The steep bill for AI tokens

A few days ago I came across an article with a title that sounded like a complaint from an angry neighbor: “Turn that algorithm off, you’re using too much!”. It talked about tokens, the electricity bills of data centers, and companies that ended up with eye-watering invoices after usage-based billing for artificial intelligence kicked in. I read it with a certain personal interest: I pay for an annual Claude subscription, I often use the more capable models, and every so often I take advantage of free credits when they’re available. The trouble is that those credits vanish in a flash, and there’s nothing more annoying than being left halfway through a task because the credit ran out.

What really happens when a model processes a token

A token isn’t a word. It’s a chunk of text, on average four characters in English, and it can also be a punctuation mark or a space. Language models break every request into these chunks, turn them into numbers, and try to guess which one comes next, until they build a complete answer. It’s an elegant mechanism on paper, but a computationally heavy one in practice. And it becomes even heavier when a system “reasons”: before answering, it splits the request into several internal steps, generates text that readers never see, and produces longer replies than usual. That’s why the most advanced reasoning models, the ones that promise more accurate results, end up consuming far more tokens than traditional ones.

Token prices are falling, but total spend isn’t

Between 2020 and 2026 the average price per token dropped roughly 600-fold. On paper that should be good news. In practice it isn’t, not entirely, because while the unit price shrinks, overall consumption multiplies: users grow, systems get more complex, workloads get heavier. OpenAI alone went from about 6 to more than 15 billion tokens processed per minute between autumn 2025 and spring 2026. It’s a bit like wider roads that still end up jammed: more available capacity, more demand ready to fill it.

When usage-based billing reveals the real bill

In the first half of 2026, OpenAI and Anthropic introduced billing models based on actual consumption, and many companies discovered a cost that had until then stayed hidden behind flat-rate plans. Uber burned through its entire annual AI budget in just four months. One company, which stayed anonymous, spent 500 million dollars in a single month. Meta, Amazon, and AT&T had to introduce internal usage caps to contain spending. According to a Deloitte report, in some organizations AI ends up covering as much as half of the technology budget. On the other side of that commercial relationship, Anthropic’s revenue rose from 4.8 to 10.9 billion dollars between the first and second quarter of 2026: client companies’ spending turns, almost as a mirror image, into suppliers’ revenue.

What a subscription like mine would really be worth

Here’s where things get personal. SemiAnalysis calculated what an intensive use of subscription plans would actually cost in terms of tokens consumed. A 20-dollar-a-month Claude plan, pushed to its full potential, would be worth around 400 dollars in real tokens; the 200-dollar plan would come close to 8,000. For ChatGPT the figures are even higher: 700 and 14,000 dollars respectively. Anyone paying a flat fee is therefore receiving an enormous subsidy from the company offering it. Reading these numbers, I understand better why free credits run out so quickly: behind that small counter emptying itself are figures that no single subscription could ever fully cover. It’s no surprise that, as I wrote in 5 struggles when using AI, one of the most concrete difficulties for anyone working with these tools every day is dealing with a limit that always shows up at the worst possible moment.

The energy it takes to run all these calculations

Behind every token there’s also an electricity bill. AI-dedicated data centers are expected to consume around 1,050 terawatt-hours by the end of 2026, more than triple Italy’s annual electricity demand, growing at 30% a year. Companies in the sector have planned 190 gigawatts of new capacity, with investments that, according to McKinsey, could reach 7 trillion dollars by 2030. In the United States, Virginia’s power grid, home to the world’s highest concentration of data centers, saw rates rise 35% in a year purely from AI demand pressure. In Italy, as I wrote in AI and its sustainability, the national recovery plan focused on fiber optics and new facilities but neglected the power grid: over the past three years, the connection cost per megawatt has risen 80%.

Who’s winning the race to use fewer tokens

In this scenario, Chinese models such as Xiaomi’s MiMo, MiniMax, and DeepSeek dominate the rankings for tokens processed as of June 2026. The advantage comes from several directions at once: cheaper energy, less expensive data centers to build, state support, aggressive domestic competition, and architectures designed from the start to use fewer resources per response. For anyone building applications with artificial intelligence too, as I wrote in SaaS, AI and Vibe coding, efficiency in managing tokens is becoming as much of a competitive edge as the quality of the generated answers.

A token bubble, or the new normal

OpenAI and Anthropic face an uncomfortable choice: lower prices to avoid losing customers, deepening already massive losses, or raise them and risk ceding ground to competitors. Meanwhile, the billions poured into infrastructure could prove excessive if real demand doesn’t keep pace, feeding fears of an AI bubble. As a user, my take stays simple: I’ll keep preferring a flat fee, even knowing I’m paying less than I should, because at least I know how much I’m spending each month. As for free credits, I’ve learned not to rely on them too much: better to treat them as an occasional bonus than as the foundation for work that can’t afford to stop halfway.