LLM usage and GPU rental prices strengthened in tandem, continuing to support AI infrastructure demand

Overview

This report is J.P. Morgan's monthly tracking of AI data center infrastructure demand, focusing on LLM token usage, pricing, and spending on OpenRouter, GPU rental prices among non-hyperscale cloud providers, and spot prices for DRAM and NAND memory. June data overall point to an improving AI infrastructure demand environment: LLM usage and spending re-accelerated, and GPU rental prices continued to rise month over month; however, memory showed internal divergence, with DRAM continuing to rise while NAND posted consecutive modest declines.

Core views

The core view is that the expansion in LLM usage is sufficient to offset the pressure from year-over-year declines in token prices, and that token economics for model providers remain favorable under conservative pricing assumptions; meanwhile, GPU rental prices rose across A100, H100, and B200, further supporting the view of strong AI compute demand. On the memory side, DRAM prices continued to rise but at a slower pace, while NAND, after a sharp increase in the first quarter, posted modest declines for three consecutive months, indicating inconsistent pricing momentum across the storage chain.

Analysis framework

The report uses a high-frequency data tracking framework: a one-week monthly sample of about 300 LLMs on OpenRouter to observe monthly trends in token volume, pricing, and spending; Bloomberg-related indices to track GPU rental prices among non-hyperscale cloud providers; and Bloomberg memory spot prices to observe changes in DDR5 DRAM and NAND 1Tb pricing.

Methodology notes

• Data TrackingLLM Token Tracker

  OpenRouter monthly sample

  Each month, one week of usage and pricing data from about 300 LLMs on OpenRouter is collected as a monthly snapshot of LLM demand and commercialization trends. The sample is skewed toward developers, startups, and agentic coding traffic, and does not include direct first-party API volume from OpenAI, Anthropic, etc., or traffic from hyperscaler-hosted endpoints.

• Price IndexGPU Rental Price Tracker

  GPU rental prices among non-hyperscale cloud providers

  Relevant Bloomberg data are used to track hourly rental prices for A100, H100, and B200 GPUs, as well as price ratios across GPU types, to assess the tightness of compute supply-demand conditions among non-hyperscale cloud providers.

• Spot Price TrackingMemory Price Tracker

  DRAM and NAND spot prices

  Bloomberg Finance L.P. data are used to track spot prices for DDR5 16Gb DRAM and NAND 1Tb, in order to observe upstream storage costs and pricing momentum in AI infrastructure.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AI infrastructure and data center hardware

  Positively correlated with demand

  Strengths

  LLM token volume, spending, and GPU rental prices all strengthened simultaneously, supporting improved AI infrastructure demand.

  Weaknesses

  The OpenRouter sample does not cover first-party APIs or hyperscaler-hosted traffic and therefore cannot fully represent total market demand.

  Comparison

  June growth in token volume and spending was clearly higher than in April and May, indicating renewed acceleration in the trend.

  Risks

  If token prices continue to decline or the sample mix changes, conclusions about demand and commercialization may weaken.

• Nvidia GPU rental capacity (A100, H100, B200)

  Price increases reflect tight supply-demand conditions

  Strengths

  Rental prices for all three GPU tiers rose month over month, with A100 and H100 increasing for multiple consecutive months.

  Weaknesses

  The relative premiums of B200 over H100 and H100 over A100 have declined consecutively, suggesting marginal narrowing in high-end GPU premiums.

  Comparison

  B200 is about 1.96x H100, while H100 is about 1.67x A100.

  Risks

  New supply additions, cloud provider pricing changes, or lower utilization could pressure rental prices.

• DRAM

  Price positive

  Strengths

  DDR5 16Gb spot prices rose for the third consecutive month, up 10% month over month and 740% year over year in June.

  Weaknesses

  The report notes that DRAM is still rising, but at a slower pace.

  Comparison

  This contrasts with NAND's consecutive modest declines.

  Risks

  If AI server demand slows or supply increases, spot prices may fall back.

• NAND

  Neutral to cautiously negative

  Strengths

  NAND 1Tb prices were still up 412% year over year, with absolute levels significantly above the same period last year.

  Weaknesses

  June prices fell 0.3% month over month, marking the third consecutive modest decline.

  Comparison

  Compared with DRAM, NAND lost momentum after a strong rise in the first quarter.

  Risks

  Consecutive month-over-month weakness may reflect short-term inventory or demand pressure.

• U.S. model service providers (OpenAI, Anthropic, Google, xAI, etc.)

  Clear advantage on the spending side

  Strengths

  Although U.S. models' usage share fell to 35%, their spending share exceeded 85%, and their volume-weighted price rose 77% year over year.

  Weaknesses

  Their usage share remained below the prior two months, indicating stronger traffic-side competition from non-U.S. or lower-priced models.

  Comparison

  U.S. model spending and pricing performance were significantly stronger than the overall model sample.

  Risks

  If demand for high-priced models shifts toward lower-priced models, spending share and pricing advantages may come under pressure.

Key data

• LLM token volume+70% m/m, about 20x y/yToken volume in the June OpenRouter sample re-accelerated, stronger than +33% m/m in May and +5% m/m in April.
• U.S. model token volume+30% m/m, about 8x y/y; 35% share of total volumeU.S. models include OpenAI, Anthropic, Google, xAI, etc., and their share was below 46% in May and 56% in April.
• Average token price+7% m/m, -5% y/yThe year-over-year decline narrowed significantly from -21% in May and -24% in April.
• U.S. models volume-weighted average price+27% m/m, +77% y/yU.S. model pricing trends were stronger than the overall sample and were an important reason for the narrowing year-over-year decline.
• LLM token spending+70% m/m, about 16x y/yJune spending growth re-accelerated after slowing in the prior two months.
• U.S. model spending+65% m/m, about 14x y/y; more than 85% share of total spendingDespite a decline in usage share, U.S. models still dominated the spending side.
• A100 rental price$1.63/GPU-hour, +6.3% m/mJune marked the fifth consecutive month of month-over-month increases.
• H100 rental price$2.72/GPU-hour, +3.7% m/m; H100/A100 at 1.67xJune marked the seventh consecutive month of month-over-month increases, while the H100 premium versus A100 continued to edge lower.
• B200 rental price$5.33/GPU-hour, +2.7% m/m; B200/H100 at 1.96xB200 pricing remained about 2x that of H100, but the relative premium continued to decline.
• DDR5 16Gb DRAM spot price$43.14, +10% m/m, +740% y/yJune marked the third consecutive month of month-over-month increases.
• NAND 1Tb spot price$27.03, -0.3% m/m, +412% y/yAfter a sharp rise in the first quarter, it fell modestly month over month for the third consecutive month.
• Top five models by usage in JuneDeepSeek V4 Flash, MiMo-V2.5, MiniMax M3, Hy3 preview, Claude Opus 4.7The top five models together accounted for about 45% of total token volume.
• Top five models by spending in JuneClaude Opus 4.7, Claude Opus 4.8, GPT-5.5, Claude Sonnet 4.6, GLM 5.2The top five models together accounted for about 75% of total token spending.

Impact & implications

The report's implications for the AI infrastructure chain are moderately positive: accelerating LLM usage and spending imply continued expansion in inference demand; rising GPU rental prices indicate ongoing supply-demand support for compute among non-hyperscale cloud providers; sustained DRAM price increases are favorable for memory-related supply chains, but consecutive declines in NAND suggest diverging conditions within storage categories.

Risks

• The OpenRouter sample is skewed toward developers, startups, and agentic coding traffic and cannot represent the entire LLM market.
• The sample excludes direct first-party API volume from OpenAI, Anthropic, etc., as well as hyperscaler-hosted endpoints, which may understate or distort the true demand mix.
• Token economics judgments are based on the sample and simplified assumptions, not a full financial model.
• Average token prices are still down year over year; if usage cannot continue offsetting price declines, revenue momentum for model providers may slow.
• GPU rental prices reflect only non-hyperscale cloud provider capacity and may not be equivalent to full-market compute pricing.
• DRAM and NAND prices are highly volatile, and changes in inventory, supply releases, and end-demand could quickly alter trends.

What to watch

• Whether OpenRouter token volume, pricing, and spending in July and beyond continue the June re-acceleration.
• Whether U.S. models can still maintain more than 85% of spending share after their usage share declined.
• Whether A100, H100, and B200 rental prices continue to rise month over month, and whether relative premiums for high-end GPUs continue to narrow.
• Whether DRAM price gains continue to slow, and whether NAND ends its run of consecutive modest declines.
• Changes in the usage and spending rankings of leading models such as Claude Opus, GPT, DeepSeek, and MiniMax.
• Whether any new data on first-party APIs and hyperscaler-hosted traffic validate the OpenRouter sample trend.