Fault-Tolerant Quantum Computer
A quantum computer with hundreds to thousands of error-free logical qubits that solves tasks classical computers cannot.
Open in the interactive tree →Useful quantum algorithms (chemistry, materials, factoring) need millions of gate operations at extremely low error rates, which is only possible with continuous error correction. Today logical qubits exist in the dozens, but not yet in the size and quality these applications require. IBM promises a first fault-tolerant machine, Starling, for 2029.
As of October 2026
As of October 2026 the reported records are a few dozen up to 96 logical qubits (company-reported) with errors of roughly 1e-3 to 1e-5 per cycle; useful tasks need many orders of magnitude lower. Estimates for breaking RSA-2048 fell from 20 million physical qubits (2019) to under 1 million (Gidney, May 2025) and, per an Iceberg Quantum preprint (February 2026, simulation only, not independently confirmed), under 100,000; Google estimates under 500,000 for elliptic curves (March 2026). IBM plans “Starling” (200 logical qubits, 100 million operations) for 2029, and DARPA is checking viability by 2033.
What is missing
- Hardware scaling to 100,000-1,000,000 physical qubits (fabrication yield, wiring, cryogenics, helium-3 supply for superconducting chips)
- Real-time decoders that evaluate errors on microsecond timescales
- Logical gates with low overhead, especially T gates/magic states (shown only at small scale)
- Logical error rates many orders of magnitude below today’s 1e-3 to 1e-5 per cycle
- Proof of economically useful algorithms beyond factoring and simulation
Becomes possible once solved
- Quantum-level calculation of catalysts and drug candidates
- Designing new materials (batteries, superconductors) by simulation
- Breaking RSA/ECC (Q-Day)
- Linking quantum computers into a network
Open steps
- Real-time decoders at scale High AI leverageDecode errors for thousands of logical qubits within microseconds on dedicated hardware, beyond today's distance 5-11 demonstrations.
- Low-overhead error-correcting codes Medium AI leverageFind codes and logical-gate schemes that need far fewer physical qubits per logical qubit than the surface code and fit real chip connectivity.
- Cheap T gates and magic states High AI leverageCut the qubit and time cost of non-Clifford (T) gates, which dominate useful algorithms; only small-scale demonstrations exist.
- Auto-calibration of huge chips High AI leverageKeep the thousands of control parameters of a 100,000-qubit machine tuned while it runs, without stopping the computation.
- Qubit fabrication yield and defects Low AI leverageMake very many near-identical qubits with few defects and uniform frequencies, plus the wiring and cryogenics to run them.
Where AI could help
Medium AI leverage. AI already improves error decoders, circuit compilation and code search, but the core bottleneck is building 100,000+ low-error physical qubits.
- Train neural decoders on simulated and real syndrome data to lower logical error rates
- Search for lower-overhead error-correcting codes and gate sequences, for example fewer T gates
- Calibrate and drift-correct thousands of control parameters while the chip keeps running
- Optimize compilation of algorithms to cut physical qubit and gate counts
Shown so far
- In November 2024 Google's AlphaQubit neural decoder made 6% fewer errors than tensor-network decoding and 30% fewer than correlated matching on Sycamore data, but is still too slow for real-time use. source
- In July 2026 Google Quantum AI reported that reinforcement learning kept logical error rates about 20% lower and 3.5 times more stable than periodic recalibration on its Willow chip (secondary report). source
- In March 2025 AlphaTensor-Quantum (Nature Machine Intelligence) cut T-gate counts in quantum circuits and matched the best human-designed solutions for arithmetic used in Shor's algorithm and chemistry. source