When the trillion-dollar bet on giant models unwinds, the firms left standing won’t be the ones who bet biggest. They’ll be the ones running AI they actually own.
I think the AI bubble is going to pop, and I think your business will be better off when it does. Not because AI is worthless but because the version being sold to you right now, billion-dollar models humming away in distant data centres, is the most expensive and least controllable way to get the benefit. When the money tightens, the firms left standing won’t be the ones who bet biggest. They’ll be the ones running AI they own. You don’t have to wait for the crash to make that move.
The brute-force ceiling
The tools you’ve heard of — ChatGPT, Claude, Gemini — are large language models. Strip away the mystique and they do one thing: predict the next chunk of text, having been trained on an enormous pile of data. They’re good at it, and getting useful work out of them is part of my week. But the method has a ceiling. You can only throw so much data and so much computing power at a prediction engine before each extra billion buys you less. A lot of serious people now think we are close to that point — the latest models aren’t improving the way bigger ones used to, and the labs are quietly changing course. The economics underneath the hype depend on that never happening.
That matters to you for one plain reason: someone has to pay for all that computing power. Right now it’s investors, subsidising the price you pay to keep the growth story alive. That is not a price you can build a business plan on.
The bill nobody’s putting on the invoice
When one data centre draws as much electricity as a small town, something is out of balance. These facilities burn power and use water at an incredible scale, and the communities living next to them are starting to push back. The International Energy Agency has flagged the climb in data-centre energy use as a real concern, and US government researchers put the demand on the scale of small cities.
I use these tools knowing all of this. I’m not here to tell you AI is the devil’s work — turning your back on it would be its own mistake. The point is narrower, and it’s about who holds the dial on your costs. The tools people build on are already starting to charge far more, raising their floors and moving to usage-based billing. Anyone whose business runs on top of them is exposed: Webflow, Lovable, every company whose product is really a layer over a model it doesn’t own. Today’s low price is the bait. The hike lands once you’re locked in and leaving is painful, a bait-and-switch with your business plan as the mark. You don’t have to be the one holding it.
The cracks are already showing
You don’t have to take my word on the timing. Look at what the people closest to it are doing. The hyperscalers spent two years racing to build data centres, and now they are quietly easing off: Microsoft has cancelled leases for AI capacity it had signed up for, and Amazon has paused data-centre deals it was negotiating. When the firms laying the foundations start slowing down, that tells you more than any press release.
The markets are noticing. This week Oracle posted record results, then watched its shares fall more than 8% in a day, not because it sold too little, but because investors did the sums on its AI spending and the cash going out the door, and didn’t like the answer.
Then there’s the bill you actually pay. The price of a single AI query keeps falling, which sounds like good news until you notice how much more of it everyone now uses. Companies are blowing through their budgets: Uber burned its entire 2026 AI budget in four months. The cheap prices are propped up by investor money, that prop is being pulled away, and the real cost is heading toward whoever is left holding the contract.
And for all that spending, the returns are thin. An MIT study found 95% of corporate AI pilots delivered no measurable return. Millions poured into tokens, very little to show a board for it. A tool can be genuinely useful and still be priced like a mania. Both are true at once, and that is exactly how bubbles work.
The future runs on your own machine
Here is what I think happens once the dust settles. The winners won’t be the companies building ever-larger models that need their own power station to run. The winners will be local models: smaller, focused tools that run on your own hardware. Modern computers increasingly ship with an NPU, a chip built specifically for AI work — a specialist brain alongside the general one. These smaller models don’t try to do everything. They are tuned for a job, they do it well, and they do it without sending your data to a server farm you’ll never see.
For a small business, that last part is the whole game. The single biggest unease I hear from owners about AI is loss of control: your data, and your customers’ data, routed through a model you don’t own, run by a company whose incentives aren’t yours. Local, open models answer that directly. The information never leaves the building. The running cost is predictable instead of hostage to a vendor’s next price rise. And you, not a supplier, decide how your data is used. That is not just cheaper. It is better governed.
What to do now, while it’s still early
You can get ahead of this without much outlay. Pick a machine that can run local models and start small, on a narrow job using data you already control. A local model set loose on your own documents, contracts, customer notes, the internal wiki nobody can ever find anything in, can search and organise all of it without a byte leaving your network.
One firm rule, and I mean firm: be very careful what you connect these things to. The technology is still young. You do not want an AI agent anywhere near your bank, your payments, or anything where a mistake costs real money. Keep a human between the model and any decision that’s hard to undo. Start where a wrong answer is cheap, learn what the tool is actually good at, and widen the circle slowly. That restraint isn’t timidity. It’s the difference between AI that helps you and AI that embarrasses you in front of a client.
The honest part
I might be wrong about the timing. The bubble could limp on for another two years, or deflate slowly instead of bursting. I’m far less worried about being wrong on the direction. The trillion-dollar bet on bigger-and-thirstier is the fragile thing here. Small, local, open tools that respect your data, your budget and your planet are the durable one. The smart money ends up there either way.
So the question isn’t whether to use AI. It’s whose AI you want to depend on when the music stops. Which job in your business would you trust a model you own to do first — and what’s stopping you trying it this month? Tell me, and I’ll help you think it through.
Be braver now.
Sources
AI scaling laws showing diminishing returns; labs changing course — TechCrunch (Nov 2024): https://techcrunch.com/2024/11/20/ai-scaling-laws-are-showing-diminishing-returns-forcing-ai-labs-to-change-course/
Data-centre and data-transmission energy use — International Energy Agency: https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
Data-centre electricity demand on the scale of small cities — Lawrence Berkeley National Laboratory / US Department of Energy: https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/
Hyperscalers pausing and cancelling data-centre leases (Microsoft, then AWS) — ITPro: https://www.itpro.com/infrastructure/data-centres/aws-data-center-lease-spending
Oracle shares fall 8.5% after Q4 results on AI spending and cash-flow concerns (11 June 2026) — The Motley Fool: https://www.fool.com/coverage/stock-market-today/2026/06/11/stock-market-today-june-11-oracle-falls-after-ai-spending-guidance-sparks-cash-flow-concerns/
AI token pricing, subsidies ending, Uber’s AI budget burned in four months — Investing.com: https://www.investing.com/analysis/the-ai-token-pricing-crisis-behind-openai-and-anthropics-revenue-race-200680777
95% of corporate generative-AI pilots show no measurable return (MIT NANDA) — Fortune: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/