As companies including Uber and Microsoft have looked to curtail unconstrained AI use among their employees, a voice from Silicon Valley has said that enterprises should be getting access to cheaper AI.

Nikesh Arora, CEO of Palo Alto Networks and a board member at Uber, took to X to lay out what he sees as a structural problem with how AI companies are currently approaching enterprise monetization. His argument is pointed: high token pricing for enterprises, while consumers get AI for free, is a trap that will ultimately drive businesses toward open-source alternatives — and away from the frontier models that AI companies are betting their futures on.

The Consumer-Enterprise Pricing Paradox

Arora frames the tension as a byproduct of how AI labs are funding themselves. The leading models need enormous compute to stay ahead in the race toward AGI, and free consumer AI has become a tool to feed post-training data pipelines. Labs can’t easily stop offering AI free to consumers — that usage is how they improve their models. So the pressure to monetize gets redirected squarely at enterprises, which end up paying the bill.

“The risk — high token pricing for enterprises while consumers for free,” Arora wrote. For consumer-distribution businesses like Google, Meta, and Apple, subsidizing free AI makes strategic sense — it protects distribution. But for enterprises trying to figure out whether AI can genuinely transform their operations, the math doesn’t work the same way. High token prices create a chilling effect. CIOs end up focused on restricting AI use and making deployments more efficient, rather than actually experimenting with what the technology can do.

This is a dynamic that’s already playing out in the real world. Uber burned through its entire 2026 AI coding budget in roughly four months, with per-engineer monthly API costs running between $500 and $2,000. Usage metrics looked impressive on paper — 95% of engineers using AI tools monthly, 70% of code commits AI-driven — but Uber’s COO Andrew Macdonald admitted the link between those numbers and actual consumer value “is not there yet.” The cost-value equation is broken when token prices are high enough to burn annual budgets in a quarter.

Two Phases of Enterprise AI — And Where It Gets Hard

Arora maps enterprise AI adoption into two distinct phases. Phase one — coding — has been relatively tractable. It’s a bottom-up motion with low customization per customer, developers naturally adopt it, and the use case is well-defined. That’s borne out in market share data, with Anthropic commanding roughly 54% of the enterprise coding API market by 2025.

Phase two is where the difficulty lies. Showing real enterprise value — efficiency gains, accuracy improvements, resource elimination — requires a fundamentally different approach. It means building depth through context, memory, and domain-specific skill libraries. It means solving for edge cases with deterministic guardrails rather than hoping a general-purpose model handles them gracefully. Arora calls the practitioners who do this work FDEs — a class of specialists who train enterprise AI systems the way Waymo trains autonomous vehicles: carefully, iteratively, with failure modes accounted for.

This kind of deployment is inherently expensive to build and token-intensive to run. When the per-token cost is high, enterprises either don’t attempt it, or they scope the project so conservatively that the results are underwhelming. Neither outcome helps the AI companies trying to prove enterprise ROI.

The Open-Source Exit Ramp

Arora’s sharpest warning is about what happens if frontier AI companies don’t move on pricing: enterprises route their workloads to secure open-source models, and the proprietary labs end up with friction-filled routing layers between them and actual enterprise usage.

That scenario is already taking shape. The top open-source AI models in mid-2026 are dominated by Chinese labs, with eight of the top ten coming from companies like MiniMax, DeepSeek, Moonshot AI, and Z.AI. The performance gap between these models and the frontier proprietary offerings has narrowed to roughly four months. For many enterprise workloads, that gap is small enough to live with — especially when the cost differential exceeds 4x. OpenRouter data shows open-source holding around 30% of total token volume, with Chinese models having displaced Western open-source alternatives substantially within that share.

The trajectory matters. Enterprises that migrate to open-source deployments for cost reasons don’t necessarily come back when models improve. They build infrastructure around what they’re running, and switching costs accumulate. Arora is essentially telling AI companies that they’re pricing themselves out of the market that matters most for long-term revenue — while giving away the consumer market for free.

What Arora Would Do

His prescription has three parts. First, cut token pricing now. Forward-priced, cheaper tokens for enterprises would unlock the kind of experimentation and workflow reimagination that currently gets squeezed out by budget anxiety. Second, show enterprises how their own context, training, and data can become a competitive advantage — not just using the AI company’s model, but building something proprietary on top of it. Third, build tools for rapid edge case learning and reducing false positives, so the work of deploying AI in complex enterprise environments becomes less of a research project and more of an operational capability.

The underlying paradox Arora identifies is that enterprises haven’t fully understood the value of AI — and the pricing structure is partly responsible. When the cost of experimentation is high, you get conservative deployments with narrow use cases and underwhelming results, which in turn makes it harder to build the internal conviction needed for broader investment. Cheaper tokens don’t just reduce costs; they change the type of bets enterprises are willing to make.

AI companies that figure this out first stand to lock in the enterprise relationships that will matter when the model race eventually slows down. Those that don’t may find themselves winning the technology competition while losing the market.