AI demand is diverging: Token usage is surging, H100 falls while B200 rebounds, and DRAM prices fall for the first time after five consecutive days of gains.

AI demand is diverging: Token usage is surging, H100 falls while B200 rebounds, and DRAM prices fall for the first time after five consecutive days of gains.

Overall demand for AI infrastructure remains strong, but price signals across different market segments are showing significant divergence. According to JPMorgan's latest data center watchdog report, both token usage and spending accelerated in September, but GPU leasing prices diverged, and memory spot prices saw their first decline after a period of continuous increases.

According to TrendFocus, a report released on September 29 by JPMorgan analyst Joseph Cardoso pointed out that in September, the usage of tokens on the OpenRouter platform surged by 71% month-on-month and 30 times year-on-year, while overall spending increased by 28% month-on-month and 14 times year-on-year.

Meanwhile, the Nvidia GPU rental market showed a clear divergence: rental prices for both the A100 and H100 declined month-over-month, while the B200 saw a slight rebound after its first drop in August. In the memory market, DRAM spot prices saw their first slight month-over-month decline in September after five consecutive months of increases, while NAND prices remained largely unchanged.

These signals have complex implications for AI infrastructure investors. The continued acceleration in token usage confirms the genuine demand for AI applications, but downward pressure on GPU leasing prices and a pause in memory price increases suggest that supply-side expansion and changes in demand structure are constraining hardware pricing.

Token usage is accelerating, with inexpensive open-source models driving growth.

Token usage growth accelerated significantly in September. According to data tracked by JPMorgan Chase on the OpenRouter platform, token usage increased by 71% month-over-month in September, higher than August's 47% and July's 18%, and the year-over-year growth rate reached 30 times.

The core driver of growth comes from open-source models. Usage of open-source models (including DeepSeek, Moonshot AI, etc.) surged 83% month-over-month and a staggering 101-fold year-over-year, significantly outpacing the 39% month-over-month and 10-fold year-over-year growth of closed-source models (OpenAI, Anthropic, etc.). The share of closed-source models in total usage has decreased from 33% in August to 27%.

The top five models by usage are: DeepSeek V4.1 Flash, GLM 5.3 Flash, Tencent Hy4 preview, GPT-5.6 Luna, and DeepSeek V4 Flash, accounting for over 53% of the total token usage. Among them, DeepSeek V4.1 Flash and GLM 5.3 Flash contributed approximately 41 trillion tokens in a single month, accounting for about 30% of the total supply, and their pricing was significantly lower than the average of previous open-source models.

Price-volume divergence: Spending growth relies on increased usage rather than price increases

The explosive growth in token usage has not led to an increase in the unit price, showing a significant divergence between volume and price. The volume-weighted average price (VWAP) in September decreased by 25% month-on-month and 55% year-on-year, mainly driven by two factors: First, usage is concentrated in cheaper, high-volume models, with the large-scale adoption of DeepSeek V4.1 Flash and GLM 5.3 Flash lowering the overall average price; second, models such as Kimi K3, GLM 5.3, and GPT-5.6 Sol have experienced similar price reductions.

Despite this, overall token spending still accelerated, increasing by 28% month-over-month in September, higher than the 7% growth rates in July and August, and expanding 14 times year-over-year. The main contributor to the spending growth came from the open-source model, whose spending increased by 64% month-over-month and 132 times year-over-year, and its share of total spending rose from 22% in August to 28%.

It's worth noting that the top five models ranked by expenditure overlap with only one of the top five models ranked by usage. The top five in terms of expenditure are GPT-6 Astra, Tencent Hy4 preview, Claude Fable 5.1, Claude Opus 5, and GPT-5.6 Sol, accounting for 50% of total expenditure, indicating that high-priced, closed-source flagship models still dominate revenue.

GPU Leasing: H100 Under Pressure, B200 Stabilizes and Rebounds

The GPU leasing market for non-hyperscale cloud service providers showed significant divergence in September. According to Bloomberg data, the average rental price for A100 was $1.59 per GPU hour, a 2.8% decrease month-over-month, with the decline widening compared to the 0.7% drop in August; the average rental price for H100 was $2.64 per GPU hour, a 2.6% decrease month-over-month, while August saw a 0.4% increase.

The B200 trend is the opposite of the previous two. In September, the average rental price of the B200 rebounded to $5.70 per GPU hour, a 1.3% increase month-over-month, reversing the 1.5% month-over-month decline in August. The price ratio of B200 to H100 rose to 2.16 times (2.08 times in August), and the price ratio of H100 to A100 also rose slightly to 1.66 times (1.65 times in August).

JPMorgan Chase points out that the B200 is priced at about 2.2 times that of the H100, and the H100 is priced at about 1.7 times that of the A100. Both ratios have increased month-on-month, reflecting that the market's relative premium for the new generation of computing power is still maintained, but the supply pressure of older GPU models is intensifying.

Memory prices: DRAM prices fell for the first time after five consecutive increases, while NAND prices stabilized.

The upward trend in the memory spot market paused in September. According to Bloomberg data, the spot price of DDR5 16Gb was $49.70 in September, down about 1% month-on-month, marking the first decline after five consecutive months of increases, but still up significantly by 614% year-on-year (compared to $6.96 in the same period last year).

Regarding NAND, the spot price of 1Tb was $30.54 in September, a slight increase of 0.1% month-on-month, essentially flat, and a year-on-year increase of approximately 470% (compared to $5.36 in the same period last year). JPMorgan Chase pointed out that NAND prices turned positive for the first time last month after four consecutive months of slight declines, and remained stable in September.

Whether the month-on-month decline in DRAM prices signals the end of the upward trend remains to be seen, but the year-on-year increase of more than 6 times indicates that the expansion of AI-driven memory demand has provided substantial support for prices over the past year.

~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Zhuifeng Trading Platform .

For more detailed analysis, including real-time updates and firsthand research, please join the [ Trading Channel Annual Membership ].

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.