Palo Alto Networks chief executive Nikesh Arora has issued a stark warning to the artificial intelligence industry: the cost of running AI must plunge by as much as 90% before businesses can deploy it at scale. In a recent interview, Arora stated that token prices — the fundamental unit of cost in AI model inference — need to fall dramatically to unlock widespread enterprise adoption.
Arora was responding to OpenAI’s announcement that its new GPT-5.6 model delivers a 54% improvement in token efficiency for agentic coding tasks. While he acknowledged this as “a good start,” Arora made it clear that far more aggressive price reductions are necessary. He predicted that efficiency gains would need to continue over the next year and accelerate the year after before large corporations find AI economically viable for broad use cases.
The demand for AI, Arora argued, is effectively unlimited. “The demand continues to be infinite,” he said, explaining that with an infinite demand curve, costs inevitably “will rationalize over time.” His logic is straightforward: either the market will grow into the spending, forcing prices down as competition intensifies, or budgets will ease as underlying technology becomes more efficient. But the path to that equilibrium is not guaranteed, and the current trajectory presents a genuine puzzle for enterprise buyers.
Palo Alto Networks, a cybersecurity giant with a market capitalization exceeding $100 billion, is itself a major consumer of AI technology. The company uses AI for threat detection, automated response, and security analytics. Arora’s comments reflect the frustrations of a buyer who sees enormous potential in AI but is constrained by the current cost structure. His message to AI vendors is clear: your product is still too expensive to use everywhere we want to use it.
The paradox behind the plea
Arora’s complaint captures a genuine paradox in enterprise AI. On the surface, per-token prices have collapsed dramatically. According to industry analysis, the cost per token has fallen by approximately 98% over the past two years. Yet total enterprise AI spending has tripled during the same period. This disconnect arises because usage grows faster than prices fall. As AI becomes cheaper per unit, companies use it more, often in ways that multiply consumption exponentially.
The primary culprit is agentic AI — systems that repeatedly call a model to complete a complex task. Unlike traditional single-query use cases, agentic AI orchestrates chains of reasoning, often running dozens or hundreds of inference calls for a single objective. A single ambitious project can burn through a fortune in tokens. One developer recently reported that his AI agents ran up a $1.3 million token bill in a single month, highlighting how quickly costs can spiral.
This phenomenon means that headline price reductions do not automatically translate into lower overall costs for enterprises. Instead, companies find themselves in a situation where cheaper prices encourage more usage, and the total bill continues to rise. Arora’s call for a 90% price drop is an attempt to break this cycle, hoping that such a dramatic reduction would make the economics palatable even with increased usage.
Background on Nikesh Arora and Palo Alto Networks
Nikesh Arora became CEO of Palo Alto Networks in June 2018, bringing with him extensive experience in technology and finance. Prior to joining the cybersecurity firm, he served as senior vice president and chief business officer at Google, where he was responsible for revenue and operations across multiple regions. Before Google, he held leadership roles at SoftBank Corp., including president and chief operating officer, and at T-Mobile as chief marketing officer. His career also includes senior positions at British Telecom and a stint at McKinsey & Company.
Under Arora’s leadership, Palo Alto Networks has expanded its portfolio beyond core firewall products to include cloud security, threat intelligence, and AI-driven security operations. The company has aggressively adopted AI and machine learning to enhance its offerings, making it a significant consumer of AI inference. In fiscal year 2024, Palo Alto Networks generated over $8 billion in revenue, with a substantial portion allocated to R&D and AI-related expenses. Arora’s perspective on AI pricing is therefore not theoretical; it is grounded in the real-world budgeting of a major enterprise.
The company’s AI strategy centers on its Precision AI platform, which combines machine learning, deep learning, and generative AI to detect and respond to threats in real time. As cyberattacks become more sophisticated and numerous, Palo Alto Networks relies on AI models to process vast amounts of data, identify patterns, and automate responses. Each security event can trigger multiple AI inference calls, adding to the token consumption. For a firm handling billions of events daily, even small inefficiencies in token cost can translate into millions of dollars.
The price war and market dynamics
The good news for Arora and other enterprise buyers is that a price war is already underway in the AI industry. Chinese startup DeepSeek recently made headlines by offering a 75% discount on its API pricing, making it permanent rather than a promotional offer. Rivals like OpenAI, Google, and Anthropic have been forced to match or undercut these reductions to retain market share. A wave of startups is also emerging, focused on cheaper inference solutions that squeeze more output from every chip.
DeepSeek’s aggressive pricing is part of a broader trend of commoditization in AI model access. As open-source models improve and proprietary vendors compete, the cost of a single token has become a key battleground. Some analysts predict that inference costs could fall by another order of magnitude within two years, driven by hardware improvements, model quantization, and architectural innovations like mixture-of-experts.
However, efficiency gains can be swallowed by ever-heavier usage. The same forces that drive down per-token costs also enable more complex and resource-intensive AI applications. For example, agentic AI workflows that require multiple reasoning steps are becoming more common, and their token consumption can dwarf that of simple question-answering tasks. Arora’s bet is that scale eventually wins, and the economics will settle at a point where enterprises can safely budget for AI without fear of runaway costs.
Enterprise adoption challenges
The strain of rising AI bills is already changing enterprise behavior. Some firms have implemented caps on how much AI their employees can use each month, while others are limiting AI access to specific departments or high-value use cases. A survey by Gartner found that over 60% of enterprises cited cost as the primary barrier to expanding AI adoption, despite enthusiasm about the technology’s potential.
Arora’s comments align with this sentiment. By publicly calling for a 90% price reduction, he is signaling to AI vendors that the current pricing model is unsustainable for broad deployment. He is effectively arguing that the industry needs to think differently about pricing — perhaps moving away from per-token billing to subscription models or usage-based tiers that cap total spending.
In the cybersecurity sector specifically, AI is critical for staying ahead of adversaries. Cybercriminals themselves are increasingly using AI to launch sophisticated attacks, forcing defenders to accelerate their AI adoption. Palo Alto Networks’ competitors, such as CrowdStrike and SentinelOne, face similar cost pressures. If AI prices remain high, smaller security firms may struggle to compete, potentially widening the gap between large and small players.
Beyond cybersecurity, enterprises across industries — from healthcare to finance to manufacturing — are grappling with the same cost dilemma. While large language models have demonstrated remarkable capabilities, their operational costs often eclipse the value they generate in low-margin use cases. Arora’s call for a 90% decline is not just a negotiating tactic; it reflects a genuine need for the technology to become cheap enough to embed in everyday business processes.
Historical context and future outlook
The history of enterprise technology adoption offers some reassurance that costs will indeed decline over time. Mainframe computing, personal computers, cloud services, and even earlier AI systems all followed a trajectory of falling prices and increasing usage. The cost of a unit of compute has dropped by roughly 50% every two years for decades, a trend that has been accelerated by Moore’s Law and now by advanced chip manufacturing.
However, AI inference presents unique challenges. The rapid growth in demand, driven by agentic and multi-modal applications, may outpace the rate of cost reduction for the foreseeable future. Arora’s 90% target is ambitious, but not impossible. If DeepSeek and other players continue to drive competition, and if hardware vendors like NVIDIA and AMD release more efficient chips, the next two years could see dramatic price drops.
For Palo Alto Networks, the stakes are high. If AI prices fall significantly, the company can deploy intelligence across its entire product line, enabling real-time threat prediction and automated remediation at a fraction of current costs. If prices remain high, the company may have to make difficult trade-offs, prioritizing only the most critical AI applications.
In the end, Arora’s message is a reflection of the market’s growing pains. The AI industry is still in its infancy, and the infrastructure for cost-effective inference is still developing. As the technology matures and the competitive landscape intensifies, prices will likely continue to fall. Whether they reach the 90% threshold remains uncertain, but the direction is clear: cheaper AI is not just desirable but necessary for the next wave of enterprise innovation.