Written by: Rita

Trends Guide

On July 1, JPMorgan Chase released a data center observation report that tracked three core indicators of AI infrastructure demand: LLM token usage and prices, GPU rental prices, and spot prices for storage chips. The conclusion is clear: the AI demand environment remains positive.

LLM token usage on OpenRouter surged 70% month-over-month in June and increased 20 times year-over-year. GPU rental prices have risen for seven consecutive months, with B200 prices now double that of H100. Spot prices for DRAM have increased more than eightfold year-over-year. Triple data cross-verification shows that the demand for AI computing power is far from reaching its peak.

LLM usage rises by 70%, prices only drop by 5%, doing the math for model providers

Data from June on the OpenRouter platform accelerated overall. Token usage increased by 70% month-over-month and 20 times year-over-year. The price per invocation slightly rose month-over-month and only decreased by 5% year-over-year, while the pricing for domestic models in the U.S. even increased year-over-year. Total consumption amount increased 70% month-over-month and 16 times year-over-year.

Looking at these three sets of numbers together, the conclusion is obvious: the growth rate of usage far exceeds the decline in prices. Even assuming token prices continue to fall at a rate of 5% to 10% per year, as long as usage maintains a doubling pace, the inference revenue for model providers will still soar. JPMorgan Chase's judgment is that the direction of token economics is favorable for model providers.

Share data is even more interesting. U.S. models account for only 35% of usage but capture more than 85% of the consumption amount. Anthropic’s Claude Opus 4.7 is the only model to simultaneously enter both the top five in usage and the top five in consumption, with pricing power held in the U.S.

GPU rental prices have risen for seven months, B200 prices are double that of H100

For non-hyperscale cloud providers, GPU rental prices increased across the board in June. A100 rose by 6.3%, marking five consecutive months of increases; H100 rose by 3.7%, marking seven consecutive months of increases; and B200 rose by 2.7%, having climbed continuously since its launch nine months ago.

The price difference between the three GPU types is narrowing. The premium of H100 over A100 decreased from 1.77 times in April to 1.67 times in June, while the premium of B200 over H100 dropped from 2.58 times to 1.96 times. The scarcity of high-end computing power is easing, but prices are still rising, indicating that demand growth is keeping pace with supply release.

The rapid rise of A100 is worth highlighting. The fact that the price of the previous generation chip has not only failed to drop but has increased means that budget-constrained customers are scrambling for cost-effective computing power. This is not a signal of oversupply.

DRAM rises while NAND falls, the two storage giants are moving in opposite directions

In June, the spot price of DRAM rose by 10% month-over-month to $43.14, increasing 740% year-over-year and continuing to rise for three consecutive months. The spot price of NAND fell slightly by 0.3% month-over-month to $27.03, with a small decline over three consecutive months, but still increased by 518% year-over-year.

The figures of 740% and 518% indicate that the overall storage market remains at historical highs. The directional divergence is also very obvious: DRAM is still rising, while NAND is beginning to weaken.

The historical pattern indicates that when the storage cycle peaks, NAND will weaken first, followed by DRAM. The continuous softening of NAND over three months is a signal that the cycle is entering its later stages. JPMorgan Chase did not directly make this determination in the report, but the data itself speaks.

The narrative of AI infrastructure is still intact, but the logic is shifting gears

With LLM usage rising by 70%, GPU rental prices increasing for seven months, and DRAM rising by 740% year-over-year, these three sets of data point to the same conclusion: the demand for AI infrastructure is far from peaking.

However, the structure is changing. U.S. models are capturing 85% of the consumption amount, with pricing power in the U.S. The premium of B200 over H100 has decreased from 2.58 times to 1.96 times, indicating that the scarcity of high-end computing power is easing. NAND is beginning to soften, and the storage cycle is shifting.

Favorable conditions are ongoing, and the logic is dynamic.

Trends Perspective

The biggest blind spot in this report lies in the bias of data sources. OpenRouter primarily targets developers, startups, and AI programming scenarios, not including first-hand API traffic from OpenAI and Anthropic, nor the internal usage from hyperscale cloud providers. This means the report captures "long-tail AI demand," rather than "total AI demand." If the long tail is surging, the head will be even hotter, but it is also possible that the head has already cooled while the long tail is still catching up—these two investment implications are entirely different.

GPU rental prices track third-party capacity from non-hyperscale cloud providers and do not include the official rental prices from AWS, Azure, or GCP. If third-party prices are rising, it indicates that the computing power shortage has overflowed from the head to the long tail, which is favorable for AI hardware, but it cannot assess the utilization rate of hyperscale providers themselves. If their self-utilized computing power is already in surplus but has not been released to the rental market, the real supply and demand situation could be weaker than what the report presents.

The signal of price differentiation in storage is worth breaking down separately. With NAND softening for three consecutive months while DRAM continues to rise, this conforms to typical characteristics of the later stage of the storage cycle. If NAND prices continue to decline, the income expectations for storage chip manufacturers will be revised downwards, thereby affecting the overall capital expenditure expectations for semiconductor equipment.

Disclaimer

This article is a整理 of third-party broker research reports by Trends Guide. The ratings, target prices, profit forecasts, and related judgments cited herein are the views of the analysts of that broker and only represent their organization's position, not representing the views of Trends Guide, nor does it constitute any investment advice.

The market has risks, and decisions should be made independently. This article should not be used as the basis for buying or selling any securities.