April 3, 2029
QuantoCorp applied laboratory
Palo Alto, California
It was a beautiful spring morning in Palo Alto. The scientist heard neither the birdsong nor felt the sun’s early rays cutting through the fog.
He was buried deep in the bowels of the lab, badged into a secure containment area, working on an airgapped system controlling the most delicate piece of hardware ever assembled: a 480-logical-qubit quantum computer, supercooled to 5 millikelvin, significantly colder than outer space. The machine, Astral IV, was a glittering chandelier, cooled with liquid helium in three stages, each one colder than the next.
At the base of the chandelier, where temperatures were so low even normally excitable atoms became torpid, sat a sapphire slab the size of a postage stamp. On this base, microscopic aluminum films etched into circuits. In each circuit, a minute weak link a few atoms thick. At these junctions, electrons no longer behaved like particles at all, but as a single coherent wave, flowing without resistance and occupying multiple states at once. It was here, in these engineered imperfections, that classical certainty broke down and quantum information came into being — fragile, probabilistic, and exquisitely sensitive to the slightest disturbance.
A few years earlier, most physicists had thought scaling beyond a few dozen fragile logical qubits would be impossible. There was no clear roadmap to keeping the logical qubits shielded from sources of interference and stable long enough to run a meaningful computation. Some even doubted that at-scale quantum computing was possible at all. Perhaps there was some sort of fundamental reason grounded in laws of physics why it couldn’t work. But this all changed in the mid 2020s when a series of breakthroughs in error correction made people realize scaling was possible. AI was an unexpected but meaningful accelerant. Models trained on experimental data learned how to tune pulses, schedule measurements, and route error correction cycles efficiently. A lot of what had looked like random decoherence turned out to be predictable enough to manage. Once that clicked, scaling was inevitable.
The engineering of the building itself was a testament to the delicate and occult task going on inside. It sat on a massive foundation, away from the rumble of the city. The QC chamber sat on a single block of concrete, hundreds of tons poured in one uninterrupted cast, decoupled from the surrounding structure by a narrow gap. The chandelier was anchored to the block through a lattice of vibration isolators and compliant mounts, so that the slow breathing of the building, the distant rumble of traffic, even the footfall of a careless technician would die out long before reaching the qubits below.
Deep inside this multi-billion-dollar marvel stood the scientist, alone. It was 6:30 am and no one was there aside from the heavily armed security. As the second most senior scientist at the firm, he had earned the right to indulge himself in the occasional bespoke run. For months QuantoCorp had been tediously solving smaller-scale problems that were tractable with their limited set of logical qubits. Logistics routing and optimization problems for multinationals. Trawling around various chemical catalysts to develop better superconductors or fertilizers.
The scientist had been seized with the urge to do something more interesting. The night before, on a whim, he had used the firm’s AI cluster to re-examine the resource costs of an elliptic curve discrete logarithm solution. OpenMind, the $10 trillion AI conglomerate, had offered them early access to a specialized model for quantum circuit optimization, and he was eager to try it out. Just a few months prior, AI had shown signs of recursive self-improvement. It wasn’t AGI exactly, but it had become extraordinarily good at certain kinds of tasks. Math and physics in particular. The remaining Erdős problems had fallen quickly. OpenMind had solved the Hodge conjecture and proved the Riemann hypothesis. It was an exciting and terrifying time to be a physicist. The scientist wasn’t sure how much longer his skills would matter for. But he was enchanted by the pace of discovery nonetheless.
The night before, he’d asked the AI model how many Toffoli gates he would need for an elliptic curve break under Shor, using the known parameters of Astral IV. The AI model had finished its run at 4:13 am. When he powered up his workstation, it read:
Shor-ECDLP Resource Estimate (secp256k1)
Logical qubits: ~0.5k | Toffoli gates: 3.52e7 | Depth: 9.6e6
Status: Feasible on Astral IV (est. runtime: minutes to hours)
He toggled the assumptions panel — logical error rates, surface-code distance, coherence margins, physical-to-logical overhead. All within spec. No relaxed tolerances. No exotic error model. The AI had stayed inside the constraints. The reduction hadn’t come from faster hardware, but from the circuit. The model had reworked the elliptic-curve arithmetic itself — collapsing layers of modular addition, reusing ancillae, trimming depth in the inversion steps that dominated older designs. Entire sections that used to run sequentially had been flattened or parallelized at the logical level.
He couldn’t quite believe it. The state of the art requirement for solving ECDLP was a circuit with over 100m Toffoli gates, meaning over 1500 error-corrected qubits using standard assumptions. Now the AI was telling him the cryptographic primitive could perhaps be broken with the hardware sitting 15 feet away from him, with a third of the qubits he had thought were required.
“Fuck it,” he thought to himself. “I’ll give it a shot.”
He was almost certain that something had gone wrong, that the AI had hallucinated an unrealistically compact quantum algorithm, but he didn’t have enough time to invoke the cluster again before his colleagues arrived and started asking awkward questions.
This wasn’t supposed to be possible yet. None of the top labs were anywhere near having the firepower for an ECDLP break under known assumptions. He told himself that mathematics didn’t care about roadmaps or preprints or social consensus. A curve was either vulnerable or it wasn’t.
The glittering, golden chandelier waited in the next room, perfectly still.
Four hundred eighty logical qubits online. Error-corrected. Stable. Boring, even. The machine had been idle for six hours, waiting for work.
He wavered, thinking about the uncomfortable consequences if the test somehow worked. Then he reminded himself that US firms can’t just go around stealing other people’s property. And besides, this would be a legal matter before it became a practical reality. The test was just a benchmarking exercise, he told himself.
He gave it his instructions. He chose a full 256-bit elliptic curve keypair on secp256k1.
Given the standard generator point G and the compressed public key
02A2B898E63BFC5439B63516550FBF95A115E983E961CE4B961F72F04D2DC08D60
recover the scalar k such that P = k*G.
In other words: given the public key P, compute the private key k – the elliptic curve discrete logarithm.
He initiated the run.
The progress bar appeared, crawling forward at first, then accelerating as the scheduler parallelized the circuit across the array. Estimated completion time flickered, recalculated, then stabilized.
Under the hood, the machine wasn’t guessing keys or trying combinations. It was doing something stranger and far more efficient. It prepared a quantum state that spanned an enormous space of possibilities and let the curve’s arithmetic act on all of them at once.
As the computation unfolded, most of those possibilities destroyed each other. Their phases drifted out of alignment, interfering destructively as the machine stepped through the math. What survived was a structure: a faint, repeating pattern encoded in the way the curve folded back on itself.
A final transform collapsed that pattern into a handful of sharp peaks. From those, the private key could be reconstructed — something no classical machine could have isolated, no matter how long it ran.
After 14 minutes and 47 seconds, the run ended with a soft chime. On his screen, a readout:
Shor-ECDLP job finished
Classical post-processing: COMPLETE
Curve: secp256k1
Recovered discrete log k (hexadecimal)= 050A17DFA0E4C57ABF4B2C9FAF00B0D49EA6BB854BC5CCFC7BD21F6196AD9C0A
Measurement cycles: 1.2e9
Logical error rate (est.): 7.4e-10
Result validated against curve arithmetic.
It was the private key corresponding to the public key he had given Astral IV. Sweating, he re-ran the computation in the other direction — the ordinary direction, the one you’re meant to go in — even though his system had already done those checks.
He left his workstation and rushed out through the secure doors. He burst out into the lobby fumbling with his phone and called his boss on Signal. His boss answered blearily, barely awake.
“I need you to come in right now. It happened. We’re way ahead of schedule.”