Nikesh Arora

Jason Calacanis says AI found two critical exploits, nicknamed PHP to shell, in PHP, the open source language that powers about 70% of websites, including WordPress and Wikipedia. The PHP security team rated them 9.2 out of 10.

The plan was one major patch on October 22, the day of the recording. The details leaked, so PHP moved the release up to Wednesday, October 21. Jason Calacanis says AWS then reported a jump in hacking activity, and the Cloudflare CEO said they saw millions of hacking attempts. Governments in France, Germany, Netherlands, Denmark, and Finland started taking websites down. Wikipedia took some systems down to patch them. In the United States, the NSA and the FBI put out

Nikesh Arora says companies avoid upgrading old software because technology leaders are rewarded for low cost and for keeping everything available, so an upgrade risks an outage, user complaints, and more spend.

Cyber security is usually 4% to 10% of the information technology budget, and Nikesh Arora says that share is going to have to go up. Financial services and the federal government sit near 10% because they want systems secure and robust. Companies already on the cloud can land at 4% to 5%, because a hyperscaler covers infrastructure security and the company mainly handles application and software security. He puts the band from 4% to 5% on the low end up to 8% to 10% on the high end

Nikesh Arora says AI with no guardrails is dangerous because safety depends on the weakest link, and a company that gives out the raw power of AI with no controls hands a nuclear bomb to anybody who wants it.

He points to debates in the White House on responsibility, alignment, and testing models beforehand. At one level, he says, people want global treaties to manage the pacing and development of AI for good. At the same time, a company can still offer the most powerful tool with no control, no guardrails, and no governance. He compares the offer to a do-it-yourself aspirin kit for building medication, or to handing someone powerful tools, and says he does not understand the business model

Nikesh Arora says the AI compute profit pool lives in NVIDIA and the memory industry and is shifting toward hyperscalers that already hold large amounts of compute. Availability of compute will matter more than the chips. Demand is insatiable, with more demand than Nikesh Arora and Chamath Palihapitiya can imagine.

Eight years ago, when he came to Palo Alto, he shut down all the data centers. The company could not run a global footprint at the scale, speed, and reliability he could rent from hyperscalers. Compute is now highly scarce, so he takes terms rather than dictates them. Building his own data centers is an empty threat. Nikesh Arora says Jensen cannot give him chips for 2028, and la

Nikesh Arora says Palo Alto Networks prefers open source, open weight models because the company can modify the weights, train them, and interrogate the chain of thought.

Palo Alto Networks wants to embrace AI as quickly as many startups. After comparing open source models with cloud models, the company chose open weight models. Over time, enterprises will need explainability and auditability, 100%. In financial services and healthcare, a team cannot tell a boss that nobody knows what happened and the IT group simply put an AI model in place, because a regulator will not accept an answer with no explanation. Financial services and healthcare want explainability, verifiability, and auditabil

Nikesh Arora says there is going to be a huge disruption in the SaaS space over the next few years. Arvind Krishna builds on that point. Krishna says the average enterprise uses anywhere from 500 to 5,000 distinct software vendors, and that stack is a massive attack surface, not only a cyber problem.

Krishna asks whether 490 of those vendors could be replaced with minor AI capabilities, leaving 5 to 10 main platforms. The consolidation is where the spend goes. If distribution cost has to stay low and the value of each product is falling, a company cannot afford independent distribution for every vendor. Krishna says the change plays out somewhere between six months and six years.

Nikesh Arora says a model that has learned what good code looks like can also point at bad code. Attackers only need one real vulnerability. Defenders have to find every one and patch it.

When Mythos first looked for bad code, Arora says it was hallucinating 30 percent of the time. He guesses that among people who have tested those findings, roughly 60 to 80 percent have been real and 20 percent have been hallucinations. He does not think anyone will patch from a flawed suggestion without testing it again.

Nikesh Arora says that in the experimentation phase a company should use the best model it can find, in the easiest way. He says not to restrict the team, not to cost-optimize, and not to stitch a private stack yet.

Two teams at his company, twenty people rather than twenty thousand, have just discovered AI in the last twelve to eighteen months, and adoption is not fully diffused. Making the tools harder slows that down. Arora calls the easy-adoption stretch phase one. After that, different tasks need different models. Inspecting all traffic at an endpoint costs twenty dollars a year, or twenty to twenty-five dollars a year. Sending that traffic through a token-priced model is not practical

Arvind Krishna puts a working quantum computer that breaks today's cryptography at 2030. The debate he describes is 2028, 2029, or 2033, and he takes the middle of that range.

A machine on that timeline will break today's cryptography. Krishna says the break is not a supposition. If an encrypted hard disk is stolen today, somebody could keep it for five years and break it then. Post-quantum cryptography is understood, the code is written, and he says people have been too lazy to move to it.

DARPA's Quantum Benchmarking Initiative started with a couple of hundred companies. Krishna says about 13 are left. Some of that work is real and some of it is not quite real. The goal is a quantum comp

Arvind Krishna says replacing a database lookup for retail banking, or the work of closing the books, with AI-generated software is going to be hard. He puts the reliability bar at 10 to the minus 15. The examples already on the table were interaction rich, and he treats a lookup or a close as a different case.

Krishna says there is going to be a distinction in how easy replacement is. A digital twin would have to run for years before a cutover. All-In's reply is that there is no cost advantage to doing that.

Interaction-rich work sits on the easier side of Krishna's distinction. Retail banking lookups and closing the books stay harder to hand to AI-generated software when the reliability

Arvind Krishna says there is no chance all 50 states will ban data centers. A ban in 30 states is possible, and the remaining 20 would rub their hands in glee.

Krishna says the regulation question on AI and data centers is driven by populism. A politician who can win more votes by looking ready to act will feel that pressure, and the pressure is real. Enough people also care deeply about the long-term economic advantage, so policy is a balancing act. Krishna says the country is on the correct balance right now, though an election can change where it falls.

The question has been active since 2022. The Senate held closed hearings in 2022 and open hearings in 2023. Congress held hearings too,

Arvind Krishna says agent-written code will be superior to human-written code. Coding agents now write probably half the code where he works, and he says that code is already superior to what was written before.

The advantage is evals. Agents find flaws and fix them before the code goes out. They still only catch flaws people already know how to look for, so the code is not fully verifiable. Krishna says full verification is almost computationally infeasible right now. A way to check correctness might appear when the inputs are limited. When an input can be anything, he thinks the check is really hard.