Illustrative image: servers in a data centre — the hardware that produces every AI token. Token spending is a data story, not a product, so no product photo exists for it. Image: Wikimedia Commons / BalticServers, CC BY-SA 3.0.
TL;DR — AI token spending surge explained in one line: venture firm a16z reports that average AI token prices have fallen 67% since May 2026, yet total token spending has risen 122% to 160% — because AI agents now burn through tokens far faster than humans do, and cheaper tokens simply unlocked more use.
📌 KEY HIGHLIGHTS
- The paradox: average AI token prices fell 67% since May 2026, but total token spending rose 122%–160% over the same period, per a16z's Charts of the Week (30–31 October 2026).
- The driver: AI agents consume nearly five times as many tokens as human users, and agent token usage has grown roughly 14 times since February 2026.
- The catch: the 122%–160% range is two estimates, not one audited figure — a16z itself calls the exercise napkin math. Treat the direction as solid and the exact size as approximate.
- The hardware twist: while token prices fell, Nvidia GPU rental prices climbed — B300 rentals up 106% since mid-July 2026 — squeezing the providers selling those cheap tokens.
- What it is not: this is about inference tokens, the billing unit for using AI models — not cryptocurrency "AI tokens", which are a different thing entirely.
AI token spending surge explained: what the numbers say
First, what a token is. A token is the small chunk of text — roughly three-quarters of a word — that an AI model reads or writes. Providers charge by the token: you pay for the input tokens you send and again for the output tokens the model returns, with output usually priced higher. Your AI bill is, at bottom, a token count times a token price.
That is what makes a16z's numbers so striking. In its Charts of the Week post at the end of October 2026, the venture firm reported that average token prices have dropped 67% since May 2026. If nothing else changed, bills should have fallen by two-thirds. Instead, total token spending rose somewhere between 122% and 160%. Usage grew so fast it swamped the price cut entirely.
The range deserves an honest explanation, because it is not a single measurement. Two different estimation methods give the two ends: Ramp's token-spend data points to about +122%, while multiplying a blended token price by OpenRouter's usage volume gives about +160%. The datasets behind both capture only part of total demand, and a16z describes the exercise as napkin math rather than a rigorous study. The durable takeaway is the direction — spending up sharply while prices fell by two-thirds — not the precise percentage.
| Figure (a16z Charts of the Week, Oct 2026) | Change | Period |
|---|---|---|
| Average AI token price | −67% | Since May 2026 |
| Total token spending (Ramp data) | +122% | Since May 2026 |
| Total token spending (blended price × OpenRouter volume) | +160% | Since May 2026 |
| Agent token usage vs a human user | ~5× more | Current |
| Agent token usage growth | ~14× | Since February 2026 |
| Nvidia B300 GPU rental price | +106% | Since mid-July 2026 |
| Nvidia B200 GPU rental price | +66% | Since mid-July 2026 |
Why agents burn so many tokens
A person asking a chatbot a question uses tokens twice: once for the question, once for the answer. An agent doing a job uses them dozens of times. It reads a long brief, plans steps, calls tools, reads the results, checks its own work, and often retries — and every step carries the full context along with it. That is why a16z puts agent consumption at nearly five times a human user's, and why agent usage has multiplied about 14 times since February 2026 as automated workflows replace one-off prompts.
Economists have a name for this pattern: the Jevons paradox. In 1865, William Stanley Jevons noticed that more fuel-efficient steam engines did not cut Britain's coal use — they raised it, because cheaper-to-run engines got installed in more factories, ships and railways. Tokens are tracing the same curve. When a task cost a dollar in API fees, marginal uses were not worth trying. When it costs a fraction of a cent, it becomes worth running speculatively, running several times to compare outputs, or embedding into a workflow that verifies its own answer before replying. Cheaper inference did not shrink the market; it expanded the set of things worth doing with AI.
Illustrative image: a graphics card of the kind rented out for AI workloads. Image: Wikimedia Commons / ITEagle Europe – Sebastiaan Broekhoven, CC BY-SA 4.0.
The rack spread: cheaper tokens, pricier GPUs
There is a twist on the supply side, and it is bad news for the companies selling cheap tokens. While token prices fell, the hardware underneath them got more expensive to rent. Nvidia's B300 rental prices rose 106% since mid-July 2026, B200 rentals rose 66%, and even the older H100 rose 17%. a16z's charts label the gap between the two curves the "rack spread": providers sell an output that keeps getting cheaper while renting the machines that produce it at prices that keep rising.
That squeeze helps explain the week's other big AI story — why investors punished Alphabet after its search advertising disappointed even as its AI spending soared. The token data and the earnings data point the same way: AI demand is real and growing, but the economics of supplying it are getting harder, not easier. A16z's charts add one more concentration warning: a relatively small group of heavy users accounts for a significant share of total token spending, so the market is sensitive to the decisions of a few big buyers.
Why it matters for India readers
No India-specific figure appears in the a16z data, and this post claims none. The relevance is structural. Indian startups and developers building on the same APIs face the same bill dynamics: falling list prices that do not translate into falling bills once agents enter the workflow. Any Indian team budgeting an AI feature as "tokens × today's price" will under-budget if its product succeeds, because success means more agentic usage — and agentic usage is exactly what multiplies the token count.
The hardware side matters for India too. GPU scarcity and rising rental prices feed into what Indian cloud providers and AI startups pay for compute, and into how quickly affordable inference reaches Indian users. Our DeepSeek super agent explainer shows where the agent workloads driving this consumption are heading, and our Mistral Large 4 open weights explainer covers the open-model route that lets teams run models themselves instead of paying per token. For the document-parsing side of the same shift, see our MiniCPM-OCR open-source document parsing explainer.
Should you care?
✅ Verdict: Yes, if you build with AI or pay an AI bill — budget for usage growth, not just price cuts. Token prices will probably keep falling, but your bill follows your token count, and agents push that count up fast. ⏳ Wait if: you only use chatbots casually — this changes nothing about your experience, and the 122%–160% figures are estimates, not a bill arriving at your door.
FAQs
Why is AI token spending rising when token prices are falling?
Because usage is growing faster than prices are falling. a16z reports token prices down 67% since May 2026 while spending rose 122%–160%, driven mainly by AI agents that use nearly five times the tokens of a human user. This is the Jevons paradox: making a resource cheaper expands the uses worth doing with it, so total consumption rises.
What is an AI token?
A token is the billing unit for AI model use — a small chunk of text, roughly three-quarters of a word, that a model reads or writes. Providers charge for input tokens you send and output tokens the model returns, with output usually priced higher. Your bill is your token count times the token price.
How many more tokens do AI agents use than humans?
Nearly five times as many, according to a16z's data, because an agent carries long context through many steps — planning, tool calls, result checks and retries — where a human asks one question and reads one answer. Agent token usage has grown roughly 14 times since February 2026.
Why are GPU rental prices rising while token prices fall?
Because demand for the hardware outruns supply. Nvidia B300 rentals rose 106% and B200 rentals 66% since mid-July 2026, even as token prices fell. The gap between falling token prices and rising GPU rents is called the "rack spread", and it squeezes AI providers whose revenue per token is collapsing while their hardware costs climb.
TAKEAWAY: AI got two-thirds cheaper per token in five months — and the world answered by using so much more of it that total spending more than doubled. Watch your token count, not just the price list.
Follow World Tech on TechByCountry for tomorrow's update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- CryptoBriefing — AI token spending jumps 122% to 160% as prices fall 67%, a16z data shows
- CryptosBank — AI Token Spending Surges Over 120% Despite Price Drops
- RuntimeWire — a16z says token spending rose more than 120% despite a 67% price drop
- Aera Technology — Tokenomics for Agentic AI (background: what a token is)
Comments
Post a Comment