Token spending accelerates as GPU rental and memory prices diverge
J.P. Morgan’s tracking shows that Token spending on OpenRouter rose 40% month over month in October, but usage growth slowed; prices for different GPUs and memory products followed divergent paths over the same period.
Original source：J.P. MorganFirst published on Unhidea：Original source date：
Key takeaway
J.P. Morgan believes growth in Token spending indicates that demand for AI infrastructure remains strong, but a rising share of higher-priced models accelerated spending growth, while signals from GPU rental and memory prices are more complex.
Why did Token spending accelerate as usage growth slowed?
The OpenRouter sample tracked by J.P. Morgan shows that Token spending rose 40% month over month in October, faster than the 28% recorded in September; Token usage, however, grew 25% month over month, below the 71% recorded in September.
The pricing mix pushed spending higher: the usage-weighted Token price rose 12% month over month in October, with usage of the higher-priced Claude Opus 5.5 model increasing by approximately 7.9 trillion Tokens; some existing models also became more expensive. For example, the price of GPT-5.6 Sol doubled after its half-price promotion ended.
Prices did not rise across the board: the simple average Token price fell 5% month over month in October, and although the usage-weighted price rebounded month over month, it remained down 38% year over year.
What demand does this Token data cover?
The bank samples one week of data each month, tracking approximately 300 large models on the OpenRouter platform. Platform traffic skews toward developers, startups and agentic coding, and excludes traffic calling OpenAI, Anthropic and hyperscaler-hosted endpoints directly.
Closed-source models accounted for 28% of Token usage in October, but 77% of spending; their spending grew 49% month over month, above the 15% for open-source models. The difference between models’ shares of usage and spending helps explain how higher-priced models affect total spending.
What signals are GPU rental prices sending?
Among GPU computing capacity provided by non-hyperscalers, the average A100 rental price fell 3.5% month over month in October to 1.53 dollars per GPU hour; H100 rebounded 6.5% to 2.83 dollars.
The average B200 rental price rose 4.7% month over month to 6.00 dollars per GPU hour, a larger increase than the 1.8% in September; but its price multiple relative to H100 fell from 2.16 times in September to 2.12 times. Rental price directions and relative prices differed between new and older models.
Why are some memory prices weakening while others remain supported?
In October, the DDR5 16Gb spot price rebounded 4.4% month over month to 51.80 dollars, ending a slight decline in September; the NAND 1Tb spot price fell 2.6% month over month to 29.40 dollars, declining for a second consecutive month.
TrendForce expects conventional DRAM and NAND contract prices to rise 10%–15% and 15%–20% respectively quarter over quarter in the fourth quarter of 2026, below the third-quarter increases of 16% and 21%. Previous price increases, pressure on consumer demand, ample client SSD buyer inventories and some long-term server DRAM pricing agreements combined to slow the increases.
Meanwhile, AI server demand and a shift in capacity toward server products continue to support prices; TrendForce expects enterprise SSDs to be the only memory category with faster price growth in the fourth quarter.
