DARPA’s Most Rigorous Quantum Evaluation Is Led by Its Least Proven Contenders

Introduction

The current scoreboard of DARPA’s Quantum Benchmarking Initiative looks like it was assembled by someone who mixed up the columns.

In Stage C, the program’s final verification-and-validation phase, sit PsiQuantum and Microsoft. Neither company has publicly demonstrated a logical qubit on the hardware architecture DARPA is evaluating. Microsoft has demonstrated logical qubits using Quantinuum and Atom Computing machines, but its own topological qubit claim is under active challenge in Nature. In Stage B, still working through a year-long R&D audit, sit Quantinuum (some of the strongest public logical-qubit results in the field), IBM (three quantum-advantage papers in a single month), QuEra (whose research collaboration has demonstrated programmable logical-qubit circuits, with its Libra fault-tolerant system planned for 2028), IonQ, and seven others. The companies with the strongest public hardware evidence are a stage behind the companies with the weakest.

That outcome is not a flaw in DARPA’s process. But understanding why requires tracing three years of institutional history that most reporting on QBI glosses over.

Two Programs, Not One

The critical fact: PsiQuantum and Microsoft are in Stage C because they were in a different, earlier program. They did not leapfrog IBM and Quantinuum through the same funnel. They were never in the same funnel.

In February 2022, DARPA announced the Underexplored Systems for Utility-Scale Quantum Computing program, or US2QC. The name said it all. Joe Altepeter, the program manager, told Breaking Defense in February 2025 that DARPA’s mission was to “reduce the danger of strategic surprise from underexplored quantum computing systems.” The emphasis on underexplored was the design constraint, not an accident. DARPA was deliberately screening for approaches that the broader industry was not rigorously evaluating.

In January 2023, DARPA selected three companies for US2QC’s initial phase: Microsoft, PsiQuantum, and Atom Computing. A fourth company, the photonics startup QC82, received a separate preliminary award later that year. Each presented design concepts to a government evaluation team drawn from the Air Force Research Laboratory, Johns Hopkins APL, Los Alamos, Oak Ridge, and NASA Ames. Over the next year, the team assessed whether each company’s approach had a plausible path to a fault-tolerant, utility-scale system.

Two survived. In January 2024, PsiQuantum announced it had advanced to US2QC Phase 2; Microsoft advanced alongside it. Atom Computing and QC82 were eliminated.

Then the situation changed. In July 2024, DARPA expanded US2QC into the Quantum Benchmarking Initiative, a much broader program open to all comers. QBI launched with three stages (A, B, C) and began accepting proposals from companies using any qubit modality. By April 2025, 18 companies had entered Stage A, including IBM, Quantinuum, IonQ, QuEra, and Xanadu. Google Quantum AI joined Stage A later, in September 2025.

Here is where the institutional history matters. When QBI launched, PsiQuantum and Microsoft were already in US2QC Phase 2. DARPA explicitly stated that “the final Phase of US2QC has the same technical goals as Stage C of QBI.” The two US2QC survivors were grandfathered into QBI’s final stage.

Meanwhile, the QBI entrants started at Stage A in spring 2025. Eleven advanced to Stage B in November 2025. They are now in a year-long R&D audit, with Stage C advancement decisions expected in late 2026. Even if IBM or Quantinuum advances to Stage C on the fastest possible timeline, the US2QC survivors will have begun government evaluation roughly two years earlier.

The head start is real. It is also fair, in the sense that US2QC’s design intended it. But anyone who reads QBI Stage C status as a verdict on which quantum technology is most mature is reading the wrong signal entirely.

What “Underexplored” Actually Meant

DARPA did not pick PsiQuantum and Microsoft because they had the best qubits. DARPA picked them because they had approaches that nobody else was rigorously evaluating.

The question US2QC asked was specific and unusual: “Is there an underexplored approach to quantum computing capable of achieving utility-scale operation much faster than conventional predictions?” Altepeter framed it in terms of DARPA’s core mission: “DARPA’s mission is to create and prevent strategic surprise. If there’s an underexplored area of quantum computing showing promise for a faster breakthrough than we previously expected, we want to explore it immediately.”

This filter created a selection bias that is important to understand. By definition, the most well-demonstrated approaches were excluded. Superconducting transmon qubits (IBM, Google) and trapped-ion systems (Quantinuum, IonQ) already had large corporate R&D programs, government lab partnerships, and extensive academic scrutiny. Whether or not those approaches would reach utility scale first, they were emphatically not underexplored. DARPA was looking elsewhere.

Both approaches that survived US2QC shared a specific profile: theoretical architectures promising dramatic scaling advantages, backed by less public experimental evidence than the established modalities. PsiQuantum’s photonic approach offered room-temperature-compatible photonic processing (coupled to cryogenic superconducting photon detectors), compatibility with existing semiconductor fabs, and native optical networking between chips, but had produced minimal public results. Microsoft’s topological approach promised hardware error rates orders of magnitude below conventional qubits (~10⁻⁶ versus ~10⁻³), which would drastically reduce the physical-qubit overhead for error correction, but the underlying physics was (and remains) contested.

This is not accidental. If you ask “show me the approaches nobody is thoroughly testing, with credible but unverified claims about performance at scale,” you will get companies with less demonstrated hardware. The filter selects for exactly the gap that critics point to.

The Atom Computing Puzzle

The most instructive data point in the US2QC story is the company that got cut.

Atom Computing entered US2QC with a functioning neutral-atom quantum computer. During the US2QC evaluation period, in October 2023, Atom Computing announced a 1,225-site neutral-atom array populated with 1,180 qubits, among the highest physical-qubit counts in any platform at the time. They had demonstrated basic quantum operations, published results, and had a credible team. They were also the only US2QC performer whose technology was not truly “underexplored” in any conventional sense: by 2023, neutral atoms were receiving substantial attention from QuEra, Pasqal, and academic labs worldwide.

DARPA eliminated them anyway.

This tells you something about what the evaluation was actually measuring. Having qubits today did not earn a passing grade. Atom Computing’s architecture, while functional, apparently did not convince DARPA’s IV&V team that it offered a path to utility-scale operation that was both faster than conventional predictions and truly underexplored.

Atom Computing has since re-entered through QBI Stage A and advanced to Stage B as of November 2025, now on the broader QBI track alongside its competitors. But the US2QC rejection is revealing. DARPA’s evaluation criteria were not “show us your best hardware.” They were “convince us your approach can reach a utility-scale system by 2033 faster than conventional roadmaps suggest, and that nobody else is seriously stress-testing that claim.”

QC82, the other company eliminated from US2QC, had proposed a photonic approach using room-temperature photon detection. Even less is publicly known about why their design did not survive DARPA’s scrutiny.

The Evidence Gaps

If the procedural history explains the head start, it does not settle the question of whether PsiQuantum’s and Microsoft’s approaches deserve the status DARPA has given them. The public evidence is thin for both.

PsiQuantum

PsiQuantum has never publicly demonstrated a logical qubit or a small-scale error-corrected computation. Its February 2025 Nature paper reported impressive component-level results: dual-rail photonic qubits with 99.98% state preparation and measurement fidelity, two-qubit fusion with 99.22% fidelity, and a chip-to-chip qubit interconnect with 99.72% fidelity. These are real numbers on real silicon photonic chips manufactured by GlobalFoundries. They are also conditional on successful photon detection and do not account for photon loss, which in a photonic architecture is one of the principal determinants of fault-tolerance overhead, not a peripheral caveat.

These remain component-level benchmarks, not an integrated system running quantum error correction. No public result demonstrates a below-threshold logical qubit or a fault-tolerant subsystem at useful scale. PsiQuantum has avoided the conventional path of building intermediate NISQ products, orienting its architecture toward direct utility-scale construction, which widens the gap between component performance and system-level proof relative to companies building incrementally.

I covered PsiQuantum’s $125 million expanded DARPA QBI Stage C agreement in July 2026. That agreement buys verification, not qubits: DARPA is funding component tests, design audits, and resource-estimate checks. In my CRQC Quantum Capability Framework, PsiQuantum scores well on manufacturability (commercial semiconductor foundry manufacturing processes) and connectivity (optical interconnects between modules). It carries public evidence gaps on quantum error correction and every demonstrated capability above the component level.

Microsoft

Microsoft’s evidence gap is more fundamental. Microsoft has demonstrated logical qubits on partner hardware (four on Quantinuum’s H2 in April 2024, twelve in a subsequent Quantinuum collaboration, and 24 entangled logical qubits on Atom Computing hardware in November 2024). But those results used Microsoft’s error-correction software on someone else’s qubits. The question for DARPA’s evaluation is whether Microsoft has demonstrated a topological qubit on its own hardware. That remains unresolved.

Henry Legg, a physicist at the University of St Andrews, published a formal critique in Nature’s “Matters Arising” in June 2026, arguing that Microsoft has not demonstrated the basic physics needed for even a single topological qubit. According to Legg, the signals Microsoft attributes to its device may be consistent with quantum dot effects rather than genuine topological states. The critique targets previously unpublished transport data underlying Microsoft’s results, arguing these data fail to show clear evidence of the superconducting gap required to support a topological qubit claim. Legg also identified what he described as coding errors in Microsoft’s Topological Gap Protocol, including a hardcoded filter and an array reversal, which Microsoft was forced to address point by point in a formal Nature reply published alongside the critique.

This is not Legg’s critique alone. As I covered in my analysis of expert reactions to Microsoft’s Majorana claims, skepticism runs deep. At the March 2025 APS meeting, physicists were broadly skeptical of Microsoft’s claims. John Preskill noted publicly that Microsoft had described a protocol for demonstrating a topologically protected qubit but had provided no publicly available evidence that the test had been conducted successfully. Microsoft has a specific history here: a Microsoft-affiliated research team’s 2018 Nature paper claiming evidence of Majorana zero modes was retracted in 2021 after problems were identified in the analysis and presentation of the data.

Microsoft has since unveiled the Majorana 2 chip (June 2026), with improved parity lifetimes and some new supporting data. Some researchers are positive about specific results, such as probing the nonlocal properties of the Majoranas. But the Majorana 2 preprint reports Z-basis parity readout on the new platform without experimentally demonstrating the complementary X-basis measurement needed for a complete qubit-level case. The debate is ongoing and the scientific community is divided, though the skeptics hold the stronger position by the standard metric of what has been independently demonstrated.

Microsoft is hedging in practice. Its partnership with Atom Computing to deliver the Magne system (50 logical qubits from roughly 1,200 neutral-atom physical qubits) to Denmark by early 2027 is a pragmatic acknowledgment that topological qubits will not deliver commercial quantum computing on Microsoft’s Azure timeline. Microsoft continues to fund the topological program through DARPA US2QC and its own roadmap, but it is building its near-term quantum business on neutral atoms.

The Nayak Defense and Why It Fails

When the Legg critique was published, Microsoft’s Chetan Nayak responded by pointing to DARPA: “We are confident in our ability to execute against our roadmap and proud of our continued engagement with DARPA, which moved Microsoft into the final phase of its Quantum Benchmarking Initiative after independently evaluating our results — those in the public realm and proprietary — with a team of highly qualified experts.”

This defense, which I expect to be widely repeated, is a category error.

Stage C is the verification phase. Its purpose is to test claims. DARPA advancing Microsoft to Stage C means DARPA judges the approach worth the cost of rigorous, independent evaluation. Being admitted to the exam is not the same as passing it. Nayak is treating admission as validation.

Consider the alternative explanations for why DARPA advanced Microsoft to Stage C, none of which require the underlying physics to be settled:

DARPA may have seen proprietary results that are more convincing than what’s public. Over 50 experts from DARPA’s IV&V team have been working with both companies since early 2023, with access to internal data. This is possible. But “DARPA may have seen something” is precisely the kind of claim that cannot be evaluated by anyone outside DARPA, which makes it a weak basis for public scientific debate. If the results exist and are convincing, publishing them would settle the question.

DARPA may have advanced Microsoft because the scalability argument works if the physics is real, and determining whether the physics is real is exactly what Stage C is designed to do. This reading treats Stage C as a test rather than an endorsement, which is consistent with how DARPA itself describes the phase.

US2QC’s mandate required DARPA to carry forward underexplored approaches that showed plausible theoretical paths. Topological quantum computing, whatever the state of its experimental evidence, qualifies as underexplored by any definition. Eliminating Microsoft from a program designed to evaluate underexplored approaches, on the grounds that the approach is underexplored, would defeat the program’s purpose.

Missing a genuine topological breakthrough costs more than continuing to evaluate a long shot. This is standard DARPA logic. The agency exists to fund high-risk, high-payoff research where the expected value justifies the probability-weighted downside. An approach with a 20% chance of working but a payoff that rewrites the scaling math if it does is exactly DARPA’s portfolio sweet spot.

Any one of these explanations, or a combination, justifies Stage C advancement without validating the topological qubit claim. Nayak’s defense collapses every possible reason for advancement into a single one: “DARPA believes our physics works.” That reading is convenient but unsupported by anything DARPA has said publicly. Altepeter’s descriptions of the program consistently frame Stage C as evaluation, not endorsement.

What QBI Advancement Actually Signals

For CISOs, investors, and policymakers reading QBI as a signal, the crucial distinction is between what QBI tells you and what it does not.

What Stage C status tells you: DARPA’s IV&V team — roughly 200 experts from across government laboratories and research centers — judged that these companies’ approaches have a plausible (not certain, not even probable) path to utility-scale operation by 2033. The approach survived a series of filters designed to eliminate implausible designs. This is meaningful, because DARPA did eliminate companies (Atom Computing and QC82 from US2QC; seven of 18 Stage A entrants were not in the November 2025 Stage B cohort, though DARPA has stressed that evaluations are staggered and additional companies could still advance). The filters have teeth.

What Stage C status does not tell you: That these approaches are more likely to succeed than the Stage B companies. That the underlying physics is settled. That these companies will build a quantum computer before Quantinuum, IBM, or QuEra. That being further along in the DARPA pipeline correlates with being further along in hardware development.

Procedural, not technological. Several Stage B companies have substantially stronger public logical-qubit evidence on their own hardware. Quantinuum’s Helios reports extensive error-corrected logical-qubit operation. IBM’s recent quantum advantage papers describe results on hardware that exists and runs. QuEra-affiliated researchers have demonstrated programmable logical-qubit circuits involving up to 48 logical qubits. If any of these companies advances to Stage C, they will arrive with demonstrated hardware rather than designs under evaluation.

What Stage B status tells you: DARPA’s IV&V team found these companies’ architectures plausible enough to warrant a year-long deep technical audit. IBM, Quantinuum, IonQ, QuEra, and seven others cleared this bar. Stage B is itself a significant endorsement of a company’s engineering seriousness, and several Stage A participants (including Rigetti and HPE) were absent from the November 2025 cohort, though Google Quantum AI had only joined Stage A in September 2025 and was likely on a later evaluation schedule.

Altepeter himself was surprised any company reached Stage C. He told Breaking Defense: “I didn’t think anybody was going to make it to Stage C. But that’s why we do this, to get surprised.” That framing is important. DARPA designed US2QC with the expectation that most or all approaches would be eliminated during evaluation. Two surviving to Stage C was an upside scenario, not a baseline expectation.

DARPA’s Bet Structure Is Defensible, With One Caveat

This is a fair question and I think the answer is: probably yes, with a serious caveat.

DARPA’s bet structure is designed for a specific purpose. The agency exists to prevent strategic surprise. If the U.S. government is going to be blindsided by a quantum breakthrough, DARPA wants it to come from an approach it has already evaluated, not one it ignored. Running a deep evaluation of photonic and topological quantum computing in parallel with the broader industry’s evaluation of superconducting and trapped-ion systems is a hedge. The cost of the hedge ($125 million for PsiQuantum’s Stage C, plus Microsoft’s undisclosed contract) is small relative to the cost of missing a viable path to utility-scale quantum computing.

The fact that US2QC survived into QBI and its performers were grandfathered into Stage C is the institutional artifact that creates the optical paradox, but the underlying logic is defensible. DARPA started evaluating PsiQuantum and Microsoft earlier because those approaches were earlier in their evaluation cycle. Starting them earlier was the point.

The caveat is about signal clarity. When QBI Stage C status gets cited in press releases, investor presentations, and government testimony as evidence that an approach works, the program’s careful evaluation logic gets flattened into a soundbite: “DARPA selected them.” Nayak’s defense of Microsoft’s topological qubit is the example I’ve already discussed, but PsiQuantum uses the same framing in its investor materials and government engagements. Stage C becomes a credibility badge rather than what it actually is: an ongoing evaluation.

DARPA could mitigate this with clearer public communication about what Stage C means and what it does not. But unclear optics around a well-designed program is a minor complaint. The program itself does what it should: it uses the most rigorous quantum evaluation team any government has assembled to stress-test approaches that the private sector’s incentive structure might otherwise leave under-examined.

The companies in Stage B should be evaluated on their demonstrated capabilities, not penalized for being later entrants to a broader program. And the companies in Stage C should be evaluated on whether their approaches survive the verification that Stage C is designed to provide, which is a question that has not yet been answered.

For CRQC timeline watchers, the QBI funnel carries a subtler signal than the scoreboard suggests. DARPA’s managing director for QBI, Micah Stoutimore, said in March 2026 that “it now seems likely that someone will build a utility-scale quantum computer by 2033.” That assessment reflects the aggregate picture across all stages, not just Stage C. The fact that 11 companies cleared Stage A and entered Stage B means DARPA’s evaluation team sees multiple plausible paths, spanning superconducting, trapped-ion, neutral-atom, silicon-spin, and photonic architectures.

The race to a cryptographically relevant quantum computer will not be decided by which company is furthest along in DARPA’s pipeline. It will be decided by which company demonstrates fault-tolerant operation at a scale that matters. As of August 2026, the companies with the most demonstrated progress toward that goal are not the ones in Stage C. Whether that changes depends on what DARPA’s evaluation teams find when they open the hood on PsiQuantum’s photonic architecture and Microsoft’s topological claims. That is, after all, what Stage C is for.