AI infrastructure demand remains broadly strong, but price signals across various sub-markets are showing clear divergence. According to JPMorgan's latest data center observation report, September saw both Token usage and spending accelerate, yet GPU rental pricing trends diverged, and memory spot prices posted their first pullback after consecutive gains.
According to the tracking trading desk, JPMorgan analyst Joseph Cardoso noted in a report released on September 29 that Token usage on the OpenRouter platform surged 71% month-over-month in September, reaching 30 times the level year-over-year, while overall spending grew 28% month-over-month and expanded 14 times year-over-year.
Meanwhile, Nvidia's GPU rental market showed notable divergence: A100 and H100 rental prices both declined month-over-month, while B200 rebounded slightly after its first drop in August. On the memory side, DRAM spot prices posted their first slight month-over-month decline in September after five consecutive months of gains, while NAND prices remained roughly flat.
These signals carry mixed implications for AI infrastructure investors. The sustained acceleration in Token usage confirms the authenticity of demand at the AI application layer, but downward pressure on GPU rental prices and the pause in memory price gains suggest that supply-side expansion and shifts in demand structure are constraining hardware pricing.
Token Usage Accelerates, Cheap Open-Source Models Drive Growth
September saw a marked acceleration in Token usage growth. According to JPMorgan's tracking of OpenRouter platform data, September Token usage rose 71% month-over-month, exceeding August's 47% and July's 18%, with year-over-year growth reaching 30 times.
The core driver of growth came from open-source models. Open-source models, including DeepSeek, Moonshot AI and others, saw usage surge 83% month-over-month and a staggering 101 times year-over-year, significantly outpacing closed-source models (OpenAI, Anthropic, etc.), which grew 39% month-over-month and 10 times year-over-year. Closed-source models' share of total usage fell from 33% in August to 27%.
The top five models by usage were: DeepSeek V4.1 Flash, GLM 5.3 Flash, Tencent Hy4 preview, GPT-5.6 Luna and DeepSeek V4 Flash, together accounting for over 53% of total Token usage. Among them, DeepSeek V4.1 Flash and GLM 5.3 Flash together contributed approximately 41 trillion Tokens in a single month, representing about 30% of the total, with pricing far below the average of previous open-source models.
Volume-Price Divergence: Spending Growth Relies on Usage Expansion Rather Than Price Increases
The explosive growth in Token usage did not drive unit prices higher, with volume-price divergence being a prominent feature. September's volume-weighted average price (VWAP) fell 25% month-over-month and 55% year-over-year, driven primarily by two factors: first, usage concentrated toward cheaper high-traffic models, with the large-scale adoption of DeepSeek V4.1 Flash and GLM 5.3 Flash pulling down the overall average price; second, models such as Kimi K3, GLM 5.3 and GPT-5.6 Sol saw price cuts on comparable offerings.
Nevertheless, overall Token spending still achieved accelerating growth, rising 28% month-over-month in September, higher than the 7% growth in both July and August, and expanding 14 times year-over-year. The primary contribution to spending growth came from open-source models, whose spending grew 64% month-over-month and 132 times year-over-year, with their share of total spending rising from 22% in August to 28%.
It is worth noting that among the top five models by spending and the top five models by usage, only one overlaps. The top five by spending were GPT-6 Astra, Tencent Hy4 preview, Claude Fable 5.1, Claude Opus 5 and GPT-5.6 Sol, together accounting for 50% of total spending, indicating that high-priced closed-source flagship models still dominate the revenue side.
GPU Rentals: H100 Under Pressure, B200 Stabilizes and Rebounds
The GPU rental market among non-hyperscale cloud service providers showed clear divergence in September. According to Bloomberg index data, the average A100 rental price was $1.59 per GPU-hour, down 2.8% month-over-month, a larger decline than August's 0.7%; the average H100 rental price was $2.64 per GPU-hour, down 2.6% month-over-month, compared with a 0.4% month-over-month increase in August.
B200 moved in the opposite direction. In September, the average B200 rental price recovered to $5.70 per GPU-hour, up 1.3% month-over-month, reversing August's 1.5% month-over-month decline. The B200-to-H100 price ratio rose to 2.16 times (from 2.08 times in August), and the H100-to-A100 price ratio also edged up to 1.66 times (from 1.65 times in August).
JPMorgan noted that B200 pricing is approximately 2.2 times that of H100, and H100 pricing is approximately 1.7 times that of A100, with both ratios rising month-over-month, reflecting that the market's relative premium for next-generation compute remains intact, though supply pressure on older GPU models is intensifying.
Memory Prices: DRAM Posts First Decline After Five Consecutive Gains, NAND Stabilizes
The rally in the memory spot market paused in September. According to Bloomberg data, the DDR5 16Gb spot price was $49.70 in September, down about 1% month-over-month, marking the first decline after five consecutive months of gains, though still up 614% year-over-year (compared with $6.96 in the same period last year).
On the NAND side, the 1Tb spot price was $30.54 in September, up 0.1% month-over-month, essentially flat, with a year-over-year gain of about 470% (compared with $5.36 in the same period last year). JPMorgan noted that NAND prices turned positive last month for the first time after four consecutive months of slight declines, and remained stable in September.
Whether the month-over-month decline in DRAM prices heralds the end of the rally remains undetermined, but the absolute level of year-over-year gains still exceeding sixfold indicates that AI-driven memory demand expansion has provided substantial support for prices over the past year.