Published 1 week ago

Palo Alto CEO: AI Token Pricing Must Drop 90% for Enterprise Adoption

Enterprise AI budgets are straining under the weight of token-based pricing. The cost of running large language models at scale has become a silent blocker—quietly eroding the business case for AI adoption even as capabilities improve. Palo Alto Networks CEO Nikesh Arora now puts a number on the problem: token costs need to fall by 90% before most enterprises will fully commit.

Arora’s statement, made in response to OpenAI’s claim of a 54% token efficiency gain in its latest model, cuts through the hype. “I think 54% is a good start,” he said. “I think we probably need another turn at it.” The message is clear: incremental efficiency isn’t enough when the baseline cost is still too high.

106

4 mins read

5 sections

Key Highlights

Nikesh Arora sets a 90% AI token price drop as the bar for true enterprise-scale adoption
Open-weight and self-hosted models emerge as a cost-control response to per-token pricing
AI tools that decouple capability from per-use cost are best positioned to win enterprise budgets

The 90% Target

Arora outlined a two-step timeline. Token efficiency needs to drop to roughly 20% of current levels within 12 months, and then to 10%—a 90% total reduction—by the following year. That’s not a prediction; it’s a condition for enterprise adoption at scale.

The math is simple. If a single agentic coding task costs dollars in API calls, running thousands of such tasks daily across a development team becomes unsustainable. The current pricing model forces businesses into uncomfortable tradeoffs: limit usage, accept lower-quality models, or delay deployment.

Open-Weight Models as a Pressure Valve

Arora isn’t alone. Palantir CEO Alex Karp recently criticized the token model used by Anthropic and OpenAI, calling open-weight models a potential solution. “Something has gone completely wrong,” Karp said, describing a sentiment among enterprises that they’ll “chillax and waste my time with tokens.”

The practical response is already visible. More businesses are turning to open-weight alternatives—models they can run on their own infrastructure, with predictable costs and no per-token meter running. Chinese open-weight models are closing the capability gap with American labs, adding competitive pressure that proprietary model providers can’t ignore.

This shift isn’t about ideology. It’s about cost control. When a model’s performance is good enough, the absence of token fees becomes the deciding factor.

The Spending Paradox

While enterprises push back on token costs, AI infrastructure spending is accelerating to record levels. Tech giants are funding massive build-outs through debt and bond sales, betting that demand will eventually justify the investment. Arora acknowledges the tension but sees a rationalization ahead: “As long as you have an infinite demand curve that you’re facing, I think all these things will rationalize over time.”

That rationalization could take two forms. Either token pricing drops dramatically as competition and efficiency improve, or enterprises permanently shift toward self-hosted and open-weight models, reducing their exposure to per-token pricing altogether. The market is currently testing both paths.

What This Means for AI Tool Adopters

For anyone evaluating AI tools—whether for coding, content, or automation—the pricing signal is worth watching closely. A tool that looks affordable in a pilot can become a budget problem at scale. The current environment suggests three practical moves:

Prioritize tools that offer fixed-cost or self-hosted deployment options, especially for high-volume use cases.
Monitor the open-weight ecosystem. Models that were “almost good enough” six months ago are now competitive for many enterprise tasks.
Build cost-tracking into AI workflows early. Token expenses compound silently, and the teams that measure them are the ones that can negotiate or switch before budgets break.

The 90% price reduction Arora calls for may or may not arrive on his timeline. But the direction is already set: enterprise AI adoption will not scale on the current pricing curve. The tools that win will be those that decouple capability from per-use cost.