Goldman Sachs warns falling AI token prices risk compute oversupply

Goldman Sachs The Goldman Sachs Group, Inc. ↗ 4.67% is warning that a rapid decline in AI token prices may be outpacing growth in usage, raising the prospect of an oversupply of computing capacity that could weigh on equity valuations across the AI infrastructure sector.

In a Delta One note circulated this week, the bank flagged that Silicon Data's benchmark price for large-language-model usage — the LLM Token Expenditure Index, tracked on Bloomberg as SDLLMTK — recently fell to roughly $0.97 per million tokens after dropping 29% in August alone. The index stood at $0.98 as of September 2. CNBC reported earlier this week that the measure had touched fresh lows amid intensifying competition.tipranks+2

The Price-Volume Mismatch

Goldman's concern centers on a straightforward arithmetic problem: if token prices fall faster than total consumption grows, the industry could generate less revenue from the infrastructure it has built. The bank said that unless future AI models produce improvements large enough to encourage much heavier usage, the risk of periods of excess compute capacity becomes increasingly realistic.tipranks

Rich Privorotsky, head of Goldman's Delta One trading desk, has been raising related concerns for months, noting as early as June that major AI companies were selling more compute hardware, directly challenging the long-held assumption that shortages would persist indefinitely. In a June note shared by ZeroHedge, the desk warned that the AI trade was shifting "from the church of scarcity to the bazaar of abundance".thedarksideoftheboom.substack+1

Competition Driving the Decline

Price cuts to proprietary models are the biggest driver of the drop, according to the Goldman note. Meta ↗ 5.12% has released multiple iterations of its Muse Spark model this year, most recently version 1.2 in August, which Meta said costs about a quarter of what Anthropic's leading model charges. Meanwhile, OpenAI is preparing to release its Astra model, which the company described this week as its most capable yet.nytimes+3

The broader trend is steep. The BenchLM Token Price Index shows frontier LLM prices are now 84% below their March 2023 baseline. Separately, Goldman Sachs Research has projected that AI token consumption could multiply 24 times by 2030, driven by the rise of autonomous AI agents — a scenario that would offset falling unit prices. But the bank's Delta One desk is less sanguine about the near-term trajectory, cautioning that the demand surge may not arrive quickly enough to prevent a glut.goldmansachs+1

What It Means for Investors

The warning adds a layer of caution to a sector that has attracted enormous capital expenditure. Goldman projects AI capital spending could more than double from $765 billion in 2026 to $1.6 trillion by 2031. If revenue per unit of compute continues to erode faster than volume grows, returns on that infrastructure spending could disappoint, creating headwinds for companies across the AI supply chain — from chipmakers to data center operators to cloud providers.threads