Palo Alto CEO: AI token values need to drop by as much as 90%.

      Palo Alto Networks CEO Nikesh Arora expressed to CNBC that for widespread enterprise adoption of AI to occur, token prices might need to decrease by as much as 90%. He described OpenAI’s assertion of a 54% improvement in token efficiency with its GPT-5.6 model as “a good start” but insufficient. Arora emphasized that the demand for AI is "infinite" and that costs “will rationalize over time.” His statement presents a notable paradox: although per-token prices have significantly dropped, total enterprise AI expenses continue to rise, largely due to agentic usage.

      During Thursday's interview with CNBC, Arora highlighted the necessity for dramatic reductions in AI operational costs before businesses can effectively implement AI on a large scale. He responded to OpenAI's announcement regarding its GPT-5.6 model's improved efficiency, stating, “I think 54% is a good start," but made it clear that much more is needed.

      Arora hopes for continued advancements in efficiency over the coming years, believing that only then will mass enterprise adoption become financially feasible. Despite the current high costs, he remains optimistic about demand, asserting it is “infinite” and predicting that costs “will rationalize over time.” His perspective indicates that the market will either expand to accommodate spending or press prices downward. He suggested that as the technology becomes more efficient, budgets will likely become more manageable.

      The underlying issue Arora highlights reflects a genuine challenge in enterprise AI: while per-token prices have fallen dramatically, overall costs are still increasing. In fact, prices dropped by 98%, yet enterprise AI costs have tripled. This phenomenon is attributed to agentic AI, which repeatedly calls a model to complete tasks. A single ambitious project can incur exorbitant costs; for instance, one developer had their agents accumulate a $1.3 million token bill within a month.

      Consequently, lower headline prices do not necessarily equate to reduced expenses. The growth in usage often surpasses the decline in prices, resulting in higher bills regardless.

      Buyers are feeling the pressure, leading to changes in behavior, including some companies imposing limits on the amount of AI usage by staff as costs escalate. From the buyer's perspective, Arora articulates a common frustration shared by many enterprises.

      On a positive note for Arora, a price war is currently in progress, with DeepSeek offering a permanent 75% discount and competitors striving to keep pace. Numerous startups are also pursuing more affordable inference solutions to maximize output from their technology.

      Whether these developments will achieve Arora's envisioned 90% reduction is uncertain, given that efficiency improvements could be offset by increased usage. Nonetheless, he bets that scaling will ultimately prevail, and the economics will stabilize.

      At this moment, the CEO of a major cybersecurity firm is essentially conveying to AI vendors that their offerings remain too costly for the extent of usage he envisions. This message, originating from such a significant customer, is likely to resonate with the model creators.