Goldman Sachs analysts warn that the rapid decline in AI token prices threatens the profitability of massive infrastructure investments made by major technology companies. According to the firm's Delta One trading desk, the return on invested capital for hyperscalers has reached an all-time low.
The Silicon Data LLM Token Expenditure Index fell by 29% in August 2026, reaching a record low of $0.97 per million tokens. This figure represents a decline of more than 50% from the peak of approximately $2.05 recorded in May 2026.
Rich Privorotsky, head of the Delta One desk, identified the improvement of proprietary models and the growth of open-source alternatives as primary factors driving down inference costs. These developments allow users to run models locally or on cheaper cloud instances.
Companies such as Meta, Microsoft, Amazon, and Alphabet face significant pressure as their unprecedented capital expenditure programs for AI infrastructure struggle against falling revenues. Goldman Sachs noted that a scenario where prices drop by 30% while usage grows by only 10% is already manifesting.
The market expects further intensification of the pricing war with the release of new models, including Meta's Muse Spark 1.3 and the upcoming OpenAI Astra. Hyperscalers must now determine if they can stimulate sufficient demand to offset the ongoing collapse in token pricing.
