Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse
Krishna and Arora argue that quantum will break today's cryptography around 2030 so migration must start now, open-weight models will take most tokens, and AI will drive cyber arms races and SaaS consolidation.
Watch on YouTube ↗Charlie's summary
- ▶ 0:28
Quantum will break today's cryptography around 2030, so waiting is already too late
Arvind cites DARPA's quantum benchmarking initiative, which has shrunk from a couple of hundred companies to about 13. With a midpoint of 2030, he says breaking today's cryptography is not a supposition, it is reality. Stolen encrypted disks can be held and cracked later. Post-quantum cryptography is well known, but people have been “too lazy to get to it.” Nikesh adds that you cannot rebuild $5 trillion of IT in one year, so replace your stack now and wrap post-quantum keys around traffic as a stopgap.
- ▶ 2:30
Watch two milestones: error correction on a running machine, and gate counts adding a zero each year
Arvind says today's quantum computers are toys, rife with errors. Milestone one is someone demonstrating real error correction or mitigation, with claims likely next year and progress by 2027. Milestone two is scaling gate operations from about 10,000, and adding a zero per year would reach 10 million by 2029. He calls that remarkable. Third parties should check any claims. He accepts logical qubits as a fair way to describe effective capacity after error correction, though he doesn't love the term.
- ▶ 6:33
AI can discover quantum algorithms but cannot run them efficiently on classical hardware
Asked whether AI could break cryptography before quantum machines, Arvind argues that you cannot simulate a different architecture and expect an efficient result. AI could find a new quantum algorithm, but running it on classical hardware would be slower by a factor of a thousand or 10,000. A classical route, such as better prime-number prediction, is possible, but if it takes 10 years to break one key it is not a practical threat.
- ▶ 9:39
Use frontier models to experiment, then match each task to the cheapest intelligence that works: “you don't need a Ferrari for every task”
Nikesh says teams new to AI should first be given the best models so adoption isn't slowed. Then work out which task needs what level of intelligence. Palo Alto inspects endpoint traffic for about $20 a year per endpoint, and running that through frontier tokens would be absurd. Instead you compress a small, task-specific model and train it to 99.9% accuracy. Frontier models are reserved for cases like an active breach, where the business impact justifies it. For most enterprise uses he expects open-weight models trained on proprietary data.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.