The token is not the new dollar. It is the new kilowatt-hour
by Cătălin Popa · updated 2026-09-21
On 19 August 2026, Patrick Collison, the CEO of Stripe, wrote in a press release: “Tokens are the central currency for companies building with AI.” The line went round the business press. It is a good line, from a man who was at that moment buying a bureau de change for that “currency”. It deserves to be taken seriously, and it deserves to be checked.
Will AI tokens become the currency of the future?
Not in the sense in which the dollar is a currency. A token is a piece of text, about three quarters of a word, and it is the unit in which every major AI provider bills its work. As money it fails two of the three tests: it does not store value, because its price falls tens of times a year, and it is not a medium of exchange, because you cannot hand it on to anyone. It passes the third: it has become a unit of account. Invoices, budgets and pay packages are already drawn up in tokens, and futures contracts on the compute they are made from are scheduled for October 2026. The right comparison is the kilowatt-hour: a commodity that is metered, priced and traded, and that nobody keeps in a wallet. For a business, the practical consequence is one few people think about: every page you publish has a cost in tokens for the machine that reads it, and that cost can be measured.
What a token is, and why it became a unit of measure
A language model does not read words but pieces of words, called tokens. In English, a token is on average three quarters of a word. Romanian comes out somewhat worse, because diacritics and inflections break into more pieces.
Everything a model does is counted in tokens: how many it read, how many it wrote. OpenAI, Anthropic, Google and xAI all bill the same way, per million tokens. Everything follows from that. When a whole industry sells its product in the same unit, that unit starts to behave like a commodity: it has a market price, it has indices, it has volume.
Volume, to give a sense of scale: at the I/O conference on 19 May 2026, Sundar Pichai said Google had been processing roughly 480 trillion tokens a month a year earlier, and now processes more than 3.2 quadrillion. Seven times more, in one year.
What already exists: seven things, with dates
None of these is a forecast. All are announcements or products from 2026.
The factory. Jensen Huang, Nvidia's CEO, no longer talks about data centres but about “AI factories” that turn electricity into tokens. At GTC Taipei, on 1 June, he put it bluntly: if you have one gigawatt of power, then throughput per watt is your revenues, “because every token is profitable”. He also gave the price of such a factory: 50 to 60 billion dollars per gigawatt, and rising.
The exchange. CME Group, the world's largest derivatives exchange, announced on 11 August, together with Silicon Data, futures contracts on the rental price of Nvidia H100 and B200 cards, with the first trading day scheduled for 5 October 2026, subject to review by the US regulator (CFTC). ICE, its rival, announced something similar in May. Mind the nuance: these are contracts on GPU-hours, not on tokens. Contracts directly on tokens are, for now, being studied by the Shanghai Futures Exchange, with no date.
The bureau de change. Stripe is buying OpenRouter, an intermediary through which a developer reaches more than 400 models from more than 80 providers through a single door. The price is not in the press release; the New York Times reported 7.5 billion dollars. Why would a payments company pay that much? Because in 2026 the price of a token stopped being fixed. Some providers price by time of day, by subscription tier, by promotion. When the price moves, choosing a model becomes an exchange-rate problem, and whoever runs the counter wins.
The salary. At GTC, in March, Huang said every Nvidia engineer would receive an annual token budget worth about half their base salary. The investor Tomasz Tunguz did the sums on Levels.fyi data: for a top-quartile engineer on a 375,000-dollar salary, around 100,000 dollars in tokens comes on top. One dollar in five is already compute.
The bookkeeping. On 4 August, the Linux Foundation launched the Tokenomics Foundation, with 30 founding members, among them JPMorganChase, IBM, SAP and Oracle. The aim is dull, and important for exactly that reason: a common standard by which a company measures its cost per token and what it got in return. When banks sit down to standardise a cost line, that line is no longer an experiment.
Agents that pay on their own. Protocols already exist through which an AI agent pays without a human: x402 from Coinbase, AP2 from Google, the Agentic Commerce Protocol from OpenAI, and in June Mastercard launched “Agent Pay for Machines”. The numbers are still small. For x402 we found a single source, which puts the cumulative total at about 50 million dollars by April. The point to keep is a different one: agents pay in dollars and in stablecoins. Not in tokens.
Content that gets paid. Since 15 September 2026, Cloudflare splits automated traffic into three categories and, for domains newly onto the platform, blocks agents and training crawlers by default on pages that carry ads. In the same move, the programme that paid publishers for crawler access was replaced with one that pays them when the content is actually used in an answer. It launched with two partners, so it is a principle, not a market. Cloudflare also says that more than half of AI crawler traffic re-fetches pages that have not changed. Waste that somebody pays for.
The paradox: the price collapses, the bill grows
The price per token, at equal quality, is falling at an unprecedented rate. Andreessen Horowitz measured roughly 10 times a year. Epoch AI found between 9 and 900 times a year depending on the task, and 40 times a year for GPT-4-level performance on PhD-level science questions. With a caveat they write themselves: the fastest drops are recent, and it is not clear they will last.
You would expect bills to fall. McKinsey says enterprise spending on models tripled in twelve months and that 93% of companies overran their AI budget. Goldman Sachs estimates that token consumption will grow 24-fold by 2030, to around 120 quadrillion a month, and that nearly half of it will come from enterprise agents, meaning programs that work on their own, continuously, without anyone asking them anything.
It is the same mechanism as with electricity. When the light bulb became cheap, we did not pay less for light. We lit everything.
Why “currency” is the wrong word
Money does three things. Let us take them in turn.
Store of value
The token failsSomething whose price falls ten to forty times a year is the opposite of a reserve. A token budget left unspent this year buys, next year, a fraction of what it was worth. That is also exactly the criticism of the “salary in tokens”: it does not vest, it does not appreciate, and you cannot take it with you into the next negotiation. A company can hold salaries flat and pad the package with something it would have had to buy anyway.
Medium of exchange
The token failsYou cannot hand your tokens from one provider to someone else, and nobody accepts them in payment. Not even AI agents pay in tokens.
Unit of account
Passes, with an asteriskInvoices, budgets, performance indicators and price indices are drawn up in it. The asterisk: a token from one model does not equal a token from another, and the same task can consume up to 30 times more or less depending on how it is solved. It is a unit of account roughly as the “working hour” would be, if hours came in different lengths.
That leaves the kilowatt-hour. Electricity is metered, priced in tiers, traded on exchanges with futures contracts, and subsidised by governments for the vulnerable. Nobody saves in kilowatts. The token is going down the same road, only faster.
This is also where Sam Altman's idea belongs. On 30 April 2026 he said he no longer believes in universal basic income as much as he once did, and that he is more interested in some form of “collective ownership”, in compute or in equities. It is an idea from an interview, not a programme. We note it here so that it is not mistaken for a fact.
The part that concerns you: what it costs to be read
Everything above happens far away, between banks and chip factories. The part close to home is this: when an AI assistant or an agent opens your site, somebody pays for the reading in tokens. We wanted to see how much, so we measured our own pages.
Method: one download per page, on 21 September 2026, counted with OpenAI's public tokenizer (o200k_base). We counted the HTML as it comes from the server, then only the text a person sees.
| Page | Tokens, raw HTML | Tokens, visible text | How many times more |
|---|---|---|---|
| gr-ai.ro, home page | 58,648 | 8,629 | 6.8× |
| gr-ai.ro, the Jurilovca article | 50,167 | 7,090 | 7.1× |
| gr-ai.ro/llms.txt | 4,396 | 4,396 | 1.0× |
| bikestylish.ro, home page | 246,726 | 7,254 | 34× |
The last row is our own shop, and it does not flatter us. Its home page, taken raw, runs to almost 247,000 tokens for 7,000 tokens of content. That is more than fits in a 200,000-token context window, which is what many models have. The structured data alone weighs 60,000 tokens, eight times the text. At an average price of about one dollar per million tokens, the difference is between 24 cents and under one cent per read. It looks like nothing. Multiplied by every agent comparing prices, every day, it stops being nothing.
What these numbers mean, and what they do not
- The big engines clean the HTML before handing it to the model, so they do not always pay for the raw version. But somebody does that cleaning, by their own rules, and what they do not understand they throw away. Agents built with generic tools often receive the page truncated at a character count. If the first 50,000 characters of your page are menus and scripts, the agent never reached you.
- This is one tokenizer. Other models count slightly differently. The order of magnitude holds.
- These are four pages, all our own. It is a measurement, not a market study.
- llms.txt is in the table only as a reference point: the same content, said in 4,400 tokens. It is not a ranking factor and no major engine has confirmed using it that way. We are not changing our minds just because here it would suit us.
What you can do, in order of effort: keep the important content in the HTML the server sends, high up in the page, not assembled by JavaScript. As much structured data as needed, not the whole catalogue on the home page. And a conscious decision about who reads you for free. If you sit behind Cloudflare, you now have three categories of automated traffic that you can allow or block separately. It is worth knowing which boxes you ticked.
What we will be able to do in five years
We split this into three tiers, by how much each leans on what already exists. None of it is our forecast. The only number, the 24-fold one, belongs to Goldman Sachs.
Likely
Because it is only an extension of what we have described: companies will have a token budget per department, tracked like the energy bill. The large ones will hedge the price of compute with futures, as an airline does with kerosene. Agents will have a wallet with a cap and will buy small things on their own. Tokens will have day and night rates. Good content will be paid on use, not only on download.
Possible
Service contracts written in tokens. Compute as an ordinary employee benefit, next to the meal vouchers. States allocating compute to universities and small firms the way they allocate funds today. And the counter-trend, equally possible: outcome pricing swallows token pricing, and the end customer stops seeing the token at all, just as you do not see the kilowatts in the price of bread.
Speculative
Compute as a universal entitlement, in Altman's version, and tokens that people can sell to one another. For that, the token would have to become transferable and comparable across providers. Today it is neither.
For an ordinary business, two things on that list matter. First, AI will show up in the budget as a utility, with variable consumption, and it is better to meter it now than to discover it at year end. Second, a growing share of those reading your site will be programs with a budget. A patient person scrolls. An agent with a spending cap does not.
What we could not demonstrate
- That agents prefer pages that are cheap to read. It is a reasonable hypothesis, not a finding. We measured the cost, not the behaviour.
- The McKinsey figures. We took them from a press synthesis dated 16 September, not from the original report.
- The Goldman Sachs report. We could not open it at source. The figures appear identically in several outlets, but they all report the same document, so it counts as one source.
- The volume of payments made by agents. The x402 figure comes from a single source.
- A figure we left out. The claim that “over 40% of tech companies offer AI credits as a benefit”, attributed to TechCrunch, is widely repeated. We read the TechCrunch article. The figure is not in it.
Where the data comes from
Read directly on 21 September 2026, unless stated otherwise.
- Google, Sundar Pichai's I/O keynote, 19 May 2026
- Stripe, the OpenRouter press release, 19 August 2026
- paddo.dev, the “Stripe bought the exchange rate” reading
- CME Group, the 11 August 2026 press release (title and summary; the page did not load in full)
- Silicon Data, how GPU-hour contracts work
- TechCrunch, 28 May 2026, token futures and the Shanghai exchange
- TechCrunch, 21 March 2026, token budgets as a job perk
- TechCrunch, 1 July 2026, the Cloudflare policy
- SiliconANGLE, 1 June 2026, Jensen Huang's GTC Taipei keynote
- NVIDIA, 25 March 2026, performance per watt
- Epoch AI, 12 March 2025, the fall in inference prices
- Andreessen Horowitz, “Welcome to LLMflation”
- PYMNTS, coverage of the Goldman Sachs Research report
- Stacker, 16 September 2026, for the McKinsey figures and the Silicon Data index
- Linux Foundation, 4 August 2026, the Tokenomics Foundation launch
- Yahoo Finance, 30 April 2026, the Sam Altman interview
- Mastercard, June 2026, “Agent Pay for Machines”
- arXiv, “AI Token Futures Market”, a token contract design
Share this article
Further reading
- Does the Cloudflare change on 15 September affect me?The three categories of automated traffic, and the Googlebot trap
- How to check whether an AI can read my siteBefore what reading costs, the question is whether it is possible
- All the questions explainedIn the order it makes sense to read them
- If you want the measurement run on your own siteIt takes a few minutes and tells you how many tokens of packaging you carry per token of content