Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse

Summary · 3 min read

Krishna discusses quantum computing milestones and IBM’s vertically integrated mainframe history. Arora describes his organization’s cybersecurity red teams and endpoint defense work, and how it tests AI models for vulnerability discovery.

Krishna says a quantum computer with error correction and sufficient scale could break today’s cryptography, so organizations should begin adopting post-quantum protections before such a machine arrives. Arora favors matching AI model capability and cost to each task, while Krishna expects open-weight models to take much of enterprise AI use because they can lower costs and give companies more control over their data and models. The guests say AI will intensify cyber offense and defense, but favor collaboration and accountability over trying to stop development globally. They also foresee more vertical integration and software consolidation, while Krishna says highly reliable systems may be harder to replace with AI-generated software.

Minute by minute

1. 0:00

   Quantum could break current cryptography. Krishna says DARPA’s Quantum Benchmarking Initiative has narrowed from a couple of hundred entrants to perhaps 13, and puts the possible arrival of a cryptography-breaking machine between 2028 and 2033, with 2030 as a midpoint. Current devices remain error-prone and difficult to scale; he says error correction and increasing gate capacity are the milestones to watch.

2. 5:00

   Quantum-safe migration cannot wait. Arora says organizations should begin replacing vulnerable infrastructure now, since upgrading an entire IT stack could take years; he points to post-quantum keys as a stopgap. Krishna argues that AI might discover a quantum algorithm but classical hardware could not efficiently deliver its quantum advantage. Arora says classical AI could still aid code-breaking, but questions the practicality of a method that takes ten years to break one key.

3. 9:01

   Match model capability to the task. Arora advises employees to experiment with the best available models before optimizing costs, then match model size to each task; he says endpoint security can inspect hundreds of megabytes of traffic for $20-$25 per endpoint per year. Krishna says open-weight models can do 95% of the work for $1-$2 per million tokens after about a week of engineering, and estimates that 75-90% of token use could shift to open weights for cost, IP and control.

4. 15:16

   AI finds flaws, but defenders must verify. Arora says models trained on code can find vulnerabilities, but attackers need only one flaw while defenders must find every one. He estimates that tests in the room found roughly 60-80% real vulnerabilities and 20% hallucinations, and says defenders still need to test AI-generated patches.

5. 18:03

   Cyber defense depends on collaboration. Arora proposes that trusted organizations share vulnerabilities and distribute patches rather than making every company repeat the work. Krishna says effective defense needs alliances, lower defense costs and technological advantage. He says agents write about half his organization’s code and outperform prior work when they also run evaluations, though correctness is not fully verifiable.

6. 21:05

   AI development will not stop globally. Arora says countries are unlikely to halt AI development and argues that companies remain accountable for products they release, including software written by agents. Krishna says regulation debates reflect populism as well as concerns about long-term economic advantage; he sees no chance all 50 states ban data centers, but says 30 could.

7. 27:04

   Compute access may divide markets. Arora says the larger access risk is compute supply: companies farther from the main centers of development may struggle to get enough, even when they can afford it. Krishna argues that capital-intensive nations have tended to win past labor-capital shifts and says social choices about protecting existing jobs could determine winners; he views job displacement as potentially leading to better jobs.

8. 30:04

   Integrated stacks create strategic leverage. Krishna says IBM’s mainframe has lasted 65 years in part because of vertical integration; he describes IBM’s 1960s stack, which extended from wafers and fabs through chips, and says the company now makes everything except the fab. Arora points to NVIDIA’s integrated stack and says a company that controls the layers can shift profit between them. He also argues that an application combining reasoning, memory and context can create a moat.

9. 34:45

   AI could consolidate enterprise software. Arora says cyber customers should not have to stitch together 40 separate solutions and expects consolidation as tools become more alike. Krishna says enterprises use 500 to 5,000 software vendors and may replace 490 with minor AI capabilities and rely on 5 to 10 main platforms. He expects consolidation over a period of six months to six years, while noting that highly reliable database and accounting tasks may be harder to replace.

Published Oct 21, 2026.