Quantum AI
Error-corrected quantum hardware is now an engineering programme, not a physics bet, but quantum machine learning remains a research field. The urgent, budgeted work is migrating the world's public-key cryptography before the machines, and the regulators, arrive.
The sovereign stake
Three regulatory clocks are already running. NIST will disallow RSA and elliptic-curve cryptography in US federal systems after 2035, deprecating them from 2030; the UK NCSC requires a full cryptographic inventory and migration plan by 2028; the EU wants critical infrastructure quantum-safe by 2030. Harvest-now-decrypt-later means diplomatic traffic and citizen records intercepted today become readable the day a cryptographically relevant machine exists, and most expert estimates cluster in the early-to-mid 2030s. For Gulf states investing at sovereign scale, including Qatar's up-to-$1bn Quantinuum venture and Aramco's 200-qubit deployment, the harder question is whether procurement builds migration capability, or just laboratories.
The board-level stake
For boards, the quantum question is not "when do we buy a quantum computer". It is "which of our secrets must still hold in 2035, and what encrypts them today". Certificates, HSMs, VPNs, code-signing keys, supplier firmware: most enterprises cannot list them, and migrating a large estate takes the better part of a decade. Regulators have converted a physics forecast into compliance deadlines, and counterparties are already writing post-quantum clauses into contracts. On quantum machine learning: fund a small watching brief, demand classical baselines from any vendor, and spend the real money on cryptographic discovery.
The hardware story changed in late 2024, and it changed in one specific way: error correction started working. Google's Willow showed logical error rates falling as code size grew: the below-threshold result the field had chased for two decades. Quantinuum's Helios now runs 48 error-corrected logical qubits from 98 physical ones. IBM has committed, in public and in silicon, to 200 logical qubits running 100 million gates by 2029. These are engineering milestones, not physics hopes. But keep the denominator honest: breaking RSA-2048 needs thousands of logical qubits and days of stable runtime. The machines are real; the cryptographically relevant machine is not yet. Treat vendor roadmaps as credible and every date on them as soft.
These are engineering commitments with dates and silicon behind them, not physics hopes, which is what makes them plannable. The number that matters for procurement is the logical qubit, not the physical one: a single logical qubit today is assembled from roughly a hundred physical ones, and it is the logical count that decides what a machine can actually run.
On quantum machine learning we will say what vendors will not: there is no evidenced near-term quantum advantage for learning on classical data. Serious benchmarks through 2025-26 show well-tuned classical baselines matching or beating variational quantum models, and the trainability problems have not been engineered away. The traffic currently runs the other way: AI is transforming quantum computing, not the reverse. DeepMind's AlphaQubit outperforms conventional error-correction decoders; NVIDIA's NVQLink now couples GPUs to Quantinuum hardware for real-time decoding. The first real quantum-AI value chain is classical machine learning making quantum machines viable. Where quantum wins first is simulating quantum systems (chemistry, materials, spectroscopy), not your recommendation engine, and not your language model.
Post-quantum migration is the item that belongs on ministerial and board agendas this quarter, because the deadline is set by regulators and data lifetimes, not qubit counts. NIST finalised FIPS 203-205 in August 2024 and will disallow RSA and elliptic-curve cryptography in US federal systems after 2035. The UK NCSC requires discovery and a migration plan by 2028, high-priority upgrades by 2031, completion by 2035. The EU wants critical infrastructure quantum-safe by 2030. Those dates are closer than they look: large estates take eight to ten years to migrate. And harvest-now-decrypt-later means any secret that must hold into the 2030s is, if intercepted today, already compromised. This is not a cryptography project; it is an asset-discovery project.
The Gulf has moved from observer to participant. Qatar's Al Rabban Capital signed a joint venture with Quantinuum worth up to $1 billion; Saudi Aramco and Pasqal put a 200-qubit neutral-atom machine into industrial operation; the UAE has folded quantum into a billion-dollar AI-and-quantum allocation. The test we apply to these programmes is the Complex AI test: does the system govern as well as it demos? Quantum computing inside a national infrastructure is complex AI territory: multi-objective, high-stakes, and worthless without traceability. A programme that buys hardware but cannot execute a cryptographic inventory of its own ministries has the priorities inverted. The machines can be rented. The migration discipline cannot.
- Whose clock governs your planning, IBM's 2029 fault-tolerance roadmap or NIST's 2035 ban on RSA, and what happens if the first slips while the second does not?
- How does a government that has never inventoried its own cryptography complete discovery by the NCSC's 2028 milestone?
- Which quantum-ML claims survive a well-tuned classical baseline, and how should public procurement be written to force that test?
- Do sovereign quantum investments in the Gulf buy durable capability (talent, error-correction expertise, migration discipline), or hardware photographs?
- Can a post-quantum migration be made auditable end-to-end (certificates, HSMs, firmware, supplier chains), rather than merely declared complete?
Fellows for this frontier are being appointed.
People who have built, governed or operated real systems in this domain. The first five fellows are named, and the rest of the founding cohort follows in September.
Have you done this at scale in Quantum AI? We want to hear from you.
Put yourself forward