JPMorgan Research Note Analysis: October Token Spend Rises 40% MoM; GPU Rental and Storage Prices Diverge

Written by: Rita

Large language model (LLM) token spend rose 40% MoM in October, versus 28% in September and 7% in August. JPMorgan believes AI infrastructure demand remains strong, but price signals for GPU rentals and storage spot prices have diverged.

In its Data Center Observations report published on October 26, 2026, JPMorgan noted that accelerating token spend was driven by both volume growth and a rebound in weighted prices. Weighted prices rose 12% MoM, mainly due to a higher volume share of the premium-priced model Claude Opus 5.5 and a price increase for the same model. On GPU rentals, A100 prices continued to decline, H100 prices rebounded, and B200 gains expanded. On storage, DRAM spot prices rebounded, while NAND fell for a second consecutive month. Contract price increases are expected to moderate in Q4.

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

Token Spend Growth Accelerates

On OpenRouter, October token usage rose 25% MoM and was up 33x YoY. Average price fell 5% MoM and rose 9% YoY. Weighted average price rose 12% MoM and fell 38% YoY. Total spend rose 40% MoM and was up 20x YoY.

The rebound in weighted prices had two causes. The premium-priced model Claude Opus 5.5 added about 7.9 trillion tokens of usage, lifting the overall average price. Price increases for models also contributed: after GPT-5.6 Sol's 50% discount expired, its price doubled, while Kimi K3 and DeepSeek v4 Pro raised output prices.

Closed-source models accounted for 77% of spend, rising 49% MoM and up 16x YoY. Open-source models accounted for 23% of spend, rising 15% MoM and up 121x YoY. JPMorgan noted that closed-source models remain the main driver of spend, while open-source growth has eased from a high base. The top five models by usage accounted for 47% of total tokens, down from 53% in September. The top five models by spend accounted for 54%, up from 50% in September. Concentration in usage and spend diverged, with premium-priced models capturing a larger share of spend.

GPU Rental Price Trends Diverge

The average A100 rental price in October was 1.53 USD/GPU-hour, down 3.5% MoM. In September it fell 3.0% MoM and in August it fell 0.7%. The average H100 price was 2.83 USD/GPU-hour, up 6.5% MoM. In September it fell 2.0% MoM and in August it rose 0.4%. The average B200 price was 6.00 USD/GPU-hour, up 4.7% MoM. In September it rose 1.8% MoM and in August it fell 1.5%.

The price ratio of H100 to A100 rose to 1.85x, from 1.67x in September and 1.65x in August. The price ratio of B200 to H100 fell to 2.12x, from 2.16x in September and 2.08x in August. JPMorgan noted that the H100 price rebound reflects recovering demand for mid-tier compute, while the A100 is gradually losing its primary position. B200 price gains have expanded, but its premium relative to H100 has narrowed, indicating improving supply of high-end compute.

GPU rental price divergence is more pronounced among non-hyperscaler providers. A100 supply remains ample, putting pressure on prices. H100 demand is stable, supporting a price rebound. B200 supply is gradually increasing, so price gains have expanded while the premium has narrowed. JPMorgan believes the divergence in GPU rental prices reflects an adjustment in the compute mix: the A100 is gradually losing its primary position, while H100 and B200 absorb incremental demand. The divergence in rental prices also suggests that the structure of compute investment is changing, with supply-demand conditions for mid-tier and high-end compute not moving in sync.

Storage Spot Prices Moved in Different Directions

Dynamic random access memory (DRAM) spot prices rebounded in October. The average price of DDR5 16Gb was $51.80, up 4.4% MoM, compared with $49.60 in September and $50.20 in August. It rose 227% YoY from $15.85 in the same period last year. NAND flash spot prices fell for a second consecutive month. The average price of NAND 1Tb was $29.40, down 2.6% MoM, compared with $30.20 in September and $30.50 in August. It rose 298% YoY from $7.38 in the same period last year.

J.P. Morgan noted that the DRAM spot rebound was supported by AI server demand, while high-bandwidth memory (HBM) crowded out traditional DRAM capacity, leaving supply tight. The decline in NAND spot prices reflected weak consumer demand and ample buyer inventories; enterprise solid-state drives (SSDs) were the only category expected to see accelerating price gains in Q4. The divergence in storage spot prices reflected the difference between the AI server and consumer electronics demand tracks. DRAM demand from AI was more concentrated, while NAND was more consumer-oriented, so their price cycles were not synchronized.

Narrowing contract price gains

DRAM contract prices rose 16% QoQ in Q3, compared with 56% in Q2. NAND contract prices rose 21% QoQ in Q3, compared with 58% in Q2. For Q4, DRAM contract prices are expected to rise 13% and NAND contract prices 18%. The narrowing gains were attributed to prior price increases, consumer demand pressure, ample customer SSD inventories, and long-term pricing agreements from some server DRAM suppliers.

J.P. Morgan noted that AI server demand and capacity tilting toward server products continued to support pricing. Enterprise SSDs were the only category expected to see accelerating price gains in Q4. Narrowing contract price gains did not indicate falling prices, only a slower pace of increases. Storage prices remained elevated, with AI-related demand as the main support. The divergence between contract and spot price trends indicated different degrees of supply-demand tightness between long-term contract orders and the spot market, with AI server customers relying more on long-term contracts to lock in prices.

J.P. Morgan believed that accelerating token spending confirmed robust demand for AI infrastructure. The price divergence between GPU rentals and storage spot prices reflected differences in compute mix and demand tracks. Contract price gains narrowed, but prices were still rising. Token usage growth, GPU rental price trends, and storage contract price negotiations were key points to watch.

Disclaimer

This article is a compilation and interpretation by Chaoxiang Research of a third-party brokerage research report (J.P. Morgan, October 26, 2026), combined with publicly available market information. The ratings, target prices, earnings forecasts, and related judgments cited in this article are the views of analysts at that brokerage and represent only the positions of their affiliated institutions. They do not represent the views of Chaoxiang Research and do not constitute investment advice.

Markets involve risk, and decisions should be made independently. This article should not be used as a basis for buying or selling any securities.

Source: TechFlow