Skip to content

Post-quantum · Guide

How close is a quantum computer that breaks encryption?

What the hardware can do today, how fast the estimates are falling, and why the exact date matters less than it seems.

Reviewed
October 2026
Reading time
2 minutes

Key takeaways

  1. 1No quantum computer today can break RSA or elliptic-curve cryptography at real key sizes.
  2. 2The estimated size of the machine needed has fallen roughly two hundredfold in seven years, from about 20 million noisy qubits in 2019 to proposals near 100,000 in 2026.
  3. 3Hardware roadmaps from IBM, Google and others aim at fault-tolerant machines around the end of this decade. Planning should assume the 2030s, not the 2040s.

Where the hardware is

Today's machines have from around a hundred to a few thousand physical qubits, and the hard problem is errors, not count. Recent milestones are about error correction: Google's Willow chip (December 2024) showed logical error rates falling as more qubits were added, and IBM's roadmap targets a fault-tolerant machine, Starling, with around 200 logical qubits by 2029. Quantinuum, IonQ, neutral-atom companies and others publish similar roadmaps.

How the estimates have moved

Breaking RSA-2048 needs a few thousand error-corrected logical qubits running for hours to days. How many physical qubits that takes depends on error rates and architecture, and the estimates keep falling.

  • 2019: Gidney and Ekerå estimate about 20 million noisy qubits running for about eight hours.
  • 2025: Gidney revises this to under one million noisy qubits running for under a week.
  • 2026: preprints propose architectures needing around 100,000 physical qubits, trading qubits for longer run times. These are not yet peer reviewed.
  • Elliptic-curve cryptography (ECDSA, ECDH) is generally estimated to need fewer resources than RSA at comparable security, so it should not be treated as the safer classical option.

Why the date matters less than you think

Nobody can give you Q-Day. What you can know is how long your data must stay secret and how long your migration will take. If those add up to more than about ten years, you are already inside the window for harvest-now, decrypt-later attacks, whichever forecast proves right.

There is also an asymmetry: if you migrate early and the machine is late, you have spent money on better cryptography you needed anyway. If you migrate late and the machine is early, the data is gone.

Let's talk

Start with two weeks.

A fixed-fee readiness sprint: an exposure heatmap, a Mosca analysis and your top ten priorities.