Palo Alto Networks CEO Nikesh Arora has delivered a stark message to the AI industry: the cost of running artificial intelligence must drop by roughly 90% before enterprises can comfortably deploy it at scale. In an interview with CNBC on Thursday, Arora said that token prices—the unit cost of processing AI queries—need to fall dramatically, far beyond the incremental improvements seen in recent models.
Arora was reacting to OpenAI’s announcement that its new GPT-5.6 model is 54% more token-efficient for agentic coding tasks. While he described that as “a good start,” he made clear it was nowhere near enough. He called for the trend to accelerate, with efficiency gains continuing over the next year and then significantly more the year after. Only when prices reach that level, he argued, will mass enterprise adoption become truly affordable.
The CEO’s comments reflect a broader tension in the AI marketplace: even as per-token costs have collapsed, total enterprise AI bills have surged. According to industry data, per-token prices fell 98% over the past year, yet enterprise AI spending tripled over the same period. The culprit, Arora noted, is “agentic” AI—systems that call a model repeatedly to complete multi-step tasks. A single ambitious project can burn through millions of dollars, as evidenced by one developer’s agents that racked up a $1.3 million token bill in a single month.
Despite the sticker shock, Arora is not bearish on demand. “The demand continues to be infinite,” he said, arguing that with an infinite demand curve, costs “will rationalize over time.” His logic is that the market will either grow into the spending or force prices down through competition and technological advancement. He suggested that budgets should ease as the underlying technology becomes more efficient, but the timeline remains uncertain.
The paradox Arora describes has deep implications for both AI vendors and enterprise buyers. While cheaper headline prices look appealing, usage grows faster than prices fall, leaving companies with larger bills than expected. This is why some firms have already started capping how much AI their employees can use, as highlighted in recent industry reports. The strain is already changing behavior, with enterprises increasingly demanding more predictable pricing models and better return on investment.
Palo Alto Networks itself is a major consumer of AI for cybersecurity. The company uses machine learning to detect threats, automate responses, and analyze vast amounts of network data. For Arora, cheaper AI would enable broader deployment across more security use cases, from real-time threat hunting to automated incident response. He envisions a future where AI is embedded in every layer of cybersecurity, but only if the economics make sense.
Industry observers say a price war is already underway, driven by both established players and startups. Chinese AI firm DeepSeek recently made a 75% discount permanent, and rivals are racing to match. A wave of startups is also focusing on cheaper inference—the process of running a trained model—to squeeze more output from every chip. These developments could push prices down significantly, though Arora’s 90% target remains ambitious.
Arora’s career gives weight to his warnings. Before joining Palo Alto Networks in 2018, he served as president and COO of SoftBank Group and as chief business officer at Google. At Google, he oversaw major revenue growth and helped shape the company’s cloud and AI strategies. He holds degrees from the Indian Institute of Technology and Northeastern University, and his leadership of a $100 billion cybersecurity firm gives him a unique vantage point on enterprise technology costs.
Historical parallels exist in other tech sectors. The cloud computing revolution, for example, saw costs drop by over 80% over a decade, enabling widespread adoption. Similarly, the cost of storage fell by 99% from 2000 to 2020, fueling the data explosion. AI may follow a similar trajectory, but the pace of usage growth is unprecedented. Analysts at McKinsey estimate that AI could add $13 trillion to global economic output by 2030, but only if costs become manageable for enterprises.
Nikesh Arora’s message is a double-edged sword for AI vendors. On one hand, it signals enormous pent-up demand. On the other, it pressures them to innovate on cost efficiency as fast as they innovate on capability. For now, the CEO of a cybersecurity giant is effectively telling the AI industry that its product remains too expensive for the full scope of enterprise deployment. Coming from a customer of that size, it is a message the model makers will take seriously.
The debate over AI pricing is likely to intensify in coming months. With OpenAI, Google, Anthropic, and others releasing increasingly powerful models, the race to lower costs is heating up. But as Arora points out, efficiency gains can be swallowed by ever-heavier usage. The ultimate solution may require not just cheaper tokens, but fundamentally new architectures that reduce the number of calls needed per task. Until then, enterprises will continue to grapple with the paradox of falling unit prices and rising total costs.