The Great AI Swindle: Why the Math Simply Doesn't Math 🧮
OpenAI – the most famous, most funded, most talked-about AI company on the planet – lost over twenty billion dollars last year. That's billion with a b. Google, Amazon, Microsoft, Anthropic – every major player is in the same boat, burning through cash reserves and investor money with no clear map to when any of it turns a profit. The AI industry is the most expensive experiment in business history. And the results are still very much pending.
While the economics of AI remain unresolved, businesses everywhere are making disruptive and irreversible decisions based on the assumption that it will. Restructuring teams. Gutting talent. And perhaps most surreally: AI adoption written into employee performance reviews. Gone are the days of the good old fashioned 'be a team player' or 'have attention to detail'. Which is unfortunate, given what's coming in the next few sections.
Some have already made the cuts and replaced headcount with subscriptions. Others are still in the mandating phase – get on board, embrace the tools, and don't ask too many questions. Both are bets on the same unproven outcome. And neither group is entirely sure they've called it right.
We've all been at that frustrating Town Hall. Leadership have watched the demos, attended the summits, seen their peers centring entire businesses around AI, and decided that they should too. Output targets double. Hiring freezes. And you are informed, warmly but firmly, that everyone needs to get on board. No discussion. No debate. Just figure it out and fall in line. Lovely.
What nobody in that Town Hall will say out loud – possibly because nobody in that room actually knows – is what happens next. Which AI companies survive and which fold when the money runs out. What these tools actually cost once they're priced to make a profit rather than dominate market share. Whether the productivity gains will ever outweigh the disruption of getting there. Worth bearing in mind that the people making the loudest predictions have the most money riding on being right.
The confidence in AI isn't based on evidence. It's based on momentum and hype. History has repeatedly demonstrated that these things have a rather unfortunate tendency to end in a car park full of journalists and a lot of explaining to do. The Metaverse was going to reorganise human civilisation around virtual meeting rooms where everyone had no legs. Crypto was going to replace money itself. Both attracted billions in capital and a wave of people who'd never previously cared about technology suddenly explaining it to you at dinner parties. We know how those ones landed.
So, If you have an office job, keep reading. Decisions are being made on your behalf, with considerable confidence, by people who don't yet know the outcomes. Leadership have already drunk the Kool-Aid. So let's get into it.
Heavy Metal Masquerading as Magic
When you fire a prompt into these models, something happens on the other end that is almost comically expensive. Your question gets processed by an enormous network of specialised computer chips – thousands of them, drawing extraordinary amounts of electricity, housed in data centres the size of several football pitches. This happens every single time, for every single user, for every single prompt. No shortcut. No bulk discount.
Most successful technology businesses get cheaper to run as they grow. Build a piece of software once and the ten-thousandth person to use it costs you almost nothing. That's historically why software companies have been such extraordinary businesses – the margins improve significantly as the customer base grows.
AI doesn't work like that. Serving a million users costs exactly the same per user as serving a hundred. Every additional customer is another mouth sucking at the AI teat. The infrastructure required is so power-hungry that decommissioned nuclear power plants are being switched back on just to keep up with demand. This isn't weightless digital magic floating in the cloud. It's heavy industry – physical, high-maintenance, and thoroughly unimpressed by your quarterly targets.
So how are these companies keeping the lights on whilst charging you less than it costs to serve you? Investors are covering the gap, on the understanding that the numbers will eventually work out. It's the oldest commercial strategy in business: give away the razor almost for free, then charge a fortune for the blades once everyone is hooked and the competition has been cleared.
The Hallucination Tax
Assume, generously, that the economics eventually sort themselves out. The power consumption gets resolved, the investors stay patient, the pricing somehow lands. What have you actually bought?
In factual work – research, legal copy, financial analysis, anything where accuracy genuinely matters – AI has a well-documented problem. It produces wrong answers with exactly the same confidence it produces correct ones. Not occasionally. Regularly, unpredictably, and without so much as a cheeky grin.
The result is that for anything requiring real accuracy, you need a human expert checking everything the machine produces. You've hired an articulate colleague with a complicated relationship with the truth, and then employed a senior person full time to follow them around with a clipboard. If those efficiency gains are in there somewhere, nobody's found them yet.
This is the Hallucination Tax. It's not a bug the industry is close to fixing – it's a fundamental characteristic of how these systems work. And even the people behind it all aren't sure it's fixable.
The World's Beigest C-Grade Student
In work that involves creative thinking, the problem shifts entirely. Here, AI isn't as unreliable, but it's extraordinarily, consistently average. It will always produce something. Swiftly, cheerfully, and without a single complaint about the brief changing for the fourth time.
These models are built by processing enormous quantities of existing human work and identifying the patterns within it. The output is a distillation of what already exists. It can't depart meaningfully from the pattern, take a genuine creative risk, or be memorably irrational in the way that actually captures attention.
In creative terms, it's a consistent C student with the personality of lukewarm tap water on a hot day. Homework always in on time. Never anything worth sticking on the fridge.
It's worth being clear about something here. There is a meaningful difference between a task and a job. AI can automate plenty of tasks – summarising a document or turning gibberish into a professional email – and that's genuinely useful. But automating tasks is not the same as replacing a job, and conflating the two is how you end up with hasty decisions made by the higher-ups who haven't been in the trenches delivering billable work for decades.
The actual job is something else entirely. It's decoding feedback that contradicts itself and reading the room well enough to know which bit to action. It's knowing when to push back and when to let something go. It's navigating a meeting where three people want different things and finding a solution that works without anyone feeling they've lost. It's talking a nervous stakeholder down from a terrible decision at the eleventh hour and somehow making them feel like it was their idea all along.
All that without even scratching the surface of how a business truly runs – teamwork, mentorship, and most importantly top-tier banter.
When the Free Champagne Runs Out
The subsidised era will end. Investors will eventually stop funding losses at scale without a credible return. When that happens, pricing resets to reflect what AI actually costs to deliver. Flat-rate subscriptions give way to token-based pricing, where the meter runs every time the machine generates a word, an image, or an opinion. The tools that currently feel as affordable as a Netflix subscription will look considerably different when the real bill drops on the table.
Nobody knows who or what will win. The current market leaders could be displaced by something cheaper or better. The industry is still very young. The use cases that seem obvious now may turn out to be quite niche. The ones that ultimately matter most probably haven't been identified yet. The loudest voices are always the ones with the largest financial stake in the outcomes. Worth keeping that in mind before taking their predictions as gospel.
Your taste, your judgement, your ability to produce work that actually connects with other human beings – these aren't inefficiencies waiting to be optimised. They're the thing the machine is unable to replicate. As the world fills with frictionless content, the genuinely human stuff will ultimately become more valuable, so don’t lose your grasp.
The true cost is coming. When it arrives, the businesses that gutted their human talent and bet the whole operation on subsidised software will be left holding it. And their blank face will say it all – what the fuck just happened and the CFO is going to murder me.
Until then – your instincts, your judgement, and your beautiful human brain are still the most valuable thing in your toolkit. We're complicated, we need rest, and we can't run twenty-four hours a day. But at least we don't fabricate information like sociopaths, burn through the energy reserves of a small nation on each minor thought, or exist solely to make a handful of greedy investors astronomically rich.