I came to the United States with two suitcases, $200, and I was willing to do anything, anything at all, to make sure that I made a life for myself because there was no way to go back. I was a security guard, I took notes for the disabled, I flipped burgers at Burger King.
The breadth versus depth tension is the way Nikesh frames the entire conversation. Frontier models β OpenAI, Anthropic, Google β compete on consumer breadth. They tolerate false positives. You can write an investment memo in 4 minutes using ChatGPT and that's amazing. But that's a consumer experience.
Enterprise is different. Enterprise tolerance for false positives is close to zero. For agentic use, you need enormous context and edge-case training. That's the Waymo lesson β billions of dollars spent training every edge case until full autonomy. The frontier models can't get there without depth-specific work.
I think the long-term token pricing should be 1/10 of what it is today.
The conclusion: frontier models are competing on the consumer brand layer, but the actual enterprise revenue will come from use cases that need deep context β and Palo Alto's wager is that cybersecurity is exactly that kind of depth-context problem.
Most companies today are layering AI marginally on top of existing work β scanning an invoice with AI to extract data 20% faster. That's not transformation. The real opportunity is reimagining the workflow around AI.
SaaS applications have no opinion. AI applications will have opinions. That's a fundamental rethink we need from a workflow perspective. The AI marketing assistant is going to say: "I looked at your copy, it sucks. It's not consistent with tone of voice. Here's what I would recommend." That has an opinion.
The prediction: in the next 3 years, G&A functions β marketing, finance, HR β will have half the people they have today. But technical resources will need to grow: more people who know how to prompt frontier models, build harnesses, bring proprietary data into play.
Palo Alto has 600 people in marketing. Nikesh: "It's not going to be 600. Probably half." The biggest problem isn't headcount β it's the inconsistency of brand voice across 600 people. AI solves that.
We don't have a free-for-all model for tokens. We have a "use judiciously" model. If somebody's using it well, we won't constrain them. If somebody's gone over the top, we'll find a way to cap it.
Ben Eoff's number: $300M on Anthropic for devs = 3.8% of developer salary spend. Brandon at Macquarie's scenario: if it stays at 3.8%, Anthropic/OpenAI are grossly overvalued. If it goes to 20%, very undervalued. If it equals salary, grossly undervalued. The question: where does it land in 3 years?
Nikesh: compute is scarce. Compute costs 2-4x more than 2 years ago. And critically, more than half the compute is feeding the consumer β which is a fundamentally loss-making entity. ChatGPT, Claude, Gemini β billions of free queries every day. That's sucking away half the compute with no return.
Which means the pressure goes onto the other half β enterprise coding and applications. They pay until consumer transaction models or advertising models come together. The problem: today's compute cost is 10x what YouTube faced in the analogous "expensive free product" era. Token prices are forced up.
The long-term token pricing should be 1/10 of what it is today. In the next 3 to 5 years we will see reduction in token pricing. At some point in time the consumer use of AI will get constrained by these frontier AI companies because they have enough post-training data and each user is inherently unprofitable.
On advertising as a fix: Google started at 2% of global advertising in 2004 when total global ad spend was $500-600B. Online now captures ~70% of that pie. The total pie isn't going to suddenly explode. So you can't fund consumer AI by carving more out of advertising β most of that share is already taken.
The opening: transaction revenue. Online ad conversion rates are 1-2% average, 7-10% best-of-breed β 85-90% of marketing is wasted. AI with memory and context can target Harry exactly when he's about to buy. Consumer goods today: 5-8% is product cost, 92% is distribution and marketing. AI compresses that 92%.
Why aren't frontier model prices coming down? "All frontier model companies are value maxing, not token maxing. They've built a sustainable business model by charging more for the fastest growing thing in their portfolio." Plus the R&D cost of new frontier capabilities gets recovered through token pricing.
And on the question of who captures value: memory becomes the moat. Frontier models are aggressively integrating memory and context because they understand that's the durable competitive advantage. The risk for the orchestration layer: you can't be model-agnostic if the model holds all your context.
Mythos, I think it ends up being an accelerant to cybersecurity.
When Mythos came out, it demonstrated that all the training given to these models on how good code is written β well, the models could turn around and find bad code too. Point the gun the other way and the model says: "Look at all these flaws you have."
Palo Alto ran it against their own code. They found in 6 weeks what would have taken them 5 to 6 years to discover. But here's the catch: defenders can't fully automate the patch. False positives mean roughly 30% of the patches would be wrong. Each patch needs human evaluation, testing, sandboxing β only then can production deploy.
The asymmetry: offensive actors can daisy-chain vulnerabilities they find from the outside in. Defensive actors can't blindly patch. So Mythos creates urgency. Enterprises must improve their cybersecurity posture β which is great for cybersecurity companies broadly.
Palo Alto's defense architecture: (1) 150 million sensors at the gate infusing AI to stop bad things before they get in. (2) Intrusion response β when something gets through perimeter defense, finding and removing the bad actor quickly becomes an AI task requiring context, intelligence, enterprise-specific knowledge.
On government intervention: guardrails haven't been built robustly. People used to jailbreak models with different phrasings. The challenge is making sure the guardrails are real. National security framing is appropriate β but the solution is technical, not legislative.
Nikesh runs a meeting twice a week called AI EIO β "like Old MacDonald had a farm, E-I-E-I-O." The top 15-20 technical leaders show up. Everyone shares what they've done with AI in the last 3 days. Why this works:
When they watch their peers around them do cool [stuff], they want to show up with cool [stuff] the next time. So it creates a little bit of Darwinian competition. It creates this urge to embrace this new technology.
The two approaches to transformation: Brian Armstrong / Jack Dorsey style β decimate the org, hire 30-40% from scratch. Or Palo Alto's gradual approach: hire only through hackathons, natural 2%/month attrition replaced by AI-savvy people. 12 months β 20-25% transformed. 3 years β fully transformed.
The token-budget risk: smart employees who know how to use AI well could use 20x the tokens of an average employee. If you start whack-a-moling token spend, you hurt the best AI-savvy people more than the average employee. Use judiciously, track patterns, find the talent through their consumption.
The "Chief Internet Officer" cautionary tale: in 2004 European CEOs would hire 24-year-old "web sherpas" to handle the internet so they wouldn't have to. They washed their hands of it. The sherpas got nothing done because no one paid attention. Same risk with Chief AI Officers today.
Harry frames the debate: "I've had guests say you cannot do enterprise adoption without FTEs. Then Factory came on and said: if you need FTEs, you have a [bad] product. Shyam from Palantir chimed in too. What's true?"
We've only been chasing the enterprise dream for AI for the last 12 months. Enterprise AI products at the application layer don't exist in their entirety because they haven't been tested against the enterprise ask. FTE is shorthand for "my product's not fully there because it's evolving."
True FTEs bring code back from a customer side and incorporate it into the product because everyone will need it. That's product evolution masquerading as services. Nikesh's view: FTEs are needed for the short term β 12 to 24 months out, people will switch from one set of products to another as the market matures.
The coding market is the canonical case. Wind Surf, Devin β early players, mostly sold or pivoted. Now Codex, Claude, Antigravity. Factory and Cognition doing SDLC. The product keeps re-forming. In 2-3 years, who knows who the leader will be?
On platformization and venture returns: a billion-dollar exit doesn't move the needle for Palo Alto anymore. "Please buy any of my companies for a billion in cash. I'll give you the catalog. But I'd pay 10." Venture returns in cybersecurity still exist because attackers keep innovating β there's still 60% of cybersecurity market cap waiting in companies not yet built.
On Chinese open source models: thought experiment β take "China" out. Are open source models dangerous? No. Then it matters where they come from. Backdoors and sleeper agents are real risks, but specific to nation-state sponsorship. "That's why you come to Palo Alto to help you secure the models."
On having money β Harry: "I've become much more impatient." Nikesh's reframe: it's not about money, it's about Maslow's hierarchy. He came to America with $200 and two suitcases. Was a security guard, took notes for the disabled, flipped burgers at Burger King. "It's karma. It's destiny. There was no way to go back."
Eastern philosophy: you believe in karma, destiny. How do you manage billions of people in the world? You make sure they believe in destiny. "I tried my best. I gave it everything I had, but perhaps this is what was intended for me." It keeps you centered.
The willingness to walk away: doesn't make you softer β it optimizes the outcome. When you're fully vested in an outcome, you fold. When you're willing to walk away, it becomes a battle of wits β who wants it more.
Parenting: there's no AI training you on how to be a good father. 20-30 billion humans have lived, too many variables. Kids absorb by watching β your work ethic, your values, how you interact in micro-situations. "Your child believes that you have the best intention for them β that goes a long way."
Organizations take on the form of the leader. If my CEO is impatient, exacting, ambitious, suffers no fools, gets stuff done β that must be what they want to reward. If the leader has the right values, people will watch your behavior and emulate it.
Biggest "oh sh*t" in a board meeting: almost $1B acquisition. Took months of effort to get the founders to the table. He called a board member for advice.
Go for a long walk. Ignore all the effort you put in. Sometimes you confuse effort with wanting to get the outcome. You haven't spent a dollar yet. If this walked in the door right now with zero effort and all you had to do was write the check β would you take it?
Best advice ever received β from an old man on a flight:
If you wake up in the morning really excited about going to do what you do for a living, you're blessed. And if you're done after a long day and really excited to go home to your family, you're blessed.
Most contrarian belief about Silicon Valley today: FOMO + euphoria coupled. You had 20 years to invest in SpaceX, but only 3 to invest in Anthropic. The window from idea to $100M ARR has shortened dramatically. Many people now assume every new company is the next Anthropic. That's the risk.
Most excited for in the next 5-10 years: not specific outcomes. "The only way I've been able to do everything I do is not to get too hung up on what's going to happen a year from now or 5 years from now. Can you get your state of mind to be optimistic, positive, one of gratitude every day? Then everything else is amazing."