AI infrastructure and data center demand: September token-spending acceleration supports AI-infrastructure demand, but GPU rental and memory pricing diverge

Overview

This Data Center Watch tracks AI-infrastructure demand through LLM token activity, non-hyperscaler GPU rental pricing, and memory spot prices. J.P. Morgan finds that September token spending and usage strengthened materially, but hardware-service and memory-price signals did not move uniformly higher.

Core views

J.P. Morgan’s central conclusion is that AI-infrastructure demand remained strong in September, led by a marked acceleration in LLM token spending. Its OpenRouter tracker showed token volume up 71% month on month, compared with 47% in August and 18% in July, and up 30x year on year. Average token pricing increased 4% month on month and 20% year on year, helping overall token spending rise 28% month on month and 14x year on year, versus 7% month-on-month growth in each of July and August. The report interprets the spending acceleration as evidence of robust demand even though the pricing picture depends substantially on model mix. Volume-weighted average token pricing declined 25% month on month and 55% year on year. The report attributes this to a shift toward newer, cheaper, high-volume models and to price compression for the same models. DeepSeek V4.1 Flash and GLM 5.3 Flash alone added about 41 trillion tokens, or 30% of total volume, at prices below the prior open-weight average. The report also cites price compression for Kimi K3, GLM 5.3, and GPT-5.6 Sol. Open-weight model usage grew 83% month on month and 101x year on year, sharply outpacing closed-weight model usage growth of 39% month on month and 10x year on year. Open-weight models therefore gained importance in the tracked volume mix, while closed-weight models accounted for 27% of token volume but 72% of spending. The report distinguishes between usage leadership and spending leadership. The five largest models by token volume—DeepSeek V4.1 Flash, GLM 5.3 Flash, Tencent’s Hy4 preview, GPT-5.6 Luna, and DeepSeek V4 Flash—represented more than 53% of total volume, up from more than 50% in August. By spending, the leading models were GPT-6 Astra, Tencent’s Hy4 preview, Claude Fable 5.1, Claude Opus 5, and GPT-5.6 Sol, which represented 50% of spend, up from 48% in August. Only one model overlapped between the two top-five lists, underscoring that high-volume activity and spending are being driven by different model sets. J.P. Morgan cautions that its LLM tracker is a monthly one-week snapshot covering roughly 300 models listed on OpenRouter. OpenRouter aggregates API access through a single endpoint, but its traffic is skewed toward developers, startups, and agentic-coding use cases and excludes first-party API volumes such as direct OpenAI or Anthropic usage and hyperscaler-hosted endpoints. The tracker is therefore presented as an indicator of selected LLM-market trends rather than a complete measure of all model consumption. GPU rental pricing provided a less uniformly positive demand signal. Based on Bloomberg indices for non-hyperscaler capacity, A100 rental pricing averaged $1.59 per GPU-hour in September, down 2.8% month on month, while H100 pricing averaged $2.64 per GPU-hour, down 2.6%. The H100-to-A100 price ratio nevertheless edged up to 1.66x from 1.65x in August. B200 rental pricing rebounded 1.3% month on month to $5.70 per GPU-hour after its August decline, and its premium to H100 increased to 2.16x from 2.08x. Thus, older-generation rental prices weakened while B200 pricing recovered modestly and maintained a substantially higher price tier. Memory spot-price trends also stalled after prior strength. DDR5 16Gb spot prices fell 1% month on month to $49.70 in September, following five consecutive monthly increases, although the price remained 614% above $6.96 in September 2025. NAND 1Tb spot prices were effectively flat, rising 0.1% month on month to $30.54, while remaining 470% above $5.36 a year earlier. The report therefore characterizes memory pricing as no longer accelerating sequentially, despite still-extraordinary year-on-year increases.

Analysis framework

The report combines three monthly indicators: a one-week OpenRouter sample of LLM token usage, prices and spending; Bloomberg indices of non-hyperscaler GPU rental rates; and Bloomberg memory spot-price data. It compares month-on-month and year-on-year movements, separates open-weight from closed-weight models, and uses price and volume mix to explain changes in token spending.

Methodology notes

• Other

  Monthly tracker of OpenRouter token volume, pricing, and spending

  J.P. Morgan samples one week each month across roughly 300 OpenRouter models to monitor changes in usage, average prices, volume-weighted prices, and spending. The report notes that this sample excludes direct first-party and hyperscaler-hosted API volumes.

• Industry AnalysisVolume-price decomposition

  Token-spending analysis using volume, average price, and volume-weighted pricing

  The report explains spending growth by separating token-volume expansion from pricing effects and model mix. High-volume, lower-priced open-weight models reduced volume-weighted prices even as average token prices rose.

• Industry AnalysisSupply-demand framework

  GPU rental and memory spot prices as AI-infrastructure demand indicators

  The report uses sequential movements in GPU rental rates and DRAM and NAND spot prices to assess whether pricing conditions across important AI-infrastructure inputs are strengthening or weakening.

Key data

• OpenRouter token volume growth+71% m/m; 30x y/ySeptember growth accelerated from +47% m/m in August.
• OpenRouter token spending growth+28% m/m; 14x y/yCompared with +7% m/m in both July and August.
• Volume-weighted token pricing-25% m/m; -55% y/yDeclined because of cheaper model mix and same-model price compression.
• A100 rental price$1.59 per GPU-hour; -2.8% m/mNon-hyperscaler average rental rate in September.
• H100 rental price$2.64 per GPU-hour; -2.6% m/mH100-to-A100 price ratio reached 1.66x.
• B200 rental price$5.70 per GPU-hour; +1.3% m/mB200-to-H100 ratio rose to 2.16x from 2.08x in August.
• DDR5 16Gb spot price$49.70; -1% m/m; +614% y/yFirst month-on-month decline after five consecutive monthly gains.
• NAND 1Tb spot price$30.54; +0.1% m/m; +470% y/yEffectively flat sequentially after the prior month's increase.

Impact & implications

The report argues that the acceleration in token spending supports continued AI-infrastructure demand, particularly as open-weight model usage expands rapidly. At the same time, falling A100 and H100 rental rates and stalled sequential memory-price momentum indicate that infrastructure pricing conditions remain uneven rather than broadly accelerating.

Risks

• The OpenRouter sample is skewed toward developer, startup, and agentic-coding traffic and excludes direct first-party and hyperscaler-hosted API volumes, limiting its representation of total LLM demand.

What to watch

• Whether token-volume and token-spending growth remains elevated after September’s acceleration.
• The continuing shift between open-weight and closed-weight model usage and spending.
• Sequential rental-price trends for A100, H100, and B200 GPUs.
• Whether DDR5 and NAND spot prices resume sequential increases or remain flat to lower.