Lower-Overhead Error Correction Puts IBM’s 2029 Quantum Roadmap Within Reach

Oliver Dial draws a sharp line between three milestones that are often blurred together: quantum utility, quantum advantage, and fully error-corrected quantum computing. IBM says it reached quantum utility in 2023. Dial defines that as the point where a quantum computer can run something complex enough that simulating the quantum computer on a classical machine becomes challenging or impossible.

The stakes are narrower, and more concrete, than the usual claim that quantum computing is “arriving.” In Dial’s account, advantage is becoming a public benchmarking fight: quantum systems will have to produce results, classical researchers will try to match them, and the claim will depend on verification rather than branding. The longer 2029 claim depends on a different question: whether IBM can turn lower-overhead error correction into deployable logical-qubit systems.

Utility is not the same as producing a useful answer. In Dial’s framing, it is “table stakes”: until a quantum machine crosses the threshold where its quantum state cannot be efficiently simulated classically, anything done on it “really should have just [been] done on a classical computer instead.” The threshold depends on both the number of qubits and the number of operations the machine can run before noise makes the output unusable. Too few qubits, and the state fits within classical memory. Too shallow a circuit, and classical simulation techniques remain efficient.

Quantum advantage is the stronger claim Dial says IBM is targeting in 2026. He defines it as using a quantum computer to do something “better, faster, cheaper” than would have been possible on a classical computer. He separates the strict scientific version of the claim from the practical one. The strict version asks whether one can prove the quantum computer produced the right answer and did so faster than any possible classical route. The practical version is heuristic: a quantum computer produces a better answer than the available classical methods, and the user is satisfied with that result.

Dial does not present advantage as a declaration IBM can make unilaterally. He says some people have already claimed “nitty gritty heuristic advantage” over the past year. But the stronger process he describes is adversarial: IBM brings hardware; scientists bring problems; other scientists try to beat the results on classical machines. When a quantum result appears, classical computer scientists will try to replicate it, improve the classical method, or reduce the claimed gap. Dial says that back-and-forth has to “come to a bit of a close” before the claim can really settle.

That is why verifiability matters. Craig Smith raises Google’s earlier quantum-advantage claims as an example of why the term confuses non-experts. Dial’s answer is that benchmarks can mislead, and the hard question is whether the output can be verified. In Google’s random-circuit-sampling work, as Dial describes it, the machine generated a distribution of outputs that was hard to compute classically. For smaller circuits, one could compare against classical simulation and measure fidelity. For larger circuits, Google extrapolated that the machine was still producing the right answer some fraction of the time. Dial says “honestly, it’s probably true,” but he also says the extrapolation makes the strict scientific claim “a little bit dicey.” He adds that later replicas have made the claim stronger in some respects, with larger gaps between quantum and classical runtimes at smaller circuit sizes where simulation is possible, while still leaving the verifiability issue central.

IBM’s response, as Dial describes it, is the quantum advantage tracker, which he compares to Hugging Face for quantum. It is a leaderboard-style public venue where people can post specific problems, quantum performance, and classical performance. IBM initially set it up, but Dial says additions are made through GitHub issues rather than a closed IBM submission process. The point is to make the contest visible: quantum claims and classical counterclaims in the open.

The result is a more cautious claim than the phrase “quantum advantage” can suggest. Dial is not saying every useful computational problem is about to move to quantum hardware. He is saying IBM hopes the hardware, algorithms, and noise-correction techniques have advanced to a point where advantage-like demonstrations may be possible for a wide variety of users on IBM systems today, and that the debate now needs public problems, public classical baselines, and public verification.

Error correction is the core of Dial’s 2029 claim. He starts with classical error correction: send redundant information so the receiver can detect or correct mistakes. A simple repetition code sends three copies of a bit; if the receiver sees 1, 1, 0, they infer the intended bit was probably one. More sophisticated codes use parity checks. If eight bits are sent along with information about whether the number of ones is even or odd, a receiver can detect certain errors. Additional parity information can support correction, not just detection.

Quantum error correction cannot simply copy and inspect qubits. Looking at the qubits destroys the quantum state. But parity-style codes can work because one can design circuits that reveal the parity of a set of qubits without revealing what any individual qubit was. Dial calls this a kind of blind error correction.

Quantum systems also have two classes of errors to correct. Bit-flip errors turn one into zero or zero into one. Phase errors alter the angle of a superposition, which Dial describes as something like a timing error. Correcting one logical qubit therefore requires enough machinery to handle both kinds of error. The correction operations themselves are noisy, including the operations used to prepare and measure parity checks. If the physical error rates are too high, the code can do more harm than good by damaging information faster than it protects it.

Before IBM’s recent work, the most commonly discussed path was the surface code. Surface-code approaches can use nearest-neighbor connections between qubits laid out like a checkerboard, which makes them attractive for hardware. But their overhead is large. Dial says it would be typical to talk about using around 100 physical qubits to make one error-corrected logical qubit good enough for far-future algorithms. That means IBM’s best present processor would yield roughly one such logical qubit, plus the associated correction apparatus. Dial says that pushed system size and cost toward something IBM felt was beyond engineering reach.

The change, according to Dial, came from low-density parity-check codes. These were known from the communications world and, in principle, could offer much lower overhead. The missing piece was practical implementation: how to map them to hardware, extract syndromes, measure parity checks, and infer errors from those checks. Dial says IBM filled in the details for a code it calls the Gross code and showed an error-correcting code that is an order of magnitude more efficient than what had previously been known as practical.

The Gross code did not come free. Surface codes require only nearest-neighbor connections. High-rate codes require non-local connections. Dial says they need a way to “pop out of the plane” and create something like a highway overpass, rather than staying in a flat two-dimensional checkerboard. That is where hardware-software codesign becomes decisive in his account. Dial’s hardware team told the error-correction team what seemed buildable: perhaps six connections per qubit, with wires able to span about 10 qubits. The error-correction team designed a code around those constraints.

IBM’s 2026 experimental device for this work is not intended as a client cloud machine. Dial says it implements two logical qubits, which is not computationally useful. Its importance, as he frames it, is that it demonstrates the features needed to implement the Gross code on a single chip. Dial describes it as the prototype for systems IBM wants to build in 2029, 2030, and 2031 — systems capable of attacking the “big chalkboard problems” that today remain out of reach.

This is the crux of Dial’s timeline. He says that the newer, more efficient error-correction codes have “rewritten what’s possible” in the next decade. The gap between today’s physical qubits and useful logical qubits has not disappeared. Error rates still need to improve by roughly a factor of three to ten on IBM’s best processors, according to Dial. But in his account, the overhead problem has changed enough that IBM can treat a computationally useful error-corrected machine as an engineering target rather than a lifetime-scale question.

Oliver Dial describes IBM’s quantum roadmap as two roadmaps running in parallel. The development roadmap covers devices IBM thinks will be computationally interesting and available to clients. For 2026, 2027, and 2028, that means improved versions of Nighthawk: better gate fidelities, fewer errors, and larger devices. The materials and physics work is aimed at lower error rates. The engineering and systems work is aimed at scaling.

Alongside that is an innovation roadmap for research demonstrations that are not yet computationally useful. In 2026, IBM aims to demonstrate logical qubits. In 2027, it aims to demonstrate connecting two modules of logical qubits, because the target system cannot be one huge monolithic chip. In 2028, it aims to demonstrate universal computation on these error-corrected systems. In 2029, if the roadmaps converge, IBM hopes to bring error-corrected devices to clients.

Dial is careful about the uncertainty. He has been at IBM for more than a decade and says he used to wonder whether he would see an error-corrected quantum computer in his lifetime. Now IBM thinks it may build one large enough to be computationally useful within four years. He calls that both amazing and terrifying.

The “terrifying” part is not mainly social harm. Craig Smith asks whether Dial means the power of quantum computing could have dangerous effects. Dial says he expects the impacts to be “overwhelmingly positive,” especially for materials science. His anxiety is operational. IBM talks internally about cycles of learning: build a device, learn what is needed for the next one, and repeat. Four years is short enough that Dial can count how many learning cycles remain before IBM has to deliver. Science on a schedule is difficult. Things will go wrong. The question is whether too many things go wrong and the 2029 target slips by a year.

That sense of schedule pressure is tied to IBM’s public posture. Dial says IBM publishes and updates its roadmap annually, showing both what it expects and what it has done. The reason, he says, is that the quantum field contains a lot of hype and optimistic projections. IBM’s answer, as Dial presents it, is to say it has mostly stuck to its projections for years, with some changes, and that this track record should make its future targets more credible.

The roadmap is therefore not just a planning document. It is part of Dial’s claim for IBM’s seriousness in a field where many claims are hard to evaluate. His argument is that IBM has crossed utility, is close to advantage, has a plausible error-correction path, and has a public sequence of milestones against which the claim can be judged.

Oliver Dial explains a quantum computer as an accelerator attached to a classical computer. Users begin with classical data, send a classical description of a problem to the quantum device, manipulate the quantum state inside the processor, and then measure the qubits to convert the result back into classical data. Humans, he notes, only have access to classical data at the beginning and end.

The number of qubits determines the size of the quantum state that can be manipulated. Dial describes the machine as doing something like linear algebra in an enormous state space: a complex vector space with 2^n dimensions, where n is the number of qubits. That exponential growth is the reason qubit count matters. Four qubits can be simulated on a phone. Around 50 to 100 qubits, depending on the circuit and simulation method, classical simulation becomes essentially impossible. IBM’s current flagship devices in its fleet, Dial says, have 156 qubits, well past the scale of Google’s 53-qubit supremacy experiment.

But Dial’s account of IBM’s 1,000-qubit Condor machine is meant to undercut simple qubit-count scorekeeping. IBM built Condor “just to prove that we could,” then took it apart almost immediately. Its error rates were not low enough to use all thousand qubits productively. Dial calls it an expensive technology demo, not a machine useful to IBM’s clients.

IBM’s next direction is not merely larger chips. Nighthawk, a processor Dial says is coming online this year, is smaller than Heron by qubit count — 120 qubits rather than 156 — but he says it has significantly better connectivity. Previous IBM devices had at most three connections from each qubit. Nighthawk has four. That sounds incremental, but for quantum circuits it changes how efficiently computations can be laid out. If two qubits that need to interact are far apart, the state has to be moved through intermediate swaps. More connectivity reduces that overhead.

The physical qubits themselves are superconducting resonators. Dial describes them as small LC oscillators: electromagnetic analogues of pendulums built from superconducting metal, capacitors, and Josephson junctions. IBM uses transmon-style qubits, where the two computational states are zero microwave photons and one microwave photon stored in the resonator. The resonator frequency is designed to be about five gigahertz, a microwave frequency. Dial emphasizes that “photon” here means microwave photons in a superconducting circuit, not visible light traveling through free space.

The Josephson junction is critical because it acts like a highly nonlinear inductor. A normal oscillator would have equally spaced energy levels, making it difficult to isolate just two states. In IBM’s qubits, Dial says, the first photon might correspond to five gigahertz, while the second would require a different frequency, such as 4.8 gigahertz. That separation lets the system isolate the lowest two states as the qubit’s zero and one.

IBM uses superconducting qubits partly because they fit the company’s manufacturing base. Dial says they can be made with the same broad tooling, clean rooms, nanofabrication methods, and packaging approaches used for classical digital logic. The output is not a laptop chip: it is a chip that has to be cooled to near absolute zero. But the manufacturing culture and skills are closely aligned with IBM’s strengths.

The qubits are also physically large by semiconductor standards. Dial says IBM’s qubits are about three quarters of a millimeter on a side — large enough to see with the naked eye if the chip were handed to someone. What is strange, in his view, is not that they are tiny, but that entanglement and superposition can be engineered in objects big enough to see.

The difficulty of scaling IBM’s superconducting machines is not one bottleneck. Oliver Dial describes it as a stack of engineering constraints that compound. Each qubit needs microwave control signals. Each tunable coupler needs control. Groups of qubits need measurement wiring. Nighthawk, the 120-qubit device, has on the order of 480 control signals going into and out of the processor.

Those signals are not digital bits. They are analog microwave signals. Each needs a precise, low-noise generator, which Dial compares loosely to a cellphone transmitter but with much more demanding requirements. That is the control electronics layer. Then the signals have to be delivered through wiring into the refrigerated environment without corrupting one another or heating the system.

Dial’s most concrete systems example is the thermal gradient. One end of a wire sits at room temperature, where people live. The other end sits at 0.02 degrees above absolute zero, colder than deep space by his comparison. If the wire conducts too much heat, it undermines the refrigeration. He likens it to trying to air-condition a house with all the windows open. The wiring has to carry signal while shedding heat correctly along its path.

The extreme cold is not primarily to make the Josephson junction superconduct. Dial says the materials IBM uses become superconducting below roughly one degree. The much lower temperature is needed because of thermal photons. Anything warm emits radiation. If the chip is “glowing” at the five-gigahertz frequency of the qubits, it adds unwanted photons and damages the quantum information. The machine is cooled so far down because the qubits must be protected from heat radiation at microwave frequencies.

That systems framing also explains why IBM’s hardware changes quickly. Dial says IBM has two classes of client-accessible quantum computers. One is IBM’s own cloud-accessible data centers, where users can access machines remotely; he says users get 10 free minutes per month. The other is a fleet of systems installed at client sites under managed-service agreements. IBM still owns and operates those machines, keeps them calibrated, and upgrades them. Dial says there are 13 such systems around the world, with four more being installed toward the end of 2026.

These machines typically get a mid-term upgrade after about two years, because the hardware is moving quickly enough that clients will want the latest device by then. IBM is working on multiple versions of the Nighthawk processor in parallel, and Dial says whichever versions look good become the next deployed machines.

The fabrication pipeline has also shifted. Until a few years ago, Dial says, every IBM quantum processor was built in the building where the interview took place, using IBM’s research fabrication facility. More recently, most fabrication moved to the Albany Nanotechnology Center, a 300-millimeter research and pilot-production facility co-run by IBM and NY CREATES. The reason is that the processors are becoming more complicated, with extra layers and more demanding process requirements. Packaging still happens at IBM’s site. Control electronics are designed by people who also design IBM’s System Z systems, though the final build is done by subcontractors. Assembly and test come together at IBM’s local site or in Poughkeepsie.

Dial estimates IBM’s quantum effort as roughly one-third hardware and two-thirds software. That split matters because the roadmap is not just about making qubits less noisy or chips larger. Better compilation, better mapping of problems onto devices, error mitigation, and early forms of error correction all determine how much useful computation can be extracted from a given processor.

Craig Smith presses Dial on whether quantum computers’ relationship to linear algebra makes them relevant to AI, given that modern AI systems depend heavily on matrix operations. Oliver Dial answers cautiously. AI on quantum hardware is still in the realm of toy problems and open questions. Chemistry, materials science, and some optimization work are closer to the frontier of practical demonstrations.

Dial divides applications into near-horizon and long-horizon categories. The near-horizon applications are those that can run, or nearly run, on current quantum hardware. The long-horizon applications require much larger, error-corrected machines. The near-horizon applications he names most clearly are chemistry and materials science: using a quantum computer to solve quantum mechanics. That mapping is natural because the target problem is itself quantum. Dial says some chemistry demonstrations are approaching the accuracy of the most up-to-date classical chemistry techniques. He does not claim broad superiority, but he does describe the area as close.

AI on quantum hardware is much less clear. Current quantum machines are not large enough to represent meaningful AI models or deep enough to run the necessary circuits. There have been proof-of-concept demonstrations around kernel estimation: finding a basis in which information is sparse, a technique relevant to some AI models. If researchers deliberately design a problem with a feature quantum computers are known to recognize efficiently, such as periodicity, quantum machines can win on toy examples. But Dial stresses that these are “totally just toy small problems,” and that no one knows how often such structures appear in real-world AI problems because they would be hard to detect classically in the first place.

His broader category for many near-term quantum approaches is “heuristic algorithms.” Smith asks him to define the term. Dial contrasts algorithms one can prove on a chalkboard with methods that often work in practice but do not come with a proof that they will produce the desired answer. Matrix multiplication and the fast Fourier transform are his clean examples of algorithms whose correctness can be shown mathematically. Shor’s algorithm, the famous quantum factoring algorithm, also belongs in the chalkboard-proof category in Dial’s explanation, even though no one can run it today on a useful-size problem. By contrast, he names beam search, genetic algorithms, gradient descent on neural networks, traveling-salesman solvers, and SAT solvers as heuristic approaches: methods that can make sense and work well in practice without a guarantee that they will return the desired answer.

AI, in Dial’s phrase, is “the king of heuristic algorithms.” That makes quantum AI doubly hard to reason about. The classical baselines are often heuristic. The quantum methods are often heuristic. And without sufficiently large hardware, many claims can only be tested on deliberately constructed examples. Dial says optimization is more encouraging, and he connects it to AI as something like a training phase, but again with the same warning: the best classical optimizers are also heuristic, so quantum work naturally ends up in that same empirical territory.

Dial also discusses famous long-horizon quantum algorithms to place the AI question in context. Shor’s algorithm can factor numbers and is believed to attack a problem hard for classical computers. It is also far beyond today’s hardware for useful problem sizes. Grover search can search an unsorted list in time proportional to the square root of the list size, but Dial argues it is often misunderstood because it is easy to explain. In practice, the overhead is so large that one would likely be better off building a massively parallel classical computer for relevant cases. He says Grover is unlikely to matter technologically in either his or Smith’s lifetime.

Some future quantum linear-algebra routines could in principle offer speedups for extremely large problems, assuming input states can be prepared efficiently. But Dial says the prefactor — the overhead cost of running on a quantum computer, which is slow and expensive relative to classical operations — is so large that the quantum machine would only win for “absolutely enormously large problems.” That world is still far off.

Craig Smith asks whether enterprises should start training people now in quantum algorithms and quantum timelines, so that they have a workforce ready when the machines arrive. Oliver Dial says it is not too early.

His first reason is problem mapping. Every enterprise has its own problems, and the hard step is not merely learning a generic quantum programming interface. It is figuring out whether a real business, scientific, or engineering problem maps to a quantum computer at all — either through a near-term heuristic algorithm or through a provable algorithm that may require an error-corrected machine. People who understand both the enterprise problem and quantum computing well enough to perform that mapping take time to develop.

His second reason is timing. Dial uses IBM’s System/360 as an analogy: if an enterprise could have known four years in advance that a major new computing platform was coming, and could have prepared to use it on day one, that would have been a significant advantage. For organizations already planning on four- and five-year horizons, he says, quantum preparation now makes sense.

Dial also points back to near-term possibilities. He says optimization and chemistry are getting close enough that additional thresholds may be crossed in the next couple of years, though he stresses again that heuristic results cannot be promised. The practical recommendation is not that every enterprise should expect immediate quantum advantage. It is that the time required to build literacy, identify candidate problems, and develop hybrid classical-quantum approaches may be comparable to the time remaining before useful error-corrected machines, if IBM’s roadmap holds.

The central tension is that the field remains uncertain while the preparation window may already be open. Dial’s account does not erase the reasons for skepticism: present machines are noisy, AI applications remain toy-scale, famous algorithms like Shor’s require hardware nobody yet has, and advantage claims depend on verification. But it also gives a concrete reason not to dismiss the timeline as permanently five to ten years away. In Dial’s case for IBM’s roadmap, the error-correction overhead has changed enough to make 2029 an engineering target rather than an indefinite research hope.