I gave a talk recently at South Coast Technology Leaders called “Stop Chasing AI: Start Solving Problems.” One slide got more reaction than the rest combined. It made a simple point: the AI subscriptions you’re paying for right now are heavily subsidised, and that won’t last.
If you’re experimenting with AI in your business, this matters. Not because you should stop, but because the assumptions you make today about cost are almost certainly wrong.
The economics nobody is talking about
OpenAI generated around $13 billion in revenue in 2025. In the same year, it spent close to $22 billion. That’s $1.69 spent for every $1 earned. Internal documents reported by the Wall Street Journal show the company expects operating losses of roughly $74 billion in 2028 alone, before turning profitable somewhere around 2030.
Anthropic, which makes Claude, spent around $6.8 billion on compute in 2025 against total spending of $9.7 billion. That’s around 70% of every pound going on the infrastructure to train and run the models. The company is still running at a loss.
These are the two biggest names in AI, and they’re both spending far more than they earn. That money has to come from somewhere, and right now it’s coming from investors prepared to bet that scale today equals dominance tomorrow.
What that means for your monthly subscription
Your £20 ChatGPT Plus, your £20 Claude Pro, your Copilot, your Cursor account. None of them reflect the real cost of running the models behind them.
Take Claude Code, which a lot of developers use for writing software. A heavy user can easily generate $60,000 to $90,000 of actual compute cost in a year, against a subscription that costs around $2,400. The provider is absorbing that gap because they want you locked in. Once your workflows depend on the tool, switching becomes painful. That’s the bet.
It’s the same playbook Uber ran with cheap rides, WeWork ran with cheap office space, and Amazon Prime ran with free shipping. Subsidise growth, build dependency, raise prices later.
The signs are already there
Anthropic has changed how it bills business customers, moving toward charging based on actual usage rather than flat fees. Rate limits have tightened across the major providers in 2026. Claude Code users have publicly complained about quotas running out faster than expected, and Anthropic’s own CEO has said the company is compute-constrained.
OpenAI doubled the price of GPT-5.5 earlier this year. None of this is a crisis. It’s a slow recalibration toward something closer to real cost.
Within 12 to 18 months, expect three things: free tiers will tighten, paid tiers will get more expensive, and pricing models will shift from flat subscriptions toward metered usage. The cheap, all-you-can-eat era is ending.
What this means if you’re using AI in your business
Two things matter.
First, experiment now. This is the cheapest these tools will ever be, and the gap between what you pay and what you get is genuinely extraordinary right now. If you’ve been hesitating about trying AI in your business, hesitate less. You will never get more capability per pound than you do today.
Second, design with cost in mind. The application of AI you build today might work commercially because the cost is artificially low. Ask yourself the harder question: would it still work if your AI bills doubled? Trebled? Went up tenfold? Because that’s not a hypothetical. That’s the trajectory.
This is where most businesses are getting it wrong. They’re sprinkling AI on everything, often where it isn’t even the right tool.
The RPA problem
RPA stands for Robotic Process Automation. It’s been around for years. It’s the sort of automation that handles structured, repetitive tasks: copying data between systems, processing invoices, moving information through known workflows. It’s deterministic, which means it does the same thing every time. It’s auditable. It’s cheap to run.
There’s a lot of business activity that genuinely doesn’t need AI. It needs RPA, or a script, or a rules engine. Things where the logic is stable and the inputs are predictable. AI is the wrong tool for those jobs. It’s slower, more expensive, and less reliable. You don’t need a language model to move a number from column A to column B.
Where AI genuinely earns its place is the messy stuff. Reading unstructured information at scale, like contracts or support tickets. Letting non-technical staff query data in plain English. Spotting patterns in chaotic inputs. Summarising long documents. Drafting content that a human will review.
The smart move is to use AI to build your automations, not to be your automations. Get Claude or ChatGPT to help you write the script, design the RPA workflow, or build the rules engine. Then let cheap, deterministic infrastructure run it day to day.
McDonald’s learned this expensively. They ran a three-year partnership with IBM to put AI in the drive-thru, taking voice orders. It was a mess. Customers got 200 chicken nuggets they didn’t ask for. Bacon ended up on ice cream. Water turned into Coke. They quietly shut the whole thing down in July 2024. The truth is, that problem didn’t need AI. Better menu design and standard speech recognition would have done most of the job at a fraction of the cost.
The opportunity that’s hiding in plain sight
Here’s the part most people are missing. Because AI tooling is cheap right now, something has shifted underneath the surface.
Three of my own clients are currently building things that wouldn’t have been viable two years ago. One is replacing a CRM that never quite fit the business. Another is cutting expensive licensing costs by building a custom platform from existing intellectual property. A third is building a multi-tenant system to sell their own approach to other businesses.
None of these projects would have made commercial sense before. They’d have needed a £200,000 consultancy engagement and a year of build time. Now they’re being done by small in-house teams with help from Claude Code, Cursor, and GitHub Copilot, in weeks rather than months, at a fraction of the historic cost.
This is the genuinely interesting consequence of cheap AI. It’s not that AI is going to replace your developers, or your service business, or your job. It’s that the things you previously couldn’t afford to build are now within reach.
If you’ve been quietly fed up with software that doesn’t fit how your business actually works, this is your window. Build the thing that does fit. Just design it knowing that the underlying AI costs will rise.
What to do this week
Three practical actions if any of this lands.
One: look at where you’re using AI today and ask whether it would still be viable if your AI costs went up significantly. If the answer is no, you’ve got an economics problem coming.
Two: look at where you’re using AI for things that RPA, scripts, or rules engines could handle better and cheaper. Move those off AI now.
Three: look at the things you’ve wanted to build for years but couldn’t afford to. The maths might have changed.
If you want to find out where your business sits on the spectrum from AI-vulnerable to AI-resilient, the assessment below takes about three minutes and gives you a personalised report.
Leave a Reply