Source: raw/Arvind_Krishna_Nikesh_Arora_-_Quantum_Threat_Open_vs_Frontier_AI_AI_Cyber_Warfare_SaaS_Collapse.md — Creator: All-In Podcast, on-stage interview with Arvind Krishna (IBM CEO) and Nikesh Arora (Palo Alto Networks CEO), moderated by the All-In hosts · URL: https://www.youtube.com/watch?v=yzk44lz-RGI · Platform: YouTube (auto-captions, no speaker labels; fetched 2026-10-21; recorded at an All-In event in front of an enterprise audience, date not stated).
Two enterprise CEOs set out how they decide which model runs which job. Their shared answer: experiment on the best frontier model without cost controls, then move production work to small or open-weight models trained for the task, and keep frontier models for cases where the money at stake justifies them. Krishna predicts 75-90% of all tokens will go to open-weight models. The episode also covers how reliable AI vulnerability findings are, post-quantum cryptography, and why they expect enterprise software vendors to consolidate. These are executive predictions and room estimates, not measured data.
Key Takeaways
- Phase one: no cost optimization. Arora: “Don’t restrict your teams in terms of capability in the experimentation phase. Don’t start trying to cost optimize.” Use “the best model you can find the easiest way to use it” until the use case is proven.
- Production: size the model to the task. Palo Alto inspects endpoint traffic for about “$20 a year” per endpoint with on-device machine learning, which frontier tokens could not match. Its answer is a small language model, made task-specific and trained “until it gets to 99.9% accuracy”. “You don’t need a Ferrari for every task and by the way the customers don’t want to pay for a Ferrari for every task.”
- Frontier models for high-stakes moments. During an active breach, where losses run to billions, Arora would run “agent swarms” on “the most up-to-date frontier LLM”. Peacetime defender models will be open-weight models trained on Palo Alto’s own data, with weights it manages, for “auditability” and “explainability”.
- The token arithmetic (Krishna). About “2 for a million tokens”, a “40 times cheaper answer for effectively what can do 95% of the work.”
- Three reasons tokens move to open weights (Krishna): economics; IP, because a logged chain of thought means “you’re giving away your IP”; and geopolitics. He sees “very little danger” in Chinese open-weight models “if you’re putting them in a sandbox and running them by yourself”, and expects Western open-weight parity within “6 to 12 months at the most”. Prediction: “75 to 90%” of all token usage.
- Put a control on defaults. “Developers are lazy they’re going to use the most expensive … answer unless you put a control on it.”
- AI vulnerability findings need retesting. Arora says Mythos “was hallucinating 30% of the time” when it first looked for bad code; his show-of-hands estimate for the room was “roughly 60 to 80% have been real and 20% have been hallucinations”, and nobody would let Claude Code patch “without having to test it again.”
- Agent-written code at IBM. Krishna: coding agents write “probably half the code” at IBM and it is “superior to what was being written before” because “the agents can also do eval”, but “we’re only finding all the flaws that we know how to find.”
Details
Zero data retention has gaps (moderator). A moderator notes that ZDR terms do not say providers won’t generate synthetic data from your traffic: “There’s all these outs. There’s all these trap doors.” Arora adds that many people use Claude Code without ZDR, which is part of why models got good at spotting bad code.
Collective defense. Arora proposes pooling vulnerabilities across companies through trusted parties, patching once and distributing the fix, citing a program Krishna runs. Krishna’s framing of cyber as warfare: alliances, making defense as cheap as attack, and technological superiority.
Quantum. Krishna says DARPA’s Quantum Benchmarking Initiative is down to “maybe 13” of a couple of hundred entrants; estimates of a cryptographically relevant machine range over 2028-2033, “let’s call it 2030”. An encrypted disk stolen today could be broken in five years. Post-quantum cryptography is understood: “the code is written but people have been too lazy to get to it.” Arora: replace infrastructure with quantum-safe equipment now, since he guesses about 50% of this year’s infrastructure spend in the room is not quantum-safe. Krishna doubts AI on classical computers will break cryptography in practice.
Application-layer lock-in (Arora). Context and memory cannot be separated from reasoning, so routers like OpenRouter become harder to use and a trained personal agent is a moat. “In the next three months, every frontier lab is going to have a full-on full court press on the Muse concept”, and “every app that exists on the iPhone … has to become headless.”
SaaS consolidation (Krishna). “The average enterprise uses anywhere from 500 to 5,000 distinct software vendors.” He expects consolidation to “5 to 10 main platforms”. Interaction-heavy software can be replaced with AI; systems that must be reliable “at 10 to the minus 15”, such as retail-banking lookups or closing the books, will be hard to replace.
A parameter count to treat with care. Krishna’s example model is “a Kim K3 something 800 billion parameters”. The wiki’s Kimi K3 article records 2.8 trillion parameters from two sources and says self-hosting it needs about a full NVL72 rack, so his figure is probably a slip or refers to a smaller model.^[ambiguous] His per-token cost estimate should be read as a general open-weight claim, not a Kimi K3 benchmark.
Try It
- Write a two-phase model policy. Phase one: any frontier model, no cost gates, while a use case is being proved. Phase two: once a task is stable and high-volume, test whether a small or open-weight model reaches your accuracy bar, and move it if it does.
- Set a default model below the top tier and require a reason to escalate, per Krishna’s “put a control on it”.
- Treat every AI-reported vulnerability as a hypothesis. Reproduce it and retest after the patch, whichever model found it.
- Read your provider’s data terms for synthetic-data and training carve-outs, not only for retention.
- For marketing teams: high-volume routine work (tagging, first-pass drafts, classification) is the candidate for a cheaper model; keep frontier models for work where a miss is expensive.^[inferred] See Cost and Intelligence Levers.
Open Questions
- The 75-90% open-weight share and the 6-12 month parity gap are Krishna’s predictions with no stated basis beyond his three drivers.
- The 60-80% real-findings figure is a show of hands in one room, not a study.
- Which model Krishna meant by “Kim K3 something 800 billion parameters” is unclear.
Related
- Cost and Intelligence Levers — routing and model-tier choice
- Chinese Models Decision Framework — when Chinese open-weight models are acceptable
- Kimi K3 (Moonshot) — the open-weight model Krishna names
- Qwen 3.8 27B Local Inflection — the small-model side of the same argument
- Epoch: Mythos Cyber Capabilities Overhyped — another view on AI vulnerability-finding claims
- The Find-Fix Asymmetry — attackers need one flaw, defenders must patch all
- Mythos Preview Cryptanalysis — AI and cryptography from the model side