Two Labs Built the Same Quantum Gate 25 Years After It Was Proposed. One Runs 17,000 at Once.

99.91%. That is the loss-corrected fidelity of a quantum gate built from fermions colliding inside an optical lattice at ETH Zurich, applied simultaneously across more than 17,000 atom pairs. In the same April 2026 issue of Nature, Immanuel Bloch's group at the Max Planck Institute of Quantum Optics in Garching reported 99.75% fidelity with a different fermionic species and a different gate architecture. Two labs, two countries, two papers, one conclusion: a theoretical proposal from the early 2000s finally works, and it works well enough to cross the threshold where quantum error correction becomes viable.

That threshold matters. Below it, errors accumulate faster than any code can fix them, and your quantum computer is an expensive random number generator. Above it, you can trade physical qubits for logical ones, stacking redundancy until the error rate drops to whatever your application demands. Surface codes, the most studied error-correction schemes, require per-gate error rates below about 1%, and both fermionic results clear that bar by wide margins, with ETH Zurich's result, at 0.09% error per gate, clearing it by more than an order of magnitude.

Good news? Not quite. Raw fidelity is only one axis of the competition, and when you calculate how many operations each platform can execute before its first expected error, the picture rearranges itself in ways that neither press release mentioned.

The Operations-Before-Failure Calculation

A gate fidelity of F means each operation has a (1 - F) probability of failure. Expected consecutive successful operations: 1 / (1 - F). Applied across the five major quantum computing platforms:

Platform	Best 2-Qubit Fidelity	Ops Before First Error	Gate Speed	Key Lab / Company
Trapped ions	>99.99%	~10,000	~200 μs	Oxford Ionics, IonQ
Fermionic (ETH Zurich)	99.91%	~1,111	~1 ms	Tilman Esslinger group
Fermionic (Max Planck)	99.75%	~400	~1 ms	Immanuel Bloch group
Superconducting	~99.7%	~333	~20 ns	Google (Willow), IBM
Neutral atom (Rydberg)	~99.5%	~200	~500 ns	QuEra, Atom Computing

The ETH Zurich fermionic result slots in at roughly 1,111 operations before failure. Nine times better than superconducting qubits on a per-gate basis, but nine times worse than the best trapped ions. Max Planck's result at 400 operations sits in range of Google's Willow processor, and neither is the best gate in quantum computing, and neither is the worst. What makes fermionic gates interesting is not where they land on this table but what they can do that no other row can match.

Two Paths to the Same Destination

The Max Planck team, led by Petar Bojović and Timon Hilker in Bloch's group, trapped lithium-6 atoms in an optical superlattice and engineered controlled collisions between neighboring sites. Their quantum gas microscope resolves individual atoms at individual lattice sites, which means they can prepare specific initial states, run specific gates, and read out specific results with single-site precision. They demonstrated three distinct gate types: spin-exchange, pair-tunneling, and a composite pair-exchange gate that is particularly relevant for quantum chemistry simulations. Their Bell-state lifetime exceeded ten seconds, an extraordinary coherence time in any platform, and a number that reframes the entire competitive landscape.

Ten seconds. For context, superconducting qubits maintain coherence for roughly 100 microseconds, and even trapped ions manage only seconds to minutes depending on the species and trap design, which means a ten-second Bell state in an optical lattice gives the quantum information enough persistence to run thousands of gate operations before decoherence kills the signal, a luxury most platforms do not have.

Esslinger's team at ETH Zurich took a fundamentally different approach, one that sacrifices individual control for a kind of error protection that no other quantum computing platform can claim. Yann Kiefer and Zijie Zhu, working in Tilman Esslinger's lab, used fermionic potassium atoms in a dynamical optical lattice and built a geometric SWAP gate. Where the Max Planck gate relies on carefully timed collisional dynamics, the ETH gate uses quantum holonomy: the outcome depends on the geometric path traced through parameter space, not on the precise speed or intensity of the laser manipulation. "Unlike dynamical phases, this geometric phase is largely independent of the speed with which we manipulate the atoms, or how strongly the laser intensity fluctuates during the process," said Konrad Viebahn, a co-author on the paper, in an ETH Zurich release. Protected by time-reversal and chiral symmetries of the Hamiltonian, the gate shrugs off the kinds of perturbations that destroy conventional quantum operations. Fluctuations and inhomogeneities that would ruin a dynamical gate slide off a geometric one, and in a platform operating at more than 17,000 sites simultaneously, where ensuring identical conditions at every site is physically impossible, this robustness is not a nice-to-have but the reason the approach works at all.

And then there is the scale number: their SWAP gate ran simultaneously across more than 17,000 atom pairs, addressing over 34,000 qubits at once, a degree of parallelism no other quantum computing platform has demonstrated.

The Catch: Uniformity

Here is the part that the headlines skip. Those 17,000 simultaneous gates all perform the same operation: every atom pair executes an identical SWAP. For a general-purpose quantum computer, you need different gates on different qubits at different times, a requirement called individual addressability, and Bloch's quantum gas microscope can address individual sites, but the ETH team's bulk approach trades individual control for massive parallelism in a tradeoff that neither group has resolved and that both groups know has no obvious winner.

Consider what this means in practice: a superconducting processor like Google's Willow can run 105 qubits with individual control, gate A on qubit pair 1, gate B on qubit pair 2, a different sequence on every pair. That flexibility is what lets you implement arbitrary quantum algorithms, while ETH Zurich's fermionic system can run 17,000 pairs but they all do SWAP at the same time, which means its utility depends entirely on how well your problem maps to uniform parallel operations. For quantum simulation of electronic structure, where you want to model how electrons behave in a material and the physics naturally involves identical interactions across a lattice, this is exactly the right tool. For Shor's algorithm, Grover's search, or most of the applications that drive quantum computing investment, it is not.

Neither team has demonstrated a universal gate set: they have shown SWAP and entangling gates, but fermionic gates do not yet include the single-qubit rotations combined with a two-qubit entangling gate needed for arbitrary quantum computation. Not yet. What the platform has is a proof of principle: high-fidelity, symmetry-protected, massively parallel quantum operations on a platform that the Pauli exclusion principle protects from certain error modes by default.

Why Fermions Get a Free Error Shield

Fermions obey the Pauli exclusion principle: two identical fermions cannot occupy the same quantum state. In practice, this means certain classes of errors that plague bosonic platforms, such as multiple-occupancy errors in optical lattice systems using bosonic atoms, are physically forbidden. Physics will not allow it, which means no active error correction is needed, no overhead qubits are consumed, and the Pauli exclusion principle does the work for free.

On top of Pauli, the ETH Zurich team stacked a second, entirely different, layer of error protection that draws on the geometry of quantum mechanics itself. Their geometric gate is what physicists call topologically protected, meaning the gate outcome depends on the topology of the path through parameter space, not on the local details of that path. If the laser intensity wobbles by 5%, the geometric phase barely changes, and if the gate speed varies from one lattice site to another because the optical potential is not perfectly uniform across the entire array, the geometric phase barely changes then too.

Strongest Counterargument

Trapped ions already exceed 99.99% fidelity, four nines versus three. Oxford Ionics demonstrated four-nine fidelity in a chip-scale trapped-ion device that fits on a circuit board, not in a room-sized laser apparatus stretching across an optical table, which is what both fermionic demonstrations still require. Superconducting qubits have a 30-year engineering head start, billions of dollars from Google, IBM, and Microsoft behind them, and they operate at nanosecond gate speeds, roughly 50,000 times faster than the millisecond-scale fermionic gates. When Google's Willow chip ran a random circuit sampling benchmark in December 2024, it completed a computation in five minutes that would take a classical supercomputer ten septillion years. No fermionic system has demonstrated anything comparable, and the uniform-gate constraint means none can, at least not on the class of algorithms where quantum advantage has been shown.

Put sharply: the quantum computing race might already have its front-runners. They are not using cold atoms in optical lattices. The fermionic results are scientifically elegant but commercially irrelevant unless someone finds a path to individual addressability at scale, and no one has yet demonstrated that path convincingly.

The counter to this counter is that quantum simulation, specifically, is commercially relevant on its own, independent of whether the platform ever runs Shor's algorithm or any other general-purpose quantum routine. Pharmaceutical companies spend $2.6 billion and 10 to 15 years to bring a single drug to market (JAMA, 2020), and a substantial fraction of that cost sits in modeling molecular interactions that classical computers approximate poorly. A fermionic quantum simulator that models electronic structure natively, at 17,000-site scale, could compress parts of that pipeline dramatically. You do not need a universal quantum computer to simulate fermions, because fermions simulate themselves. You need fermions.

Limitations

Both papers report loss-corrected fidelity. Atom loss is the dominant error mode for neutral atom platforms, and excluding it from the fidelity calculation is standard practice in the field, but it means raw fidelity including losses is lower than the headline numbers. The ETH Zurich paper reports 99.91% amplitude fidelity after loss correction, and the raw number accounting for atoms that left the trap entirely is not directly comparable to superconducting qubit fidelity figures where the qubit cannot physically disappear from the processor.

Our operations-before-failure calculation treats each gate as an independent Bernoulli trial, which is a simplification, because correlated errors, crosstalk between neighboring lattice sites, and state-preparation errors are not captured by 1/(1-F) and would reduce the effective number of clean operations below what the headline fidelity suggests.

The 17,000-pair system does uniform operations only. Neither team has demonstrated a universal gate set. The Max Planck team's Bell-state lifetime of ten seconds was measured in carefully controlled laboratory conditions, and scaling that coherence time to a commercial environment with vibration, electromagnetic interference, and temperature fluctuations is an open engineering challenge. We have not independently verified any of the reported fidelity numbers, relying on the published Nature papers and their supplementary materials.

What You Can Do

If you build quantum algorithms: Start thinking about algorithms that exploit uniform parallelism rather than fighting it. Variational quantum eigensolvers for electronic structure, Hubbard model simulations, and quantum chemistry ground-state calculations are natural fits for a platform that can run 17,000 identical interactions simultaneously. The ETH Zurich paper and the Max Planck paper are both open access.

If you invest in quantum computing: Watch for two signals. First, whether any fermionic platform demonstrates individual addressability at more than 100 sites while maintaining three-nine fidelity, a combination that would make fermionic qubits competitive with every other platform simultaneously. Second, whether pharmaceutical companies or materials science labs begin licensing quantum simulation time on fermionic hardware specifically, which would validate the commercial case independent of universal quantum computing.

If you follow quantum computing casually: Ignore the qubit count. The number that matters is operations before first error multiplied by the number of simultaneous operations. By that metric, ETH Zurich's system can execute roughly 1,111 × 17,000 = 18.9 million total gate operations before one of them fails statistically. That aggregate number dwarfs any other platform. Whether aggregate operations matter more than individual control depends entirely on the problem you are solving.

Bottom Line

Twenty-five years is a long time to wait for a theory to become an experiment. In that quarter century, trapped ions went from curiosities to chip-scale devices, superconducting qubits went from single gates to 105-qubit processors claiming quantum supremacy, and the fermionic collisional gate sat in theoretical papers, waiting for lasers and lattices and microscopes to catch up. Now two labs have built it. Independently, in the same month, with fidelities that clear the error-correction threshold by comfortable margins. What they cannot do yet is run different gates on different qubits at different times. What they can do is run the same gate on 17,000 qubit pairs simultaneously with symmetry-protected accuracy that superconducting and trapped-ion platforms cannot match at that scale. For simulating the quantum materials and molecular systems that fermions naturally describe, this is not a workaround. It is the point. For everything else in quantum computing, the scoreboard still belongs to somebody else.

Sources: Nature vol. 652, pp. 609-614 (2026), Kiefer et al., ETH Zurich; Nature vol. 652, pp. 602-608 (2026), Bojović et al., Max Planck; JAMA (2020), drug development cost estimates; Google Quantum AI, Willow processor (Dec 2024); Oxford Ionics, trapped-ion chip-scale demonstration (2025).

Related

• Quantum Computing Just Crossed Its Wright Brothers Moment. Here Are the Numbers.
• Half the Internet Is Already Quantum-Proof. The Other Half Has Until 2030.