JPMorgan Research Report Analysis: October Token Spending Up 40% Month-over-Month, GPU Rental and Storage Prices Diverge

Written by: Rita

Large language model (LLM) Token spending increased 40% month-on-month in October, up from 28% in September and 7% in August. JPMorgan considers that AI infrastructure demand remains strong, but price signals for graphics processing unit (GPU) rentals and storage spot markets show divergence.

In its data center observation report released on October 26, 2026, JPMorgan pointed out that the acceleration in Token spending was driven by both volume growth and a rebound in weighted prices. The weighted price increased 12% month-on-month, mainly due to an increased share of usage from high-priced models like Claude Opus 5.5 and price hikes within those same models. Regarding GPU rentals, A100 prices continued to fall, H100 rebounded, and B200 saw widening gains. In storage, DRAM spot prices rebounded, while NAND fell for a second consecutive month. Contract price increases are expected to narrow in the fourth quarter.

OpenRouter is an aggregation platform that routes API requests to approximately 300 large language models through a single endpoint. The sample skews toward developer, startup, and agent coding traffic, excluding first-party API volume from AI model companies such as OpenAI and Anthropic, as well as endpoints hosted by hyperscalers. This limitation should be taken into account when interpreting Token data.

Accelerating Growth in Token Spending

On the OpenRouter platform, Token volume increased 25% month-on-month and surged 33 times year-on-year in October. The average price decreased 5% month-on-month but rose 9% year-on-year. The weighted average price increased 12% month-on-month, down 38% year-on-year. Total spending grew 40% month-on-month and expanded 20-fold year-on-year.

There are two reasons for the rebound in weighted prices. High-priced model Claude Opus 5.5 added approximately 7.9 trillion Token usage, pulling up the overall average price. Price increases within the same models also played a role; after the expiration of its 50% discount, GPT-5.6 Sol saw its price double, while Kimi K3 and DeepSeek v4 Pro increased their output pricing.

Closed-source models accounted for 77% of spending, up 49% month-on-month and 16 times year-on-year. Open-source models accounted for 23%, up 15% month-on-month and 121 times year-on-year. JPMorgan noted that closed-source models remain the main driver of spending, while open-source model growth rates have cooled from their peak. The top five models by volume accounted for 47% of total Tokens, down from 53% in September. The top five models by spending accounted for 54%, up from 50% in September. Concentration in volume and spending diverged, with high-priced models capturing a larger share of spending.

Diverging Price Trends in GPU Rentals

The average price for A100 rentals was $1.53/GPU hour in October, a 3.5% monthly decline (down 3.0% in September and 0.7% in August). H100 averaged $2.83/GPU hour, up 6.5% month-on-month (down 2.0% in September and up 0.4% in August). B200 averaged $6.00/GPU hour, up 4.7% month-on-month (up 1.8% in September and down 1.5% in August).

The price ratio of H100 to A100 rose to 1.85x (1.67x in September, 1.65x in August). The price ratio of B200 to H100 fell to 2.12x (2.16x in September, 2.08x in August). JPMorgan noted that the H100 price rebound reflects recovering demand for mid-range compute power, while the A100 is gradually stepping back from its dominant position. Although B200 price increases widened, its premium relative to the H100 narrowed, indicating improving supply of high-end compute.

Price divergence in GPU rentals is more pronounced among non-hyperscaler vendors. A100 supply is ample, keeping prices under pressure. H100 demand remains stable, driving a price rebound. B200 supply is gradually increasing, with wider price gains but a narrowing premium. JPMorgan believes that the divergence in GPU rental prices reflects adjustments in compute architecture, as the A100 recedes from its core role and H100s and B200s absorb incremental demand. This rental price divergence also signals shifting structures in compute investments, with supply and demand dynamics for mid-range and high-end compute operating out of sync.

Divergent Spot Storage Prices

Dynamic random-access memory (DRAM) spot prices rebounded in October. DDR5 16Gb averaged $51.80, up 4.4% month-on-month (September: $49.60, August: $50.20). Year-on-year, prices surged 227% compared to $15.85 last year. Flash (NAND) spot prices fell for a second consecutive month. NAND 1Tb averaged $29.40, down 2.6% month-on-month (September: $30.20, August: $30.50). Year-on-year, prices were up 298% against $7.38 last year.

JPMorgan pointed out that the DRAM spot price rebound is supported by AI server demand, with high-bandwidth memory (HBM) cannibalizing traditional DRAM capacity, leading to tight supply. NAND spot price declines were impacted by weak consumer demand and ample buyer inventory. Enterprise-grade solid-state drives (SSDs) are the only category expected to see accelerating price increases in Q4. The divergence in storage spot prices reflects differences between two demand lines: AI servers and consumer electronics. DRAM demand is more AI-centric, while NAND is more consumer-driven, meaning their price cycles are not synchronized.

Narrowing Increases in Contract Prices

DRAM contract prices rose 16% quarter-on-quarter in Q3 (vs. 56% in Q2). NAND contract prices rose 21% qoq (vs. 58% in Q2). For Q4, DRAM is projected to rise 13%, and NAND 18%. The narrowing price increases are attributed to prior price hikes, consumer demand headwinds, ample client SSD inventory, and long-term pricing agreements held by some server DRAM suppliers.

JPMorgan noted that AI server demand and production capacity skewing toward server products continue to support pricing. Enterprise SSDs remain the only category expected to see accelerating price increases in Q4. A narrowing increase in contract prices does not mean falling prices, only a deceleration in the pace of rises. Storage prices remain at elevated levels, with AI-related demand serving as the primary support. The divergent trends between contract and spot prices indicate differing supply-demand tightness between long-term agreements and the spot market, suggesting AI server clients rely more heavily on forward contracts to lock in prices.

JPMorgan believes the acceleration in Token spending confirms strong AI infrastructure demand. The price divergence in GPU rentals and storage spot markets reflects differences in compute architecture and demand lines. While contract price increases are narrowing, prices continue to rise. Future monitoring points include Token volume growth, GPU rental price trends, and negotiations over storage contract prices.

Disclaimer

This article compiles and interprets a third-party broker research report (JPMorgan, October 26, 2026) by Chaowang Research, combined with public market information compilation. The ratings, target prices, earnings forecasts, and related judgments cited in the article are solely the views of the respective broker's analysts, representing only the stance of their affiliated institutions, not those of Chaowang Research, and do not constitute any investment advice.

Markets carry risk, and decisions must be made independently. This article should not serve as the basis for buying or selling any securities.