PANW CEO: AI Token Costs Must Fall 90%

Palo Alto Networks CEO Nikesh Arora says enterprise AI adoption hinges on a 90% price reduction in token costs as agentic workloads push total bills higher despite per-unit declines.

Lead

What Happened

Arora's remarks came directly in response to OpenAI CEO Sam Altman's claim that the lab's newest model delivers 54% greater token efficiency for agentic coding tasks. Arora acknowledged the progress — "I think 54% is a good start" — while making clear it is insufficient. "I think we probably need another turn at it," he said, indicating enterprise buyers require a sustained, multi-year compression in pricing before committing to large-scale rollouts.

The comments reflect a widening chorus among enterprise technology leaders. Palantir CEO Alex Karp separately criticized the per-token pricing model employed by Anthropic and OpenAI, suggesting open-weight models could offer a structural alternative for cost-conscious buyers.

The Cost Paradox

The frustration Arora articulated is grounded in a counterintuitive market dynamic. Per-token prices have collapsed approximately 98% since 2023, yet total enterprise AI spending has tripled over the same period. The driver is agentic AI: rather than processing a single user prompt, agentic systems call a model repeatedly — often five to thirty times per task — to reason, retrieve, and execute across multi-step workflows. That usage pattern turns a falling per-unit cost into a rapidly climbing aggregate bill.

Industry data shows that a complex orchestrated agentic interaction now costs roughly $1.20 per session, compared with approximately $0.04 for a standard chatbot exchange in 2023. Average annual enterprise AI budgets have grown from roughly $1.2 million in 2024 to $7 million in 2026. Some organizations have reported exhausting their annual AI budgets within a single quarter.

The phenomenon mirrors Jevons paradox: as each token becomes cheaper, enterprises consume exponentially more of them, erasing the per-unit savings in aggregate.

Strategic Context

For Palo Alto Networks, the economics of AI token costs are not peripheral — they are central to the company's core security business. PANW has positioned itself at the intersection of cybersecurity and AI, integrating large language models into threat detection, automated response, and security operations center workflows. High token costs create a direct friction point for customers evaluating whether AI-driven security tools deliver sufficient return on investment.

Arora's public call for steeper price reductions carries weight because PANW is both a buyer of AI infrastructure and a reseller of AI-augmented security capabilities. His two-year benchmark — a 90% cost reduction — effectively signals to model providers where enterprise demand inflection points are likely to occur.

Market Reaction

PANW shares were trading at approximately $325.77 on July 10, within a session range of $323.85 to $342.00. The stock reached an all-time high of $368.17 on July 6, 2026, and carries a market capitalization of approximately $265 billion, ranking it among the largest dedicated cybersecurity companies by valuation. The stock's recent retreat from its peak reflects broader market recalibration around AI monetization timelines rather than company-specific concerns.

Outlook

Arora's two-year cost target sets a visible benchmark for the AI infrastructure sector. If token costs follow the trajectory he outlined — 80% lower within 12 months and 90% lower within 24 — the economics of deploying agentic AI at enterprise scale shift materially. That would broaden the addressable market for AI-native security platforms, cloud providers, and enterprise software vendors. Conversely, if cost compression stalls, corporate adoption curves remain constrained, and near-term revenue projections for generative AI-enabled services will likely require revision. The gap between falling per-unit prices and rising aggregate bills remains the defining tension in enterprise AI spending for the remainder of 2026.

Technology }}