The Delft spinoff applies reinforcement learning to magic state distillation — the process of preparing the fragile quantum states that useful algorithms require on error corrected hardware, a subroutine industry analysis pegs at up to 90% of a
MagiQware, a TU Delft spinoff, has closed a €575K (~$658K USD) pre-seed round to apply reinforcement learning to one of the most expensive steps inside an error-corrected quantum computer.
The target is "magic state distillation": a subroutine that prepares the special quantum states most useful algorithms require on error-corrected hardware. The states are fragile, so the machine has to manufacture fresh ones throughout any meaningful computation. Industry analysis widely cited in the FTQC community puts the cost of those factories at up to 90% of a full-stack machine's physical qubit footprint, making them the dominant resource bottleneck in any commercial-scale roadmap.
MagiQware's pitch is software-only. Reinforcement-learning agents discover and optimize distillation circuit architectures inside the FTQC compilation stack, without touching the underlying hardware. The company says its automated approach has demonstrated up to a 40% reduction in circuit length for target magic state factories. That figure is company-stated and has not yet been independently benchmarked.
The round was led by LUMO Labs via the TTT.AI programme, with follow-on capital from Graduate Ventures and Delft Enterprises B.V. (TU Delft's holding) to take the total to €575K. An earlier TechScout Venture Fund investment was announced by TU Delft Campus on 25 January 2026.
The founding team is an all-PhD group out of TU Delft: CEO Arash Ahmadi, CTO Shakeeb Majid, Head of Device Sahar Hejazi, and Head of Theory Ali Moghaddam. The company's homepage positions the work as compiler-level optimization, feeding automated agents into the FTQC toolchain rather than proposing a new qubit architecture. YES!Delft, the university's startup incubator, lists MagiQware among its quantum cohort.
It's a small bet on a specific technical hypothesis: that classical reinforcement learning, applied at the compilation step where distillation schedules get laid out, can cut overhead in ways hand-tuned methods have not matched. There's no public benchmark yet, and no peer-reviewed paper. No customer has been named. If the 40% reduction holds up under independent testing, the lever it pulls is on the costliest line item in any error-corrected machine: not on qubit count or gate fidelity.