Quantum computers need error correction. Today’s leading method, the surface code, spreads one clean “logical” qubit across many noisy physical qubits. More protection means more hardware.
On October 2, a Google Quantum AI team posted a preprint with a new chip design. Their circuit-level simulations suggest Google’s modular error correction approach could cut the physical qubit count by roughly tenfold or more. The authors also extrapolate to a cut of more than 30 times, but they did not simulate those larger codes. If the design holds up on real hardware, it could change how soon a useful error-corrected machine gets built, and what it costs to run. That “if” is a big one, and we cover it below.
How the modular design works
Authors Oscar Higgott, Craig Gidney, and colleagues split a processor into flat modules. Step by step, the idea runs like this:
- Inside each module, qubits mostly talk to their neighbors, much like current chips.
- The modules connect through sparse, fixed links at their boundaries.
- Those boundary links turn flat modules into hyperbolic surface and color codes, which pack more protection into fewer qubits.
- In one simulation, 120 modules and 13,920 physical qubits stored 146 logical qubits, with a simulated logical error rate of about one in ten billion per logical qubit per round.
- The simulations let the module links be noisier than the gates inside modules: 1 percent error on links against 0.1 percent inside.
That works out to roughly 95 physical qubits per logical qubit in the headline simulation. PostQuantum, which analyzed the paper, says the tenfold saving “comes from simulations of logical qubits held in memory.”
Why fewer qubits could shorten the wait
Every physical qubit needs its own wiring, cooling, and control electronics. Fewer qubits per logical qubit would ease all three. It won’t hand anyone a finished quantum computer next week, but it could shrink the gap between today’s noisy machines and tomorrow’s useful ones.
It fits a pattern we have tracked this month. Oak Ridge National Laboratory’s LuGo algorithm reports a 95 percent cut in one software bottleneck. D-Wave built an error-aware quantum simulator, now in beta.
Why It Matters to You
- Nothing is available to use. This is a preprint of simulations, and the design has not been built.
- If it works, cheaper error correction could one day let providers sell more machine-hours for less. Picture a materials scientist we’ll call Dana, an illustrative example, not a real person, who works at a small battery startup and waits for scarce, costly quantum simulation time. That day is not close, and the paper sets no date.
- The nearer-term action is on the security side: if you manage IT at a credit union, clinic, or city office, put post-quantum encryption upgrades on the planning calendar.
The encryption question
Co-author Craig Gidney published an estimate in 2025, covered by The Quantum Insider. It said breaking RSA-2048 encryption could take under a week on a machine with fewer than one million noisy qubits. That hardware does not exist today, and RSA-2048 still protects a lot of online traffic.
PostQuantum ran an illustrative check. Swapping Gidney’s memory layout for the modular one would roughly halve his 897,864-qubit total, to about 481,000. PostQuantum is clear that this is its own sensitivity check, not a Google estimate. The Google paper has no resource estimate for breaking RSA.
This is not a “your password is unsafe tomorrow” result. The paper is a preprint, and nobody has built these module links at scale on real hardware.
Ask your software vendor one direct question this week: do you have a post-quantum migration plan, and when does it ship?
What the numbers don’t show
The tenfold figure comes from simulations, and no team has run this design on a physical chip. The “more than 30 times” figure is an extrapolation to codes too large to simulate, and PostQuantum says it has the least evidence of any number in the paper because those codes have no access scheme yet. The 1 percent link error rate is an assumption that PostQuantum says is not validated on Google hardware. Real-time decoding across hundreds of modules is unsolved, and no decoding speed is reported for the hyperbolic matching decoder. The paper also does not compare against yoked surface-code storage, the method Gidney’s 2025 estimate uses, so the gain over that baseline is unmeasured. Finally, it has not been peer reviewed, and Gidney is a co-author of both the new paper and the 2025 estimate.
| Item | Figure | Basis | Source |
|---|---|---|---|
| Physical qubit saving vs. surface code | About 10x or more | Circuit-level simulation | arXiv preprint |
| Projected saving for larger codes | More than 30x | Extrapolation, not simulated | arXiv preprint, PostQuantum |
| Headline simulation | 13,920 physical, 146 logical qubits | Simulation, 120 modules | PostQuantum |
| Largest real surface-code memory (Willow, 2024) | 101 physical qubits, distance 7 | Hardware run | PostQuantum |
Quick answers
- Did Google build this chip? No. The paper reports simulations. The design has not been built or tested as a full system.
- Is my encryption about to break? No. The paper makes no estimate for breaking RSA. PostQuantum’s illustrative check suggests the qubit count for Gidney’s 2025 attack estimate could roughly halve, but the needed hardware still does not exist.
- Logical qubit: the clean, error-protected qubit a program actually uses. It is built from many noisy physical qubits working together.
- Who wrote the paper? Researchers from Google Quantum AI and Google DeepMind, led by Oscar Higgott and Hasan Sayginel, with Craig Gidney among the co-authors. It was submitted to arXiv on October 2, 2026.
Sources
- arXiv: Low-Overhead Quantum Error Correction with Boundary-Connected Planar Modules (Oct 02, 2026)
- PostQuantum: Google Modular Error Correction (Oct 06, 2026)
- The Quantum Insider: Google Researcher Lowers Quantum Bar to Crack RSA Encryption (May 24, 2025)