Average prices for artificial intelligence tokens have fallen 67% since May 2026, but total spending on tokens has climbed more than 120%, according to data published by venture capital firm Andreessen Horowitz, widely known as a16z. The figures appeared in the firm's Charts of the Week, a recurring data publication, released on October 31, 2026, and show demand for AI inference expanding quickly enough to overwhelm the steepest price declines. Estimates of the increase in total token spending range from 122% to 160% over the roughly five-month period, underscoring how falling per-unit costs have coincided with an overall expansion in AI usage. One note for readers accustomed to "token" meaning a crypto asset: in this data, the word refers to units of AI computation, not digital assets.

What the Numbers Show

In this context, a token is the billing unit for AI model inference: a small chunk of text that a model reads or writes, typically a few characters to a full word. Inference is the act of running a trained AI model to produce an output, and providers charge for it by the token.

The divergence in the data is stark. Per-token prices dropped 67% since May, while total spending over the same stretch rose somewhere between 122% and 160%. The unit got cheaper, but the receipt got longer. Usage grew fast enough to swamp the price cut entirely.

The biggest driver appears to be AI agents. According to the a16z data, agents consume nearly five times as many tokens as human users. A person typically asks a chatbot a question and reads the answer. An agent runs a longer task from start to finish, carries more context between steps, calls tools and models repeatedly, and takes more steps overall — and every step burns tokens.

Agent token usage has also grown roughly 14 times since February 2026, which points to a shift toward larger, more automated workflows rather than one-off prompts.

The Jevons Paradox, Now With GPUs

a16z ties the trend to the Jevons paradox: the observation, attributed to 19th-century English economist William Stanley Jevons, that when a resource becomes more efficient or cheaper to use, total consumption can rise instead of fall. In the 1860s, Jevons noted that more efficient coal use in Britain coincided with rising, not falling, coal consumption. The concept has enjoyed a second life in modern technology debates, where it is routinely invoked whenever efficiency gains appear to fuel, rather than curb, demand. In the same way, cheaper inference makes new use cases economical, and tasks that were too expensive to automate at May prices suddenly pencil out at today's rates.

There is a twist on the supply side. While token prices kept dropping, GPU rental prices moved sharply in the opposite direction. Rental costs for Nvidia's B300 chips rose 106% since mid-July 2026, according to the data, while B200 rental prices climbed 66% over the same period. The B300 and B200 are high-end Nvidia data center GPUs used to run AI workloads at scale.

That is an awkward combination for anyone selling inference. The output they sell is getting cheaper while the hardware they rent to produce it is getting pricier. The gap between the two curves has been labeled a "rack spread" — a nod to the trader's term for the distance between two linked prices, in this case between what a token sells for and what the compute behind it rents for.

The a16z charts also point to concentration among buyers. A relatively small group of top users accounts for a significant share of total token spending. In practical terms, that means the headline totals can be swayed by the habits of a handful of accounts rather than by broad-based demand across the entire user base.

What It Means

For AI model providers, the data is a mixed blessing. Volume is surging, which is the good part, but revenue per token is collapsing and the cost of the GPUs underneath is rising. The rack spread is that squeeze expressed in numbers.

For GPU owners and cloud operators, the picture looks friendlier. The 106% jump in B300 rental rates since mid-July suggests demand for compute is outrunning supply, at least for the newest chips.

For companies building on AI, each individual call is cheaper than it was in May. Yet agent-heavy products can rack up bills quickly, since agents use nearly five times the tokens of a human user.

The concentration point deserves attention as well. When a small group of heavy users drives most of the spending, the market becomes sensitive to their decisions.

For now, a16z's charts tell a straightforward story with a counterintuitive ending. AI inference got two-thirds cheaper in roughly five months, and the world's response was to use so much more of it that total spending more than doubled. Subsequent updates to the firm's data will show how the three moving parts — per-token prices, GPU rental rates, and the mix of spenders behind the totals — evolve from there.

Source: CryptoBriefing. The original report was edited by Diego Almada Lopez.