Nikesh Arora on Enterprise AI, Cybersecurity and Model Economics

Palo Alto Networks chairman and CEO Nikesh Arora tells Harry Stebbings that his operating question is how to make things incrementally better today and radically better over several years. The answer is not always success, he says, but repeated attempts can improve the odds.

Stebbings argues that a founder's personal brand can accelerate a business. Arora places product and execution first: a strong brand cannot indefinitely compensate for a weak product. He acknowledges that branding matters more when a product is difficult to distinguish from alternatives.

Arora frames the frontier-model problem as breadth versus depth. Consumer models can answer a wide range of questions, but enterprise agents acting independently may need much more context, testing and domain-specific information than a general model has.

He argues that people sometimes tolerate errors in consumer interactions because a human can review the answer, whereas an autonomous enterprise workflow may have less room for incorrect outputs. That contrast is his framing, not a finding that consumers are safe from model mistakes.

He uses autonomous driving as an analogy for depth: a system needs extensive training on unusual situations before it can perform a bounded task reliably. His cost estimates and comparisons are illustrative, not an independently verified account of any provider's development spending.

Arora says many enterprises are adding AI to existing processes instead of redesigning them. He expects the larger opportunity to come from rethinking workflows, though he acknowledges that this requires decisions about where humans retain control and where systems may act.

In hiring, he imagines a system helping staff review applications, suggest interview questions and coordinate decisions. This is a proposed use case, not proof that an AI system would be accurate or fair enough to make hiring decisions without appropriate oversight.

Stebbings raises employee resistance to monitoring. Arora distinguishes broad data collection from selectively redesigning enterprise workflows and says people need to decide which decisions they are willing to delegate. His distinction does not resolve the privacy or labor questions involved.

He sees marketing as an example where models could help check whether content matches a company's tone and message. He predicts that general and administrative functions could require fewer people over three years, while expecting greater demand for technical and sales resources. These are his forecasts, not announced staffing outcomes.

Arora says teams need people who can work with models, company data and product infrastructure. His concern is not just replacing an existing tool; it is finding resources to build new workflows and deciding which capabilities are particular to the organization.

On token budgets, he says many employees are still learning and warns that a blanket cap could constrain the most productive users. His estimates of employee AI proficiency and attrition are his own company or market observations, not independently verified workforce statistics.

Palo Alto Networks, he says, uses tokens judiciously rather than operating an unrestricted free-for-all: usage is monitored, useful work is allowed and excessive use may be limited. He prefers building proprietary capabilities where the company has distinctive data and waiting for generic applications to mature.

Asked what share of developer pay will be spent on AI, Arora points to compute scarcity and says current pricing makes the answer difficult to predict. His claims about consumer usage, profitability and the allocation of compute are estimates and are not verified financial disclosures from model providers.

He predicts token prices could fall substantially over the next three to five years as compute and model efficiency improve. He also expects use to rise as prices fall. Neither the magnitude nor timing is guaranteed, and individual model prices may move differently.

Stebbings asks whether advertising could finance consumer AI. Arora doubts that the total advertising market will grow enough on its own and sees possible transaction-based revenue instead. His advertising and conversion figures are illustrative estimates, not an audited market analysis.

Arora says model providers must fund research, training and service costs, which influences current prices. He expects technical advances and smaller models to handle many tasks more efficiently, while acknowledging that demand for compute could still increase.

Asked about an infrastructure bubble, he says he does not know when physical constraints on data centers, energy and equipment might cause a pause or adjustment. His scenario distinguishes demand for compute from the ability to build capacity at the desired speed.

He expects value to be shared between frontier models and the enterprise context built around them. Memory and access to relevant organizational information could make applications more useful, but maintaining that context also requires storage, compute and governance.

Arora predicts that model providers will invest in memory because contextual familiarity can make a service more useful and sticky. The resulting allocation of value and customer dependence is a business forecast, not a settled architecture.

Stebbings asks whether a powerful cyber-focused model threatens Palo Alto Networks or helps it. Arora believes it accelerates demand for cybersecurity, while acknowledging that models can produce false positives and that defensive automation requires testing. As the CEO of a security vendor, he has a commercial interest in that assessment.

Arora says his team used a model to identify issues in its own code and claims it found in six weeks work that might otherwise have taken five to six years. He describes testing and human evaluation before fixes were deployed; this is a company account, not an independently verified comparison.

He distinguishes blocking known threats from detecting and responding to an intrusion after preventive controls fail. Palo Alto Networks, he says, intends to add AI to its defenses and use enterprise context to support investigation. No product or model is shown to prevent every attack.

On government intervention, Arora says current model guardrails need improvement and hopes the industry can make them more robust. He presents this as an ongoing discovery process; he does not establish that any safeguard can guarantee intended use.

Asked how he would start a security company today, Arora compares approaches to autonomous driving: designing for autonomy from the start versus gradually automating parts of an existing product. He uses the analogy to ask how an incumbent can move quickly without abandoning customers who depend on current reliability.

He favors an incremental but serious transition for an established enterprise rather than a superficial layer of AI. He says existing customers will not accept a security product that simply works only some of the time, so the company must balance new capability and operational dependability.

Arora says he does not plan to dismantle the organization to build generic marketing, HR or enterprise software from scratch. He hopes suppliers will develop common tools while his company concentrates on distinctive security data, context and products.

He describes twice-weekly meetings where technical leaders explain product changes, agent integration, token use and resource needs. The purpose is to align leaders and spread learning across a large organization while also allowing experimentation from employees.

He recalls earlier executives delegating the internet to a specialist while remaining detached themselves. He worries an equivalent approach to AI could leave a talented team unable to change the business unless senior leaders understand and support the transition.

Arora says existing product roadmaps can consume attention even when technology is changing rapidly. He asks teams to free resources for new work and uses regular reviews and peer comparison to encourage leaders to learn. He presents this as his management approach, not a universal organizational formula.

On forward-deployed engineers, he says enterprise AI products are still evolving and some customers need engineers on site while requirements become clearer. He distinguishes adoption support from engineers who build capabilities at a customer and bring reusable improvements back to the product.

He expects products and vendors to change as enterprise requirements become better understood, citing the rapid evolution of coding tools. He declines to predict a specific winner and says his own priority is finding tools his teams can use.

Arora says he contacts founders working on ideas he does not yet understand in order to learn which technologies could affect his products. He argues that an acquirer cannot simply wait for a winner without understanding why a company or approach succeeded.

He says Palo Alto Networks acquired an agent gateway because he expects companies to need visibility into agent traffic and a way to govern actions. This is the reasoning of a vendor with an interest in the product; routing through a gateway alone does not guarantee complete visibility or security.

He says the speed of technological change makes continuous learning essential: missing one shift may be survivable, but repeated missed shifts can make a company obsolete. That concern explains his interest in both internal development and acquisitions.

On the sell-off in software companies, Arora says he does not know the correct valuations. He sees uncertainty around seat counts, changing analytics and redesigned workflows as reasons investors are reassessing businesses, not proof that any particular vendor will fail.

He describes traditional analytical products built on systems of record and expects some analysis to move toward models working with enterprise data. The availability, quality and security of that data remain important practical conditions.

Asked about Palo Alto Networks' downside, Arora says it could fail to make the transition to AI-enabled security quickly enough. His optimistic case is that the company combines security products and enterprise context effectively; both cases are his forward-looking assessments.

He argues that customers may prefer fewer security vendors and describes his company's platform strategy. He also says innovative startups can still emerge because threats change. His market-share estimates and predicted opportunity are company-leader claims, not an independent market measurement.

Stebbings asks about Chinese open models and security. Arora first separates model openness from country of origin, then considers potential risks for models from any state. He raises questions about hypothetical hidden behavior; the conversation does not establish that any named model contains a backdoor.

He expects a mix of specialized models and an orchestration layer to choose among them. He also says memory embedded in a provider's model could make switching harder. These are architectural predictions, and security depends on the actual model, hosting and permissions.

Turning to wealth and success, Arora recalls arriving in the United States with $200 and taking jobs including security work, note-taking and restaurant work to help pay tuition. He says that experience shaped his willingness to work and his appreciation of having choices later in life.

He argues that the ability to walk away can improve a negotiation by allowing a person to assess whether the outcome is worthwhile. He presents that as a personal lesson, not a guarantee of a better outcome in every negotiation.

Asked how to balance parenting and a demanding career, Arora says there is no universal formula. He believes children learn by watching a parent's behavior and values, including how that parent treats people in ordinary situations.

He draws a parallel with organizations: employees may imitate the behavior they see rewarded by leaders. That makes a leader's conduct and values consequential, he argues, although outcomes depend on the people and context involved.

In the quick-fire round, he warns that fear of missing out can make investors assume every new company will repeat the success of a small number of breakout businesses. His caution is a perspective on uncertainty, not a forecast for a named company.

He recalls a board member advising him not to confuse the effort already spent on a possible acquisition with its merits. After a long walk, he asked whether he would make the same purchase if it arrived without the preceding months of work.

His favorite advice is to value both enthusiasm for work and the pleasure of returning to family. Looking ahead, he favors gratitude and finding something to enjoy each day rather than assuming a particular health or family outcome will occur.

Stebbings thanks Arora for returning to the show, and the two close the conversation. Arora's broader theme is that leaders must learn quickly while acknowledging uncertainty about how AI products, organizations and markets will evolve.