AI infrastructure demand remains strong, B200 rental pricing is strengthening, and memory price inflation is starting to cool

Overview

The report tracks three sets of high-frequency data for AI infrastructure demand: LLM token usage and pricing on OpenRouter, non-hyperscaler GPU rental pricing, and DRAM/NAND memory pricing. The overall conclusion is that the AI infrastructure demand environment remains healthy: LLM usage and spending continue to expand, and rental pricing growth for the latest-generation B200 GPU is accelerating; however, contract memory price increases are expected to slow materially in 3Q26.

Core views

The core views include: first, LLM token growth slowed in July versus June, but remained in a high year-over-year growth range, indicating that AI inference demand continues to expand; second, closed-source models still dominate spending, but pricing and usage of open-weight models are growing rapidly and increasing their share of spending; third, GPU rental pricing is showing generational divergence, with B200 stronger relative to H100 and A100, which may reflect demand migrating toward the latest GPUs; fourth, while DRAM and NAND spot prices are still up sharply year over year, contract price increases are expected to pull back from 1H26 highs.

Analysis framework

The report uses a monthly tracking framework that combines a one-week usage and pricing sample for about 300 LLMs on OpenRouter, Bloomberg GPU rental price indexes, Bloomberg memory spot data, and TrendForce contract price forecasts to assess marginal changes in AI infrastructure demand, compute pricing, and memory costs.

Methodology notes

• Demand trackingLLM Token Tracker

  Observe changes in LLM token volume, pricing, and spending through the OpenRouter sample.

  This tracker collects one week of usage and pricing data each month, covering about 300 LLMs listed on OpenRouter; the sample is skewed toward developers, startups, and agentic-coding traffic, and does not include direct first-party API traffic from OpenAI, Anthropic, and others, or hyperscaler-hosted endpoints.

• Price trackingGPU Rental Price Tracker

  Use non-hyperscaler GPU rental pricing to measure the tightness of external compute supply and demand.

  The report compares monthly average rental pricing and generational spreads for A100, H100, and B200 to observe whether demand is migrating toward the latest GPUs.

• Cost trackingMemory Price Tracker

  Use DRAM and NAND spot and contract pricing to assess AI hardware cost pressure.

  The report combines Bloomberg spot pricing with TrendForce contract price expectations to distinguish short-term spot strength or weakness from quarterly contract pricing trends.

Asset mapping & comparison

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

• AI infrastructure

  Benefiting from growth in LLM usage and spending

  Strengths

  Token volume grew 21x year over year and spending grew 11x year over year, showing that inference demand continues to expand.

  Weaknesses

  The month-over-month growth rate in July slowed significantly versus June, indicating some near-term cooling in momentum.

  Comparison

  Compared with June, both July token volume and spending slowed, but year-over-year strength remained high.

  Risks

  The OpenRouter sample does not cover first-party APIs or hyperscaler-hosted endpoints, so it may not represent total market demand.

• Nvidia B200 GPU rental market

  Benefiting from demand migration toward the latest-generation GPUs

  Strengths

  Average rental pricing rose 7.3% month over month in July, and the B200/H100 price ratio increased to 2.13x.

  Weaknesses

  Pricing performance may be affected by the sample mix of non-hyperscaler capacity.

  Comparison

  B200 outperformed A100 at +0.9% and H100 at -1.1%.

  Risks

  If supply increases or demand weakens, the rental price premium may narrow.

• DRAM

  Still supported by AI and the hardware cycle, but the slope of price increases is slowing

  Strengths

  DDR5 16Gb spot pricing rose 890% year over year, and 3Q26E contract pricing is still expected to rise 16% quarter over quarter.

  Weaknesses

  The 3Q26E contract price increase is significantly slower than the 56% seen in 2Q26.

  Comparison

  DRAM spot pricing is still rising month over month, while NAND spot pricing is down slightly month over month.

  Risks

  A high base, weaker consumer demand, and higher inventory after long-term agreements and advance purchases in 1H26 could cap price increases.

• NAND

  Pricing remains above last year's level, but near-term momentum is weaker than DRAM

  Strengths

  NAND 1Tb spot pricing rose 417% year over year, and 3Q26E contract pricing is expected to rise 13% quarter over quarter.

  Weaknesses

  Spot pricing fell 0.5% month over month in July, marking the fourth consecutive month of slight declines.

  Comparison

  NAND 3Q26E contract price growth is below DRAM's 16%.

  Risks

  Higher inventory and weak consumer demand may cause prices to cool further.

• Open-weight LLM ecosystem

  Its spending share is rising and it is eroding the share of closed-source models

  Strengths

  In July, spending on open-weight models grew 71% month over month, their spending share rose to 20%, and Kimi K3 entered the top five by spending.

  Weaknesses

  Its overall spending share still remains below that of closed-source models.

  Comparison

  Closed-source models contribute 80% of spending, but the share of open-weight models increased clearly versus May and June.

  Risks

  Changes in model performance, pricing, and the pace of developer adoption could cause share volatility.

Key data

• July token volume+18% m/m; 21x y/yIn the OpenRouter sample, July token volume growth slowed month over month versus 70% in June, but remained at a high year-over-year growth level.
• July average token price-3% m/m; +2% y/yAverage token pricing declined slightly month over month, but shifted from a prior year-over-year decline to slight growth.
• July total token spending+7% m/m; 11x y/ySpending growth slowed versus June, but still increased significantly year over year.
• Closed-source model spending share80%Closed-source models account for only 29% of token volume, but still contribute most of the spending.
• Open-weight model spending share20%Open-weight model spending share rose from 10% in May and 12% in June to 20% in July.
• A100 rental pricing$1.66/GPU-hour; +0.9% m/mA100 pricing rose month over month for the sixth consecutive month in July, but the increase was modest.
• H100 rental pricing$2.68/GPU-hour; -1.1% m/mH100 pricing declined month over month for the first time after seven consecutive months of increases.
• B200 rental pricing$5.72/GPU-hour; +7.3% m/mThe B200 price increase was significantly larger than for A100 and H100, with the B200/H100 price ratio rising to 2.13x.
• DDR5 16Gb spot price$50.83; +17% m/m; +890% y/yDRAM spot pricing continued to rise in July.
• NAND 1Tb spot price$26.90; -0.5% m/m; +417% y/yNAND spot pricing declined slightly month over month for the fourth consecutive month, but remained sharply higher year over year.
• 3Q26E DRAM contract pricing+16% q/qThis slowed materially from +56% q/q in 2Q26.
• 3Q26E NAND contract pricing+13% q/qThis slowed materially from +58% q/q in 2Q26.

Impact & implications

In terms of investment implications, the report supports the view that demand across the AI infrastructure chain remains resilient, especially in high-end GPUs and inference-traffic-related segments; at the same time, memory pricing is shifting from extreme increases to more moderate gains, which may ease some downstream hardware and data center cost pressure. The rising spending share of open-weight models also suggests that the competitive landscape of the model ecosystem is changing, which could affect compute consumption patterns and model service pricing.

Risks

• The OpenRouter sample is skewed toward developers, startups, and agentic-coding traffic, and does not include direct first-party API traffic or hyperscaler-hosted endpoints.
• July month-over-month growth in LLM token volume and spending slowed significantly versus June, which may signal volatility in near-term growth momentum.
• GPU rental pricing is based on non-hyperscaler capacity and may not fully represent internal supply and demand at large cloud providers.
• DRAM and NAND contract price increases are expected to slow; if consumer demand weakens further or inventory continues to rise, memory pricing could come under pressure.
• The report is not company-specific research and does not provide company ratings, target prices, or a valuation framework.

What to watch

• Whether OpenRouter token volume, average pricing, and total spending re-accelerate in subsequent months.
• Whether the spending share of open-weight models continues to rise, and whether models such as Kimi K3 can maintain high spending rankings.
• Whether the B200/H100 and H100/A100 rental price ratios continue to diverge.
• Whether actual 3Q26 DRAM and NAND contract pricing matches TrendForce expectations of +16% and +13%.
• The impact of consumer electronics demand, CSP long-term agreements, and buyer inventory changes on memory pricing.