EXECUTIVE SUMMARY · UNGATED
Can quantum computing accelerate LLM training?
Every few months a headline says quantum computers are about to make AI training cheap. This is a reality check, then a map — grounded in our own experiment, not vibes. The full read is the whitepaper; the data lives in the repository; the studio is Advance Labs.
Quantum computing cannot accelerate large-scale LLM training today. The most rigorous analyses put meaningful impact a decade or two away — into the 2040s on the pessimistic end. Three walls stand in the way. But a real, dated road exists, and quantum-inspired math already pays off on ordinary GPUs.
Four takeaways
The hype is wrong about timing.
An estimated ~10¹³ aggregate hardware handicap, the unsolved data-loading problem, and dequantization push practical quantum LLM training well past the 2030s. Our own 6-qubit hybrid lost to its parameter-matched classical control.
But quantum-inspired methods already pay off — on classical hardware.
Tensor-network compression (with quantization) shrank a 7B model's memory ~93% and halved its post-compression recovery-retrain. No quantum computer involved.
The roadmap is concrete, not hand-wavy.
Fault-tolerant milestones now carry named processors and dates through 2033+ — enough to reason about when narrow quantum speedups could enter the pipeline.
The winning posture is optionality.
Don't wait for fault-tolerance, don't dismiss it. Harvest what works now; architect so the hardware can slot in later.
The milestone map
Google Willow
“Below threshold” error correction — error rate falls as the code grows.
IBM Kookaburra
First fault-tolerant module (vendor roadmap).
IBM Starling
~200 logical qubits, 100M+ operations (vendor roadmap).
IBM Blue Jay
2,000+ logical qubits, billion-gate scale (vendor roadmap).