts RSA-2048 within reach in under a week with fewer than a million noisy qubits, assuming a 0.1% gate error, a 1 µs surface-code cycle and a 10 µs reaction time. The Pinnacle architecture, a February 2026 preprint revised in May, claims fewer than 100,000 physical qubits using qLDPC codes (error-correcting codes with sparse parity checks). It has not been peer reviewed either. For elliptic curves, a 2026 estimate puts ECC-256 under 1,200 logical qubits and 90 million Toffoli gates, with runtimes of minutes on fewer than half a million physical qubits.
Put a four-qubit demo next to those figures and the distance is several orders of magnitude, even if every Majorana 2 claim holds. Microsoft has not published a physical or logical qubit total for its 2029 goal.
What does move your risk is the trend on paper. The estimate for RSA-2048 fell from about a million qubits to about a hundred thousand in nine months, without any new chip. Algorithms and error-correction schemes improve faster than the press cycle, and that is what a long-lived secret has to plan around.
Your deadline is your data's shelf life
The threat that applies today is harvest-now-decrypt-later: someone records encrypted traffic now and decrypts it once the hardware exists. So the question is not "when does the machine arrive?" but "how long must this data stay secret, and how long does migrating take?" Say you keep encrypted backups, a password-manager vault and family documents that should stay private for ten years. If migrating takes you two years, the working deadline is eight years out, whatever a chip announcement says.
The standards are already there. NIST published FIPS 203, 204 and 205 in August 2024, and NIST IR 8547, its transition guidance, was finalised on 27 October 2026. A secondary report says it deprecates classical public-key algorithms by 2030 and disallows them by 2035; NIST's own summary page names no years, so treat those dates as one outlet's reading.
The UK's NCSC sets 2028 for full discovery, 2031 for priority migration and 2035 for completion. Its reasoning rests on the threat being "well understood" and on ecosystem readiness, and it names no date for a quantum computer. Germany's BSI TR-02102, in version 2026-01, recommends hybrid key agreement. None of these calendars is tied to a hardware announcement, which is the point.
A 15-minute audit
One scope limit first. Hybrid ML-KEM, the NIST lattice-based key-encapsulation standard combined with classical X25519, protects key exchange, and key exchange is the harvest-now-decrypt-later part. Certificates and signatures stay classical. That is a separate migration.
Say you run a Debian or unRAID box with a reverse proxy in front, an SSH jump host, and WireGuard to a VPS. Four checks, in order:
- SSH. OpenSSH 10.0 (April 2025) made
mlkem768x25519-sha256the default, and 10.1 warns when a non-post-quantum key agreement is negotiated, citing store-now-decrypt-later. Runssh -vv yourhostand look at which key-exchange algorithm was picked. If you want it pinned, so a later config edit cannot quietly drop it:
# /etc/ssh/sshd_config.d/10-pq.conf KexAlgorithms mlkem768x25519-sha256,sntrup761x25519-sha512,curve25519-sha256 All three names are in the sshd_config manual. Ending on curve25519-sha256 keeps old clients working but also allows a classical fallback. Drop it once every client you use is current.
- TLS endpoints. OpenSSL 3.5.0 prefers hybrid groups by default. Against your own endpoint, with OpenSSL 3.5 or later on the client:
openssl s_client -connect host:443 -groups X25519MLKEM768 A hybrid handshake prints Negotiated TLS1.3 group: X25519MLKEM768. Force -groups X25519 instead and that line does not appear, which is a useful control.
The reverse proxy. nginx's
ssl_ecdh_curve autofollows OpenSSL's built-in list, and Cloudflare's table says nginx offers X25519MLKEM768 by default when built against OpenSSL 3.5 or later. Caddy'stlsdirective documentsx25519mlkem768and recommends keeping the defaults, but does not document what the default set is. Go 1.24 enabled the hybrid by default, andtlsmlkem=0in GODEBUG reverts it. Whatever you run, use step 2 to see what it actually negotiates, and check the hop from proxy to backend separately. That hop often negotiates its own, frequently classical, key exchange.Long-lived secrets. Inventory by shelf life, not by protocol: backups, vault, archived mail, anything that should still be private in 2036.
Clients matter too. Per Cloudflare's support table, Chrome 131+, Safari 26+ and desktop Firefox offer the hybrid by default. Apple notes that some legacy servers fail on the larger ClientHello this produces, so a failure after an update can have this cause. Your origin is also the likelier weak link: Cloudflare reported in October 2025 that most human traffic to its edge was post-quantum, while origin support was only 3.7%.
WireGuard needs more than a config line
WireGuard's handshake accepts an optional 32-byte pre-shared key, and its protocol page names post-quantum resistance as a use case. A static PSK only helps if you distributed it securely, which means out of band.
Automation exists. Rosenpass is a post-quantum key exchange that feeds fresh PSKs into WireGuard and refreshes the symmetric key every two minutes. Mullvad exchanges a PSK using post-quantum KEMs, on by default on desktop since January 2025. NetBird, on the other hand, has deprecated its Rosenpass support and advises against new deployments that depend on it. That is one vendor's position, but it shows the tooling has not settled.
Large keys cost something. A Classic McEliece public key is about 524 KB, per an April 2026 mailing-list proposal for a Noise extension that is not a shipped protocol. A study of large KEMs in IKEv2, which is not WireGuard, found handshakes taking up to nearly two minutes on average on lossy links. Treat that as indicative only.
My advice: if the tunnel carries data with a long shelf life, add a PSK distributed out of band or a post-quantum-capable layer on top. Otherwise it is reasonable to wait for the tooling to settle.
What this means for you
First, the date is your data's shelf life minus your migration time, not a chip. Second, update OpenSSH and OpenSSL and the defaults do most of the work. Third, verify the hops that defaults do not cover: proxy to backend, and tunnels. Finally, watch for the things that would actually change the picture: a journal publication of the Majorana 2 results, independent replication, a published qubit count with error rates, and new resource estimates. Press releases are not on that list.
The alert can stay unread
Either Majorana 2 holds up or it does not, and the to-do list is the same. A contested, preprint-level chip changes the headline. The falling resource estimates and the finalised standards are what change your work, and both were true before 2 June.
If you want the audit in a form you can reuse, subscribe to the newsletter. Subscribers get a one-page post-quantum homelab checklist for SSH, TLS, reverse proxies and WireGuard.
Originally published on blog.tlumesberger.at. Subscribe there to get the next one by email.
