AI infrastructure and data center demand: September token spending accelerated sharply, but GPU rental and memory pricing gave a more mixed AI-infrastructure signal.

Overview

JPMorgan tracks monthly indicators for AI infrastructure demand through LLM token activity, non-hyperscaler GPU rentals, and memory spot prices. Its September reading points to strong and accelerating token demand, but less uniform pricing support across compute rental and memory markets.

Core views

JPMorgan's September Data Center Watch argues that the latest token-spending data continue to support a strong AI-infrastructure-demand backdrop, although GPU rental and memory spot-price trends are more mixed. The report's LLM Token Tracker uses a one-week monthly snapshot of usage and pricing across roughly 300 models on OpenRouter. It cautions that this dataset is weighted toward developer, startup, and agentic-coding traffic and excludes first-party API volumes from direct OpenAI, Anthropic, and hyperscaler-hosted endpoints. Token volumes accelerated materially in September. Overall volume grew 71% month over month, versus 47% in August and 18% in July, and reached 30x year-over-year growth, versus 28x in August and 21x in July. Closed-weight models represented 27% of volume, down from 33% in August, and their growth moderated to 39% month over month and 10x year over year. By contrast, open-weight-model volumes accelerated to 83% month over month and 101x year over year, compared with 39% month over month and 56x year over year in August. The report identifies the stronger open-weight contribution as a key driver of the aggregate acceleration. Pricing moved in opposite directions depending on the measure. Volume-weighted average token pricing declined 25% month over month and 55% year over year, reflecting a mix shift toward newer, cheaper high-volume models and same-model price compression. DeepSeek V4.1 Flash and GLM 5.3 Flash added about 41 trillion tokens, around 30% of total volume, at prices below the prior open-weight average; Kimi K3, GLM 5.3, and GPT-5.6 Sol also saw price compression. Yet simple average token pricing rose 4% month over month and 20% year over year, partly because more expensive image models entered the catalog. This distinction means rapidly growing lower-priced usage can depress volume-weighted pricing even as the average listed price rises. Combined volume and pricing produced stronger token spending: aggregate spending rose 28% month over month, versus 7% in both August and July, and 14x year over year, versus 12x in August and 11x in July. Closed-weight-model spending, 72% of total spending versus 78% in August, increased 18% month over month and 10x year over year. Open-weight-model spending accelerated more sharply, rising 64% month over month and 132x year over year. Usage and spending leadership also differed: the five largest models by token volume represented more than 53% of volume, while the top five by spending represented 50% of spend, with only Tencent's Hy4 preview appearing in both groups. Non-hyperscaler Nvidia GPU rental pricing was mixed. A100 rental pricing averaged $1.59 per GPU-hour in September, down 2.8% month over month, while H100 pricing averaged $2.64 per GPU-hour, down 2.6%. The H100-to-A100 ratio nevertheless edged up to 1.66x from 1.65x in August. B200 rental pricing rebounded 1.3% month over month to $5.70 per GPU-hour after an August decline of 1.5%; its price premium to H100 increased to 2.16x from 2.08x. Thus, newer B200 pricing recovered modestly while legacy A100 and H100 rental rates softened. Memory spot-price momentum also slowed. DDR5 16Gb DRAM declined 1% month over month to $49.70 after five straight monthly increases, but was still up 614% year over year from $6.96 in September 2025. NAND 1Tb pricing was effectively flat, increasing 0.1% month over month to $30.54, after a first monthly increase following four modest declines; it remained up 470% year over year from $5.36. The report therefore characterizes memory as no longer providing the same month-to-month upward price signal, despite exceptionally elevated year-over-year comparisons.

Analysis framework

The report triangulates AI-infrastructure conditions using three monthly trackers: OpenRouter token volume, spending, and pricing; Bloomberg non-hyperscaler GPU rental indices; and Bloomberg DRAM and NAND spot-price data. It compares sequential and year-over-year changes, separates open-weight from closed-weight models, and distinguishes average token pricing from volume-weighted pricing to explain how model mix affects spending.

Methodology notes

• Industry AnalysisVolume-price decomposition

  Token spending analysis based on token volumes and pricing

  The report explains spending growth by separating changes in token usage from changes in pricing, including the effect of cheaper models gaining a larger share of volume.

• Other

  Monthly OpenRouter usage-and-pricing sample

  The tracker takes one week of monthly data across roughly 300 models on OpenRouter, providing a directional snapshot rather than a measure of all first-party LLM API activity.

Key data

• Overall token volume growth+71% m/m; 30x y/ySeptember growth, versus +47% m/m and 28x y/y in August
• Overall token spending growth+28% m/m; 14x y/yVersus +7% m/m and 12x y/y in August
• Volume-weighted token pricing-25% m/m; -55% y/yDeclined due to lower-priced model mix and same-model price compression
• A100 rental price$1.59 per GPU-hour; -2.8% m/mSeptember non-hyperscaler average
• H100 rental price$2.64 per GPU-hour; -2.6% m/mH100-to-A100 price ratio was 1.66x
• B200 rental price$5.70 per GPU-hour; +1.3% m/mB200-to-H100 price ratio increased to 2.16x
• 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 sequentially

Impact & implications

The report interprets accelerating token usage and spending as evidence that AI-infrastructure demand remains strong. However, declining A100 and H100 rental rates and stalled memory-price growth indicate that pricing signals across the infrastructure stack are not uniformly strengthening.