JPMorgan Report: October token spending rises 40% MoM, GPU rental and storage prices diverge

Written by: Rita

LLM token expenditure increased by 40% month-over-month in October, compared to 28% in September and 7% in August. JPMorgan believes demand for AI infrastructure remains strong, but price signals for GPU rentals and spot storage have diverged.

JPMorgan's data center insights report, released on October 26, 2026, notes that the acceleration in token spending is driven by increased usage and a rebound in weighted prices. The weighted price rose 12% quarter-over-quarter, primarily due to higher usage share of the premium model Claude Opus 5.5 and price increases for the same model. In GPU leasing, A100 prices continued to decline, H100 prices rebounded, and B200 price increases widened. In storage, DRAM spot prices rebounded, while NAND prices fell for the second consecutive month. Contract price increases are expected to narrow in the fourth quarter.

OpenRouter is an aggregation platform that routes API requests through a single endpoint to approximately 300 large language models. The sample is biased toward developers, startups, and agent coding traffic, and excludes first-party API volumes from AI model companies such as OpenAI and Anthropic, as well as endpoints hosted by hyperscale providers. This limitation should be considered when interpreting token data.

On the OpenRouter platform, token usage increased by 25% month-over-month and by 33 times year-over-year. The average price decreased by 5% month-over-month and increased by 9% year-over-year. The weighted average price increased by 12% month-over-month and decreased by 38% year-over-year. Total spending increased by 40% month-over-month and by 20 times year-over-year.

The weighted price rebound is due to two factors. The high-end model Claude Opus 5.5 added approximately 7.9 trillion token usages, raising the overall average price. Additionally, price increases for the same models are contributing: the 50% discount for GPT-5.6 Sol has expired, doubling its price, while Kimi K3 and DeepSeek v4 Pro have raised their output prices.

Closed-source models accounted for 77% of spending, up 49% month-over-month and 16-fold year-over-year. Open-source models accounted for 23% of spending, up 15% month-over-month and 121-fold year-over-year. JPMorgan noted that closed-source models remain the primary driver of spending, while the growth rate of open-source models has slowed from its peak. The top five models by usage collectively accounted for 47% of total tokens, down from 53% in September. The top five models by spending collectively accounted for 54%, up from 50% in September. A divergence has emerged between usage and spending concentration, with higher-priced models capturing a larger share of spending.

The average rental price for A100 was $1.53 per GPU hour in October, a 3.5% decrease month-over-month. In September, it decreased 3.0% month-over-month, and in August, it decreased 0.7%. The average price for H100 was $2.83 per GPU hour, up 6.5% month-over-month. In September, it decreased 2.0% month-over-month, and in August, it increased 0.4%. The average price for B200 was $6.00 per GPU hour, up 4.7% month-over-month. In September, it increased 1.8% month-over-month, and in August, it decreased 1.5%.

The H100-to-A100 price ratio rose to 1.85x, up from 1.67x in September and 1.65x in August. The B200-to-H100 price ratio declined to 2.12x, down from 2.16x in September and 2.08x in August. JPMorgan noted that the H100 price rebound reflects renewed demand for mid-tier computing power, as the A100 gradually loses its dominant position. While B200 prices continue to rise, the premium over H100 has narrowed, indicating improving supply of high-end computing power.

GPU rental prices among non-hyperscale providers show more pronounced divergence. A100 supply is ample, putting downward pressure on prices. H100 demand remains steady, leading to a price rebound. B200 supply is gradually increasing, with price gains expanding but premium narrowing. J.P. Morgan believes that the divergence in GPU rental prices reflects an adjustment in compute infrastructure: A100 is gradually stepping down from its primary role, while H100 and B200 are absorbing incremental demand. This price divergence also signals a shift in compute investment structure, as supply and demand dynamics for mid-tier and high-end compute are not aligned.

DRAM spot prices rebounded in October. The average price for DDR5 16Gb was $51.80, up 4.4% month-over-month from $49.60 in September and $50.20 in August. Year-over-year, prices rose 227%, compared to $15.85 in the same period last year. NAND flash spot prices declined for the second consecutive month. The average price for 1Tb NAND was $29.40, down 2.6% month-over-month from $30.20 in September and $30.50 in August. Year-over-year, prices increased 298%, compared to $7.38 in the same period last year.

JPMorgan noted that the rebound in DRAM spot prices is supported by demand from AI servers, with high-bandwidth memory (HBM) consuming capacity traditionally allocated to DRAM, leading to tight supply. NAND spot prices have declined due to weak consumer demand and sufficient buyer inventories; enterprise SSDs are the only segment expected to see accelerated price increases in the fourth quarter. The divergence in storage spot prices reflects the differing dynamics between AI server demand and consumer electronics demand. DRAM demand driven by AI is more concentrated, while NAND has stronger consumer-oriented characteristics, resulting in unsynchronized pricing cycles.

In Q3, DRAM contract prices rose 16% quarter-over-quarter, compared to 56% in Q2. NAND contract prices increased 21% quarter-over-quarter, following a 58% rise in Q2. For Q4, DRAM prices are expected to rise 13% and NAND prices 18%. The narrowing gains are attributed to prior price increases, pressure on consumer demand, sufficient SSD inventories among customers, and long-term pricing agreements with some server DRAM suppliers.

JPMorgan noted that demand for AI servers and production capacity continuing to shift toward server products are sustaining pricing. Enterprise SSDs are the only category expected to see accelerated price increases in the fourth quarter. A narrowing of contract price increases does not indicate price declines, but rather a slowdown in the rate of growth. Storage prices remain high, with AI-related demand serving as the primary support. The divergence between contract and spot price trends reflects differing levels of supply-demand tightness between long-term agreements and the spot market, with AI server customers relying more heavily on long-term contracts to lock in prices.

JPMorgan believes that increased token spending signals strong demand for AI infrastructure. The price divergence between GPU rentals and spot storage reflects differences in computing power structure and demand profiles. Although the rate of increase in contract prices has slowed, prices continue to rise. Key areas for continued monitoring include token usage growth, GPU rental price trends, and negotiations around storage contract prices.

Disclaimer

This article is a compilation and interpretation by Chaoxiang Research of a third-party brokerage research report (JPMorgan, October 26, 2026), combined with publicly available market information. The ratings, price targets, earnings forecasts, and related judgments cited herein reflect the views of the brokerage’s analysts and represent the position of their respective institution, not the views of Chaoxiang Research, nor do they constitute any investment advice.

The market carries risks; make decisions independently. This article should not be used as a basis for buying or selling any securities.