The cost of running artificial intelligence models has fallen sharply in recent months, and Goldman Sachs believes the decline poses structural risks for the technology giants spending hundreds of billions of dollars to build the infrastructure behind it.

In a September 2026 warning, Goldman Sachs Delta One trading desk flagged mounting risks for hyperscalers — the operators of the vast cloud and data center networks that underpin AI services — as the cost of AI model usage tokens collapses faster than demand can grow. Tokens are the per-unit measure by which model usage is billed, and the Silicon Data LLM Token Expenditure Index dropped 29% in August 2026, landing at a record low of $0.97 per million tokens — more than 50% below its May 2026 peak of roughly $2.05.

Pricing pressure outpacing demand

Rich Privorotsky, head of Goldman's Delta One desk, pointed to two converging forces crushing per-token pricing for cloud inference: the rapid improvement of proprietary models, which can accomplish more with less, and the rising tide of open-source alternatives that undercut commercial offerings on price.

Goldman flagged a scenario in which token prices fall 30% while usage grows only 10% — a gap that would eat directly into revenue from AI infrastructure investments. According to the desk, August's numbers suggest roughly that dynamic is already underway. That is the crux of why the warning matters: cheaper tokens lower costs for the companies and developers using AI, but for the providers that own the infrastructure, shrinking per-unit prices erode the revenue needed to justify enormous capital spending.

The companies caught in the squeeze are among the biggest names in technology: Meta, Microsoft, Amazon, and Alphabet. All four have committed to capital expenditure programs for AI infrastructure that Goldman described as unprecedented in the tech sector's history.

Open-source models intensify the price war

Models that once would have required expensive API calls to frontier labs can now be run locally or on cheaper cloud instances using open-weight alternatives. Every improvement in open-source capability puts downward pressure on what commercial providers can charge.

Meta's Muse Spark 1.3 recently launched, while OpenAI's Astra is on the horizon. Both are expected to intensify the pricing war.

Return on invested capital for hyperscalers has already dropped to an all-time low, according to Goldman's analysis. The token pricing collapse from $2.05 to $0.97 in roughly three months is the kind of move that forces strategic recalculation.

Whether hyperscalers can engineer enough demand growth to outrun the price decline may define the next chapter of the AI investment cycle, according to the report. Future readings of the Silicon Data index — showing whether usage growth can narrow the gap left by falling prices — and the competitive effects of the newly launched Muse Spark 1.3 and the anticipated Astra are the near-term markers to watch.