Quantum Hardware in 2026: Below Threshold, Not Yet Useful, and the Estimates That Fell 200x

Yesterday I wrote about the post-quantum migration from the defender's side: which of my TLS and SSH endpoints are already hybrid, and which certificates are stuck waiting on 2027. Several readers asked the obvious follow-up. How close is the machine we are migrating away from? This post is my attempt to answer that from primary sources, as of 7 September 2026, without either the vendor gloss or the reflexive "it's always 20 years away".

I am not a physicist. I run infrastructure, and I read quantum papers the way I read a CPU vendor's roadmap: skip the adjectives, find the table, check whether the number is measured or projected. That turns out to be a useful discipline here, because the field has three kinds of numbers that get mixed together in coverage: physical qubit counts, which are easy to grow and mostly irrelevant; two-qubit gate fidelities, which are hard to grow and mostly what matters; and logical error rates, which almost nobody has yet and which are the only ones that decide whether the thing computes.

The one result everything else hangs on

In December 2024 Google published "Quantum error correction below the surface code threshold" in Nature. On their 105-qubit Willow chip they ran a surface-code memory at distance 3, 5 and 7, and the logical error rate fell each time the distance went up, by a factor Λ = 2.14 ± 0.02 per two steps of distance. The distance-7 memory used 101 physical qubits and had a logical error of 0.143% per cycle, and the encoded qubit lived 2.4 times longer than the best physical qubit on the chip.

That is the whole game. Error correction only helps if adding qubits makes the logical error go down; below threshold it does, above threshold it makes things worse. Every roadmap in this post assumes Λ stays above 1 as the code grows, and Willow is still the only superconducting demonstration of that at distance 7. Google's own milestone page has not moved past milestone 2 in the twenty-one months since, and I found no 2026 announcement of a long-lived logical qubit. Google's 2026 news was elsewhere: a neutral-atom hardware program at JILA under Adam Kaufman, targeting "thousands or tens of thousands" of qubits as a second modality, which tells you something about how they see superconducting scaling.

You can reproduce the shape of the Willow result on a laptop in ten seconds, without a quantum computer, because the interesting part is the decoder and the statistics rather than the physics. Google open-sourced the simulator (Stim) and the standard decoder is PyMatching. This is the complete script I ran this morning:

import stim, pymatching, numpy as np

def logical_error_rate(d, p, shots=20000):
    circuit = stim.Circuit.generated(
        "surface_code:rotated_memory_z",
        distance=d, rounds=d,
        after_clifford_depolarization=p,
        before_round_data_depolarization=p,
        before_measure_flip_probability=p,
        after_reset_flip_probability=p,
    )
    dem = circuit.detector_error_model(decompose_errors=True)
    matcher = pymatching.Matching.from_detector_error_model(dem)
    dets, obs = circuit.compile_detector_sampler().sample(
        shots, separate_observables=True)
    pred = matcher.decode_batch(dets)
    return np.sum(np.any(pred != obs, axis=1)) / shots

print(f"{'p':>7} " + " ".join(f"d={d:<2}" for d in (3, 5, 7)))
for p in (0.002, 0.005, 0.007, 0.01, 0.015):
    row = [logical_error_rate(d, p) for d in (3, 5, 7)]
    print(f"{p:>7.3f} " + " ".join(f"{r:.4f}" for r in row))

Output on my machine (Python 3.14, Stim 1.16, PyMatching 2.4, about eight seconds total):

      p d=3    d=5    d=7
  0.002 0.0028 0.0012 0.0004
  0.005 0.0165 0.0138 0.0103
  0.007 0.0331 0.0348 0.0327
  0.010 0.0556 0.0828 0.1072
  0.015 0.1152 0.1968 0.2701

Read it across each row. At a physical error rate of 0.2%, going from distance 3 to distance 7 cuts the logical error sevenfold. At 0.7% the three columns are equal: that is the threshold for this circuit-level noise model and this decoder. At 1% and above, a bigger code is a worse code. Willow's physical gates sit around the 0.2% row, which is why it works, and why a chip with 1% gates and a thousand qubits is a thousand-qubit random number generator. Every headline qubit count you read this year should be filtered through that table.

Where each platform actually is

Here is what I could confirm from vendor pages and papers, not press summaries. Where a number is missing, the vendor did not publish it.

Platform	Qubits	Two-qubit fidelity	Best logical demonstration	Source
Google Willow	105 superconducting	not restated in 2026	d=7 memory, 0.143%/cycle, Λ=2.14	Nature 638:920
IBM Nighthawk r2	120 superconducting, 218 couplers	not stated	qLDPC decoding in under 480 ns on FPGA	IBM blog, 31 Aug 2026
Quantinuum Helios	98 trapped Ba⁺ ions	99.921%	48 error-corrected logical qubits at 2:1	Quantinuum, 5 Nov 2025
QuEra / Harvard	3,000-atom array run continuously for over 2 hours	not stated	up to 96 logical qubits, logical magic-state distillation	QuEra 2025 summary
Atom Computing / Microsoft "Magne"	1,225 neutral Yb atoms	not stated	50 logical qubits planned; toric code demo Jun 2026, no numbers	QuNorth
IonQ	256-qubit systems targeted 2026	99.99% on lab prototypes	none reported	IonQ
Rigetti Cepheus-1	108 superconducting (12 chiplets)	99.1% median	none	secondary reports
Microsoft Majorana 2	a few devices	n/a	none; parity lifetime ~20 s claimed	Science News, 2 Jun 2026

Three things in that table deserve a closer look.

Quantinuum's fidelity is the number to watch. 99.921% two-qubit and 99.9975% single-qubit is comfortably into the "0.2% row" of my table, on all-to-all connected ions rather than a nearest-neighbour grid, which is why they can run concatenated codes at a 2:1 physical-to-logical ratio that a surface code cannot touch. The catch is speed: ion gates are microseconds to milliseconds where superconducting gates are tens of nanoseconds, and their roadmap to Sol in 2027 (192 qubits, 2D grid) and Apollo in 2029 ("fully fault-tolerant") is really a roadmap about speed and count, not fidelity. Quantinuum also went public on 4 June, 28 million shares at $60 for $1.68 billion gross, Nasdaq ticker QNT. That is the first pure-play fault-tolerance company with audited financials, which I count as progress of a different kind.

IBM's bet is a different code, not a better qubit. Their June 2025 roadmap aims at 200 logical qubits and 100 million gates by 2029 (Starling), and the way they get there is quantum LDPC "bicycle" codes: the [[144,12,12]] "gross" code holds 12 logical qubits in 144 data qubits, roughly ten times denser than a surface code, at the price of needing long-range couplers and a decoder fast enough to keep up. Nighthawk r2 delivers the couplers; the Relay-BP decoder runs a round in under 480 nanoseconds on an FPGA. The chip that puts the two together, Kookaburra, is still marked "planned" on IBM's 2026 roadmap page as of this week. I would not call the year a miss until December, but it is September.

Majorana is still a claim. Microsoft's Majorana 1 in February 2025 drew an unusual editor's note from Nature saying the data did not constitute evidence of Majorana zero modes. Majorana 2, shown at Build on 2 June 2026, reports a parity lifetime of around 20 seconds, up from milliseconds, in a Microsoft technical paper that is not peer-reviewed. Henry Legg at St Andrews points out that only Z-basis measurements are shown, and a qubit needs X and Z. I find the argument that the device is interesting persuasive and the argument that it is a qubit not yet made. I will update when a referee does.

"Advantage" now comes in three grades

On 30 July 2026 IBM and partners announced quantum advantage three times in one press release: a doped-Clifford sampling task with UChicago on 97 qubits (arXiv:2607.25941), Floquet-Ising dynamics with Qedma on up to 74 qubits cross-checked on Quantinuum hardware (arXiv:2607.24937), and an operator Loschmidt echo with Algorithmiq on 56 qubits (arXiv:2607.25998). Google's "Quantum Echoes" result from October 2025 estimated 13,000× over Frontier on a 65-qubit out-of-time-order correlator and called itself "verifiable".

Notice the adjectives. Google says verifiable. IBM's release says trusted. Neither says useful, and the three IBM papers sit at different evidentiary levels: only the first has a complexity-theoretic hardness argument, and the third lacks quantitative accuracy bounds, as an independent fact-check laid out. D-Wave's March 2025 "supremacy" claim on spin-glass dynamics was partly reproduced classically within weeks by Sels at NYU and Tindall at Flatiron, which is the fate of most advantage claims that pick a problem nobody outside the field wanted solved.

My reading is that the word has been graded down to "a task where the classical cross-check is expensive", which is a real milestone and a long way from "a task a customer would pay for". IBM's CEO told analysts in April that partners would achieve "the first examples of quantum advantage this year" and, on the field's own definition, they did. On a customer's definition, the Quantum Advantage Tracker is still a leaderboard, not an invoice.

The number that matters to me: how many qubits to break RSA-2048

This is where the year actually moved, and it moved on paper, which is the part that should worry defenders more than any chip announcement.

In 2019 Gidney and Ekerå estimated 20 million noisy qubits over eight hours. In May 2025 Gidney cut that to under one million qubits in under a week, mostly via "magic state cultivation" and yoked surface codes. In February 2026, Iceberg Quantum's "Pinnacle" architecture paper (arXiv:2602.11457) claimed under 100,000 physical qubits at a 0.1% error rate and about a month of runtime, using qLDPC codes of the kind IBM is building toward. And in March, Google with Dan Boneh and Justin Drake published an estimate for the 256-bit elliptic-curve discrete log, the one that protects Bitcoin and most of your TLS certificates, at under 500,000 physical qubits with a runtime of minutes. I only have that last one through secondary coverage, and Google notably released a zero-knowledge proof of the circuit rather than the circuit itself, so treat the figure as reported rather than checked.

Set the required column against the built column. The gap is between two and three orders of magnitude in count, and the built column is not yet at the required quality except on Quantinuum's 98 ions and Google's 105 transmons. The gap has been closing from the top, by algorithmic and architectural work, faster than from the bottom. Scott Aaronson, who spent a decade telling people to calm down, wrote in April that trusted hardware and error-correction people now tell him a cryptographically relevant machine "ought to be possible by around 2029", with explicit caveats that it might take longer and that the industry has every incentive to say sooner. He co-signed a position paper convened by Coinbase to that effect. When Aaronson moves his date, I move mine.

For what I do, the conclusion is the same as yesterday's post and I will not repeat it at length: harvest-now-decrypt-later means the key-exchange migration had to be done already (it is, in your defaults), and the signature migration has until roughly the end of the decade, which is exactly the window the estimates above now point at. The difference between "2029 with caveats" and "2035" is the difference between a certificate rotation you plan and one you do in a panic.

Who is paying, and for what

Government money moved in 2026 in a way that is worth recording, because it changes which companies survive to 2029.

• DARPA's Quantum Benchmarking Initiative selected eleven companies for Stage B in November 2025: Atom Computing, Diraq, IBM, IonQ, Nord Quantique, Photonic, Quantinuum, Quantum Motion, QuEra, Silicon Quantum Computing and Xanadu. The program's stated goal is to decide whether utility-scale is achievable by 2033; Stage C decisions are expected late this year but no date is official. Microsoft and PsiQuantum are in the final phase of the related US2QC program.
• US Commerce, on 21 May 2026, signed $2.013 billion in CHIPS R&D letters of intent with minority equity stakes: $1 billion to IBM, $375 million to GlobalFoundries, $100 million each to Atom Computing, D-Wave, Infleqtion, PsiQuantum, Quantinuum and Rigetti, up to $38 million to Diraq. The US government taking equity in quantum vendors is new and, I think, under-discussed.
• The National Quantum Initiative reauthorization (S.3597, through 2034) cleared Senate Commerce in April and is not yet enacted. DOE renewed its five research centres for $625 million in November 2025.
• Denmark is buying the Magne machine from Atom Computing and Microsoft for around €80 million, 50 logical qubits on 1,225 atoms, due end of 2026 or early 2027 depending on the source. That will be the first "level 2" machine a customer owns rather than rents, if it ships.

On the private side, the equity market did what it does. IonQ's second-quarter revenue was $80.1 million, up 287%, on a valuation reported around $16 billion; Rigetti trades near $5 billion on roughly $10 million of trailing revenue. Nvidia's April launch of "Ising", open AI models for calibration and decoding, moved the whole sector 50% in a week, fifteen months after Jensen Huang's "15 to 30 years" remark cut it in half in a day. None of that tells you anything about qubits. It tells you the sector is now priced like a 2029 outcome, which means a 2032 outcome will be very expensive for somebody.

What I would tell an engineering team

I get asked what to do about this by people who run systems, not labs, so here is the short version.

1. Ignore qubit counts. Track two numbers. Two-qubit gate fidelity above 99.9% on a scalable architecture, and a demonstrated Λ above 1 at distance 7 or more. Today that is Quantinuum on the first and Google on the second. When one vendor has both at a few thousand qubits, the timelines above become schedules.
2. The decoder is a systems problem you already understand. Google needs a decision every microsecond; IBM's qLDPC decoder runs a round in 480 nanoseconds on an FPGA; Riverlane's Deltaflow 2 averaged 16 microseconds on Google's distance-5 data. This is real-time, low-latency stream processing with a hard deadline, and the people being hired to build it are increasingly FPGA and low-latency-network engineers rather than physicists. If you run trading infrastructure, you already know this shape of problem.
3. The cryptography deadline is not a physics question any more. The estimates fell 200× on paper in seven years without a single qubit being built. Plan signatures for 2029, as I wrote yesterday, and stop treating the machine's absence as an argument.
4. Learn Stim. It is the perf of this field: free, fast, and the fastest way to tell whether a paper's numbers mean what the press release says. The script above took me longer to format for this post than to write.

I expect to revisit this in six months, and the things I will check first are whether Kookaburra shipped, whether Google's milestone page changed, whether the Majorana 2 data has been refereed, and whether DARPA's Stage C list is shorter than eleven. If all four move, 2029 will look conservative. If none do, I will say so.

---

Related: Post-Quantum Migration in 2026 and Neuromorphic Computing in 2026.