AI Token Spending More Than Doubled as Inference Prices Fell, a16z Data Show

Artificial intelligence got dramatically cheaper to use in 2026, and the market responded by using vastly more of it. According to data from venture firm a16z, published in its Charts of the Week on 31 October 2026 and relayed on 1 November, average AI token prices have fallen 67 percent since May 2026, while total token spending has climbed between 122 percent and 160 percent over the same period. The figures describe a striking divergence: the unit price collapsed, and total spending more than doubled anyway.

The story is counterintuitive only until the underlying mechanics are understood. Cheaper inputs do not always mean smaller bills. When a service becomes affordable enough for entirely new uses, demand can expand faster than price falls. That is what appears to have happened with AI inference, the process by which a trained model generates answers, code, summaries or actions in response to a prompt. The a16z figures suggest that the AI economy is not shrinking as prices fall. It is growing, and growing quickly, because lower prices have unlocked heavier, more automated use.

Before going further, a vocabulary warning is essential, because the word token carries two very different meanings in this story, and confusing them changes the conclusion entirely. In the world of AI model providers, a token is the billing unit for inference: a small chunk of text that a model reads or writes, and providers charge by the token. In the world of crypto markets, an AI token is a tradeable digital asset linked to an AI project, whose price rises and falls on exchanges. The a16z data as presented by CryptoBriefing concern the first meaning: spending on AI inference, measured in tokens consumed. They are not, on that reading, a direct measure of the market prices of crypto AI tokens, even though some secondary coverage has framed the same figures in crypto-token terms. That ambiguity matters, and it is addressed directly below, because readers who mistake one kind of token for the other will draw the wrong investment conclusion.

What the numbers actually show

The headline figures are simple to state and harder to interpret. Per-token prices are down 67 percent since May. Total spending is up somewhere between 122 percent and 160 percent. The range itself is a clue: it suggests more than one way of counting, or more than one segment of the market being measured, rather than a single precise total. A reader should therefore treat the direction as firm and the exact magnitude as approximate.

The most important driver, according to the a16z data, is the rise of AI agents. Agents consume nearly five times as many tokens as human users, the data show. The reason is structural. A person typically asks a chatbot a question, reads a short answer, and stops. An agent, by contrast, runs a longer task: it reads background material, breaks the task into steps, tries approaches, checks results, and carries a growing body of context from step to step. Every step burns tokens. A workflow that a human might complete with a handful of exchanges can therefore generate hundreds or thousands of token-consuming operations when handed to an agent.

Agent token usage has also grown roughly 14 times since February 2026, according to the same data. That is not a marginal increase. It points to a shift in how AI is used, from occasional one-off prompts toward larger, automated workflows that run with less direct human supervision. Customer support systems that resolve cases end to end, coding assistants that work through a repository, research tools that gather and synthesise sources, and trading or operations agents that monitor conditions and act on them all fit this pattern. Each is a token-hungry use case that would have been prohibitively expensive at May prices and becomes plausible as prices fall.

This is the classic Jevons paradox, named for the nineteenth-century economist William Stanley Jevons, who observed that more efficient use of coal did not reduce coal consumption, because cheaper coal made new uses economical. Applied to AI, the logic is straightforward: cheaper inference makes tasks that were previously too expensive to automate suddenly worth automating. The result is not a smaller market but a larger one, with volume growth overwhelming the price decline.

The awkward supply side: GPUs are getting pricier

There is a twist that makes the story more than a simple tale of falling prices. While the price of AI output has fallen, the price of the hardware needed to produce it has risen sharply. Rental prices for Nvidia’s B300 chips rose 106 percent since mid-July 2026, according to the a16z data, while B200 rental prices climbed 66 percent over the same period. That combination is uncomfortable for anyone selling inference: the product they sell is getting cheaper while a key input they rent is getting more expensive.

Analysts have labelled the gap between those two curves a rack spread, a shorthand for the margin pressure on inference providers caught between falling output prices and rising hardware costs. The implications differ sharply depending on where a company sits in the stack. For GPU owners and cloud operators, rising rental rates suggest demand for compute is outrunning supply, at least for the newest chips, and that is a favourable position. For model providers and inference platforms, the same trend squeezes margins unless volume growth, efficiency gains or pricing changes compensate.

The caveat that must not be dropped

All of this rests on a relayed report, and honest coverage must say so plainly. The figures reach readers through a16z’s Charts of the Week as relayed by Cointelegraph and CryptoBriefing on 1 November, not through the full a16z report itself. The underlying methodology is therefore not visible in the coverage available, and the precise definition of AI token spending is not stated in that relay. It could mean total amounts paid for inference tokens, a broader measure of AI-related spending, or a blend of measures across providers. The 122 percent to 160 percent range hints at more than one counting method, but without the source report, that cannot be confirmed.

The ambiguity is visible in the secondary coverage itself. One relay frames the figures as spending on AI inference tokens, with the token defined as a billing unit for model use. Another relay frames the same percentage range as spending tied to AI crypto tokens, whose prices fell by about two thirds. Those are not the same claim, and they point to different markets. A reader who follows only the second framing might conclude that crypto AI tokens are seeing rising on-chain spending despite falling prices, which would be a claim about token economies and user behaviour on blockchains. A reader who follows the first framing would conclude something about the economics of AI compute: inference is cheaper, agents are hungrier, and total compute spending is rising.

Which reading is correct? The CryptoBriefing account, with its detailed explanation of inference billing, agent consumption and GPU rental prices, is internally consistent as a story about AI infrastructure economics. The crypto-token framing appears to be a reinterpretation of the same headline percentages through a crypto-market lens, without the supporting detail on agents, GPUs or inference pricing. That does not prove the crypto reading wrong in every respect, but it does mean the inference reading is the better-supported one on the evidence published so far. Until a16z’s full charts and methodology are public, the durable takeaway is the direction, usage up while unit price fell, rather than any precise percentage.

Why crypto readers should care anyway

Even on the inference reading, the story matters for crypto. The narratives around tokenised AI infrastructure, decentralised compute marketplaces and AI agents that transact on-chain all depend on the same underlying trend: AI work is becoming cheaper per unit and vastly larger in total volume. If agents are the fastest-growing consumers of AI, then systems that let agents pay for compute, data or services autonomously become more relevant. That is the conceptual bridge between inference economics and crypto AI projects, even though the a16z figures themselves are not a price forecast for any crypto token.

Frequently asked questions

What is an AI token in this story?

On the best-supported reading, it is the billing unit for AI model inference: a small chunk of text a model reads or writes, charged by providers per token. It is not necessarily a crypto asset, although some secondary coverage has framed the same figures in crypto-token terms.

Why did total spending rise if prices fell 67 percent?

Because usage grew even faster than prices fell. AI agents, which consume nearly five times as many tokens as human users, expanded rapidly, and cheaper inference made new automated workflows economical. This is an example of the Jevons paradox.

How much have AI agents grown?

According to the relayed a16z data, agent token usage grew roughly 14 times since February 2026, and agents now consume nearly five times as many tokens as human users.

Why are GPU rental prices rising while inference prices fall?

Demand for the newest chips appears to be outrunning supply. Nvidia B300 rental prices rose 106 percent since mid-July 2026 and B200 prices rose 66 percent, according to the data, creating margin pressure for inference providers.

Should these figures be treated as precise?

No. They come via a relayed post, not the full a16z report, and the exact definition of AI token spending is not stated in the relay. Treat the direction, usage up and price down, as the durable signal, and hold the exact percentages loosely until the full report is available.

Source: CryptoBriefing

Related reading on MM Crypto Analysis Web

• XRP Ledger Hits 14 Million AI Agent Transactions, But Value Stays Tiny
• Quack AI Plunges 19% as Major Token Unlock Looms on November 2
• Y Combinator Bets on Stablecoins, AI Fundraising and Tokenization for 2027

More in AI Crypto on MM Crypto Analysis Web.