Goldman Warns Falling AI Token Prices Could Squeeze Hyperscaler Returns

Goldman Sachs has warned that collapsing AI token prices could weaken returns for hyperscalers pouring hundreds of billions into infrastructure. The concern is that demand may not grow quickly enough to offset the decline in what providers can charge for model usage.

The Silicon Data LLM Token Expenditure Index fell 29% in August 2026 to a record low of $0.97 per million tokens. That was more than 50% below its May 2026 peak of roughly $2.05.

Goldman’s Delta One trading desk raised the issue in September 2026. Rich Privorotsky, who heads the desk, identified two forces pushing cloud inference prices lower: proprietary models are becoming more efficient, while open-source alternatives are undercutting commercial products.

Goldman outlined a scenario in which token prices drop 30% while usage rises only 10%. That gap would directly pressure revenue generated by AI infrastructure investments, and the August data suggests a similar dynamic may already be playing out.

Meta, Microsoft, Amazon and Alphabet are exposed to the squeeze after committing to AI infrastructure capital expenditure programs that Goldman described as unprecedented in the history of the tech sector. The firm’s analysis found that hyperscaler return on invested capital has already reached an all-time low.

Open-weight models are adding to the pricing pressure by allowing workloads that previously required costly API access to run locally or through cheaper cloud instances. Meta’s Muse Spark 1.3 has launched, while OpenAI’s Astra is expected to intensify the pricing competition when it arrives.

The drop from roughly $2.05 to $0.97 per million tokens in about three months leaves hyperscalers facing a tougher economic equation. Their ability to generate enough additional usage to outrun falling token prices could shape the next phase of AI infrastructure spending.