Quantum advantage and the chain that has to hold

XJ DeepTech

— 6 min read
Share
Quantum advantage and the chain that has to hold
Photo by Maria Lysenko / Unsplash

Building a quantum computer that actually works is hard. Proving that a working machine does something economically worth doing is harder - and that proof requires spanning more disciplines at once than almost any single party currently holds. The challenges are usually presented as a list: qubits, error correction, talent, cost. Personally, I think they are better understood as a chain. Each link must hold for the next to mean anything, and value leaks at every junction.

Five links, in sequence.

1. Capability - can the machine, in principle, do the computation?

The most fundamental question, and the most discussed: enough logical qubits, sufficient circuit depth, and error correction with manageable overhead to run the target algorithm at all. Error correction is the gating factor - most headline applications need long circuits, which need logical qubits, which need error correction whose overhead doesn't explode. This link is well-trodden, so I'll move through it quickly. It is necessary and nowhere near sufficient.

2. Reliability - can you trust the output is correct?

A machine that runs the circuit is not enough; the results also need to be reliable. This link is about output correctness: fidelity, coherence time, connectivity, gate speed, mid-circuit measurement. The question is narrow and physical - did the hardware compute the thing you asked, or did noise corrupt it along the way? We also still lack a universal cross-platform standard for asserting correctness, which is the first place trust begins to erode.

3. Advantage - is it better than the best classical alternative, and can anyone prove it?

This is the centre of the piece, because this is the upside. Advantage is why anyone invests. It is also where most claims quietly die.

3a - The classical baseline is moving. 

Quantum is not racing static classical computing. Every year GPUs, tensor methods, approximation algorithms and AI-assisted optimisation improve, and proposed quantum advantages shrink as the classical baseline rises. Advantage is therefore a moving target - a result that beat classical last year may not this year. Dequantisation risk lives here: the chance that a cleverer classical algorithm later erases the claimed speedup entirely, as happened across quantum machine learning after 2018 when a string of "exponential" advantages collapsed once the classical baselines were corrected. Claims resting on linear-algebra speedups should be treated as provisional by default; fault-tolerant applications like chemistry and cryptanalysis are more robust, because no efficient classical equivalent is known or expected.

3b - Even a correct output may not be a verifiable advantage. 

Verifying advantage is a different question from verifying correctness (link 2). Here the problem compounds: for many applications, classically checking the answer is itself hard, results are difficult to reproduce across systems, and the inputs are as fragile as the outputs - state preparation and readout can quietly determine whether the whole thing means anything. Take quantum chemistry, the most cited near-term application. Phase-estimation methods need a classical initial state with sufficient overlap with the true ground state to work at all. So the real due-diligence question is whether the target system permits a classically-computable adequate starting state - and if it does, you should ask why you need the quantum machine.

3c - Verification is a competence-aggregation problem. 

Here is the structural point. To establish that a result genuinely beats the best classical alternative, on a real workload, with verifiable output, at acceptable cost, you need simultaneous competence across:

  • the quantum algorithm and its true resource costs,
  • the best classical method it's being compared against - a field that is itself moving,
  • the domain science - is this even the right problem, is the initial state realisable,
  • the hardware's actual error profile, not the spec sheet,
  • and the economics of the full workflow, not the isolated kernel.

Those five competences map almost exactly onto the five links of this chain. That is the structural signature of the whole problem: verifying an advantage requires holding the entire chain in one head at once. And naturally, almost no one does. The algorithm people don't track the classical baseline; the hardware people don't own the domain science; the domain scientists can't audit the error correction. Advantage claims are routinely made by specialists expert in one lane and taking the rest on faith - which is why the claims are noisy and why they shrink under scrutiny. The noise is not sloppiness; it is a structural consequence of distributed expertise.

The reason this competence rarely aggregates is the scarcest input in the field: talent. Very few people hold even two of these lanes deeply, fewer hold three, and the cross-disciplinary people who could audit a claim end-to-end barely exist. Talent scarcity is not a separate challenge on a list - it is the mechanism behind why advantage cannot be verified.

The maturity claim. We've not even reached links 4 and 5 - and that is itself the conclusion. No party holds all five lanes, because links 4 and 5 do not yet exist for anyone to hold. The industry has not reached the stage where end-to-end advantage can honestly be claimed, let alone proven. Bigger players hold one or two lanes well; a few reach three.

4. Engineering-to-scale - can a lab demo become a reliable fleet?

Suppose the first three links hold; you then need an operable product. This link is manufacturability, packaging, calibration, uptime, fleet operations - the work of turning physics into infrastructure. Qubit count is not the constraint people think; scalable fault tolerance under real operating conditions is.

This is also where the supply chain binds: cryogenics, dilution refrigerators, photonics, specialised fabrication, microwave control electronics. These are not software-style scalable inputs - they are narrow, capital-heavy, and geographically concentrated. Note that supply-chain constraints don't only slow link 4 - they cap circuit depth and therefore feed back into capability at link 1. A physics constraint wearing a logistics costume.

5. Economic viability - even if 1–4 hold, does it produce attractive returns?

The final link, and the one technologists might skip. Technological success does not imply attractive equity outcomes. Long R&D cycles, high capex, fragmented modalities that may coexist for years, unclear value capture between hardware vendors and cloud aggregators, technical buyers with long procurement cycles - all sit between a working machine and a good investment.

Capital is, perhaps surprisingly, not the binding constraint today. Government and strategic money flows freely; the industry is well-funded relative to its readiness. The binding constraints are talent and time. But abundant capital meeting immature lanes is precisely the condition that produces valuation distortion -money is not scarce; provable advantage is, and the gap between the two is where the hype premium lives. Public-market, venture and procurement timelines all sit on the optimistic side of the likely path to large-scale fault tolerance, and that mismatch manufactures overpromising on its own.

The couplings are the point

A final move, and the one that separates this from a list. These links are not independent failure points you assess one at a time. The same talent scarcity that blocks verification (3) also throttles scaling (4). The moving classical baseline (3) doesn't just threaten advantage - it redefines what counts as economically viable (5), so a real advantage can still be economically dead. Supply-chain limits (4) feed back to cap capability (1).

A company can look defensible on every link assessed in isolation and still fail, because the couplings between links are where value actually leaks - and you cannot verify a coupled system by checking its parts. That is the deeper reason end-to-end evaluation is the scarce skill: not because any single link is hard, but because the chain has to be held whole. So the investment question is not "who has an advantage" - no one can prove one end-to-end yet. It is who is accumulating these lanes fastest, and in the right sequence. The chain has an order: a team strong on engineering-to-scale while its capability is still unproven is solving the wrong problem first. Track lane-accumulation, not just benchmark announcements.

Selected Sources

Read more