Quantum Computing: what are the biggest challenges now?
In our quantum computing market deck, you will find everything you need to understand the market
SUMMARY
Quantum computing’s biggest challenge now is scaling reliable logical computation faster than the cost, hardware overhead and infrastructure needed to create it.
Error correction has moved from a mostly theoretical promise to an engineering problem. Google, Quantinuum and neutral-atom teams have all shown meaningful logical error suppression, but useful machines still need those gains to survive vastly longer computations.
The raw number of qubits is becoming a weaker way to judge progress. A smaller system with cleaner physical qubits, lower logical error rates and efficient codes can be more useful than a much larger noisy processor.
The physical-to-logical overhead remains brutal in some architectures. Google’s distance-7 surface-code memory used 101 physical qubits for one logical memory experiment, while other platforms have shown much denser logical encoding under different conditions.
The scaling problem is increasingly about the whole machine rather than the qubit alone. Cooling, wiring, lasers, optical systems, calibration, real-time decoding and classical control could become bottlenecks at the same time as the quantum hardware itself.
No architecture has clearly won. Superconducting qubits, trapped ions, neutral atoms, photonics and silicon approaches are all making credible progress, and DARPA’s continued support for many competing designs shows how unsettled the field still is.
Roadmaps deserve more attention than they did a few years ago, but only when they expose testable intermediate steps. IBM’s fault-tolerance milestones and DARPA’s risk-retirement process are more informative than a distant claim about a future million-qubit machine.
Near-term commercial advantage remains much less convincing than the hardware progress. Quantum systems can dominate carefully chosen benchmarks, but broad end-to-end economic advantage over the best classical alternative has still not been demonstrated.
Chemistry and materials science currently look like the strongest application candidates because the problems themselves are quantum mechanical. Generic optimization and AI face a harder test because classical solvers, GPUs, heuristics and machine-learning methods keep improving too.
The economic challenge may become as important as the technical one. Quantum systems are starting to require infrastructure that looks more like specialized data centers or semiconductor-scale projects, while most customers are still paying for access, experiments and strategic preparation rather than proven production savings.
The field is therefore more credible and more demanding at the same time. Below-threshold error correction, logical processors and modular architectures are real; the decisive test now is whether those pieces can be combined into machines that run useful workloads deeply, reliably and cheaply enough to beat classical computing where it counts.
This market map, featured in our quantum computing market deck, highlights top companies and startups in the quantum computing market
Why does quantum computing look more credible now?
Quantum computing looks much more credible today because several hardware teams have finally shown pieces of fault-tolerant computing that researchers had spent years promising but could not demonstrate at convincing scale.
Google's Willow result was one of the clearest turning points. In its surface-code experiment, Google increased the size of an encoded qubit from distance 3 to distance 5 and then distance 7, and the logical error rate fell each time. The reported suppression factor was about 2.14 whenever the code distance increased by two. That “below-threshold” behavior is essential: if adding more physical qubits makes the encoded information more reliable, error correction can theoretically keep improving as the machine grows.
Other teams have reached different milestones. Quantinuum has demonstrated logical operations that outperform their underlying physical operations and fault-tolerant teleportation with very high fidelity. Harvard, QuEra, MIT and NIST researchers have operated algorithms on as many as 48 logical qubits using neutral atoms. Microsoft and Atom Computing have demonstrated 24 entangled logical qubits and computations involving 28 logical qubits.
The newest roadmaps are also becoming more concrete. IBM currently plans Starling, a modular machine designed for 200 logical qubits and 100 million gates, while its newer Nighthawk hardware is being used to bridge the gap between today's noisy processors and fault tolerance. DARPA has moved 11 different architectures into the deeper Stage B of its Quantum Benchmarking Initiative and says it now looks likely that somebody can build a utility-scale quantum computer by 2033, although it still cannot say which architecture will get there.
We have enough experimental evidence now to take large-scale quantum computing seriously, even though the engineering problem is still far from solved.
Is quantum error correction still the biggest problem?
Yes. Quantum error correction is still the biggest challenge in quantum computing because every ambitious application eventually depends on running far more operations than today's physical qubits can survive.
Physical quantum operations remain extraordinarily unreliable compared with conventional digital logic. Error rates around one failure in every thousand operations can be excellent by quantum-hardware standards. They are disastrous if an algorithm needs millions, billions or eventually trillions of operations.
Google has shown that suppression works experimentally. Quantinuum has shown that encoded operations can beat physical ones. Neutral-atom teams have demonstrated logical processing across dozens of encoded qubits. Those results settle an important scientific question: useful error suppression is physically achievable.
The remaining problem is the cost. Google's distance-7 surface-code memory used 101 physical qubits for one logical memory experiment. Other architectures can achieve much lower ratios because their connectivity, native fidelity and error-correcting codes differ, but a logical qubit count tells us very little unless we also know its error rate and what operations can be performed on it.
A commercially useful machine also has to keep logical qubits alive while performing gates, measurements, state preparation, communication between modules and real-time decoding. Error-correction overhead therefore still sits at the center of the scaling problem.
As this chart shows, and as featured in our quantum computing market deck, search interest in quantum computing has grown significantly
Why can one good logical qubit require so many physical qubits?
A reliable logical qubit can require dozens or hundreds of physical qubits because quantum computers have to spend hardware continuously correcting the errors created by the hardware itself.
The exact ratio varies enormously. Google's distance-7 surface-code memory used 101 physical qubits to protect one logical qubit. Microsoft and Quantinuum, using a very different trapped-ion architecture and error-correcting code, demonstrated 12 logical qubits on Quantinuum's 56-qubit H2 processor. Neutral-atom experiments have encoded dozens of logical qubits inside arrays containing a few hundred atoms.
Those numbers cannot be compared as if they were RAM capacities. The logical qubits were protected to different strengths and used for different operations. A logical qubit that survives a short experiment may still be nowhere near reliable enough for an algorithm containing 100 million gates.
The required protection also grows with the computation. Suppose an algorithm is allowed only a 1% chance of failure across one billion logical operations. Very roughly, its average logical error per operation needs to land many orders of magnitude below what today's physical gates deliver.
The cryptography estimates show how extreme the overhead can become. A widely cited calculation by Craig Gidney and Martin Ekerå estimated that factoring a 2,048-bit RSA integer in about eight hours could require around 20 million noisy physical qubits under their assumptions. Better algorithms, higher fidelities and new codes can reduce that figure substantially, so 20 million should never be treated as a fixed requirement.
| Example | Physical-to-logical picture | What we learn from it |
|---|---|---|
| Google surface-code memory | 101 physical qubits for one distance-7 logical memory | Stronger protection still carries heavy overhead |
| Microsoft + Quantinuum | 56 physical qubits supporting 12 demonstrated logical qubits | Better fidelity and connectivity can cut that overhead |
| Neutral-atom experiments | Dozens of logical qubits from a few hundred atoms | Reconfigurable architectures can encode efficiently |
| RSA-2048 resource estimate | Around 20 million noisy qubits under modeled assumptions | Very long algorithms remain far beyond today's machines |
Has Google actually solved quantum computing errors?
No. Google's Willow work solved one crucial part of quantum error correction, while a useful error-corrected computer still requires several other pieces to work together for millions of operations.
The big achievement was reaching the regime where enlarging Google's surface code made the logical qubit more reliable. For years, researchers knew that should happen mathematically once hardware crossed the error-correction threshold. Demonstrating it on an actual processor removed one of the deepest doubts about the superconducting route.
A complete fault-tolerant computer has to protect quantum memory continuously while also performing logical gates, creating entanglement between logical qubits, measuring them, moving information around the machine and decoding error syndromes quickly enough to influence what happens next.
Universal computation adds another difficult layer. Many leading error-correcting codes handle a limited set of logical operations relatively cheaply, while other operations require resources such as magic states.
IBM plans to demonstrate logical memory and processing, real-time decoding, links between fault-tolerant modules and eventually magic-state distillation before Starling. Neutral-atom researchers have already demonstrated pieces of logical entangling operations and magic-state preparation. Quantinuum has pushed real-time conditional control and fault-tolerant teleportation.
Google's result is a genuine breakthrough, but it still covers only part of the fault-tolerant computing stack.
This chart, included in our quantum computing market deck, illustrates yearly VC funding for quantum computing startups
Can quantum computers really scale to millions of qubits?
Quantum computers can probably reach very large physical qubit counts, but we still do not know whether anybody can manufacture, control, cool, calibrate and error-correct millions of qubits cheaply enough to make the resulting machine useful.
Superconducting processors need dilution refrigerators operating only a few thousandths of a degree above absolute zero, together with microwave lines, amplifiers, filters, control electronics and readout hardware. A laboratory can tolerate messy wiring for 100 qubits. Scaling the same arrangement literally to one million qubits would be absurd.
IBM is attacking that problem with modular cryogenic infrastructure, long-range couplers and increasingly dense control electronics. Its System Two architecture already treats the refrigerator, classical runtime servers and modular quantum processors as one computer rather than one isolated chip.
Trapped ions avoid millikelvin refrigeration, but large systems need complicated optical control, ion transport and many stable laser channels. Neutral atoms can arrange very large arrays and move atoms with optical tweezers, which gives them an attractive scaling route, although atom loss, laser stability and gate errors remain difficult. Photonic architectures can use optical networking naturally, but they have to fight photon loss, generate enormous numbers of high-quality states and coordinate huge optical circuits.
Manufacturing makes the problem harder. A process that produces 99.9% good components sounds excellent until we apply it naïvely to one million components and get roughly 1,000 defects. Real quantum architectures can route around defects and incorporate redundancy, but fabrication quality still has to be extraordinary.
Calibration also has to become highly automated. Today's processors regularly need qubits, gates and readout channels characterized and retuned. Error correction then creates continuous streams of syndrome measurements that classical decoders must process quickly enough for the computation to continue.
PsiQuantum shows how different the end state could look from today's prototypes. Its strategy assumes fault-tolerant machines containing roughly a million physical photonic qubits installed in purpose-built facilities. That resembles a specialized data center more than a laboratory experiment.
Which quantum computing technology is actually winning?
No quantum computing architecture is clearly winning today, and the strongest evidence comes from the industry's own behavior: even the best-funded organizations are still keeping multiple hardware paths alive.
Superconducting qubits have one of the strongest track records. Google and IBM have demonstrated increasingly complex processors, fast gates and important error-correction results using technology that benefits from semiconductor-style fabrication. Their weaknesses include cryogenic complexity, relatively local connectivity and the engineering burden of routing control signals into very cold environments.
Trapped ions give Quantinuum and IonQ exceptionally clean qubits with flexible connectivity. That makes efficient error-correcting codes possible and helps explain why Quantinuum can create many logical qubits from relatively few physical ones. Gate operations tend to be slower, and scaling ion traps into very large modular machines creates its own transport, optics and manufacturing problems.
Neutral atoms are improving unusually quickly. They offer large naturally identical qubit arrays, flexible geometry and massive parallelism. Google recently made a revealing move by adding neutral-atom computing to Google Quantum AI while continuing its superconducting program.
Photonic companies such as PsiQuantum and Xanadu take another route. Photons are excellent for communication, potentially making modular networking easier, and photonic components can use semiconductor manufacturing. The challenge shifts toward optical loss, single-photon sources, detectors, switching and the very large number of components required for fault tolerance.
Silicon spin qubits, bosonic codes and topological approaches add still more possibilities. Microsoft's Majorana work remains especially ambitious because topological protection could dramatically reduce error-correction overhead if the underlying physics and devices reach the required standard.
DARPA currently has 11 Stage B teams spread across neutral atoms, superconductors, trapped ions, photonics and several flavors of silicon qubits. Two additional approaches from its earlier program have reached Stage C validation.
| Architecture | Big advantage today | Main scaling headache |
|---|---|---|
| Superconducting | Fast gates and mature fabrication | Cryogenics, wiring, connectivity and error rates |
| Trapped ions | Very high fidelity and strong connectivity | Speed, optics and large-system engineering |
| Neutral atoms | Huge arrays, parallelism and flexible geometry | Atom loss, laser control and gate fidelity |
| Photonics | Networking and semiconductor compatibility | Photon loss and very large component counts |
| Silicon spins | Tiny devices and semiconductor manufacturing potential | Uniform high-fidelity control at large scale |
If you want more recent data on this point, please see our latest quantum computing market report.
This chart, included in our quantum computing market deck, looks at IonQ’s strategy in quantum computing
Can we trust the big quantum computing roadmaps?
We can take quantum computing roadmaps more seriously now, but demonstrated intermediate milestones still deserve much more trust than the final numbers printed at the end.
IBM currently has one of the easiest roadmaps to test because it exposes several steps before its big claim. Starling is planned as a 200-logical-qubit machine capable of 100 million gates. Before that, IBM says it needs real-time error decoding, logical memory and processing, modular entanglement and magic-state distillation.
IBM has also pushed the noisy-computing side forward with Nighthawk r2. The new processor keeps 120 programmable qubits but dramatically increases throughput: IBM says it can execute more than 100,000 circuits per second, about 25 times the rate of previous Heron systems, while reducing reset time to around one microsecond. It is designed to support circuits beyond 7,500 gates.
Other roadmaps are harder to judge. IonQ has published targets running into thousands of logical qubits before the end of the decade. PsiQuantum is building directly toward million-physical-qubit utility-scale machines. Quantinuum's plan relies heavily on keeping physical fidelities extremely high while scaling modular trapped-ion hardware.
DARPA's Quantum Benchmarking Initiative requires teams to submit full utility-scale system designs, identify the risks that could break them and build prototypes intended to retire those risks. DARPA currently says a useful machine by 2033 looks likely, while remaining unsure who will build it.
Are quantum computers simply too slow?
Quantum computers are still too slow for many serious fault-tolerant workloads because logical qubit count means little if each reliable operation takes too long.
Imagine two future machines with 500 logical qubits. One completes 100 million logical operations in minutes while the other needs several days. They technically have the same number of logical qubits, yet their commercial usefulness is completely different.
Physical gate speed already varies widely by architecture. Superconducting qubits can perform very fast operations, often on nanosecond or microsecond timescales. Trapped-ion entangling gates are generally slower but achieve excellent fidelity. Neutral atoms try to compensate partly through parallel operations across many atoms.
Fault tolerance adds extra work around every useful operation. Error-correction cycles have to run repeatedly. Classical decoders process syndrome data. Some logical gates require complicated protocols, ancilla qubits or distilled magic states.
The gap between current circuits and ambitious algorithms is still enormous. IBM's Nighthawk program is pushing circuits beyond several thousand gates today, while Starling is designed around 100 million fault-tolerant gates. Moving from roughly 10,000 operations to 100 million represents four orders of magnitude in circuit depth.
If you want more recent data on this point, please see our latest quantum computing market report.
This chart, included in our quantum computing market deck, illustrates yearly funding for quantum computing startups
Can quantum computers do anything useful before full fault tolerance?
Maybe, but this remains one of the weakest parts of the commercial quantum story: nobody has yet shown a broad, economically valuable workload where today's noisy quantum computer clearly beats the best classical alternative end to end.
There are impressive demonstrations. Google's Willow completed a random-circuit-sampling benchmark in less than five minutes that Google estimated would take a leading classical supercomputer an astronomically longer time. That demonstrates an extreme computational separation on a carefully chosen problem.
Companies cannot use random-circuit sampling to design a drug or optimize their supply chain.
That distinction has repeatedly caught the industry. A quantum machine produces an impressive result on an optimization benchmark, then a stronger classical algorithm or more specialized hardware shrinks the gap. The comparison often becomes even harder once we include data preparation, repeated sampling, error mitigation and the cost of accessing the quantum hardware.
The best near-term case may come from quantum simulation. Chemistry and materials problems are themselves quantum mechanical, so the mapping between problem and machine is much more natural than forcing an ordinary business optimization problem onto qubits. Researchers are increasingly studying hybrid methods where classical HPC handles most of the workflow and the quantum processor attacks a particularly difficult part.
IBM is pushing that route aggressively. Its current roadmap explicitly targets quantum advantage using quantum hardware combined with high-performance computing before Starling reaches full fault tolerance.
Until a quantum system beats a first-rate classical system on a useful workload after every relevant cost is counted, pre-fault-tolerant commercial advantage remains unproven.
Why is it so hard to find useful quantum algorithms?
Useful quantum algorithms are rare because quantum computers only win when a problem contains structure that quantum mechanics can exploit; most ordinary computing tasks do not suddenly become faster when we put them on qubits.
Shor's factoring algorithm is famous because its advantage is unusually strong. Grover's algorithm gives a quadratic improvement for unstructured search. Quantum simulation is compelling because a quantum system can represent another quantum system in a way that becomes extremely expensive for classical computers as complexity grows.
Optimization is a good example of the limitation. Logistics, scheduling, portfolio optimization and manufacturing problems appear constantly in quantum marketing. Researchers can certainly map versions of them onto quantum machines. The important question is whether the resulting quantum method beats highly optimized classical solvers on realistic instances.
The classical side keeps getting better. GPUs have become vastly more capable. Mathematical solvers improve continuously. AI can generate better heuristics. Specialized accelerators can attack certain optimization problems very efficiently.
Chemistry and materials science currently look stronger because exact simulation of interacting quantum systems can become exponentially difficult for classical computers. Even there, quantum computers compete against density-functional theory, tensor-network methods, Monte Carlo techniques and machine-learned approximations that are improving too.
A handful of extremely valuable problems that remain stubbornly beyond classical computing could still be enough to support a major quantum industry.
If you want more recent data on this point, please see our latest quantum computing market report.
This chart, included in our quantum computing market deck, compares the main business model options for quantum computing hardware startups
Is quantum computing becoming too expensive for the customers it needs?
Quantum computing is becoming extremely expensive while customers are still mostly paying for experiments, access and future readiness rather than proven production savings.
The cost profile is already visible. Large quantum programs require custom chips, specialist fabrication, cryogenic equipment or precision lasers, complex control electronics, huge research teams and facilities that can take years to build.
PsiQuantum is perhaps the clearest example because its strategy depends on utility-scale photonic systems installed in dedicated sites rather than relatively small laboratory processors. IBM is also building toward modular fault-tolerant systems that combine quantum processors, large refrigerators, conventional servers and dedicated control hardware.
Public investment has followed. Governments in the United States, Australia, Japan, Europe and elsewhere are putting substantial amounts of money into quantum fabs, research centers and national programs. McKinsey estimated that quantum-technology startups attracted roughly $2 billion in investment in 2024, followed by a dramatically larger funding year as capital flowed into several major companies and financing rounds.
Commercial spending is real as well. IBM says its quantum program has generated more than $1.1 billion in client contracts since 2017. BCG has found quantum activity particularly concentrated in finance, insurance, pharmaceuticals, healthcare and industrial companies.
Yet those customers are usually paying for cloud access, research collaborations, algorithm development, pilot projects and strategic preparation. We still rarely see a company saying that a quantum computer has taken a production workload costing $10 million on classical hardware and reduced it to $1 million.
That's the central economic problem. The machines are starting to require semiconductor-scale capital while the customer value still looks mostly like R&D spending. DARPA's definition of “utility scale” captures the hurdle neatly: the computational value produced has to exceed the computational cost.
Is quantum computing still full of hype?
Yes, quantum computing still has a hype problem because companies can choose from too many impressive-looking metrics that tell us surprisingly little about whether their machines can solve useful problems.
Raw physical qubit count is the easiest example. A processor with 1,000 noisy qubits may be less useful than one with 100 much cleaner qubits. Logical qubit count sounds better but creates another problem: logical qubits encoded with different codes can have radically different error rates and capabilities.
Gate fidelity can be cherry-picked as well. Excellent single-qubit fidelity means less if two-qubit operations or measurements are substantially worse. Quantum volume combines several properties but still does not tell a customer whether a specific fault-tolerant algorithm can run economically.
Roadmaps make comparisons messier. IBM talks about 200 logical qubits and 100 million gates for Starling. IonQ has published targets reaching thousands of logical qubits. PsiQuantum talks in terms of roughly a million physical qubits because its architecture and error-correction strategy are completely different. Putting those three numbers next to each other without context tells the reader almost nothing.
DARPA's benchmarking effort is useful because its evaluators ask companies to model the complete industrial machine, expose their technical assumptions, identify the prototypes still needed and eventually submit the design to independent validation. As pointed out above, 11 architectures have reached Stage B.
| Popular claim | Why we should be careful | More useful question |
|---|---|---|
| “1,000 qubits” | Says little about qubit quality | How many reliable logical qubits can those physical qubits create? |
| “99.9% fidelity” | May describe only one type of operation | What failure rate does the full algorithm experience? |
| “50 logical qubits” | Protection strength may differ greatly | How long can they compute and what gates can they perform? |
| “Quantum advantage” | Benchmark may have little commercial value | Does it beat the best classical method on a useful problem? |
| “Million-qubit roadmap” | Future engineering target | Which required milestones already work experimentally? |
If you want more recent data on this point, please see our latest quantum computing market report.
This chart, featured in our quantum computing market deck, illustrates how revenue is divided among customer segments in the quantum computing market
Could the infrastructure around quantum computers become the real bottleneck?
Yes. The control electronics, decoding hardware, cooling and power systems around quantum computers could become as difficult to scale as the qubits themselves.
Small processors hide the problem. Engineers can dedicate expensive microwave electronics, lasers, cables and calibration equipment to relatively few qubits. Million-qubit systems have to compress and automate nearly everything.
Error correction adds an enormous classical workload. Every correction cycle creates measurements that have to be interpreted as patterns of likely errors. A decoder must process them fast enough that the quantum processor can keep moving through the computation.
Google has explored machine-learning decoders, including AlphaQubit, to improve error detection. IBM's latest roadmap explicitly includes real-time decoding as a milestone before Starling. Quantinuum has increasingly integrated classical compute and GPU resources alongside its trapped-ion machines.
Superconducting systems also have a difficult energy problem because qubits operate at millikelvin temperatures. Removing heat at those temperatures is extremely inefficient. Trapped ions and neutral atoms shift more of the burden toward lasers and optical systems, while photonic processors need sources, switching, detectors and sometimes cryogenic components.
The relevant comparison will eventually be energy per solved problem. A quantum computer consuming megawatts could still make economic sense if it finishes in an hour a materials simulation that would occupy a giant classical supercomputer for months. We simply do not have machines large enough yet to measure that trade-off properly.
What could make quantum computing fail even if the hardware works?
Quantum computing could succeed technically and still become a smaller industry than expected if only a narrow group of problems gains enough quantum advantage to cover the cost of these machines.
Imagine researchers build a reliable processor with several hundred logical qubits. Its business impact would depend entirely on what those logical qubits can do.
One possibility is that quantum chemistry and materials simulation work extremely well while most optimization, machine-learning and financial applications remain better on classical hardware. The resulting market could still be valuable, especially in pharmaceuticals, chemicals, batteries and advanced materials, but it would look much narrower than the “quantum changes all computing” story.
Runtime could shrink the market further. An algorithm may provide a beautiful mathematical speedup and still require so many logical operations that one useful calculation remains painfully expensive.
Classical approximation is another threat. Companies rarely need a mathematically exact global optimum if a classical method can produce a 99.5%-good answer in seconds. AI makes this competition stronger because many businesses care about useful predictions or designs rather than exact solutions.
Quantum computing therefore still needs workloads valuable enough to make an expensive fault-tolerant machine worth running.
If you want more recent data on this point, please see our latest quantum computing market report.
This chart, included in our quantum computing market deck, shows how cloud quantum computing access technology has evolved over time
What are the biggest quantum computing challenges now?
The biggest quantum computing challenge right now is scaling reliable logical computation faster than the cost and complexity required to create it.
Error correction comes first. Recent experiments give us much more confidence that logical errors can be suppressed, but useful algorithms need protection across far longer computations than researchers have demonstrated.
System engineering comes immediately after it. Thousands or millions of qubits have to be manufactured, controlled, calibrated, connected and decoded without turning every new qubit into another infrastructure problem. Google's decision to pursue neutral atoms alongside superconducting hardware, IBM's increasingly modular architecture and DARPA's continued support for 11 very different Stage B approaches all show that nobody has settled the scaling question yet.
The application gap is just as important. Quantum hardware has produced impressive scientific benchmarks, while economically useful advantage remains scarce. Chemistry and materials currently offer the clearest route because the underlying problems are inherently quantum mechanical. Generic optimization and AI applications face much tougher classical competition.
Then comes economics. These machines are starting to look like semiconductor fabs and specialized data centers, yet customers still mostly pay for access, research and preparation rather than proven production savings.
Below-threshold error correction is real. Logical processors are real. Modular fault-tolerant architectures are being engineered rather than merely sketched. DARPA now considers a utility-scale machine by 2033 likely enough to justify full independent validation programs.
For now, raw qubit records tell us less than they used to. The companies worth watching are the ones driving down logical error rates, increasing useful circuit depth, cutting physical overhead and showing that complete workloads can beat increasingly strong classical machines.
OUR METHODOLOGY
This analysis treats “Quantum Computing: what are the biggest challenges now?” as a multidimensional evidence problem. We broke the question into the factors that determine whether quantum computers can become reliable, scalable, useful and economically viable: error correction, physical-to-logical overhead, system scaling, architecture trade-offs, runtime, applications, infrastructure and economics.
For each dimension, we prioritized the freshest evidence that could materially change the assessment. Experimental papers were given the most weight for claims about what has actually been demonstrated, technical documentation for system performance and engineering progress, and concrete intermediate milestones for company roadmaps.
We did not treat physical qubits, logical qubits, fidelity, circuit depth or runtime as directly comparable across architectures. Those metrics can describe very different levels of protection and capability, so cross-architecture comparisons were used mainly to understand trade-offs, overhead and scaling strategies rather than to build a simple leaderboard.
Roadmaps were judged by the milestones that can be tested before the final target. A distant claim about logical-qubit count or utility scale carried less weight than a sequence involving logical memory, real-time decoding, modular entanglement, magic-state preparation or other intermediate engineering steps that can be independently checked.
We also separated benchmark performance from economic usefulness. A scientific demonstration was treated as evidence for the problem it actually solved, while claims of quantum advantage were judged more strictly: the relevant question is whether a quantum system can beat a strong classical alternative on a useful workload after runtime, sampling, data preparation, error mitigation and access costs are included.
The final conclusions came from aggregating evidence across these dimensions and looking for convergence. Where several independent developments pointed in the same direction, such as the growing credibility of logical error suppression, we drew a stronger conclusion. Where the field remains fragmented, especially around the winning hardware architecture and near-term commercial advantage, the assessment stays more cautious.
Key sources used in this analysis include Google Quantum AI and Nature for Willow and below-threshold surface-code error correction; Quantinuum for logical qubits, logical teleportation and fault-tolerant control; Nature for the Harvard, QuEra, MIT and NIST neutral-atom logical-processor work; Microsoft and Atom Computing for logical-qubit demonstrations; IBM for Starling, Nighthawk r2, System Two and its fault-tolerance roadmap; DARPA for the Quantum Benchmarking Initiative, Stage B architecture selection and utility-scale evaluation framework; PsiQuantum and IonQ for architecture and scaling roadmaps; Nature for AlphaQubit; Craig Gidney and Martin Ekerå for the RSA-2048 resource estimate; and BCG and McKinsey for enterprise activity, investment and commercialization context.
In our quantum computing market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Quantum Computing: what are the biggest unsolved problems?
- Quantum Computing: what is getting real adoption now?
- Quantum Computing: what is actually working now?
- What are the latest funding developments in quantum computing?
