Utility-Scale Quantum Computing

In short

In June 2023, IBM and collaborators published a Nature paper in which a 127-qubit superconducting processor (Eagle) ran a dynamics simulation of the two-dimensional transverse-field Ising model — 60 steps of a Trotterised evolution on a 127-site lattice — and produced expectation values that agreed with classical tensor-network simulations where those simulations could still be trusted. Beyond that regime, the quantum machine kept running; the classical simulations became expensive. IBM called this "quantum utility" — a softer claim than quantum advantage. Utility means useful work at a scale where classical methods are strained. Advantage means provable speedup no classical computer can match. The distinction matters because, within six months of the paper, three separate classical groups (Tindall et al, Begušić et al, Kechedzhi et al) published tensor-network and belief-propagation methods that matched or beat the 127-qubit results on a laptop. Utility is a moving goalpost — it tracks the frontier of what classical simulation can reach, and that frontier moves every time a quantum experiment prompts classical algorithm designers to try harder. The honest 2026 view: utility-scale quantum computing is a real regime (noisy processors running real physics problems at useful sizes), but utility-scale demonstrations do not yet constitute quantum advantage. Early industrial customers — BMW, Boeing, JSR Corporation, Mitsubishi Chemical, HSBC — run small chemistry and optimisation problems on IBM, Quantinuum, and IonQ hardware as learning exercises. Expect the next five years (2026-2030) to be dominated by mid-sized utility-scale experiments on 100-1000 physical qubit machines, with the first credible quantum advantage arriving only after logical-qubit machines with 50+ logical qubits exist, around 2029-2032.

If you asked a quantum computing researcher in 2019 when quantum computers would become useful, the word they would have used was advantage — and they would have meant it in the strict complexity-theoretic sense. A problem for which a quantum algorithm runs in polynomial time while every classical algorithm provably requires super-polynomial time. Shor's algorithm against RSA-2048 is the textbook example. Nobody has a quantum machine big enough to run it, but when one exists, no amount of classical cleverness will catch up — the gap is exponential and the math prevents the gap from closing.

If you ask the same researcher in 2026, the word they are more likely to use is utility. And they will mean something softer: a quantum machine running a physics or chemistry problem at a size where classical simulation is expensive, even if it is not impossible. Utility does not ask "can a classical computer ever catch up?" — it asks "is the quantum machine earning its keep today?"

This chapter is about the gap between these two words. Where "utility" came from. What the 2023 IBM demonstration showed and why it landed on the cover of Nature. Why three separate classical groups beat the quantum result within six months. And what industrial customers — BMW running battery chemistry, HSBC running portfolio-risk toys, JSR running polymer simulations — actually get out of today's hardware. The goal is to leave you with a calibrated view: utility-scale quantum computing is real, it is interesting, and it is not quantum advantage. Distinguishing the two is the most important editorial skill you can develop before reading any press release about "quantum breakthroughs" between now and 2030.

Utility versus advantage — the definition

The two words sit on the same axis but measure different things.

The classical-difficulty axis runs from "easy" to "impossible". Utility-scale demonstrations occupy the middle — strained but not impossible — where classical algorithms still technically work but are expensive. True quantum advantage sits at the right edge, reserved for problems where classical provably cannot catch up.

Read the diagram carefully. The utility regime is bounded on the left by "classically easy" (a laptop finishes in seconds) and on the right by "classically impossible" (no classical computer can ever do this in the lifetime of the universe). The utility regime is the middle band — where classical methods exist, they work, but they get slow and expensive at the scale you care about.

The advantage regime is the right edge only. A problem is in the advantage regime if no classical algorithm, present or future, can keep up. Shor's algorithm on a 2048-bit RSA key is in the advantage regime, assuming the standard complexity conjectures hold. Simulating the ground state of a large enough strongly-correlated material is plausibly in the advantage regime. Most other things — including almost everything demonstrated on actual quantum hardware as of 2026 — are not.

The formal distinction:

Quantum advantage (strict): there is a problem P and a quantum algorithm Q running in time T_Q(n) such that every classical algorithm solving P requires time T_C(n) \gg T_Q(n), where the gap grows super-polynomially in the input size n. The gap is a statement about all possible classical algorithms, not just the ones we know today.
Quantum utility (IBM's framing, 2023): a quantum machine produces accurate results on a problem at a scale where the best currently available classical algorithms are computationally expensive. The comparison is with known classical methods, not all possible classical methods.

The difference looks small on paper. It is huge in practice. Utility is empirical and depends on what classical algorithms exist today; advantage is complexity-theoretic and depends only on what algorithms can exist. Utility is a moving target — every time someone publishes a clever classical simulation, the utility line moves to the right. Advantage, if you can prove it, stays put.

Quantum utility (IBM 2023)

A quantum processor is said to operate in the utility regime if it produces verifiable results on a scientifically meaningful problem at a scale where classical simulation using currently-known methods is computationally strained — but not necessarily impossible. Utility does not imply quantum advantage; it implies useful work at the frontier of classical tractability.

If that definition feels squishy, that is on purpose. Utility is a practical claim. The scientific community has never formally endorsed "utility" as a theorem; it is a marketing term IBM introduced in 2023 to describe what their machine was actually doing. It has since been adopted, cautiously, by parts of the research community because it names a regime that does exist and needs a name.

The Kim et al 2023 experiment

On 14 June 2023, Nature published "Evidence for the utility of quantum computing before fault tolerance" by Youngseok Kim and 14 collaborators at IBM, with contributions from Google Research and UC Berkeley. The experiment ran on IBM's Eagle processor, a 127-qubit superconducting device with heavy-hexagonal connectivity (each qubit coupled to 2-3 neighbours in a honeycomb-like lattice).

The problem they chose

The target was the two-dimensional transverse-field Ising model on a heavy-hex lattice of 127 sites. The Hamiltonian is

H = -J \sum_{\langle i, j \rangle} Z_i Z_j - h \sum_i X_i,

where J sets the strength of nearest-neighbour ZZ couplings, h sets the strength of the transverse (x-axis) magnetic field, and \langle i, j \rangle runs over every pair of neighbouring qubits in the heavy-hex lattice. Why this Hamiltonian: the transverse-field Ising model is the simplest quantum spin model with a genuine quantum phase transition; it is the "hydrogen atom" of quantum many-body physics; and it has 60+ years of classical-simulation tooling. Any quantum claim about it can be compared against serious classical competition.

What they simulated

They evolved an initial product state under H for a fixed total time, broken into N Trotter steps. Each Trotter step approximates e^{-iH\delta t} as a product of single-qubit X rotations (for the field term) and two-qubit ZZ rotations (for the coupling term):

e^{-iH\delta t} \approx \prod_{\langle i,j \rangle} e^{iJ\delta t Z_i Z_j} \cdot \prod_i e^{ih\delta t X_i} + O(\delta t^2).

Why Trotterisation: unitary time evolution under a Hamiltonian with many non-commuting terms cannot be implemented directly as a quantum circuit; the standard trick is to chop time into small steps and approximate each step as a product of commuting-block evolutions. Each block corresponds to one physical interaction. The error per step is O(\delta t^2), which you can make as small as you want by taking smaller steps — at the cost of longer circuits.

They ran up to N = 60 Trotter steps. Each step involves roughly 144 two-qubit ZZ gates (one per edge of the heavy-hex lattice) and 127 single-qubit X rotations. Sixty steps is therefore roughly 9,000 two-qubit gates — well beyond any circuit that had been run to produce accurate results before 2023.

What they measured

The quantity of interest was the magnetisation \langle X_i \rangle on specific sites and correlator \langle X_i X_j \rangle on specific pairs, as a function of Trotter step count N and interaction strength J\delta t. These are physically meaningful observables — experimentalists measuring real materials care about exactly these kinds of quantities.

Why naive execution would fail

At 9,000 two-qubit gates with a per-gate error rate of around 5 \times 10^{-3} (the IBM Eagle specification), the total error per circuit is roughly 9000 \times 0.005 = 45 — meaning the circuit is completely corrupted many times over. Running the raw circuit and averaging the measurement outcomes would give you random noise, not a signal.

The key trick — zero-noise extrapolation

The paper's central technical contribution is zero-noise extrapolation (ZNE), a form of quantum error mitigation (not correction). The idea:

Run the same circuit at the natural noise level. Call the result \langle O \rangle_{\lambda = 1}.
Artificially increase the noise by stretching gate durations or inserting identity operations. Run at noise level \lambda = 1.5 and \lambda = 2. Call those \langle O \rangle_{\lambda = 1.5} and \langle O \rangle_{\lambda = 2}.
Fit a polynomial (usually exponential decay) through (\lambda, \langle O \rangle_\lambda) and extrapolate back to \lambda = 0 — the hypothetical noise-free value.

Zero-noise extrapolation runs the circuit at the physical noise level (λ=1) and at artificially amplified noise levels (λ=1.5, 2). A fit through those points is extrapolated back to λ=0 to recover a noise-free estimate. This is error *mitigation*, not error correction — it trades circuit repetitions for bias reduction, but does not scale to arbitrarily deep circuits.

Why ZNE works at 127 qubits but fails at 500: the extrapolation fit has a finite statistical error that grows roughly as e^{T\lambda} where T is the total gate count. At 9,000 gates, the amplification is large but tolerable; at 100,000 gates, the variance of the extrapolated estimate blows up to where the signal is drowned. ZNE is a mitigation technique, not a correction technique — it reduces bias at fixed circuit depth but cannot make deep circuits tractable.

The result

For Trotter steps N \leq 15, the ZNE-corrected quantum measurements of \langle X \rangle and \langle XX \rangle agreed with classical tensor-network (MPS/TEBD) simulations to within a few percent. For N > 15, the tensor-network simulations started to struggle — the entanglement in the evolving state grew beyond the bond dimensions the classical simulation could handle in reasonable time. Kim et al continued running the quantum machine up to N = 60 and reported the corresponding expectation values as the quantum's answer in a regime where classical verification was strained.

The paper's title choice was deliberate: evidence for the utility — not demonstration of advantage. The authors were careful. They did not claim the classical simulation was impossible; they claimed it was expensive.

The counter-attack — how classical methods caught up

The paper was published 14 June 2023. By 2 August 2023 — seven weeks later — the first classical counter-attack appeared on arXiv.

Tindall, Fishman, Stoudenmire, Sels

Joseph Tindall and collaborators at the Flatiron Institute posted "Efficient tensor network simulation of IBM's Eagle kicked Ising experiment" in late June, showing that belief propagation — a classical message-passing algorithm used in error-correcting codes — combined with matrix-product-state methods reproduced all of Kim et al's 127-qubit results using a few hours of desktop CPU time. Their key insight: the heavy-hex lattice has tree-like local structure, and belief propagation is exact on trees and approximately correct when the loops are long.

Begušić, Gray, Chan

Roman Begušić, Johnnie Gray, and Garnet Chan at Caltech posted "Fast and converged classical simulations of evidence for the utility of quantum computing before fault tolerance using a classical, tensor-network algorithm" in August 2023, independently matching Kim et al's results with a different tensor-network contraction scheme running on a standard workstation.

Kechedzhi et al (Google)

Google's team, including Kostyantyn Kechedzhi, posted "Effective quantum volume, fidelity and computational cost of noisy quantum processing experiments" in June 2023, arguing that the IBM demonstration operated in a regime where the effective entanglement was low enough that the quantum state could be well-approximated by tensor networks — so classical simulation could track the dynamics.

Patra et al

Further work by Siddhartha Patra and collaborators extended the classical envelope even further, handling circuits deeper than Kim et al had attempted.

What the counter-attack showed

Within six months of publication, three independent groups had used previously existing classical methods (tensor networks, belief propagation) on standard workstations (not supercomputers) to match or exceed the accuracy of IBM's 127-qubit experiment. The experiment was not in the "classically impossible" regime; it was in the "classically expensive but doable" regime, and the classical frontier moved fast when prompted.

Hype check. If you read the June 2023 press coverage — "IBM's quantum computer beats supercomputers," "Quantum utility achieved" — and then read the August-through-December 2023 arXiv preprints, the story flipped. The quantum result was not beaten in the sense of being incorrect; it was beaten in the sense of being classically reproducible. That distinction does not show up in press releases. Utility claims in quantum computing are almost always followed by classical catch-up within 6-18 months. The right response to any such announcement is: wait a year, then check if the classical frontier has moved.

Does the IBM paper still count?

Yes, and for a subtle reason. Kim et al's contribution was not just the quantum result; it was the error-mitigation methodology (ZNE at scale), the problem-formulation template (non-trivial Trotterised dynamics on real hardware), and the benchmark against which classical methods could be tuned. The classical catch-up happened because IBM's experiment forced classical algorithm designers to think harder about tree-like lattices and bounded-entanglement regimes. That is a real scientific contribution — just not the kind "quantum advantage" implies.

IBM's own framing in 2024-2025 shifted in response: newer papers from the group use phrases like "error-mitigated quantum simulation of an Ising quench" rather than "quantum utility." The word "utility" itself has become more cautious in scientific writing, though it persists in marketing.

Other utility-scale demonstrations

The Kim et al paper was the first and most famous. Several others followed.

Quantinuum H-series (trapped ions). Quantinuum's H1 and H2 machines, running 20-56 ions with very high fidelity, have demonstrated chemistry simulations of small molecules (H_2, LiH, HeH^+) at near-chemical-accuracy using VQE and related methods. The scale is small (fewer qubits) but the errors are much lower, so the results are more trustworthy per gate. Quantinuum's 2024 demonstrations of logical qubits (covered in the logical-qubit era chapter) are arguably more scientifically significant than IBM's 127-qubit raw-physical demonstrations — they are a step toward the advantage regime rather than a claim about the utility regime.

IonQ and Rigetti. Both have run small-molecule chemistry via VQE and quantum sensing demonstrations. Neither has claimed utility in IBM's sense.

Google Quantum AI. Google has focussed on random-circuit sampling (quantum supremacy / advantage) and, after 2024, on logical qubits (Willow). Their posture on utility has been deliberately cautious — they consider the term marketing-adjacent and prefer "quantum advantage" for claims they can defend complexity-theoretically.

Industrial partnerships. BMW, Mitsubishi Chemical, HSBC, Boeing, JPMorgan Chase, and Goldman Sachs have published joint results with IBM or Quantinuum on small-scale problems — battery-electrolyte simulations, protein-folding toys, portfolio-optimisation proofs of concept. The scientific content of these papers is typically that the quantum machine produced an answer that agrees with classical calculation — not that it beat classical calculation. They are learning exercises, not demonstrations of advantage, and the customers themselves know this.

Industry use cases — the honest view

What are industrial customers actually doing on utility-scale machines as of 2026?

A calibrated 2026 view of industrial use cases on utility-scale hardware. Every tile is in the "learning" or "utility" phase. None are in the "production advantage" phase, which is reserved for logical-qubit machines that do not yet exist at the required scale.

Read the grid carefully. No tile says "production advantage". Every industrial engagement on quantum hardware as of 2026 is either:

A learning exercise — the company wants its R&D team to develop expertise before the advantage era arrives, so they start running small problems now.
A benchmarking exercise — running the same problem on both classical and quantum hardware to track when the crossover point might arrive.
A hype exercise — press releases that overstate the result for marketing.

All three are rational corporate behaviour. A 2026 hedge fund cannot yet run a useful portfolio-optimisation problem on a quantum computer faster than its classical backtesting infrastructure; it runs small QAOA problems so that in 2030, when logical qubits exist, its quantum team is not starting from zero.

What does a typical industrial run look like? A pharmaceutical company wants to compute the ground-state energy of a 12-atom organic molecule (say a drug fragment). They run the molecule on:

A classical CCSD(T) computation (gold-standard post-Hartree-Fock): takes 30 minutes on a workstation, gives the answer to chemical accuracy (\pm 1 kcal/mol).
A quantum VQE computation on IBM's 127-qubit Eagle: takes 6 hours of compute-time, 10^5 circuit repetitions, yields an answer agreeing with CCSD(T) to 2-3 kcal/mol accuracy.

The classical run is faster, cheaper, and more accurate. The quantum run is a learning exercise — the team is building internal expertise in variational methods, Pauli grouping, error mitigation, and hardware calibration. The value of the quantum run is not the answer it produces; it is the expertise the team develops.

Pretend the classical run doesn't exist and the quantum run is "the" answer, and you are misled. Understand that the quantum run is practice for the post-2029 machines, and it is a rational investment.

Example 1: Reading the Kim et al result honestly

Setup. You pick up the June 2023 Nature paper (the Kim et al IBM Ising dynamics result). You want to decide: is this quantum advantage? Quantum utility? A hype announcement? How should you calibrate?

Step 1. Identify the complexity-theoretic claim. The paper claims the 127-qubit circuit produced accurate expectation values at N = 60 Trotter steps, a regime where classical tensor-network simulation with contemporaneous methods was expensive. Why this matters: complexity theory requires you to distinguish "classical methods we used" from "every possible classical method." The paper's language is carefully about the former — it reports on contemporary MPS simulations, not on a complexity-theoretic lower bound.

Step 2. Check for error-mitigation disclosure. The paper uses zero-noise extrapolation, which is error mitigation (not correction). ZNE amplifies noise, fits a decay curve, and extrapolates. This works at 9,000 gates but would not work at 90,000 gates. Why this matters: the scaling of ZNE with circuit depth is a structural limit on how far utility-scale demonstrations can push. Any claim that utility demonstrations scale to larger problems must address the ZNE variance-blowup issue, or switch to error correction entirely.

Step 3. Check for classical counter-attacks. Search arXiv for papers citing Kim et al 2023 and posted within 18 months. You find Tindall et al (belief propagation + MPS), Begušić et al (tensor-network contraction), Kechedzhi et al (effective quantum volume argument). Three independent teams reproduce the result classically within 6 months. Why this matters: utility is defined relative to currently-known classical methods. When new classical methods catch up, the utility claim shrinks even if the quantum measurement itself remains correct.

Step 4. Identify the scientific contribution. Despite the classical catch-up, Kim et al contributed: (a) a benchmark problem that drove classical-algorithm development, (b) a practical error-mitigation protocol deployed at unprecedented scale, (c) engineering evidence that 9,000-gate circuits can produce non-random results. Why this matters: scientific contribution and advantage claim are separate axes. A paper can make a real contribution while making a contested advantage claim.

Step 5. Calibrate the final reading. Kim et al 2023 is utility-scale quantum computing: real, interesting, not-yet-advantage. The experiment demonstrates that 127-qubit processors can produce measurable non-trivial dynamics; it does not demonstrate a classical impossibility.

Result. You read the paper as a quantum-utility demonstration — a learning milestone on the road to fault-tolerant advantage, not the crossing of the advantage line itself. Press claims of "beating supercomputers" should be read as journalism, not science.

A calibration grid for the Kim et al 2023 result. Four of five rows read positively for utility; none reads as advantage. The net: a meaningful utility demonstration, not a crossing of the advantage threshold.

Realistic 2026-2030 outlook

What should you expect in the five years after this article is written?

2026 — the utility middle

Hardware roughly doubles in qubit count from 2024. IBM's 1121-qubit Condor (2023) has been superseded by tiling approaches — multiple 127/156-qubit chips linked by classical communication and delayed feedforward. Utility-scale demonstrations continue, with classical counter-attacks matching them at 6-18 month delay. No fault-tolerant machine is large enough to do quantum advantage problems.

2027 — the trough

A likely "trough of disillusionment" (Gartner's curve): utility claims continue but are largely discounted by press and investors; the scientific community focuses on logical-qubit milestones rather than noisy physical scaling. Industrial customers continue small-problem engagements but begin shifting R&D spend toward post-quantum cryptography readiness rather than quantum-advantage chasing.

2028 — first multi-logical-qubit algorithms

Quantinuum and IBM, and possibly Google, demonstrate fault-tolerant algorithms on 5-10 logical qubits. These are small — not yet enough for practical advantage — but they are the first time a quantum computer runs a non-trivial algorithm where the error is suppressed by code distance rather than mitigation. The scientific character of quantum computing papers shifts from "utility" back to "advantage" because the advantage line is finally approaching.

2029-2030 — the 50-logical-qubit threshold

IBM's announced Starling target is 200 logical qubits by 2029; Quantinuum's roadmap projects similar timescales. If any of these arrive on schedule — and "if" is the operative word — the first non-contested quantum advantages on small-molecule chemistry could appear around 2030. This would be the first time since quantum supremacy (2019) that a quantum machine does something where no known classical algorithm can catch up, except now for a scientifically useful problem rather than a random-circuit benchmark.

The Indian context

India's National Quantum Mission (₹6003 crore, 2023-2031) has utility-scale demonstration as a core deliverable for its first phase. C-DAC Pune and TCS Research are building indigenous superconducting and photonic platforms targeting 50-100 physical qubits by 2028. TIFR Mumbai and IISc Bangalore have theoretical groups contributing to the error-mitigation and utility-benchmark literature. The mission's second phase (2028-2031) is projected to shift toward logical qubits, mirroring the global trajectory. No Indian institution is yet at the utility-scale hardware frontier, but the theoretical contributions are real.

Example 2: A pharma chemistry VQE as an industrial utility run

Setup. A Bengaluru-based pharmaceutical company wants to assess whether quantum computing could help screen drug candidates. They pick a well-studied target — the energy of an active-site model for a kinase inhibitor — and decide to run Variational Quantum Eigensolver (VQE) on an IBM 127-qubit machine. The molecule has 12 electrons in 12 spatial orbitals (a 24-qubit problem in Jordan-Wigner mapping).

Step 1. Set up the VQE problem. Map the molecular Hamiltonian to a qubit Hamiltonian using Jordan-Wigner. The resulting 24-qubit Hamiltonian has roughly 4000 Pauli terms after fermionic-to-qubit mapping. Why Jordan-Wigner: it turns fermionic ladder operators into strings of Pauli operators, giving you a Hamiltonian you can measure on a qubit device. The string lengths grow as O(n), so for 24 qubits the measurements are moderate in cost.

Step 2. Choose an ansatz. Pick a hardware-efficient ansatz — alternating layers of single-qubit rotations and nearest-neighbour CNOTs — with 6 layers. This gives 24 \times 3 \times 6 = 432 variational parameters. Why hardware-efficient: chemistry-motivated ansätze (UCCSD) are too deep for current hardware; hardware-efficient ansätze are shallower and run, though they may not express the true ground state exactly.

Step 3. Run the classical benchmark. First, run CCSD(T) on the same molecule using a classical workstation (say Gaussian or PySCF). The 24-orbital problem takes about 10 minutes and produces the reference energy E_{CCSD(T)} accurate to within 1 kcal/mol of experiment. Why you always run this first: the classical benchmark tells you what "the right answer" looks like, so you can judge the quantum result. A VQE run with no classical comparison is uninterpretable.

Step 4. Run VQE on IBM Eagle. Submit the 432-parameter ansatz to IBM's cloud queue. Each evaluation requires roughly 4000 circuit repetitions (one per Pauli term) times 8000 shots for statistics — about 3 \times 10^7 shots per energy evaluation. With 432 parameters and gradient-based optimisation, a full convergence takes 6-8 hours of quantum compute-time plus queue delays. Use zero-noise extrapolation and readout-error mitigation throughout.

Step 5. Compare. The VQE result converges to an energy agreeing with E_{CCSD(T)} to within 3-5 kcal/mol — worse than chemical accuracy. The classical run took 10 minutes and produced a more accurate answer; the quantum run took 8 hours and cost more. Why the quantum result is worse: the hardware-efficient ansatz cannot express the full correlated wavefunction, and residual error after mitigation is large enough to wash out the last 1-2 kcal/mol of accuracy. This gap is exactly what logical qubits are designed to close — but those do not yet exist at the required scale.

Result. The quantum run reproduces the classical answer at reduced accuracy and higher cost. The value of the exercise is not the answer — the classical answer is better. The value is the team now has hands-on experience with Jordan-Wigner, ansatz design, Pauli-grouping, and error mitigation. When fault-tolerant 50-logical-qubit machines arrive in 2029-2032, this team is prepared to run 100-orbital problems (FeMoco, biological catalysts) where the classical CCSD(T) run genuinely fails — there, the quantum will be the only option. In 2026, they are warming up.

The industrial reality of a 2026 VQE run. At 24 qubits, classical CCSD(T) is faster, cheaper, and more accurate. The exercise is valuable anyway — as preparation for 2030, when classical methods will no longer reach the problem sizes that matter most.

Common confusions

"Quantum utility means quantum advantage." No. Utility is empirical — "the quantum machine is doing something classical machines find expensive today." Advantage is complexity-theoretic — "no classical machine can keep up, now or ever." Utility claims are regularly reversed when classical algorithms improve; advantage claims, when proved, are permanent.
"Kim et al beat supercomputers." That was press framing. The paper's own language was "evidence for the utility of quantum computing before fault tolerance." Within six months, three independent classical teams reproduced the results on workstations. The quantum machine produced correct answers; it did not produce answers classical methods could not also produce.
"Utility demonstrations prove quantum computers are commercially ready." No. Utility-scale experiments demonstrate that noisy processors can produce non-random results at previously-unreachable circuit depths. They do not show that those results are better than classical on any commercially meaningful problem. No production workload runs on quantum hardware as of 2026 where the quantum version is cheaper, faster, or more accurate than the classical alternative.
"Zero-noise extrapolation is error correction." No. ZNE is error mitigation — a statistical post-processing technique that trades runs for bias. It cannot scale to arbitrarily deep circuits because the extrapolation variance grows exponentially with gate count. Only proper quantum error correction with logical qubits (see the logical-qubit era) can scale.
"Industrial customers wouldn't pay for quantum runs if there were no advantage." They pay for workforce-readiness and R&D option value, not for production advantage. A 2026 pharma company running VQE on a quantum cloud is practising for 2030; the practice has value independent of the 2026 answer quality.
"Classical tensor networks can simulate everything, so quantum is useless." No. Tensor networks work well for low-entanglement states and short-time dynamics. They struggle — and will continue to struggle — for long-time dynamics of strongly-correlated systems, random quantum circuits, and problems with deep algebraic structure (like factoring). The utility counter-attack exists precisely because Kim et al's problem had tree-like low-entanglement structure; not every problem does.

Going deeper

You have the picture: utility is a softer cousin of advantage, introduced by IBM in 2023 to describe noisy-machine demonstrations that push the classical frontier without crossing it. The rest of this section covers the technical machinery: why ZNE works and when it stops; how tensor-network classical simulators interact with quantum noise to chase the utility frontier; the role of the Preskill NISQ framing in making "utility" a necessary term; and the open question of whether any noisy intermediate-scale machine will ever cross into genuine advantage without fault tolerance.

Why ZNE works — the Richardson-extrapolation view

Zero-noise extrapolation is a quantum-specific application of Richardson extrapolation, an 1911 numerical technique for improving the accuracy of an approximation by combining results at different step sizes. In the quantum setting, the "step size" is the noise strength \lambda, and the approximation is \langle O \rangle_\lambda — the expectation value you measure at noise level \lambda.

For many noise models, the measured expectation value satisfies

\langle O \rangle_\lambda = \langle O \rangle_{\lambda = 0} + a_1 \lambda + a_2 \lambda^2 + \ldots

If you measure at \lambda = 1, 1.5, 2 and fit a quadratic, you can solve for \langle O \rangle_0 exactly — assuming the expansion is quadratic. In practice, the expansion is often closer to exponential — \langle O \rangle_\lambda \approx \langle O \rangle_0 e^{-\gamma \lambda G} where G is gate count — and the fit is exponential. Either way, the extrapolation is an educated guess, not a measurement.

The variance-amplification limit

The statistical uncertainty on \langle O \rangle_\lambda at any single \lambda is bounded by the shot count. When you extrapolate back to \lambda = 0, the uncertainty on the extrapolated value is amplified: the variance of the estimate grows roughly as e^{2\gamma G} where G is gate count. At G = 9000 (Kim et al), this amplification is tolerable with 10^4 shots per point. At G = 100{,}000, the amplification is e^{2\gamma \cdot 100000} which is astronomical — you would need 10^{40} shots to resolve a signal. ZNE has a natural ceiling, and that ceiling is roughly the current 100-qubit, 10000-gate regime.

This is why utility-scale demonstrations saturate rather than extend indefinitely. You cannot keep running bigger and bigger circuits and fixing them with ZNE; you run into the variance wall. Crossing that wall requires error correction, not mitigation.

Preskill's NISQ coinage and the utility gap

John Preskill coined NISQ (Noisy Intermediate-Scale Quantum) in 2018 to describe the era between "small toy hardware" (pre-2018) and "fault-tolerant machines" (post-202X). The word carefully avoided "useful" — Preskill's original essay explicitly said NISQ machines would not run Shor's or break cryptography, and the useful applications were unclear.

IBM's 2023 utility claim sits inside the NISQ era. The claim is essentially: NISQ machines at 100+ qubit scale can do something meaningful, even if they cannot do Shor's. "Utility" fills the linguistic gap between "toy" and "advantage" — a word for the regime where NISQ machines earn some (contested) keep.

The "practical quantum advantage" debate

In 2024-2025, several researchers — including Scott Aaronson in his Shtetl-Optimized blog — argued that "practical quantum advantage" (a term stronger than utility, weaker than strict advantage) requires:

A clear problem specification — not a random circuit, but a scientifically or commercially meaningful task.
A well-defined classical baseline — the best classical algorithm known, running on the best classical hardware available.
A verifiable quantum result — you can check the answer, at least in a restricted regime.
A robust gap that survives 18+ months of classical counter-attack.

By this definition, no demonstration as of 2026 has cleared all four bars. Kim et al cleared (1), (2), and (3) but not (4). Random-circuit supremacy claims (Google 2019, USTC 2020, Jiuzhang 2020) cleared (4) in some sense but failed (1) — random circuits have no scientific application. The cleanest forthcoming claim is quantum chemistry on a ~100-orbital active space, where CCSD(T) starts to fail and quantum phase estimation on a logical-qubit machine would both beat classical and be commercially interesting. That target is projected for 2029-2032.

The Flammia-Nielsen paradox

A curious feature of utility-scale experiments: they often operate in a regime where the quantum state is classically approximable (tensor networks work), but the quantum circuit itself is not (direct gate-level simulation is expensive). This is not a contradiction — the state can be low-entanglement (tensor-networkable) even if its preparation circuit is deep. Kim et al's 127-qubit circuit ran for 9000 gates but produced an output state with bond dimension under 100, which is why belief-propagation methods caught up. Future utility claims will need to target problems whose final state has high entanglement, not just deep circuits — and high-entanglement final states are exactly where classical simulation genuinely fails.

Will noisy machines ever reach advantage?

The honest answer, as of 2026: probably not, for problems of scientific interest. The scaling argument is brutal: mitigation techniques (ZNE, probabilistic error cancellation, symmetry verification) have variance that grows exponentially with gate count; circuits useful for chemistry advantage require 10^6-10^9 gates; no mitigation scheme can absorb that much error. The field's consensus is that useful advantage will require fault tolerance, and the utility era is a bridge, not a destination.

The exception is boson sampling / random-circuit sampling: problems where the goal is to produce a sample from a specific distribution rather than compute an expectation value. These do not need error correction in the same way, and Google-Sycamore-class machines have credibly produced samples that classical computers struggle to verify. But sampling problems have no known commercial application, so they fit "quantum supremacy" (demonstration of a gap) rather than "quantum utility" (useful work).

Indian context — C-DAC, TCS, and NQM utility targets

India's National Quantum Mission lists "demonstrations at utility scale" as a first-phase deliverable. C-DAC Pune is building a 100-qubit superconducting machine targeting 2028. TCS Research collaborates with IIT Madras on algorithm benchmarks. IIT Bombay has a photonic platform project. The Indian utility-scale effort mirrors the global trajectory with a 3-5 year delay and roughly 10x smaller budget; by 2030, India is expected to have a 50-100 physical qubit machine capable of joining the utility-scale demonstration pool, though not leading it.

Where this leads next

Quantum advantage case by case — a problem-by-problem ledger of where quantum advantage is strong, weak, or absent.
The logical-qubit era — the post-2024 era of error-corrected machines, which is what ultimately replaces utility with advantage.
State of the art 2026 — current physical qubit counts, gate fidelities, and hardware roadmaps.
Landscape 2026 — the full map of players, funding, and institutions.
Lessons about quantum speedups — the theoretical framework for understanding when quantum is actually faster.
Reading an arXiv paper — a practical guide for evaluating quantum-computing preprints, including utility claims.

References

Kim, Eddins, Anand, Wei, van den Berg, Rosenblatt, Nayfeh, Wu, Zaletel, Temme, Kandala, Evidence for the utility of quantum computing before fault tolerance (2023) — Nature 618, 500-505.
Tindall, Fishman, Stoudenmire, Sels, Efficient tensor network simulation of IBM's Eagle kicked Ising experiment (2023) — arXiv:2306.14887.
Begušić, Gray, Chan, Fast and converged classical simulations of evidence for the utility of quantum computing before fault tolerance using a classical, tensor-network algorithm (2023) — arXiv:2308.05077.
John Preskill, Quantum Computing in the NISQ era and beyond (2018) — arXiv:1801.00862. The paper that introduced the NISQ terminology.
Wikipedia, Quantum supremacy — the encyclopaedic entry covering the advantage / supremacy / utility distinction.
Scott Aaronson, Shtetl-Optimized blog posts on quantum utility — scottaaronson.blog. The clearest running commentary on whether contemporary demonstrations clear the advantage bar.