The best quantum computer on Earth needs to get roughly 100,000 times more capable before it can break RSA-2048 encryption or model the enzyme that makes fertiliser. That is the headline finding of QUOPS, a new benchmark from Sandia National Laboratories that tested machines from Quantinuum, Google and IBM on the same yardstick for the first time.
In the same month, IonQ traded at around 57 times its sales, and quantum stocks jumped between 7% and 9% in a single session on news that did nothing to close that gap. Totally normal market behaviour, nothing to see here.
What Sandia actually measured
The paper, led by Timothy Proctor and Robin Blume-Kohout of Sandia’s Quantum Performance Laboratory with co-authors from Quantinuum and NVIDIA, was posted to arXiv on 10 September and picked up by The Quantum Insider and Phys.org the following week. It has not been peer reviewed yet.
QUOPS (Quantum Universal Operations Performance System) throws randomised circuits at a machine: layers of arbitrary single-qubit rotations and two-qubit CNOT gates, at varying widths and depths, which loosely resemble the shape of real algorithms. Two numbers come out. Q is the size of the biggest circuit the machine can run and still pass a success threshold. Omega is how many of those effective operations it gets through per second.
The clever bit is that the same test works on raw physical qubits and on logical qubits (bundles of physical qubits running error correction), so you can compare a trapped-ion box, a superconducting chip and a fault-tolerant prototype on the same footing. Vendors have spent a decade choosing their own metrics, which is a bit like letting each horse pick the length of its own race.
The scores
Figures below are from the paper as reported by Quantum Computing Report and PostQuantum.
| Machine | Type | Q (circuit size) | Speed (QUOPS/sec) |
|---|---|---|---|
| Quantinuum Helios-1 | Trapped ion | 1,504 (1,824 with postselection) | 303 |
| Quantinuum H2-1 | Trapped ion | 1,320 | 353 |
| Google Willow | Superconducting | 216 | 20 million |
| IBM ibm_boston | Superconducting | 204 | 310,000 |
| Helios-1, 8 logical qubits | Error-corrected | 40 | 4.9 |
| Needed: RSA-2048 | 250 million | 5,700 | |
| Needed: FeMoco chemistry | 340 million | 800 |
Look at the speed column for a second. Google’s chip already runs thousands of times faster than a code-breaking job would require. Speed is fine; circuit size is where every machine hits the wall, and each one sits around five orders of magnitude short on the size of circuit it can run, which the authors spell out in the abstract: “computational capability must grow by 5 orders of magnitude, motivating fault-tolerant approaches.”
The logical-qubit result is the humbling one. Running Helios-1 as eight Steane-encoded logical qubits scored 40, about 2.7% of what the same hardware manages raw. Error correction is the only credible road to useful machines, and right now it costs you roughly 97% of your capability to get on it.
Then there’s the timeline. PostQuantum’s analysis of the paper notes Quantinuum’s capability has been doubling roughly every 1.4 years and IBM’s every 2.1 years. Keep those trends and you reach the challenge problems somewhere between 2050 and 2070. The roadmaps promising scientific payoff in the early 2030s would need that rate to quadruple.
To be fair to everyone involved, the authors flag the limits themselves: random circuits are a proxy for chemistry and cryptography, results depend on the chosen success threshold, and independent testing still needs vendor access. It’s also worth noting Quantinuum co-wrote the paper and topped the table. Its machine really does win on circuit size, but a benchmark co-authored by the winner will always get a raised eyebrow, and it should.
Meanwhile, in press release land
On 24 September, Infleqtion (NYSE: INFQ) announced 30 entangled logical qubits from 80 physical qubits on its Sqale neutral-atom machine, which it called a first for a commercial neutral-atom system. “Getting 30 logical qubits to work together is hard, and our team has done it,” said CEO Matt Kinsella. It is genuinely hard. The company says it is on track for 100 logical qubits in 2028 and 1,000 by 2030.
The small print is doing a lot of lifting, though. The experiment ran about 1,000 physical operations using an [[8,3,2]] code, which can detect a single error but cannot correct one, and it relied on throwing away every run where an error was flagged. Marin Ivezic at PostQuantum pointed out that roughly 75% of outputs fell outside the ideal set, and that the share of runs you keep shrinks as circuits grow. “A company that has traded publicly since February, and whose chief executive defines logical qubits as error-corrected, should say in its headline which kind it is counting,” he wrote. Harvard hit 48 logical qubits with the same code back in 2023, for context.
Two days earlier, Microsoft opened a 15,000 sq ft centre at the University of Maryland and gave DARPA hands-on access to its Majorana 2 hardware under the agency’s Quantum Benchmarking Initiative, which aims to decide whether any approach can deliver a commercially useful machine by 2033. “This is a big deal for us because it essentially enables DARPA to kick the tires and figure out how good this machine is,” Microsoft’s Zulfi Alam told eWeek.
Credit where it’s due: that is the right move. Majorana 1, announced in February 2025, arrived with a Nature paper whose editors noted it did not by itself prove the exotic particles Microsoft’s whole approach depends on, and physicists have argued about it ever since. Letting the Pentagon’s research arm poke the hardware in person beats another blog post. Independent measurement being, as it happens, the one thing this entire industry has been short of.
The stocks versus the science
None of the three most-traded quantum stocks had a machine in the QUOPS test. D-Wave builds annealers, a different kind of device that can’t run these circuits at all. IonQ and Rigetti weren’t included. Investors, as ever, were unbothered. On 17 September, the week the benchmark landed, IonQ rose 9%, D-Wave 8% and Rigetti 7% on news of NVIDIA optimisation work.
By 24 September, per 24/7 Wall St., IonQ sat at $44.08, D-Wave at $17.23 and Rigetti at $16.28. The Motley Fool puts their market caps at roughly $18 billion, $6.4 billion and $5.3 billion. Their second-quarter revenue was $80.1 million, $3.1 million and $5.1 million respectively. D-Wave is valued at about 2,000 times its quarterly revenue. (Jesus Christ… I despair.)
Bull, bear, neutral
- Bull: IonQ grew Q2 revenue 287% year on year and holds about $3 billion in cash, so it can fund a long wait. Error correction is improving fast, and DARPA’s 2033 programme gives the serious players a public scoreboard to win on.
- Bear: IonQ trades at around 57 times sales against a tech average near 8, and its Q2 adjusted EBITDA loss widened to $120.3 million. A national lab says the useful machines are five orders of magnitude away and, on current trends, decades off. Prices assume the 2030s.
- Neutral: The companies with real cash can survive a long winter; the question is how much of today’s price survives with them. Watch for independent benchmarks like QUOPS and DARPA’s results, and weight them above press releases.
Not investment advice, just a market watcher’s notes. We ran through the sector in more detail in our sceptic’s guide to quantum stocks, and the error correction progress we covered in our piece on below-threshold error rates is the real reason for hope.
What this means
- For the first time, Google, IBM and Quantinuum machines have been scored on one test, and all of them are about 100,000 times too small for code-breaking or serious chemistry.
- Speed is already sufficient; circuit size is the bottleneck, and error correction currently eats about 97% of capability.
- “Logical qubit” now covers everything from full error correction to detect-and-discard. Ask which one.
- Your RSA-encrypted bank login is safe from quantum computers this decade. Upgrading to post-quantum encryption is still sensible, because governments will insist on it anyway.
- Researchers are making honest, measurable progress. The hype sits with the marketing departments and the ticker.
If you’d like someone reading the arXiv preprints so you don’t have to, and flagging which AI and tech tools earn their subscription fee, our newsletter does precisely that, with the hype stripped out.
Did you know: FeMoco, one of the benchmark’s two target problems, is the iron-molybdenum cluster bacteria use to fix nitrogen; the industrial version of that reaction, the Haber-Bosch process, uses an estimated 1-2% of the world’s energy.
Sources
- Proctor et al., “Benchmarking the computational power of quantum computers”, arXiv
- Quantinuum: Introducing QUOPS
- The Quantum Insider: Quantum computers need 100,000-fold performance gain
- Phys.org: New benchmark puts quantum computers to the test
- Quantum Computing Report: Sandia, Quantinuum and NVIDIA introduce QUOPS
- PostQuantum: QUOPS benchmark analysis
- The Quantum Insider: Infleqtion achieves 30 entangled logical qubits
- PostQuantum: Infleqtion’s 30 logical qubits rely on error detection
- The Quantum Insider: Microsoft gives DARPA access to Majorana system
- eWeek: Microsoft gives DARPA access to Majorana 2
- Motley Fool: Rigetti, D-Wave or IonQ
- 24/7 Wall St.: Quantum stock moves, 24 September