J.P. Morgan: LLM Tokens Jump 70% in June; GPU Rentals Extend Seven-Month Climb; DRAM Spot Prices Surge

J.P. Morgan's July 1 data-center note points to sustained strength in AI infrastructure demand, based on three indicators: LLM token usage and pricing (OpenRouter), GPU rental rates from non-hyperscale cloud providers, and spot pricing for memory chips. OpenRouter activity accelerated sharply in June. Token usage rose 70% month over month and was up 20x year over year. Pricing moved far less: cost per invocation edged up versus May and is down only 5% year over year, while pricing for U.S.-based models increased year over year. With usage rising much faster than prices falling, total spend on the platform climbed 70% month over month and 16x year over year. J.P. Morgan argues that even if token prices keep declining at a 5% to 10% annual pace, inference revenue can still expand materially as long as usage continues to scale. The report also highlights a concentration of monetization in U.S. models. They account for about 35% of usage but capture more than 85% of total spending. Anthropic's Claude Opus 4.7 is the only model appearing in both the top five by usage and the top five by spending, which the note interprets as evidence of pricing power among leading U.S. providers. On compute supply, GPU rental prices continued to rise for a seventh straight month. In June, A100 rental rates increased 6.3% month over month, marking five consecutive months of gains; H100 rose 3.7%, extending a seven-month uptrend; and B200 climbed 2.7%, continuing a steady increase since its launch nine months ago. B200 rental rates remain around double H100 levels, but the premium has narrowed: the H100-to-A100 premium eased from 1.77x in April to 1.67x in June, and the B200-to-H100 premium fell from 2.58x to 1.96x. J.P. Morgan reads the combination of easing scarcity and rising prices as a sign that demand growth is keeping pace with expanding supply. The strength in A100 pricing is also notable, suggesting budget-sensitive buyers are competing for more cost-effective capacity rather than signaling oversupply. Memory markets show a split. DRAM spot prices rose 10% month over month in June to $43.14, up 740% year over year, for a third consecutive monthly increase. NAND spot prices dipped 0.3% month over month to $27.03, extending a three-month soft patch, though they remain up 518% year over year. The divergence is consistent with historical late-cycle patterns in which NAND weakens before DRAM. While J.P. Morgan stops short of calling a cycle peak, the data suggests the storage cycle may be entering a later phase. Overall, the note frames the picture as bullish but evolving: LLM usage up 70% month over month, GPU rentals rising for seven consecutive months, and DRAM prices up 740% year over year all point to AI infrastructure demand that remains far from peak. At the same time, spending is concentrated in U.S. models, high-end GPU scarcity appears to be easing as premiums compress, and NAND softening hints at shifting dynamics in memory. Tide View A key caveat is potential data-source bias. OpenRouter primarily reflects developer, startup, and agent-building activity and does not include direct API traffic from OpenAI and Anthropic or usage tied to hyperscalers' internal deployments. As a result, the dataset may capture "long-tail" demand rather than total AI demand. A surge in the long tail could mean top-tier demand is even stronger, but it could also mean leading demand has already cooled while the long tail is catching up—two scenarios with very different investment implications. GPU rental pricing in the report also excludes official rates from AWS, Azure, and Google Cloud, focusing instead on third-party capacity. Rising third-party rates may indicate that compute tightness is spilling into the long tail, but it does not reveal utilization within hyperscalers. If hyperscalers are already saturated internally and are not releasing surplus capacity to rental markets, underlying supply-demand conditions could be weaker than these indicators imply. The storage split also warrants separate attention. Three months of NAND declines alongside continued DRAM strength fits a late-cycle pattern. If NAND weakness persists, revenue expectations for memory manufacturers could be revised lower, potentially weighing on capital-expenditure outlooks across the semiconductor equipment chain. Disclaimer This article is compiled and interpreted by Chaoxiang Research based on third-party brokerage research. Any ratings, price targets, earnings forecasts, and related judgments cited are the views of the brokerage analysts and represent their institutions' positions; they do not reflect Chaoxiang Research's views and do not constitute investment advice. Markets involve risks; make decisions independently. This article should not be used as the basis for buying or selling any securities.