AI Is Free. Until It Isn’t.
Ask anyone what AI costs and they say twenty bucks. Ask them what they paid last month and watch them open their banking app. Then ask what their last laptop cost, what their Office subscription costs now versus two years ago, and how many evenings they spent teaching a chatbot to do a thing they used to just do. This episode is about all of those numbers, and about why they only go one way.
Here in Canada, 19.2 percent of businesses were using AI in the second quarter of 2026, up from 6.1 percent two years earlier. When Statistics Canada asked what stops the rest, cost came second at 10.6 percent, right behind privacy at 13.4.[14] When American small business owners were asked what they spend, a third said nothing, 28 percent said $25 to $99 a month, 16 percent said $100 to $249, and one in ten said $250 or more.[15] That survey is what owners say. Cledara is what the invoices say. The gap between those two is the whole point of this section.
And the invoice is still not the bill. Here are the rungs nobody budgets, in order of how much they hurt.
Training hours. Among large Canadian companies using AI, 68 percent trained their staff on it. Among small ones, 24 percent did.[14] The tool is cheap. Learning it is not. And the small shop skips the learning, which is exactly why the seat goes cold after two months.
Rework. Researchers at Stanford and BetterUp surveyed 1,150 American desk workers in the fall of 2025 and gave a name to the thing you have already received in your inbox. Workslop. AI-generated work that looks fine and is not. Forty percent had received some in the previous month. Each piece took nearly two hours to sort out, and they put the cost at close to $200 US per employee per month.[16][17] For our ten-person shop that is close to $2,000 US a month of cleanup, against a subscription in the $300s. The rework bill can run five times the software bill, and it is invisible because it is paid in other people’s afternoons.
Lock-in. Copilot needs Microsoft 365. Gemini needs Workspace. The AI seat quietly becomes the reason you can never leave the suite.
Data. Where does the customer list go when your bookkeeper pastes it into a chatbot. I covered this in Episode 1 and I will not repeat it, except to say the answer is usually “a server in another country” and the price of that is not on any invoice.
The pilot that never pays. MIT’s NANDA lab looked at $30 to 40 billion US of corporate spending on generative AI in 2025. Ninety-five percent of organizations had nothing to show for it on the profit and loss statement.[18][19] Gartner predicts more than 40 percent of the “AI agent” projects being built right now will be cancelled by the end of 2027.[20]
Now the fair part, because I promised you honest and not just scary. Among small businesses that do use AI, 52 percent say the return is positive and 48 percent say it saves them four or more hours a week.[15] It works. For about half. The other half paid for the lesson, and the lesson was mostly the rungs above.
Anthropic publishes a benchmark for what a working developer burns using its coding tool: about $15 US per active day, or $150 to $250 a month, with 90 percent of people staying under $30 a day.[23] That is one person, working, with every efficiency turned on.
Now put the same meter behind a chatbot on your website. It is no longer you typing. It is your customers typing, on your bill, at 2 a.m., asking the same question four different ways.
Let me do the math for one imaginary bot. Say it handles 10,000 conversations a month, and each one, with a few tool calls and a bit of back and forth, runs about 50,000 tokens. That is 500 million tokens. On a mid-tier model that is about $1,000 US a month before you count the output. Pick the top model from the dropdown instead and the same bot costs $5,000. Same bot. Same customers. One dropdown.[22] I am rounding hard and your numbers will differ, but the shape is right.
And then there is auto-reload. Every credits system offers it. It is the feature that turns “we hit the limit” into “we found out on the statement.”
Kitchen-table version: the app your kid uses that “talks” is running a meter somewhere and somebody is paying it. If the app is free, go back to rung zero and re-read the part about ads.
The tax on software you already had. In January 2025 Microsoft raised the price of Microsoft 365 Personal from about $7 to $10 US a month and Family from about $10 to $13. That is a 30 to 40 percent jump. The reason given was that Copilot, Microsoft’s AI, was now included. Whether you asked for it or not.[24][25] There is a “Classic” plan without the AI at the old price, but you have to know it exists and go looking for it.
Then in July 2026 the business plans went up too. The basic plan up about 15 percent, the standard one about 10, and the cheapest frontline worker plan up a third. Microsoft’s justification was “everything we have shipped across AI, security and IT management.”[26]
I have watched the same move in the design tools, the note-taking apps, the meeting software and the CRMs I use with clients. An AI feature appears. A price increase follows the same year. You did not climb a rung. The rung climbed you.
The computer you will have to replace. This one surprised me. The memory chips inside your laptop and your phone are the same kind of thing the AI data centres are buying by the truckload. And the data centres are winning the bidding war. In January 2026 CNBC reported that AI memory was simply sold out.[27]
By the summer, TrendForce, the industry’s price tracker, reported memory prices up about 60 percent in a single quarter and roughly 500 percent over twelve months, as much as ten times the lowest prices ever recorded. A basic memory upgrade for a desktop PC was suddenly a few hundred dollars. The manufacturers were openly shifting production to the high-margin server chips and passing the cost down to laptops and phones.[28] The only reason the surge started cooling was that consumers hit the ceiling of what they would pay.
So the next time you price a laptop and it feels wrong, that is the AI bill. It just arrived through the back door, from a data centre in Texas, and it will never be labelled as such.
The building is the problem. The International Energy Agency counted 415 terawatt-hours of electricity going into data centres worldwide in 2024, about 1.5 percent of everything humans used. It expects that to more than double to around 945 terawatt-hours by 2030, roughly what all of Japan consumes, and AI is the main reason. By 2030 the United States will burn more electricity on data centres than on making aluminium, steel, cement and chemicals combined.[33]
And it is already leaking onto other people’s bills. American residential electricity prices rose 7.1 percent in 2025, more than 20 percent in some states. In the parts of Northern Virginia with the densest data centres, prices are up 267 percent in five years.[34]
Which brings us home. On February 19, 2026, Hydro-Québec asked the Régie de l’énergie for a new rate for data centres over five megawatts. 13 cents a kilowatt-hour, roughly double what large industrial customers pay, with crypto miners pushed to 19.5 cents. About 190 megawatts of existing data centres would be affected now, and the sector could reach 1,000 megawatts by 2035. Hydro’s stated reason was a Bloomberg analysis showing that in American regions with fast data centre growth, everyone’s bill doubled over five years.[35] In Episode 1 I said cheap hydro and cold weather are Canada’s real edge in the AI race. That is still true. It is only an edge if the data centre pays for it and your household does not.
Now the mechanism, which is the point of this whole episode.
Only about 5 percent of chatbot users pay anything at all.[36] Against that, OpenAI plans to spend $1.4 trillion on data centres over the next eight years, about $175 billion a year, on revenue of roughly $20 billion.[36] The five biggest American tech companies plan around $600 billion of capital spending in 2026, three quarters of it AI.[36] A Northeastern University economist put it in one sentence: “This is not a sustainable price.”[36]
In April the Boston Globe laid it out. Price increases are coming for ChatGPT, Claude, Gemini and Perplexity. The comparison analysts reach for is Netflix, which more than doubled its price once it had you.[36]
But here is the twist, and it is the thing I most want you to take away. The increases are already here. They are just not on the sticker. As the French analyst Bertrand Duperrin wrote this month, the quotas, the five-hour windows, the weekly caps and the $100 and $200 tiers are the price hike.
Here is how it works. The cost of computing one token keeps falling. But the number of tokens per task keeps rising faster, because of longer conversations, tools and agents. So the vendors keep the entry price at $20 for the person who checks in twice a week, and charge everyone else by intensity.[37] The ladder in the first section is not an accident of product design. It is how a finite input, power and memory and chips, gets billed to a customer who was promised $20.
Thinking. In 2025 researchers at MIT’s Media Lab put 54 students in an EEG cap and had them write essays. Some with ChatGPT, some with a search engine, some with nothing but their heads. The ChatGPT group showed the weakest brain connectivity, could not quote from essays they had written minutes earlier, and felt the least ownership of their own work. Small study, not yet peer-reviewed, and the authors say so. But it points where the teachers are pointing.[41]
On August 27, 2026, the Fraser Institute published a survey of Canadian grade 6 to 12 teachers. Almost all of them, 95.6 percent, say at least some students use AI to do their homework. Forty-two percent say more than half do. Eighty-four percent say critical thinking is weaker than it was five years ago, and 73 percent say in-class writing is weaker.[29]
Kitchen-table version: this is the bill your kids are paying, and they cannot read the statement yet.
And then the bill that closes the loop back to rung zero. If $20 a month cannot cover what the chatbot costs to run, someone is subsidising it. Right now that someone is an investor, a pension fund and, increasingly, a private credit lender. Morgan Stanley estimates $3 trillion US of data centre spending between 2025 and 2028, with about half of it borrowed from private credit.[42]
The Bank of England and the IMF have both warned of a correction. Jamie Dimon has said the odds of a meaningful drop in stocks over the next two years are higher than people think. Sam Altman himself said the word “bubble” in August 2025. And in the last week of June 2026 the market gave everyone a preview. Samsung and SK Hynix lost 12 percent in a single morning, Oracle had its worst week since the dot-com crash, and the Nasdaq sank with them.[42]
The ladder exists because the free rung loses money and every rung above it is trying to earn it back. If the earning-back fails, that is a payout too, and it lands on everyone, whether or not they ever typed a prompt.