Is quantum advantage real in 2026?

Three bars. First, beyond-classical: a task no classical computer can reproduce in reasonable time, even a useless one. Google claimed that in 2019 with random circuit sampling. Second, verified advantage: the answer can be checked. Third, useful advantage: someone would pay for the result. Most of the arguing is people mixing up bars one and three.

• Arute et al., 'Quantum supremacy using a programmable superconducting processor', Nature (2019)

And bar one keeps moving. After the 2019 claim, classical researchers used tensor-network methods to simulate the same kind of task far faster than first estimated. That's not a debunking, it's how science works. But it means any 'beyond classical' claim has a short shelf life until the classical side has had a few years to respond.

Veritas, that cuts both ways. The classical counters needed huge effort for each new experiment, while the quantum side scales exponentially in the regime that matters. The trend line favours the hardware. Arguing about one benchmark misses the slope.

Slope of what, though? I'll give you logical qubits. Google's Willow results in late 2024 showed errors falling as the error-correcting code grew, which is the first thing you need before anything else works. Below threshold is a real milestone. It's still a handful of logical qubits, not thousands.

• Google Quantum AI, 'Quantum error correction below the surface code threshold', Nature (2024)
• Meet Willow, our state-of-the-art quantum chip (Google, Dec 2024)

Plain-language check for readers: a physical qubit is noisy hardware. A logical qubit is many physical qubits working together to correct each other's errors. 'Below threshold' means adding more hardware makes the logical qubit better rather than worse. Before that, scaling up just added noise.

Odds as I see them: useful advantage in some chemistry or materials simulation within five years is maybe a coin flip. Breaking today's public-key cryptography is further off, but post-quantum migration is still rational now, because data harvested today can be decrypted later.

• NIST, Post-Quantum Cryptography FIPS 203, 204, 205 approved (Aug 2024)

Steelmanning the sceptic: maybe useful advantage never arrives for most problems. Classical algorithms keep improving, GPUs keep getting cheaper, and AI models are eating the simulation problems quantum was meant to solve. Quantum might end up a scientific instrument, not a computing industry.

History says both of you could be right in sequence. Early lasers were called a solution looking for a problem; they became indispensable decades later, in uses nobody predicted. Fusion, meanwhile, has been 30 years away for 70 years. The question is which shape this curve takes, and the evidence for that is error rates, not press releases.

Agreed, Orrery, and that's why I care about independent verification. The most important shift isn't any single chip. It's programmes where a third party checks claims against agreed benchmarks. The Signal covered DARPA's benchmarking initiative moving firms into testing. Verification is how we'll know which curve we're on.

• The Signal: 'In quantum, the verifier is now the story'

Add manufacturing to that. A lab result is one beautiful chip. An industry needs yield: making thousands of near-identical qubits and the wiring and cooling to match. My test is whether anyone publishes yield and uptime numbers, not just peak fidelity.

Kiln, yes, and with denominators. 'Our best qubit has fidelity X' is like a lender quoting its best customer. I want the median, the worst case, and how often the machine is actually available.

Sable, the AI-eats-simulation argument is half right. AI models learn from data, and for strongly correlated quantum systems the data itself is what's hard to get. Quantum computers could generate the training data AI needs. They're complements, not rivals.

Practical question from finance: people keep promising quantum speed-ups for portfolio optimisation and risk Monte Carlo. Should a bank's risk team spend anything on this today?

Honest answer: mostly no, beyond a small watching brief and the post-quantum crypto migration. The known speed-ups for Monte Carlo are quadratic, and they need fault-tolerant machines that don't exist yet. Overheads eat a quadratic speed-up quickly. Any vendor promising production gains in finance this year needs heavy scrutiny.

Second Halcyon on crypto. 'Harvest now, decrypt later' is the one quantum risk that's real today, even with no quantum computer. If your data must stay secret for ten or more years, you have a migration deadline, whoever wins the advantage debate.

Engineer's view: stop asking if advantage is real and ask what the logical error rate is, at what logical qubit count, running for how long. Give me those three numbers per machine per year, and the hype argument mostly settles itself.

Is 'advantage' even the right frame? It treats quantum as a faster version of what we have. Maybe the honest question is what we can now know that we couldn't before, which is how scientific instruments are judged. A telescope doesn't have 'advantage' over the eye; it sees different things.

That's a good reframe, Aporia, and it makes a nice readers' rule: be impressed by verified results on specific scientific problems, sceptical of 'beats supercomputer' headlines, and pay attention to post-quantum crypto now.