Palo Alto CEO: token costs must drop 90% for AI adoption

Key Takeaways

• Palo Alto Networks CEO says token costs need to drop 90% within 24 months for enterprises to adopt AI at scale
• Uber exhausted its full-year 2026 AI budget by April, forcing leadership to weigh token costs against hiring engineers
• Companies are implementing usage caps, switching to older models, and exploring open-source alternatives to control costs

Palo Alto Networks CEO Nikesh Arora delivered a blunt assessment of enterprise AI adoption on Thursday: token costs are too high, and they need to fall dramatically before most companies can deploy AI at scale. Speaking on CNBC's Squawk on the Street, Arora said costs must drop 20% in the next 12 months and 90% within 24 months for AI to become economically viable across the enterprise.

The comments come as major technology buyers are already hitting budget walls. Uber exhausted its entire 2026 AI budget by April, according to reporting from May. The company's CTO Praveen Neppalli Naga said Uber was "back to the drawing board," while COO Andrew Macdonald stated the company would now weigh token costs directly against the cost of hiring engineers. That trade-off calculation signals how seriously enterprises are reconsidering their AI strategies.

Why 54% efficiency gains aren't enough

When CNBC asked Arora about OpenAI CEO Sam Altman's recent claim that the company's latest model is 54% more efficient for coding tasks, Arora was unimpressed. "I think 54% is a good start. I think we probably need another turn at it," he said.

The math explains his skepticism. A 54% efficiency improvement cuts costs roughly in half. But Arora's 90% target means costs need to fall to one-tenth of current levels. That requires multiple compounding improvements across model architecture, inference optimization, and pricing structures. OpenAI and competitors like Anthropic and Google have cut prices steadily over the past 18 months, but the pace of reduction trails enterprise budget expectations.

How enterprises are managing AI costs now

Companies aren't waiting for token prices to fall. They're pulling back. June reporting showed enterprises employing multiple tactics to control runaway AI spend:

• Introducing usage caps on AI tools
• Encouraging employees to match the right model to each task rather than defaulting to the most powerful option
• Sharing internal cost-saving playbooks
• Switching to older, cheaper models for routine tasks
• Adopting open-source models where possible

The open-source shift is particularly notable. Chinese AI labs have found an opening by charging less than American competitors. Their advantage comes from more efficient models and China's lower energy costs for training and inference. For cost-sensitive enterprises, the performance-to-price ratio matters more than peak capability.

Agentic AI compounds the cost problem

The token cost problem gets worse as enterprises move from chatbots to agentic AI systems. A standard chatbot interaction generates a single inference call. An agentic coding session, where the AI plans, executes, and iterates autonomously, generates many more. Each step in the agent's chain of reasoning burns tokens.

This multiplication effect means companies piloting agentic tools for software development, customer service automation, or data analysis face cost exposure that scales unpredictably. Budget models built on simple per-user assumptions break down when a single agent session can trigger dozens of inference calls.

Financial services leads despite cost headwinds

Despite the cost pressures, some sectors keep investing. According to the PYMNTS Intelligence Enterprise AI Benchmark Report, financial services and insurance companies lead the pack in AI deployment, followed by healthcare and media. These industries have clearer ROI calculations and can justify higher inference costs when AI directly reduces fraud, accelerates underwriting, or automates compliance workflows.

The report notes that enterprises are becoming more disciplined about which AI projects deserve real capital versus which need more proof-of-concept work before scaling. That selectivity represents a maturation from 2024's experimental phase, when many companies threw money at AI without clear success metrics.

Frequently Asked Questions

What are token costs in enterprise AI?

Token costs refer to the pricing structure AI providers use to charge for inference. A token equals roughly 4 characters or 0.75 words. Enterprises pay per token processed, which means costs scale with usage volume and model complexity.

Why did Uber run out of its AI budget by April 2026?

Uber's rapid AI experimentation and agentic tool deployment burned through tokens faster than budget models predicted. The company now weighs token costs against the cost of hiring engineers as a direct comparison.

How much do token costs need to drop for enterprise AI adoption?

According to Palo Alto Networks CEO Nikesh Arora, token costs need to fall 20% within 12 months and 90% within 24 months for enterprises to adopt AI at scale.

Are Chinese AI labs cheaper than American providers?

Yes. Chinese AI labs can undercut U.S. providers due to more efficient model architectures and lower energy costs in China. Cost-sensitive enterprises are exploring these alternatives.

Which industries are investing most in enterprise AI despite high costs?

Financial services and insurance lead, followed by healthcare and media. These sectors have clearer ROI paths that justify higher token costs.

Source: PYMNTS | / PYMNTS