AI infrastructure and data center market conditions: AI token spending accelerated sharply in September, while GPU rental and memory pricing gave mixed signals

Overview

This data-center tracker assesses AI infrastructure demand through monthly LLM token activity, non-hyperscaler GPU rental prices, and memory spot prices. J.P. Morgan sees strong token-demand evidence but more mixed pricing signals across compute rentals and memory.

Core views

J.P. Morgan’s September Data Center Watch argues that AI infrastructure demand remains strong based on accelerating LLM token activity, although GPU-rental and memory-spot-price data were less uniformly supportive. The tracker uses a monthly snapshot of roughly 300 models on OpenRouter. Token volume rose 71% month over month, accelerating from 47% in August and 18% in July, and was 30x higher year over year. Token spending increased 28% month over month, versus 7% growth in both August and July, and reached 14x year-over-year growth. The report attributes the volume acceleration primarily to open-weight models: their volume increased 83% month over month and 101x year over year, compared with 39% month-over-month and 10x year-over-year growth for closed-weight models. The report distinguishes average token pricing from volume-weighted pricing. Average pricing rose 4% month over month and 20% year over year, helped by the addition of more expensive image models. However, volume-weighted pricing fell 25% month over month and 55% year over year because demand shifted toward newer, cheaper high-volume models and because some existing models saw price reductions. DeepSeek V4.1 Flash and GLM 5.3 Flash added about 41 trillion tokens, or 30% of total volume, at prices below the prior open-weight-model average. Same-model price compression in Kimi K3, GLM 5.3, and GPT-5.6 Sol also contributed. The five largest models represented more than 53% of volume and 50% of spending, but only Tencent’s Hy4 preview appeared in both top-five rankings, illustrating that usage leadership and spending leadership differed. GPU rental pricing for non-hyperscaler capacity was mixed. A100 prices averaged $1.59 per GPU-hour, down 2.8% month over month, while H100 prices averaged $2.64 per GPU-hour, down 2.6%. In contrast, B200 rental prices rose 1.3% month over month to $5.70 per GPU-hour after August’s 1.5% decline. The B200-to-H100 rental-price ratio increased to 2.16x from 2.08x in August, while the H100-to-A100 ratio edged up to 1.66x from 1.65x. J.P. Morgan therefore characterizes compute-rental signals as mixed rather than uniformly strengthening. Memory spot-price momentum also paused. DDR5 16Gb spot pricing declined 1% month over month to $49.70 after five consecutive monthly increases, though it remained 614% above September 2025’s $6.96. NAND 1Tb pricing was effectively unchanged month over month, rising 0.1% to $30.54, but remained 470% above the $5.36 level a year earlier. These figures show that year-over-year memory pricing remained elevated even as month-over-month momentum weakened in September.

Analysis framework

The report combines a monthly OpenRouter sample of token usage and pricing across roughly 300 LLMs with Bloomberg indices for non-hyperscaler GPU rentals and memory spot prices. It compares monthly and annual growth, separates average from volume-weighted token prices, and uses model mix, price changes, and hardware-price ratios to interpret AI-infrastructure demand conditions.

Methodology notes

• Industry AnalysisVolume-price decomposition

  Token volume, average pricing, volume-weighted pricing, and spending are examined separately.

  The report separates demand growth from pricing and product-mix effects: spending can rise rapidly as token volumes expand even when the price paid by the volume-weighted mix falls.

• Industry AnalysisSupply-demand framework

  AI infrastructure demand is assessed through token activity, GPU rental prices, and memory spot prices.

  The report uses usage growth as a demand indicator and GPU and memory pricing as market signals for compute and component availability or demand conditions.

Key data

• OpenRouter token volume growth+71% m/m; 30x y/ySeptember growth, versus +47% m/m in August and +18% in July.
• OpenRouter token spending growth+28% m/m; 14x y/ySeptember growth, versus +7% m/m in both August and July.
• Volume-weighted token pricing-25% m/m; -55% y/yDeclined because of cheaper high-volume model mix and same-model price compression.
• A100 rental price$1.59 per GPU-hour; -2.8% m/mAverage September non-hyperscaler rental price.
• H100 rental price$2.64 per GPU-hour; -2.6% m/mAverage September non-hyperscaler rental price.
• B200 rental price$5.70 per GPU-hour; +1.3% m/mAverage September non-hyperscaler rental price.
• DDR5 16Gb spot price$49.70; -1% m/m; +614% y/yFirst monthly decline after five consecutive monthly increases.
• NAND 1Tb spot price$30.54; +0.1% m/m; +470% y/yEffectively flat month over month.

Impact & implications

The report interprets accelerating token volume and spending as evidence that AI infrastructure demand remains strong. At the same time, falling A100 and H100 rental prices and stalled month-over-month memory pricing indicate that pricing conditions across infrastructure inputs were not uniformly tightening.

Risks

• The OpenRouter tracker reflects developer, startup, and agentic-coding traffic and excludes first-party API volumes, including direct OpenAI, Anthropic, and hyperscaler-hosted endpoints.