AI costs begin to bite as agents may increase token demand by 24 times, says Goldman Sachs report — Uber and Microsoft among companies feeling the bite of tokenized billing

AI token costs are emerging as a material drag, with Goldman Sachs warning agentic AI could increase token demand by more than 24x over the next few years. Uber says higher token usage has not translated into proportional consumer-feature gains, while Microsoft is revoking Claude Code access and shifting developers to internal tools, likely to reduce costs. The article suggests major tech firms are reassessing AI deployment economics as tokenized billing and agent usage pressure budgets.

Analysis

The market is starting to discover that AI monetization is not just a model-quality problem but a unit-economics problem. The first-order losers are the companies whose internal workflows are most token-intensive and hardest to measure in direct revenue terms: enterprise software, ride-hailing, and cloud ecosystems that are using AI as an operating lever rather than a billable product. That creates a second-order benefit for hyperscalers and model vendors with pricing power only if they can keep throughput growth ahead of customer pushback; otherwise the spend migrates inward toward cost control, not outward toward incremental demand.

The bigger implication is that adoption is likely to bifurcate. Light-touch copilots and bounded inference use cases should remain sticky, while agentic workflows face a near-term usage reset as CFOs impose quotas, human-in-the-loop gates, and model downgrades. That means the current risk is not a collapse in AI demand, but a mix shift away from the highest-token, lowest-ROI workloads — exactly the segment most bullish consensus narratives rely on for explosive consumption growth.

This is potentially bearish for GPU demand over the next 2-6 quarters because the constraint is budget, not capability. If customers stretch hardware replacement cycles and optimize for cost per task instead of raw output, the benefit of next-gen accelerators arrives later than bulls assume, and interim utilization may disappoint. The most vulnerable names are those priced for a rapid, uninterrupted step-up in inference spend; the more resilient names are those with software-layer control of the workflow or enough scale to absorb token inflation through platform stickiness.

Contrarian view: the selloff risk in AI infrastructure may be overdone if the market is underestimating how quickly vendors can force price discrimination. Token billing, model routing, and on-prem/private deployments can preserve gross profit even as end-user consumption is capped. The real tell over the next 1-2 quarters will be whether enterprise customers cut usage sharply, or simply become more surgical — if the latter, the narrative shifts from demand destruction to margin reallocation.

Key Decisions for Investors

• Reduce near-term exposure to NVDA on rallies; the market may be overpricing a 2025-2026 inference boom that is vulnerable to enterprise budget caps and longer hardware refresh cycles. Favor waiting for a better entry after management commentary confirms actual utilization, not just bookings.
• Go long MSFT vs long-beta AI pure plays on a 3-6 month horizon: MSFT can absorb internal AI cost discipline better than peers and may use product bundling to defend demand. Risk/reward improves if the market rewards cost control over raw usage growth.
• Short UBER or sell upside calls into strength for the next 1-2 quarters; the investment case is exposed if AI spend does not map to consumer feature monetization. Add only if management continues signaling AI ROI skepticism.
• Pair trade: long GOOGL / short NVDA into any AI capex scare, on the thesis that platform owners with diversified monetization can re-route workloads and capture spend even if token growth slows. Use a 3-5 month window and cover if inference pricing remains firm.
• Consider a tactical long GS only as a secondary beneficiary of the AI capex debate, but not as a core expression; if AI spending gets repriced, the more durable trade is on budget discipline and software monetization than on hardware throughput.