October 21 2026
(from Gemini)
- Arvind Krishna (Chairman & CEO of IBM) and Nikesh Arora (Chairman & CEO of Palo Alto Networks)
- The Quantum Threat & Post-Quantum Cryptography: Referencing DARPA’s Quantum Benchmarking Initiative (QBI)—which has narrowed a field of hundreds of companies down to roughly 13 serious contenders—Krishna estimates a 2030 midpoint for when working quantum computers will break today's standard cryptography. He outlines two critical milestones to watch: first, real-world quantum error correction on a running machine (expected around 2027), which reduces noisy physical qubits into effective "logical qubits" (e.g., 10,000 physical qubits yielding ~200 logical qubits); second, scaling gate computations by an order of magnitude each year from 10,000 today to 10 million by 2029. Because bad actors can steal encrypted hard drives today and wait five years to decrypt them ("harvest now, decrypt later"), and enterprises cannot replace $5 trillion of global IT infrastructure overnight, Arora warns that companies are already late and must begin deploying post-quantum cryptography immediately. Addressing whether classical AI models could break encryption before quantum hardware arrives (following OpenAI's recent math breakthroughs), Krishna cites Richard Feynman to argue that classical arithmetic hardware (including GPUs) cannot efficiently simulate quantum algorithms without suffering a 1,000x to 10,000x slowdown, while classical number-theory workarounds would likely take far too long per key to be practical.
- Open-Weight vs. Frontier Models & Enterprise Token Economics: Both CEOs advocate letting employees use top-tier frontier models during Phase 1 experimentation so friction doesn't slow down internal AI adoption, before shifting to ruthless optimization in production. Arora illustrates the unit economics with endpoint cybersecurity: Palo Alto Networks inspects hundreds of megabytes of traffic per endpoint for just $20–$25 a year using local machine learning on a 20–40 MB payload. Routing all employee endpoint traffic through frontier API tokens would bankrupt an enterprise; instead, companies must compress and train task-specific small language models (SLMs) to 99.9% accuracy for "peacetime" defense, reserving expensive frontier LLMs and agent swarms for high-stakes "wartime" active breaches where billions of dollars are at risk. Krishna reinforces this through three lenses: first, economics—taking an 800B-parameter Chinese open-weight model (like Kimi K3) and spending a week of engineering work on self-hosted GPUs drops inference costs from $20–$50 per million tokens down to $1–$2 per million tokens for 95% of tasks; second, IP protection—preventing third-party providers from logging proprietary chain-of-thought reasoning; and third, geopolitics—Running sandboxed open-weight models locally while Western open-weight models close the 6-to-12-month gap with Chinese models. Consequently, Krishna predicts 75% to 90% of all enterprise token volume will ultimately run on open-weight models.
- AI Cyber Warfare, "Mythos" & Agentic Code Verification: Discussing early previews of frontier coding and cyber models like "Mythos," Arora explains why AI labs won't easily displace dedicated security vendors. Because millions of developers use coding agents without Zero Data Retention (ZDR)—and many ZDR agreements contain loopholes allowing synthetic data generation—frontier models have learned what both good and bad code look like. However, when scanning codebases for vulnerabilities, these models hallucinate 20% to 30% of the time. This creates a massive asymmetry: an attacker only needs one real vulnerability to succeed, whereas a defender must verify, test, and patch 100% of real flaws while filtering out hallucinations. Applying lessons from military history, Krishna and Arora outline three pillars of cyber defense: forming alliances to pool vulnerability intelligence and distribute patches globally; eliminating cost asymmetry so defense is as cheap as offense; and achieving technological superiority. Krishna notes that agents now write roughly half of IBM's internal code, producing higher quality than human engineers because automated evals catch known flaws prior to release—though formal mathematical verification of software across unbounded inputs remains computationally infeasible.
- Geopolitics, Data Center Populism & Capital vs. Labor: Rejecting the idea of an international AI pause, both leaders note that unilateral slowdowns only hand an advantage to global competitors like China, India, or Japan, while pointing to Europe's precautionary regulatory regime as a cautionary tale that stifled its domestic tech sector. Arora argues that software creators are commercially accountable for the products they unleash—dismissing the excuse that "an autonomous agent wrote it." On Washington's regulatory mood, Krishna views threats like a rumored congressional push for a 50-state data center moratorium as political populism rather than reality; even if 30 states banned data centers, the remaining 20 states would gladly absorb the investment. Looking globally, Arora warns that the true "haves vs. have-nots" divide won't be software pricing, but physical access to compute supply outside the US and China epicenters. Krishna frames this macro shift through the historical lens of the 1700s–1900s industrial revolutions: nations willing to make massive capital-intensive bets and accept job displacement (which leads to higher-value roles) always defeat nations governed by rigid, labor-protectionist philosophies that try to freeze every existing job in place.
- Vertical Integration vs. Horizontal SaaS Consolidation: Examining how market pendulums swing back toward vertically integrated conglomerates during turbulent eras, Krishna shares the history of the 65-year-old IBM Mainframe: in the 1960s, IBM integrated "all the way to dirt"—mining sand, making wafers, running fabs, and designing chips, hardware, and software to prevent IP leakage and capture the entire profit pool. Arora points to Nvidia as the modern equivalent, funding the entire AI ecosystem from silicon to Hugging Face so Jensen Huang can dynamically shift profit pools across layers whenever competitors attack one part of the stack. At the AI application layer, Arora predicts open routers will fail because reasoning cannot be cleanly separated from persistent memory and context; as personal and enterprise agents accumulate deep context and API connections, switching stacks becomes too high-friction, creating vertically integrated end-to-end moats where brute-force distribution wins and smartphone apps are forced to become "headless." Finally, on M&A and the "SaaS collapse," both CEOs foresee massive horizontal consolidation. Just as 23 disparate CRM tools thirty years ago consolidated into Salesforce, enterprises today cannot manage the cyber attack surface or distribution overhead of 500 to 5,000 distinct software vendors. While interaction-rich UI software will be easily replaced by minor AI capabilities (consolidating spend into 5 to 10 core platforms), Krishna notes that core systems of record requiring $10^{-15}$ reliability—such as retail banking ledgers and financial close databases—will remain deeply entrenched for years.
Comments