Table of Contents
July 23, 2026 – A multi-institutional team spanning MIT, Harvard, NIST/University of Maryland, UCLA, Yale, Stanford, Chicago, and more than twenty other institutions posted a 135-page strategic plan for neutral-atom quantum computation on arXiv. The paper grew out of an NSF-funded town hall convened at MIT’s Endicott House in January 2025, and it is one of the most detailed public roadmaps yet produced for any qubit modality.
The bottom line: the authors argue that neutral atoms could reach practical quantum advantage within a decade if the scaling rates of the past decade persist. That projection is conditional on sustained scaling. The paper places cryptanalysis, including RSA-2048 factoring, on a resource-frontier map that compares experimental capability against theoretical requirements. The two fronts are moving toward each other. The gap is still enormous, but the direction of travel is clear.
Among the authors are researchers with affiliations or present addresses at PASQAL, QuEra, Infleqtion, planqc, and NanoQT, alongside academic and government research institutions. The paper discloses these affiliations and cautions that the views represent individual scientific assessments, not official positions of any company, university, or funding agency. It is funded through NSF Awards 2410716 and 2533041 under the National Quantum Virtual Laboratory (NQVL) program.
What “Practical Quantum Advantage” Actually Means
The paper opens by doing something the quantum computing field badly needs more of: defining its terms precisely. The authors propose a four-part test for “practical quantum advantage,” requiring that a quantum computation (1) correctly solves a task, (2) cannot be solved on available classical hardware, (3) has a scaling advantage over the best classical algorithms, and (4) is relevant to a community outside the hardware builders themselves.
That fourth criterion deserves emphasis. Under this definition, canonical random-circuit-sampling demonstrations would not count as practical advantage because their purpose is to establish computational separation, not to solve a task valued by an external user community.
From there, the paper maps three regimes of practical quantum advantage by computational scale, measured in “Quops.” The authors define a quop as a one- or two-qubit operation executable within one error-correction cycle. Toffoli count provides only a lower bound on quop count; total overhead may be 10 to 100× higher depending on the algorithm and architecture.
At the smallest scale, 10^3 to 10^6 Quops, we find the current generation of random circuit sampling experiments. The authors classify these as “weak, unverifiable quantum advantage” because the results cannot be efficiently checked in the regime where they claim advantage.
A second regime, at 10^6 to 10^9 Quops (megaquop to gigaquop), is where the paper sees near-term opportunity. Certified random number generation, specific quantum simulation problems, and cryptographic proofs of quantumness fall here. This is also where quantum advantage claims can first be classically verified.
At the far end, 10^9 to 10^12 Quops (gigaquop to teraquop), sit quantum chemistry, materials science, optimization, and cryptanalysis. This is where RSA-2048 factoring and ECC-256 discrete logarithms sit. The paper’s Figure 3 plots specific resource estimates from multiple research groups. For RSA-2048, three data points trace the evolution: Gidney and Ekerå 2021 at roughly 6,189 logical qubits and 3 billion Toffoli gates; Chevignard, Fouque, and Schrottenloher 2024 at roughly 1,400 logical qubits and 6.5 billion Toffoli gates (trading fewer qubits for more gates); and Gidney 2025 at roughly 1,730 logical qubits and 2 trillion T gates. For 256-bit elliptic-curve discrete logarithms, Babbush et al. 2026 report two paired trade-offs: fewer than 1,200 logical qubits with fewer than 90 million Toffoli gates, or fewer than 1,450 logical qubits with fewer than 70 million Toffoli gates. Separately, the Pinnacle architecture (Webster et al. 2026) is a qLDPC-based end-to-end design claiming RSA-2048 factoring with fewer than 100,000 physical qubits, and Cain et al. 2026 estimate that Shor’s algorithm could run on a neutral-atom qLDPC system with as few as 10,000 physical qubits.
None of these numbers suggest imminent cryptanalytic capability. They do suggest a path that is becoming progressively less impossible.
The Scaling Trends
The paper’s central empirical claim is a pair of exponential scaling fits derived from the best published neutral-atom results over the past decade. Physical qubit counts have grown at roughly 1.85× per year, from early tweezer arrays of tens of atoms to the Harvard group’s recent 3,000-atom continuously reloaded system and the Caltech group’s 6,100-qubit cesium array. Two-qubit gate errors have dropped at roughly 0.62× per year, from early Bell fidelities below 90% to the Lukin group’s April 2026 result of 99.85% raw CZ fidelity (99.94% with loss postselection).
The authors are careful to note these are trend-line fits rather than predictions. They explicitly acknowledge that gate fidelity improvement “may slow down due to either physical limits or reduced pressure once error thresholds for fault-tolerant operations are sufficiently surpassed.” And they note that best-in-class system sizes and best-in-class gate fidelities have not yet been demonstrated in the same system simultaneously, a gap that must be closed by engineering scalable gate controllers.
Several groups have already crossed the 1,000-atom mark: Atom Computing with more than 1,200 ytterbium-171 atoms, PASQAL with more than 1,100 rubidium-87 atoms in a cryogenic environment, Birkl’s group at TU Darmstadt with more than 1,300 rubidium-85 atoms, and a group at USTC with more than 2,000 rubidium-87 atoms. The authors argue that higher-power lasers and beam combining could push tweezer systems beyond 10,000 atoms, and project that kW-level systems could support as many as 100,000.
The Convergence Chart
Figure 2c places the experimental frontier and the theoretical resource frontier on the same logarithmic map. The x-axis is physical qubit count; the y-axis is two-qubit gate fidelity. Experimental demonstrations occupy the lower-left region. Theoretical resource estimates for useful computation occupy the upper-right. The two regions are moving toward one another, although the plotted points come from different systems, codes, and operating assumptions.
On the experimental side, the advancing front runs through Maller 2015, Graham 2022, Bluvstein 2024 (the 48-logical-qubit demonstration), and Bluvstein 2026 (below-threshold fault-tolerant operation with loss detection). Each step has moved both axes forward: more qubits and higher fidelity.
On the theoretical side, resource estimates for RSA-2048 factoring have been dropping across multiple architectures. Fowler’s 2012 estimate assumed surface codes with nearest-neighbor connectivity and required on the order of 10^8 physical qubits. Gidney and Ekerå’s 2021 estimate reduced that to 20 million. Gidney’s 2025 preprint cut it to fewer than one million under its stated assumptions. And the most recent qLDPC-based estimates (Cain et al. 2026 for neutral atoms, Webster et al. 2026 for the Pinnacle architecture) push the requirement down to the 10,000 to 100,000 physical qubit range.
Thousands-of-atom arrays have been demonstrated, and separate systems have produced two-qubit fidelities in the 99.5% to 99.85% range. Combining both records in one fully controlled system remains an engineering challenge the roadmap itself identifies. On the logarithmic resource-frontier map, the two fronts are closing.
Why qLDPC Codes Give Neutral Atoms an Architectural Edge
A recurring theme across the paper is the match between quantum low-density parity-check (qLDPC) codes and neutral-atom hardware. The argument runs like this: surface codes, the workhorse of most fault-tolerant quantum computing proposals, have an encoding rate that scales poorly. They need many physical qubits per logical qubit, and that overhead grows with the code distance. High-rate qLDPC codes can encode more logical qubits per block with fewer physical qubits at comparable error protection, but they require nonlocal or dynamically reconfigurable connectivity between qubits, something that superconducting processors, with their fixed 2D planar layouts, struggle to provide.
Neutral-atom arrays, by contrast, can rearrange atoms dynamically. Coherent transport moves qubits to interact with partners elsewhere in the array. The exact connectivity requirement is code- and architecture-dependent, but the ability to reconfigure atom positions at speeds compatible with computation gives neutral atoms a structural advantage for implementing many qLDPC families. The paper cites recent work showing that qLDPC families (hypergraph product, bivariate bicycle, and generalized bicycle codes) can reduce physical qubit counts by close to an order of magnitude compared to surface codes at comparable distance, with circuit-level thresholds above 0.5%.
The paper also highlights an AI-assisted decoder called Cascade (arXiv:2604.08358, a preprint by Gu et al.). In simulation, Cascade decodes both surface codes and high-rate bivariate bicycle codes near-optimally. For the [[144, 12, 12]] Gross code, it reaches logical error rates near 10^-10 per cycle per logical qubit at a physical error rate of 0.1%, and its latency falls within the real-time decoding budget of current neutral-atom cycle times.
These are simulation and theory results, not system-level hardware demonstrations. No group has yet run a large qLDPC code at scale on a neutral-atom processor. But the architectural alignment is real.
The Distance from Threshold to Megaquop
The paper anchors its near-term credibility on the Lukin group’s result (Bluvstein et al., Nature 649, 39-46, 2026): a neutral-atom processor demonstrating below-threshold error correction with more than 2× suppression of logical errors when increasing from distance-3 to distance-5 surface codes, using atom loss detection and machine-learning decoding. This is the capability I wrote about when the result first appeared, and it remains one of only a handful of demonstrations across any platform where increasing the code distance actually reduces the logical error rate.
To go further below threshold, the paper identifies two complementary strategies. One is pushing two-qubit gate fidelities beyond the current 99.5% to 99.9% range, primarily by increasing Rydberg-excitation laser power and managing the resulting off-resonant light shifts. The other is exploiting the error structure that neutral atoms offer: a large fraction of errors correspond to atom loss, which can be detected and converted into “erasures” that error-correcting codes handle much more efficiently than random Pauli errors. Exploiting erasure detection can push the effective surface code threshold from 1% to 2% or higher.
The April 2026 Evered et al. preprint reports a raw CZ infidelity of about 1.5 × 10^-3, improving to about 6 × 10^-4 after postselection on atom survival. Those are strong physical-gate results. But reaching the megaquop regime, where circuits can execute more than a million error-corrected operations before uncorrected failure, requires pushing logical error rates to the 10^-6 level. The Bluvstein 2026 experiment reported 0.62% logical error per round for its distance-5 code, approaching 0.1% on shots without qubit loss. Against a 10^-6 megaquop target, the remaining logical-error gap is roughly three to four orders of magnitude.
What the Paper Does Not Say
For all its thoroughness, the paper has gaps.
It does not provide a systematic, normalized cross-platform comparison with superconducting or trapped-ion systems. The paper mentions Google and Quantinuum results in passing, but a side-by-side comparison covering qubit count, two-qubit infidelity, QEC cycle time, connectivity, logical suppression, and demonstrated circuit depth would have made the platform case more rigorous. It would have strengthened the paper’s argument where neutral atoms lead (system size, reconfigurable connectivity, facility simplicity) and honestly acknowledged where other platforms have advantages.
The paper does not address whether the 1.85×/year qubit scaling rate and 0.62×/year error reduction rate can survive the transition from research demonstrations to engineered products. The history of quantum computing is littered with exponential trend lines that flattened when they hit engineering walls. The paper acknowledges this possibility but does not model it.
And the paper frames its “within the next decade” timeline as contingent on continued scaling, then devotes much of its 84-page main text to cataloging the unsolved problems that stand in the way: photonic control at 10,000+ qubit scale, standard fluorescence readout that routinely takes 5 to 20 milliseconds, integration of continuous reloading with fault-tolerant computation, and decoder throughput that keeps pace with progressively more complex codes. Specialized readout methods can be much faster, and continuous reloading has already been demonstrated in a coherent 3,000-atom system. What remains missing is their integration at the scale and reliability that deep fault-tolerant circuits demand.
My Analysis
I have followed neutral-atom quantum computing since the early tweezer experiments, and I have covered the modality extensively in my Quantum Systems Integration book and in the Building a Quantum Computer series on this site. This paper makes a compelling case for neutral atoms as a serious contender for practical quantum advantage. But the security implications deserve separate, careful analysis.
Start with what the convergence chart actually tells us. The experimental front is advancing. The theoretical resource requirements are dropping. These two fronts are converging on the same logarithmic map. But the remaining gap still spans orders of magnitude in both qubit count and sustained fidelity, and the final steps are always the hardest because they require multiple capabilities to work simultaneously in the same system.
In my CRQC Quantum Capability Framework, I track nine core technical capabilities plus a cross-cutting engineering-scale and manufacturability dimension that a cryptographically relevant quantum computer must demonstrate, from basic qubit connectivity and routing to magic state production, real-time decoder performance, and continuous operation over hours or days. This paper addresses several of those dimensions with concrete experimental results, particularly below-threshold operation and engineering scale. But magic state production at the rates required for cryptanalysis, real-time decoding for qLDPC codes at scale, and sustained operation over the hours needed to factor RSA-2048 are all acknowledged as open problems in the paper itself.
The qLDPC advantage is real and worth tracking closely. If high-rate qLDPC codes deliver on their theoretical promise, the physical qubit requirement for RSA-2048 drops by roughly an order of magnitude compared to surface codes. Neutral atoms’ reconfigurable connectivity gives them a real architectural edge for implementing these codes. But “reduce by an order of magnitude” applied to a number in the millions still leaves a number in the hundreds of thousands, and no group has yet demonstrated a large qLDPC code running fault-tolerantly at scale on any hardware, neutral-atom or otherwise.
Then there is the question I always apply to roadmaps like this: who is writing it and why? The authors include researchers with commercial interests in PASQAL, QuEra, planqc, Infleqtion, and NanoQT. The paper is funded by the NSF and positioned to support continued federal investment in neutral-atom quantum computing through the NQVL program. None of this invalidates the science. Many of its anchors are peer-reviewed experimental results, but several of the most consequential architecture and decoder claims remain preprints. When a community produces a document arguing that its platform should receive sustained research funding, readers should evaluate the claims with that context in mind. This is both a technical roadmap and a case for sustained investment in the platform.
