---
title: "AI costs: why this isn't a bubble but a bill"
author: Marc Diks
date: 2026-06-27
modified: 2026-06-27
category: AI & Strategy
reading_time: 11 min
url: https://www.marcdiks.nl/en/blog/ai-costs-not-a-bubble-but-a-bill
canonical: https://www.marcdiks.nl/en/blog/ai-costs-not-a-bubble-but-a-bill
language: en
---
# AI costs: why this isn't a bubble but a bill
> **TL;DR**
>
> - **Not a bubble, but a correction.** Markets are demanding proof of returns — not evidence that AI has no value. That distinction is everything.
> - **Token bills are rising.** AI agents consume 3× more tokens on average than a year ago; the more complex the task, the less predictable the cost.
> - **The best models are being gated.** Washington is blocking Anthropic's most powerful models for Europe and limiting GPT-5.6 to twenty approved partners.
> - **Governance is shifting.** The risk is no longer about content but about availability: what if your model disappears tomorrow or triples in price?
> - **Vendor risk is now a board-level issue.** Spreading across models and keeping control of token costs is the core of your AI strategy.
Last month I looked at the stock prices of the big AI companies and saw red. Oracle down 5.5 percent, the Korean market down ten percent, the same word everywhere: bubble. Yet I don't believe a word of it.
Because a few days earlier, a friend of mine — an entrepreneur — ran out of tokens right before a presentation. It wasn't the stock prices that told me whether AI is here to stay. His panic did.
## The man who ran out of tokens
He uses an AI tool like Gamma to build presentations. Not as a toy, not to impress people on LinkedIn, but as a fixed part of his workflow. Slides, structure, first drafts: the machine does the groundwork, he does the finishing touch. That's how he's been working for months.
And then the meter hit zero. Right before a deadline that mattered. What happened wasn't that a useful feature disappeared. His entire way of working stalled. He had to go back to how it used to be done, and only at that moment did he feel how slow that actually was.
That's the interesting part. Not that he got stuck, but what it revealed. The value of AI had been there all along. He only noticed it once it briefly vanished. Like electricity: you only notice it when the power goes out.
It reminded me of something I see more often in the market. People are still wondering whether AI delivers anything, while it's already deeply woven into their daily output. The question is no longer whether it works. The question is whether you can keep affording it.
## Why this is a correction, not a crash
Let's start with the markets, because that's where conversations usually begin. The drop was real. The [Nasdaq fell 2.2 percent in a single day, on top of 1.3 percent the day before, with Nvidia down 4 percent, AMD down 6.2 percent and Micron down 8.5 percent](https://informedclearly.com/en/economy/56673/ai-bubble-fears-global-stock-selloff-2026). Oracle fell 5.5 percent. In Asia it was even more severe, with the Korean Kospi down ten percent and a trading halt for the first time since March.
Sounds like a bursting bubble. But read further and you see something else. It is [widely interpreted as a correction, not a crash: investors are shifting from rewarding AI spending to demanding proof of returns](https://intellectia.ai/blog/ai-stocks-selloff-june-2026). That is a fundamental difference. A bubble bursts because there's nothing behind it. A correction corrects a price, not a promise.
And the numbers making markets nervous tell you exactly where the pain sits. Not in the question of whether AI works. In what it costs to build it.
Look at the numbers. [Alphabet is targeting 175 to 185 billion dollars in investment in 2026, Amazon around 200 billion, and together with Microsoft, Meta and Oracle the total capex exceeds 452 billion dollars for that single year](https://intellectia.ai/blog/ai-stocks-selloff-june-2026). At the same time, [Alphabet's free cash flow fell 47 percent in the first quarter of 2026, and Amazon's free cash flow over twelve months collapsed 95 percent to 1.2 billion dollars due to infrastructure costs](https://intellectia.ai/blog/ai-stocks-selloff-june-2026).
That's not a company that earns nothing. That's a company spending so much on the future that nothing is left over right now. The difference is everything.
## Token consumption: the bill nobody sees coming
Back to my friend with his presentations. What happened to him is now happening at scale, and the cause has a name: tokens. That is the unit in which AI models bill your usage. Every prompt you send and every response you receive is counted in tokens. The more the machine thinks for you, the more tokens, the higher the bill.
And that bill is rising fast. According to [tech investor Prosus, AI agents now consume on average three times as many tokens as a year ago, and the more complex the task, the less predictable that bill becomes](https://www.bernarddonners-ai.nl/post/de-oude-buy-or-build-voor-ai-is-dood). Writing code, reviewing an application, processing invoices: consumption spikes the moment you let the machine do real work.
This is the silent problem beneath the entire AI promise. Everyone wants to reach the next rung. From a chatbot that answers questions to an agent that takes over entire processes. But that next rung is precisely the most expensive one.
The image McKinsey uses in its report *AI in insurance: Understanding the implications for investors* is a staircase. At the bottom sits the predictive AI we've known for years: fraud detection, pricing, risk models. Above it, generative AI processing documents and policies. And at the top, agentic AI promising to autonomously manage entire workflows from start to finish. Each rung delivers more. And each rung costs more. The promise and the bill climb together.
For those thinking this is a far-off concern for large enterprises: the opposite is true. Smaller companies feel this first, because for them the difference between a fifty-euro subscription and a five-hundred-euro monthly bill is the difference between doing it and not. I wrote about this earlier in [getting a grip on your AI costs](/en/blog/ai-cost-management).
## The second blow: the best models disappear behind a gate
And then this month a second development lands on top, one that sharpens the cost question even further. The very best models are becoming harder to access. Not through the market, but through politics.
Earlier this month the US government [imposed export controls on Anthropic, after which the company disabled its two most powerful models, Mythos and Fable, for all customers worldwide](https://www.cnn.com/2026/06/25/tech/openai-limit-release-white-house). That is not a commercial choice by a company. That is Washington intervening.
And it didn't stop with Anthropic. On 26 June OpenAI announced its new GPT-5.6, but [at the government's request restricted the release to approximately twenty government-approved partners](https://www.axios.com/2026/06/26/openai-gpt-sol-terra-luna-trump). The same pattern, within two weeks, from the two biggest players. Washington is beginning to treat the most advanced models as products that must pass through government before the rest of the world can access them.
This is precisely the scenario a group of European AI researchers sketched in June in [Europe 2031](https://europe2031.ai/), a detailed futures narrative about how Europe becomes dependent. It was published on 11 June, one day before Washington cut European access to those Anthropic models. Fiction overtook reality within twenty-four hours.
The core of that narrative touches you directly. The EU AI Act, the authors argue, is effectively built on the assumption that American providers won't want to lose the European market. But [if compute becomes scarce, there is little incentive for those providers to keep serving European customers](https://sciencebusiness.net/news/sovereignty/european-sovereignty-ai-requires-ugly-trade-offs-say-experts). At the end of their scenario, Europe is left with three thin options: become an American protectorate, turn to China, or sink into isolation.
## Turning to China, with a new problem in your luggage
Suppose the American top tier becomes inaccessible or unaffordable for you. An alternative is ready and waiting, and it comes from China. Models like those from DeepSeek perform well and are often openly available, meaning you can run them on your own infrastructure.
Sounds like a way out. But you're trading one problem for another. With an American model you already have a dependency and a governance question. With a Chinese model you add a new layer on top: different rules, different concerns about where your data goes, different political sensitivities with your customers and regulators.
Seen this way, this isn't one problem but a pincer. On one side, costs are rising. On the other, access is becoming uncertain. And whichever way you move, your governance gets more complicated, not simpler. I wrote earlier about why open source is therefore not a hobby but a governance necessity, in [this post about the kill switch](/en/blog/ai-vendor-lock-in-kill-switch).
## The real governance question is shifting
This is where everything comes together, and here sits the point I almost never hear anyone make. We talk about AI governance as if it's about content. Can this model see this data? Does it comply with GDPR? Is it consistent with the AI Act? Important questions, but they're questions about the content of what you do.
The new risk is not about content. It's about availability. You can set up a model in perfect compliance and still find yourself without one next year, because a government decides you no longer have access, or because the price triples. That is a governance risk that is separate from GDPR and separate from the AI Act. And it's new.
For an executive, this means a different question at the table. Not only: is our AI use properly regulated? But also: what happens to our business if the model our processes run on becomes three times more expensive tomorrow, or disappears entirely? My friend the entrepreneur got a mini-version of that question when his meter hit zero. He lost a few hours. A company that has built its entire customer service or claims handling on a single model loses more than a few hours.
This is why I maintain that AI is not a bubble. A bubble would mean the value is hot air. But the value is precisely the problem. It is so real, so deeply woven into how we work, that its disappearance hurts. You don't panic when a toy breaks. You panic when a tool you depend on disappears.
## What this means for you
The potential is no longer up for debate. My friend proved it without meaning to: he was demonstrably faster with AI than without it, and only noticed when it was gone. That is not a reason to wait until the hype blows over. The hype is not the point. The bill is the point.
So the strategic imperative for 2026 is shifting. Not: are we going to do something with AI? That answer has long been yes, whether you consciously decided it or not. The question is: do we keep costs and access in our own hands? That means knowing what your AI runs on, what it costs you per month, and what you do if that one provider shuts the door tomorrow or raises its price. It means not building everything on a single model, however good that model is.
Markets will recover. Models come and go, blocked one month, available the next. But the underlying movement continues: value grows, and with it, control over costs and access is no longer a technical footnote. It becomes the core of your AI strategy.
My friend now has a second tool at hand for when the first one falters. A small entrepreneur thinking about vendor risk. Without calling it that, he's doing exactly governance. The rest of us have some catching up to do.