Palo Alto CEO: AI token pricing must fall 90% for enterprise adoption

Palo Alto Networks CEO Nikesh Arora told CNBC Thursday that AI token costs must drop as much as 90% to enable large-scale enterprise adoption. Arora called OpenAI's 54% efficiency gain for agentic coding "a good start," but said token efficiency needs to fall to roughly 20% within 12 months and 90% within 24 months to be viable for business workloads. The CEO warned that high token costs are forcing enterprises to cap AI usage instead of expanding it, while consumer AI burns money on nearly free services that fund frontier labs' post-training pipelines.

Arora is among a growing group of executives—including Palantir CEO Alex Karp—pushing back on the token model. He framed the problem as a structural bind: labs need enormous capital to train larger models, relying on free consumer usage for post-training data, but cost-recovery pressure falls almost entirely on enterprises paying frontier-model rates. Meanwhile, open-source alternatives from Chinese labs now hold roughly 30% of token volume, with eight of the top ten models coming from providers like DeepSeek and Moonshot AI—a performance gap of just four months relative to proprietary offerings.

For architects: this signals a reckoning in the unit economics of AI-powered products. Token-heavy workloads (agentic systems, knowledge-work automation) become unprofitable or require prohibitive custom guardrails if per-token costs don't fall sharply. Arora's timeline implies pricing pressure will intensify within the next quarter, favoring either lower-cost proprietary tiers, open-source switching, or models designed to operate within constrained token budgets.