JPMorgan Research Report Analysis: LLM Usage Surges 70%, GPU Rental Rates Rise for Seven Consecutive Months, AI Hardware Demand Shows No Signs of Cooling
TechFlow Selected TechFlow Selected
JPMorgan Research Report Analysis: LLM Usage Surges 70%, GPU Rental Rates Rise for Seven Consecutive Months, AI Hardware Demand Shows No Signs of Cooling
The demand for AI computing power is far from peaking.
Navigating Web3 tides with focused insightsWritten by: Rita
TechFlow Guide
JPMorgan's Data Center Observation Report released on July 1 tracks three core indicators of AI infrastructure demand: LLM token usage and price, GPU rental price, and storage chip spot price. The conclusion is clear: the AI demand environment remains positive.
LLM token usage on OpenRouter surged 70% month-over-month in June, up 20 times year-over-year. GPU rent has risen for seven consecutive months, with B200 prices now double that of H100. DRAM spot prices increased more than 8 times year-over-year. Triple data cross-validation indicates AI compute demand is far from peaking.
LLM Usage Up 70%, Price Down Only 5%: Calculating the Numbers for Model Vendors
Data on the OpenRouter platform accelerated comprehensively in June. Token usage rose 70% month-over-month and 20 times year-over-year. Price per call increased slightly month-over-month and decreased only 5% year-over-year, while pricing for US domestic models even increased year-over-year. Total consumption amount rose 70% month-over-month and 16 times year-over-year.
Looking at these three sets of numbers together, the conclusion is obvious: usage growth rate overwhelms price decline. Even assuming token prices continue to fall at a rate of 5% to 10% annually, as long as usage maintains a doubling pace, model vendors' inference revenue will still surge. JPMorgan's judgment is that the direction of token economics favors model vendors.
Share data is even more interesting. US models account for only 35% of usage but capture over 85% of consumption amount. Anthropic's Claude Opus 4.7 is the only model to enter both the top five in usage and top five in consumption; pricing power lies in US hands.
GPU Rent Up for Seven Consecutive Months, B200 Price Double That of H100
GPU rental prices for non-hyperscale cloud providers rose across the board month-over-month in June. A100 up 6.3%, rising for five consecutive months; H100 up 3.7%, rising for seven consecutive months; B200 up 2.7%, climbing continuously since launch nine months ago.
The price gap between the three tiers of GPUs is narrowing. H100 premium over A100 dropped from 1.77 times in April to 1.67 times in June; B200 premium over H100 dropped from 2.58 times to 1.96 times. Scarcity of high-end compute is alleviating, but prices are still rising, indicating demand growth has kept up with supply release.
The fact that A100 rose the most sharply deserves separate mention. Previous-generation chip prices rising instead of falling means budget-constrained customers are snapping up cost-effective compute. This is not a signal of oversupply.
DRAM Up NAND Down, Storage Duo Moving in Opposite Directions
DRAM spot price rose 10% month-over-month to $43.14 in June, up 740% year-over-year, rising for three consecutive months. NAND spot price fell slightly 0.3% month-over-month to $27.03, falling slightly for three consecutive months, but still up 518% year-over-year.
740% and 518%, these two numbers indicate storage overall remains at historical highs. Directional divergence is also clear: DRAM is still rising, NAND is starting to loosen.
Historical pattern is that when the storage cycle peaks, NAND loosens first, DRAM loosens later. NAND weakening for three consecutive months is a signal the cycle is entering the later stage. JPMorgan did not directly make this judgment in the report, but the data itself speaks.
AI Infrastructure Narrative Continues, But Logic Is Shifting Gears
LLM usage up 70%, GPU rent up for seven consecutive months, DRAM up 740% year-over-year, triple data points to the same conclusion: demand for AI infrastructure is far from peaking.
But the structure is changing. US models captured 85% of consumption amount; pricing power lies in the US. B200 premium over H100 dropped from 2.58 times to 1.96 times; scarcity of high-end compute is alleviating. NAND is starting to loosen; storage cycle is shifting gears.
Benefits are continuous; logic is dynamic.
TechFlow Perspective
The biggest blind spot in this report lies in data source bias. OpenRouter mainly targets developers, startups, and AI agent programming scenarios; it does not include first-hand API traffic from OpenAI and Anthropic, nor does it include usage deployed internally by hyperscale cloud providers. This means the report captures "long-tail AI demand," not "total AI demand." If the long tail is surging, the head will only be hotter, but it is also possible the head has cooled while the long tail is still catching up—these two have completely different investment implications.
GPU rental prices track third-party capacity from non-hyperscale cloud providers, excluding official rental prices from AWS, Azure, GCP. Third parties are all rising, indicating compute shortage has spilled over from the head to the long tail; this is positive for AI hardware in direction, but unable to judge utilization rates of hyperscale providers themselves. If their self-used compute is already oversupplied, just not released to the rental market, then the true state of supply and demand would be weaker than presented in the report.
Signals of storage price divergence deserve separate unpacking. NAND weakening for three consecutive months, DRAM still rising, this fits typical characteristics of the later stage of the storage cycle. If NAND prices continue to decline, storage chip vendors' revenue expectations will be revised down, subsequently affecting capital expenditure expectations for the entire semiconductor equipment sector.
Disclaimer
This article is a compilation and interpretation by TechFlow Research of third-party broker research reports. Ratings, target prices, earnings forecasts, and related judgments cited in the text are the views of the broker's analysts, represent only the position of their affiliated institution, do not represent the views of TechFlow Research, and do not constitute any investment advice.
Market involves risks; decisions must be independent. This article should not be used as a basis for buying or selling any securities.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@chaoxiangooo
