All-In Podcast

Krishna and Arora on why you defend what you can't stop

Arvind Krishna & Nikesh Arora· CEOs of IBM and Palo Alto Networks at IBM & Palo Alto Networks
·~40 min·English·All-In
AI SafetyAI CompanyPolicyBusiness StrategyOpen Source
TL;DR

The CEOs of IBM and Palo Alto Networks argue that quantum computers, frontier AI, and automated attackers cannot be slowed down, so the winning move is to make defense cheap, built-in, and vertically owned, starting now.

01The Quantum Clock

The 2030 Quantum Cliff

A working quantum computer around 2030 will break today's cryptography, and data stolen now can be decrypted the day it arrives.

if you have an encrypted hard disk today that gets stolen, somebody could keep it for 5 years and break it at that point.

— Arvind Krishna, All-In Podcast
Key Insight
The threat is not in 2030 — it is today. Anything long-lived that you encrypt now (records, keys, archives) is already harvestable, so the migration deadline is set by your data's shelf life, not by when the machine finally boots.
02The Hardware Limit

A Clever Algorithm Is Not a Shortcut

AI might discover a code-breaking algorithm, but you still need quantum hardware to get the speedup — a classical or GPU machine runs it a thousand to ten thousand times slower.

You cannot simulate something on a different architecture and expect a more efficient result.

— Arvind Krishna, All-In Podcast
Key Insight
This is the reassuring half of the quantum story: the threat is bounded by physics, not just by software cleverness. It also reframes the whole AI-breaks-crypto panic — proving a thing is possible and getting its practical benefit are separated by an architecture gap that does not close for free.
03Model Economics

No Ferrari for Every Task

Routine security runs on small task-specific models at tens of dollars a year, and the frontier LLM comes out only when a breach makes the business impact worth it.

you'll suddenly discover 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.

— Nikesh Arora, All-In Podcast
Key Insight
The real engineering skill here is triage, not raw capability: deciding what level of intelligence each task actually requires and what that intelligence costs. Running all endpoint traffic through a frontier API would, in Arora's framing, hand one vendor a company's entire year of revenue.
04Open vs Closed

Why Open Weight Takes the Lion's Share

Economics, IP control, and geopolitics all point the same way, pushing the large majority of token usage toward open-weight models you run yourself.

I don't know whether it'll be 75% to 90%, but that's the range I would put of all token usage will go to open weight models

— Arvind Krishna, All-In Podcast
Key Insight
Notice the three forces are independent — cost, IP leakage through logged chain-of-thought, and the fact that the strongest open-weight models are currently Chinese — yet they reinforce the same conclusion. When unrelated pressures agree, the outcome tends to be durable, not a passing swing.
05The Core Asymmetry

The Attacker Needs One Hole

An attacker wins with a single vulnerability while the defender has to find and patch every one, so defense only scales when rivals pool what they find.

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 Arora, All-In Podcast
Key Insight
This asymmetry is why AI helps the attacker first: a model that is right most of the time is enough to find the one hole, but a defender cannot ship a patch that is right most of the time. The structural fix is collective — turn competitors into an alliance that patches once and distributes to all.
06Accountability

No, You Wrote It

Shipping an AI system means owning what it does, and blaming the agent that generated the code does not transfer the liability off the company.

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 Arora, All-In Podcast
Key Insight
For agent builders this is the governance line in one sentence: autonomy in the toolchain does not dilute responsibility at the company boundary. The moment you take economic value for a product, the fact that agent number so-and-so emitted the code is an implementation detail, not a legal shield.
07Inevitability

Nobody Is Slowing Down

No country pauses AI development, so a unilateral slowdown just lets everyone else catch up and get ahead.

I think it's abundantly clear that nobody's stopping. Everybody seems to be working at a frenetic pace to try and make sure that they get a better, smarter, faster model.

— Nikesh Arora, All-In Podcast
Key Insight
The pair's point is not that regulation is wrong but that a one-sided brake is self-defeating — the same logic that made aviation add security after the planes existed, not before. If the threat cannot be paused, the only lever left is to out-build the defense, which is the whole thread of this conversation.
08Market Structure

All the Way to Dirt

Owning the full stack lets you choose which layer to commoditize and move the profit to where you keep control, which is why vertical integration is back.

if you control the vertical stack, you can decide which part of the stack to commoditize. So, you can decide to move profit between layers.

— Nikesh Arora, All-In Podcast
Key Insight
Carry this down to the application layer and it explains the agent-moat race: a personal agent loaded with your context, APIs, and memory becomes a floor no competitor can cheaply relocate you out of. Vertical integration is not nostalgia for the conglomerate — it is control over where margin is allowed to collect.