Tokens are getting cheaper, but companies are spending even more on AI as a result, top economist warns

AI token prices have fallen more than 90% since 2023, but spending on large language models has doubled since late last year, while Bain found token costs halved from December 2024 to December 2025 and token usage rose 450%. The article argues that Jevons paradox is driving higher aggregate AI spend as cheaper tokens spur heavier usage, more agents, and more complex tasks. It highlights rising AI budget pressure at companies like Uber and the risk that AI cost savings may not translate into lower opex.

Analysis

The market is still pricing AI like a software margin expansion story, but the article points to a different regime: AI is becoming a variable cost center that scales with ambition. That matters because the winners are not just model vendors; the more important beneficiaries are the infrastructure layers that monetize heavier usage per workflow—GPU supply, networking, power, and orchestration software. The losers are end-users with weak internal governance, where AI spend becomes a line-item tax on productivity rather than a lever on operating leverage.

Second-order, this is a classic budget reallocation problem: as tokens get cheaper, teams do not stop buying them, they widen the surface area of automation. That shifts demand from “training once, deploying many times” to “continuous inference everywhere,” which is structurally more favorable for the picks-and-shovels stack than for application-layer companies trying to prove ROI on a per-seat basis. It also raises the probability of margin disappointment in businesses with discretionary AI experimentation and low usage discipline—especially large consumer platforms and ride-sharing/logistics operators where marginal AI spend is easy to hide inside opex.

The near-term catalyst is not lower token prices; it is CFO intervention. Once usage caps, chargeback systems, or ROI hurdle rates get imposed, aggregate spend can decelerate even if unit economics keep improving. Over a 3-6 month horizon, the biggest risk is that investors extrapolate exploding token volumes into durable revenue for AI infrastructure without recognizing that enterprises will keep ratcheting internal controls, which could create a volatile demand path rather than a straight line.

The contrarian read is that the AI spend overhang may be less bearish for Nvidia than feared and more bearish for companies that promised immediate opex savings. If AI is replacing labor slowly but inflating consumption quickly, then the first earnings downgrades are likely to come from adopters, not enablers. That makes this a relative-value setup: long the infrastructure beneficiaries with pricing power, short the names where AI is already embedded in guidance but monetization remains unproven.

Key Decisions for Investors

• Long NVDA on 1-3 month weakness; use pullbacks to add rather than chase. Risk/reward remains favorable if enterprise inference demand continues compounding, but trim if management commentary shifts toward visible customer budget saturation.
• Pair trade: long NVDA / short UBER for 2-4 months. UBER appears most exposed to internal AI spend caps and low-visibility budget overruns; NVDA benefits from the same usage intensity. Best if held into next earnings season.
• Overweight META and AMZN on a relative basis versus consumer-facing software/app names. Both can internalize AI spend and amortize it across large user bases, but size positions modestly because the market may still punish any evidence of opex creep.
• Consider a basket short of AI-adopter margin-risk names against a long basket of infrastructure enablers: long NVDA, long selected networking/power names, short UBER or other high-volume workflow adopters. Target 3-6 months, looking for dispersion as budgets tighten.
• Avoid paying up for pure AI monetization stories with weak chargeback discipline until next quarterly commentary confirms ROI. The setup favors delay, not denial: unit costs falling faster than budgets means near-term volatility in spend growth is likely.