SKN | Goldman Sachs Flags a New AI Risk as Falling Token Prices Challenge Infrastructure Economics

Key Takeaways:

• Goldman Sachs is highlighting a new risk within the AI investment cycle: computing capacity could outpace the revenue generated by AI usage if token prices continue falling faster than demand expands.
• The bank’s Delta One desk points to a Silicon Data benchmark that fell 29% in August to $0.97 per million tokens, its first move below the $1 threshold.
• Goldman’s concern is not that AI demand is disappearing, but that falling unit prices could weaken the economics supporting enormous data-center capital expenditure.
• For Goldman, the critical variable is whether AI consumption can accelerate sufficiently to offset declining prices and sustain returns on the infrastructure being built.

Goldman Sachs is drawing attention to a less visible risk within the artificial intelligence boom: the rapid decline in the price of AI computation. Through its Delta One desk, the bank is examining whether AI usage can expand quickly enough to absorb the extraordinary amount of computing infrastructure being deployed before falling prices begin to undermine the economics behind that investment cycle.

The distinction is important. Goldman is not arguing that AI demand is weakening. Instead, the bank is questioning whether lower prices can eventually become a problem for the suppliers of computing capacity if the increase in usage does not compensate for the decline in revenue generated per unit of consumption.

Goldman Identifies a Growing Unit-Economics Problem

Goldman highlighted Silicon Data’s LLM Token Expenditure Index, which measures the usage-weighted price of AI tokens. The index fell 29% during August to approximately $0.97 per million tokens, more than 50% below its May peak of roughly $2.05.

For Goldman, this creates a critical tension. Cheaper AI can encourage broader adoption, but each unit of consumption generates less revenue. If prices decline faster than token consumption increases, overall spending may fail to grow sufficiently to support the infrastructure being constructed to meet anticipated demand.

Goldman Questions the Durability of Per-Token Pricing

The bank is also highlighting structural forces pushing prices lower. Competition among frontier-model developers, cheaper open-source models and the migration of some inference workloads toward lower-cost or local hardware are increasing pressure on the traditional cloud-based pricing model. Goldman’s Delta One desk has expressed skepticism that per-token pricing can remain a durable end-state economic model.

This matters because AI infrastructure requires substantial upfront capital. If the revenue associated with each unit of computing continues to decline, the return generated on data centers and related infrastructure could become more difficult to justify even while aggregate AI usage continues rising.

The Critical Variable Is Revenue, Not Token Volume

Goldman’s warning ultimately shifts attention from headline AI adoption toward the relationship between usage growth and pricing power. A rapidly expanding volume of AI activity is not necessarily sufficient if the price paid for that activity falls even faster.

The bank therefore sees a potential scenario in which computing capacity grows faster than customers can profitably consume it. That would not necessarily signal an immediate collapse in AI demand; rather, it could force investors and financial institutions to reassess the returns expected from the enormous infrastructure commitments supporting the sector.

For Goldman Sachs, the implication is clear: the next phase of the AI cycle will increasingly be measured by monetization, utilization and infrastructure returns, not simply by usage growth. If falling prices stimulate enough additional consumption, the model can remain viable. If they do not, excess capacity could become a material risk to the economics underpinning the AI investment cycle. For a confidential discussion regarding technology-sector exposure, cross-border portfolio structures and risk management, contact our senior advisory team.