AI Computing Demand Diverges: H100 Prices Drop, B200 Rebounds, DRAM Pulls Back After Five Consecutive Gains
nashnova research
JPMorgan's September data-center report shows token volume surging 71% month-on-month while average price plunged 25%, GPU rental prices splitting between old and new chips, and DRAM retreating for the first time after five straight months of gains — AI demand stays strong, but the money is flowing in very different directions.
Token volume jumped 70% — so why didn't revenue keep up?
September token volume on OpenRouter surged 71% month-on-month, with year-on-year growth hitting 30×. Total spending rose 28% MoM and 14× YoY.
Yet the volume-weighted average price (VWAP — the average price weighted by actual transaction volume) fell 25% MoM and 55% YoY.
This means → the explosion in usage was driven by cheap, high-throughput models, not premium pricing. Volume up, price down, but total spend still growing — a "thin margins, massive volume" dynamic is taking shape.
Open-source is crushing closed-source on volume — but who is actually making money?
Open-source models (DeepSeek, Moonshot AI, etc.) saw volume surge 83% MoM and 101× YoY. Closed-source models (OpenAI, Anthropic, etc.) managed just 39% MoM and 10× YoY.
Four of the top five models by volume are open-source. DeepSeek V4.1 Flash and GLM 5.3 Flash alone contributed roughly 41 trillion tokens in a single month — about 30% of total volume — at prices well below prior open-source averages.
But rank by spending and the leaderboard flips: GPT-6 Astra, Claude Fable 5.1, Claude Opus 5 and other high-priced closed-source flagships still account for 50% of total spend. In plain terms = open-source runs the volume; closed-source collects the revenue — the usage chart and the revenue chart are almost two entirely different lists.
GPU rentals — why are new chips rising while old ones fall?
A100 rental average: $1.59 per GPU-hour, down 2.8% MoM. H100 average: $2.64, down 2.6% MoM. Both older models are sliding, and the declines are widening.
B200 moved the opposite way: September average rebounded to $5.70, up 1.3% MoM, reversing August's dip. The B200-to-H100 price ratio rose to 2.16×.
This means → the market still pays a premium for the latest generation of compute, but supply pressure on older GPUs is intensifying. This reflects a "generational split" — new chips in demand, old chips in surplus — rather than an across-the-board move.
DRAM paused after five months of gains — has AI memory demand peaked?
DDR5 16Gb spot price came in at $49.70 in September, down roughly 1% MoM — the first retreat after five consecutive monthly gains. Year-on-year it is still up 614% (from $6.96 a year ago).
NAND 1Tb spot held at $30.54, up just 0.1% MoM, essentially flat. YoY gain: roughly 470%. NAND had fallen for four straight months before turning positive last month; September held steady.
In plain terms = a 1% MoM dip does not mean the rally is over. The absolute level — still more than 6× higher year-on-year — shows that AI-driven memory demand has structurally lifted the price floor. Whether supply expansion and shifting demand patterns can re-establish equilibrium in the coming months is the key test for storage pricing direction.
市场有风险，内容仅供研究参考，不构成投资建议。