JPMorgan: October AI Computing Chain Data Diverges, Token Spending Up 40% MoM
nashnova research
J.P. Morgan's October data-center tracker shows LLM token spending surged 40% month-on-month, but GPU rental prices split across three chip generations and memory contract price gains are slowing — money is pouring into AI faster, yet supply-side pricing signals no longer agree.
Why did token spending suddenly accelerate?
Total LLM spending on OpenRouter jumped 40% m/m in October, reaching 20× year-ago levels — up from 28% growth in September.
This means → it is not just more usage; usage is also getting more expensive. Volume-weighted token prices rose 12% m/m.
The price increase came from two directions: users shifted toward pricier models — especially Claude Opus 5.5, which contributed roughly 7.9 trillion new tokens at a premium — and models themselves repriced, with GPT-5.6 Sol doubling after its discount expired, plus output-price hikes at Kimi K3 and DeepSeek v4 Pro.
Closed-source vs. open-weight — where is the money going?
Closed-source models (OpenAI, Anthropic, etc.) captured 77% of total spending, up 49% m/m and 16× year-ago levels.
Open-weight models (DeepSeek, Moonshot AI, etc.) slowed sharply, growing just 15% m/m.
In plain terms = by usage, the top five are all cheap open-weight models, accounting for over 47% of total tokens. By revenue, the top five are premium closed-source models — Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra — capturing 54% of total spend. Being widely used and being highly profitable are not the same thing.
Three GPU generations — why are their rental prices diverging?
A100 averaged $1.53/GPU-hour, down 3.5% m/m, extending September's decline — the oldest generation is being priced out.
H100 averaged $2.83/GPU-hour, rebounding 6.5% m/m after a 2.0% drop in September. This means → H100 remains the workhorse, and demand still supports its price.
B200 averaged $6.00/GPU-hour, up 4.7% m/m, accelerating from September's 1.8% gain. This reflects a widening supply shortage for the latest-generation compute chips.
What do the price ratios between GPU generations tell us?
The H100-to-A100 price ratio rose to 1.85× (from 1.67× in September). In plain terms = the market is widening the premium it pays for "current-gen" over "last-gen."
The B200-to-H100 ratio edged down to 2.12× (from 2.16×). This means → the newest generation is still getting more expensive, but not as fast as H100 is rebounding — the "generational premium" between the two narrowed slightly.
Memory spot and contract prices — do the signals agree?
DDR5 16Gb DRAM spot hit $51.80, up 4.4% m/m, recovering from a slight September dip. Year-on-year it is still up 227%.
NAND 1Tb spot was estimated at $29.40, down 2.6% m/m — a second straight monthly decline — though still up 298% y/y.
On the contract side, TrendForce data show Q3 DRAM contracts rose 16% q/q and NAND 21% q/q, but Q4 gains are expected to slow to 13% and 18% respectively. This reflects prior large run-ups, consumer demand pressure, ample client-SSD buyer inventory, and some server-DRAM suppliers locking in long-term pricing agreements.
Across the whole compute chain — what matters most right now?
Token spending is accelerating while memory contract gains are systematically decelerating — a visible crack has opened between the two.
This means → the demand side is burning cash faster, but supply-side pricing power is loosening. Whether the memory chain can arrest its deceleration, supported by AI-server demand, is the single most important validation point next quarter.
The lone bright spot: enterprise SSDs are the only memory category where Q4 contract gains are expected to accelerate — this reflects AI servers still pulling hard on high-performance storage in that specific segment.
市场有风险，内容仅供研究参考，不构成投资建议。