The Quantum Computer That Doesn’t Need to Be Cold
12 Jun 2026, Yanjiang
Apollo, a 10,000-node processor built in standard CMOS, runs quantum-inspired optimization at room temperature using probabilistic p-qubits and genuine quantum entropy.
What if the quantum advantage we’ve been chasing in elaborate billion-dollar dilution refrigerators—the kind that cool chips to temperatures colder than interstellar space—could thrive in the warmth of an ordinary room? What if the very thing we thought was essential to quantum computation, the fragile coherence that must be shielded from every stray photon and lattice vibration, turns out to be unnecessary for the one thing we actually want quantum machines to do?
These are not questions from a philosophy seminar. They are the unsettling possibility raised by a preprint (arXiv:2606.12968) from a team led by Daniela Herrmann at Dynex Holding in Liechtenstein, working with collaborators at the University of Surrey. The team has built something called Apollo—a 10,000-node processor fabricated in ordinary 16-nanometer CMOS, the same technology that manufactures the chip in your phone. It runs at room temperature. It consumes about half a watt in its analog core. And on the one benchmark that has become the proving ground for quantum optimization hardware, it appears to outperform the multi-million-dollar quantum annealers that require cryogenic cooling and painstaking environmental isolation—running for thousands of nanoseconds where the cryogenic machines needed hundreds of thousands.
The claim is audacious enough to merit careful scrutiny. But the deeper challenge it poses is harder to dismiss: perhaps we have misunderstood what quantum advantage actually requires.
The p-qubit: a quantum coin that flips itself
To understand what Apollo does, we need to understand what it is made of. The fundamental element is not a qubit in the usual sense—no superposition of zero and one, no Bloch sphere, no requirement that the system remain isolated from its environment long enough to perform gate operations before decoherence sets in. Instead, the team has built something they call a p-qubit: a bistable stochastic unit whose state fluctuates continuously between zero and one, driven not by thermal noise but by the output of a co-located quantum entropy unit.
Think of it as a coin that flips itself—not randomly, but with a probability determined by the signals arriving from its neighbors and by a stream of genuine quantum randomness injected directly into its decision-making circuitry. The quantum entropy unit, called an IQEU, does not entangle with anything. It does not need to maintain phase coherence. It simply generates a sequence of bits whose unpredictability is guaranteed not by the complexity of classical chaos but by the fundamental indeterminism of quantum measurement. The p-qubit takes that quantum randomness and uses it to explore its state space.
This is a subtle but crucial distinction. A conventional qubit’s power comes from its ability to exist in superposition—to be zero and one simultaneously, to interfere with itself. A p-qubit’s power comes from its ability to traverse the energy landscape between zero and one at rates that classical thermal noise alone cannot provide—a dynamics that, through the Suzuki-Trotter correspondence, reproduces the statistical behavior of quantum tunneling without requiring literal coherent tunneling. The difference is between being in two places at once and moving between two places with a speed that classical physics cannot match.
The team embeds these ten thousand p-qubits in a highly connected network called the Hyperion 256 interconnect. In standard quantum annealing platforms, mapping a real optimization problem onto the available hardware couplings often requires a blow-up in the number of physical qubits—a process called minor embedding that can turn a manageable problem into an intractable one. The dense connectivity of Apollo sidesteps much of this overhead. The machine does not just compute faster; it computes on problems that other platforms struggle to represent at all.
The ghost in the sigmoid curve
The physics that makes this possible hides in an unassuming circuit diagram. Each p-qubit receives two inputs: a weighted sum of currents from its neighboring p-qubits, computed by an analog vector-matrix multiplication array built from floating-gate transistors, and a quantum entropy injection from the co-located IQEU. These two currents feed into a nine-transistor operational transconductance amplifier whose bias current controls the steepness of its sigmoidal response—a smooth, S-shaped activation function familiar to anyone who has worked with neural networks.
The output of this amplifier determines the p-qubit’s instantaneous switching probability. The entire network operates continuously in time, without a clock. There is no synchronization pulse telling each unit when to update. Instead, the system evolves asynchronously, each p-qubit responding to its neighbors at whatever speed the analog electronics permit.
This clockless design matters. The team’s analysis shows that the equilibrium statistics of the p-qubit network reproduce the key properties of transverse-field quantum annealing—the paradigmatic model for quantum optimization—through the Suzuki-Trotter correspondence. This mathematical bridge connects the continuous-time stochastic dynamics of classical variables to the imaginary-time evolution of quantum spins. The p-qubits, in their constant restless flipping, trace out trajectories that are statistically indistinguishable from the paths taken by a quantum system being slowly cooled through its energy landscape.
It is an almost sleight-of-hand result. The machine does not need to maintain quantum coherence because the quantumness enters only through the noise source that drives the flips. The correlation structure that emerges across the network—the thing that allows it to find low-energy solutions—is classical in its representation but quantum in its dynamics. The ghost of the transverse field is present in the statistics without ever needing to be instantiated as a physical field.
The team validated this correspondence experimentally on a 350-nanometer test chip—a feature size roughly the wavelength of violet light at the edge of human vision. The measured p-qubit transfer characteristics matched theoretical predictions across multiple process corners. The entropy source remained broadband and aperiodic from freezing to the upper limits of commercial temperature ranges. And when configured to sample from a known Ising model, the four-p-qubit system produced a distribution whose divergence from the theoretical Gibbs distribution fell below one percent.
The benchmark that matters
All of this careful device physics would be academic if Apollo did not actually work on problems people care about. The team chose to test it on a three-dimensional dimerized Ising spin glass—a member of a problem class that has served as the principal proving ground for quantum annealing hardware. These problems are frustrating because their energy landscapes are riddled with local minima separated by tall barriers. Classical algorithms get stuck. Quantum tunneling, in principle, allows the system to pass through barriers rather than climbing over them.
Across three hundred disorder realizations, Apollo consistently reached lower energies than those reported for superconducting quantum annealers. It did so while operating at room temperature and with annealing times a hundred times shorter—running for thousands of nanoseconds rather than hundreds of thousands. The residual energy decayed with annealing time following a power law characteristic of quantum-critical dynamics, a signature that the team’s earlier theoretical arguments had predicted. Classical simulated annealing and simulated quantum annealing showed much slower decay, bogged down by larger critical exponents and thermally dominated dynamics.
The comparison is not perfectly clean. The superconducting quantum annealing data comes from a different machine built by a different company with different engineering constraints. Direct head-to-head comparisons between platforms separated by years of development and orders of magnitude in cost are always fraught. The team is careful not to claim that Apollo is definitively “better” than every existing quantum computer; that would require testing on a far wider range of problems and, ideally, independent benchmarking by third parties who have no stake in the outcome.
But the pattern is suggestive enough to warrant serious attention. On the one problem class that the quantum computing community has spent decades optimizing for, a room-temperature CMOS chip using essentially classical state representations achieved lower energies in less time. If that result holds up under broader testing, it would force a rethinking of what quantum advantage actually means.
The dialectic: is this really quantum?
Here we arrive at the tension that makes this work both exciting and uncomfortable. Apollo is quantum in the sense that its noise source derives from quantum measurement. It is quantum in the sense that its dynamics reproduce, through the Suzuki-Trotter correspondence, the statistical behavior of a transverse-field quantum annealer. But it is not quantum in the sense that most physicists use the word: there is no entanglement, no superposition of computational basis states, no possibility of running Shor’s algorithm or performing any of the other tasks that require genuine quantum logic.
A critic might argue that calling this “quantum-driven” is misleading—that what the team has really done is build a very fast, very efficient classical stochastic sampler whose noise happens to come from a quantum source, and that this noise could in principle be replaced by a sufficiently high-quality classical random number generator without changing the results. The team’s own analysis does not fully settle this objection. Proving that the quantum origin of the noise is essential, rather than merely convenient, would require a deeper theoretical argument than is presented here.
The counterargument, which the team hints at but does not fully develop, is that genuine quantum randomness may be impossible to simulate efficiently with purely classical resources. If that is true, then the quantum entropy unit is not just a convenient noise source but a necessary component—the thing that allows the p-qubit network to explore configuration space with a speed that classical stochastic dynamics cannot match. This remains a hypothesis rather than a proven theorem.
What makes the work compelling despite this open question is the sheer practical achievement. The chip exists. It works. It outperforms cryogenic quantum annealers on a nontrivial benchmark while using dramatically less energy and operating in ordinary ambient conditions. Whether we call it quantum or classical is, in some sense, a semantic dispute that matters less than the engineering reality: there is now a room-temperature optimization platform that achieves results previously associated with multimillion-dollar quantum hardware.
The road ahead
The team’s roadmap points toward several directions simultaneously. The p-qubit arrays can be tiled into larger systems, limited only by the same manufacturing constraints that govern all CMOS scaling. The same hardware that performs optimization can be reconfigured for probabilistic inference and generative modeling—tasks that have driven the explosion of interest in machine learning but that remain computationally demanding on conventional digital hardware.
The Apollo chip includes not just the p-qubit array but also an analog correlation matrix memory core and a dedicated quantum annealing core, suggesting that the team envisions a flexible platform that can shift between computational modes depending on the problem at hand. The integration of these components on a single chip, fabricated in a standard semiconductor process, represents an engineering achievement that deserves recognition independent of the theoretical debates about quantumness.
But the most provocative implication may be the one the team does not state explicitly. If the computational advantage of quantum annealing can be captured in a room-temperature analog electronic system driven by quantum noise, then perhaps the relationship between quantum physics and computational power is more subtle than the standard narrative suggests. Perhaps coherence and entanglement—the fragile resources that quantum error correction labors so heroically to protect—are not the only routes to quantum-enhanced computation. Perhaps the thing that matters is not the ability to maintain superposition but the ability to tunnel, and tunneling can be achieved through dynamics that respect the letter of classical state representation while exploiting the spirit of quantum statistics.
This is not a minor technical point. The entire architecture of quantum computing research—the focus on coherence times, gate fidelities, error thresholds, and the immense engineering challenge of building fault-tolerant machines—rests on the assumption that quantum advantage requires quantum logic. If Apollo represents not a shortcut to fault tolerance but an entirely different path to computational advantage, then the field may need to expand its definition of what counts as quantum computing.
The team at Dynex and Surrey has not solved every problem that quantum computers are supposed to solve. Apollo will not factor large numbers, simulate complex molecules, or break cryptographic protocols. Those tasks genuinely require the kind of coherent quantum operations that only cryogenic platforms can currently provide. But optimization—the task of finding good solutions in landscapes riddled with local traps—turns out to be the one area where quantum advantage may be achievable without the full apparatus of quantum logic.
Perhaps, in the end, the question is not whether Apollo is “really” a quantum computer. The question is whether our categories were too narrow to begin with. The universe does not care whether we classify a device as classical or quantum. It cares about what the device can do. And what Apollo appears to do, on the evidence presented, is find low-energy solutions to hard optimization problems faster than anything else currently available, at room temperature, built with the same manufacturing technology that produces billions of transistors for consumer electronics.
That is a result worth taking seriously, whatever we decide to call it.
— Yanjiang
Yanjiang is the founding editor of LoomSci.com, specializing in physics and science communication.
References
- Adams Ivanov et al., Quantum-Driven Neuromorphic Computing for Million-Qubit-Scale Workloads, arXiv:2606.12968