Stolen encrypted drives today may be cracked by quantum computers in five years
IBM and Palo Alto Networks' CEOs say quantum computers will likely break today's encryption around 2030, and post-quantum cryptography solutions already exist—enterprises just haven't bothered to switch.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Quantum-resistant infrastructure must be deployed today, not after quantum computers arrive
Once quantum computers can break today's encryption, nearly every encryption key in enterprise IT systems becomes worthless. Worse: if an encrypted hard drive is stolen today, an attacker can store it now and crack it later once quantum computers mature in five years. Post-quantum cryptography algorithms are already written and publicly available; enterprises simply haven't deployed them. Arvind estimates that roughly half of enterprise IT budgets this year go to infrastructure without quantum resistance.
— Arvind KrishnaAI cannot break quantum cryptography—it is not an engineering problem but physics
Feynman's principle: you cannot simulate one architecture on another without losing efficiency. Quantum computers don't follow arithmetic rules; GPUs and classical computers do. AI might discover new quantum algorithms, but running them on classical hardware would be thousands or even millions of times slower—proving an algorithm exists is not the same as capturing quantum speedup. That said, if classical mathematics advances (say, a better model for prime distribution), theoretical classical methods could eventually break encryption, though no one knows the timeline.
— Arvind KrishnaSeventy-five to ninety percent of enterprise tokens will flow to open-weight models
Today's frontier models cost roughly $20 per million tokens, up to $50 at peak. An 800-billion-parameter open-source model, optimized on a GPU over a week, drops that to $1–2 per million tokens: ten to forty times cheaper. Enterprises also fear that offloading inference to a third party exposes internal IP, and frontier vendors won't hand you the weights. Add geopolitics—today's best open-weight models come from China—and Arvind predicts 75–90% of future token consumption will flow to open-weight models.
— Arvind KrishnaMythos hallucinates often, but attackers need only one real gap to win
When Mythos launched, 30% of its code-vulnerability assessments were fabrications. But cybersecurity has always been asymmetric: attackers need one winning exploit, defenders must find and patch every flaw. In practice, the model's reported vulnerabilities are about 60–80% genuine; the rest are hallucinations. That means even with the model, you must manually verify; you cannot deploy its patches directly.
— Nikesh AroraBuild and release an AI system, and your company owns the consequences
Against a backdrop of global AI arms-race reality—no one will voluntarily slow down—the realistic lever is not a research pause but product liability. You build it, you release it; you own it. Nikesh rejects the notion that ‘the model coded itself and spiraled out of control.’ It didn't. Your company did, and as a commercial entity collecting fees for the service, you must accept responsibility. Existing regulatory frameworks already have the tools to hold companies accountable for what they release.
— Nikesh AroraDistance from compute centers—not wealth—determines who can access AI capability
The real divide is not who can afford to pay, but proximity to the compute center. Technology gets cheaper; most enterprises that operate normally can likely afford AI safety and deployment. Today, compute is acutely scarce, concentrated in one or two superpowers. Countries and companies far from that center lack access not from poverty, but because supply simply cannot reach them.
— Nikesh AroraSixty-five years of IBM mainframe leadership came from owning the entire supply chain
In the 1960s, IBM integrated the entire supply chain from sand upward: sourcing materials, building wafers, constructing fabs, manufacturing chips, designing every peripheral component. Vertical integration blocked IP leakage and kept profit across the value chain. Today IBM no longer builds its own fabs, but chip design, engineering, system resilience, and most software remain in-house. Arvind brackets NVIDIA the same way—Jensen Huang owns the stack from hardware to applications and funds ecosystem partners to keep them within the vertical-integration orbit.
— Arvind KrishnaWhen enterprises use 500–5,000 software vendors, consolidation of platforms becomes inevitable
Security buyers now need 40 different products and must stitch them together. The security industry itself barely understands its own products, much less how to integrate a dozen vendors' wares. Nikesh draws a parallel to Fidelity: it once cobbled 23 systems into a CRM, then a unified platform replaced them all. A typical enterprise uses 500–5,000 software vendors—a massive attack surface. If AI can replace 490 of them and leave 5–10 core platforms, consolidation becomes inevitable.
— Nikesh AroraIn their own words · checked verbatim
So you should be replacing your IT stack today with quantum safe or quantum proof infrastructure right now.
Arvind Krishna5:14
You cannot simulate something on a different architecture and expect a more efficient result.
Arvind Krishna7:26
The best open weight models today are Chinese.
Arvind Krishna13:43
If you teach a model what good code looks like, it also can figure out what bad code looks like.
Nikesh Arora16:47
But the problem is the reason it's perfect for attackers is they just need one vulnerability. They don't need 10 vulnerabilities. If you're a defender, you need to find every vulnerability and patch it.
Nikesh Arora16:47
If you build something and unleash it, you can't unleash and say, oops, I don't know how to control it. That argument doesn't work.
Nikesh Arora23:17
The IBM mainframe has lasted 65 years. Why? It was vertically integrated.
Arvind Krishna30:21
The average enterprise uses anywhere from 500 to 5,000 distinct software vendors.
Arvind Krishna37:37
Figures
| Quantum computer breaks current encryption | ~2030 | 1:03 |
| Physical to logical qubit ratio | ~10,000 physical qubits yield ~200 logical qubits | 4:11 |
| Frontier model token cost | ~$20 per million tokens, up to $50 | 12:42 |
| Optimized open-source model token cost | $1–2 per million tokens | 12:42 |
| Predicted open-weight model share of token consumption | 75–90% | 13:43 |
| Mythos early-stage hallucination rate | 30% | 16:47 |
| Actual vulnerability rate in model findings | ~60–80% genuine, rest hallucinations | 17:49 |
| IBM mainframe lifespan | 65 years | 30:21 |
| Average software vendors per enterprise | 500–5,000 | 37:37 |
Glossary
- QBI (Quantum Benchmarking Initiative)
- A DARPA-led competition among companies to demonstrate usable quantum computers
- logical qubit
- A quantum bit that remains reliable after error correction on physical qubits; the effective clean unit of quantum computation
- post-quantum cryptography
- Next-generation encryption algorithms designed to resist attacks by quantum computers
- ZDR (zero data retention)
- A contractual commitment that a service provider will neither retain nor use customer data to train models
- open-weight model
- A model whose parameters are publicly available for download and deployment, though training data and code may remain proprietary
How to listen
People responsible for enterprise security architecture, evaluating open-source vs. closed-source model choices, or tracking SaaS M&A consolidation trends.
Around the 34-minute mark the transcript repeats and stutters; skip it—no substantive content.