So here's something I've been chewing on all week: the biggest companies on the planet just discovered they have no idea how much AI should actually cost them. And if they can't figure it out, the rest of us building workflows on top of their tools should probably stop and think about what exactly we're standing on.

Spoiler: it's someone else's money.

Earlier this year, the message from Big Tech to their employees was beautifully simple: use AI as much as possible. Amazon and Meta set up internal leaderboards tracking which developers consumed the most tokens. Some companies tied AI usage to performance bonuses. Others issued minimum weekly usage targets. The logic being that if AI is the future, the companies whose staff burn through the most of it will get there first.

It took approximately five minutes for Goodhart's Law to do what Goodhart's Law always does. When token volume became the target, it stopped measuring anything useful. Developers started running automated agents around the clock — not because the output was valuable, but because the activity registered on the scoreboard. Engineers constructed elaborate multi-step prompts specifically to inflate consumption. Entire side projects were spun up with no purpose beyond burning through tokens. The metric was being hit. The point was being missed entirely.

The correction arrived fast. Amazon scrapped its leaderboards, with a senior manager telling staff not to use AI for the sake of using AI. Meta capped usage after watching costs go exponential. Uber disclosed it had blown through its projected AI budget for the entire year by April — four months in. Its COO acknowledged publicly that the company was struggling to draw a direct line between its AI spending and the value it was getting back. Microsoft reportedly scaled back internal Claude Code licences after per-engineer costs hit between $500 and $2,000 monthly.

The industry has now pivoted to what's being called "tokenminning" — token minimising. Which, if you think about it for more than thirty seconds, is just the other side of the same thoughtless coin. Cutting AI spend without understanding which usage was valuable and which was waste isn't a strategy. It's a diet that starts on Monday.

The enterprise spending correction is worth paying attention to on its own. But the deeper story — the one that should matter to anyone building processes around AI tools — is the financial reality of the companies providing those tools.

OpenAI's recently leaked audited financials show a company whose revenues are growing rapidly and whose losses are growing faster. Revenue rose from $3.7 billion in 2024 to $13.07 billion in 2025, with monthly revenues approaching $2 billion by the end of last year. Impressive, right? Except R&D costs alone outpaced total revenue in both years — $7.81 billion in 2024, ballooning to $19.18 billion in 2025. Throw in the cost of actually serving the product plus sales and marketing, and OpenAI's operating loss hit $20.92 billion in 2025. The headline net loss for the year was nearly $39 billion (roughly $30 billion of that reflects a one-time accounting charge for the for-profit conversion, but still).

Here's the thing. These numbers tell you something fundamental about the price you're currently paying for AI. You're not paying market rate. You're paying an introductory rate, subsidised by companies collectively burning tens of billions per year in pursuit of market share, platform dominance, and — imminently — public market valuations.

OpenAI raised $122 billion in financing earlier this year at a valuation of $852 billion. Both OpenAI and Anthropic have filed paperwork for what are expected to be among the largest stock market flotations in history. They've told investors they intend to be profitable. The economics you're building on today are the economics of companies that haven't yet been required to make money.

(That sound you hear is every VC-backed company in history nodding along nervously.)

The natural instinct when reading about AI budgets at Uber or Meta is to file it under "big company problems." The budgets involved are orders of magnitude beyond what a mid-market brand or Amazon seller would encounter. But the structural dynamic is identical at every scale.

Every AI workflow you've built — every agent monitoring competitors, every automation generating listing copy, every process routing customer queries or analysing ad performance — runs on pricing that reflects a land-grab, not a mature market. When frontier AI companies begin the transition from growth-at-all-costs to demonstrable profitability, pricing will adjust. It's already started. Anthropic's newest model is twice as expensive as its predecessor. The direction is clear even if the timeline isn't.

This isn't an argument against using AI. It's an argument for using this window — while the economics are favourable — to build systems that are efficient, intentional, and durable. The distinction matters enormously.

An ecommerce operation that builds a token-hungry workflow requiring the most expensive frontier model for every task is building a cost structure that scales in exactly the wrong direction. An operation that designs lean processes — using frontier models only where complexity demands it and routing simpler tasks to cheaper alternatives — is building something that actually survives a pricing correction.

AT&T's chief AI officer offered a useful framing recently: his engineers use the most powerful models for tasks that require them and less expensive models for everything else. The result is savings of up to 90 per cent on tasks where frontier capability is unnecessary. That kind of model-routing discipline sounds obvious in principle. In practice, as the tokenmaxxing phenomenon demonstrated, most organisations defaulted to using the most powerful model for everything. Breaking that habit requires deliberate process design, not just a memo telling people to spend less.

This newsletter has covered the scissors gap before — the structural disadvantage smaller sellers face when algorithmic campaign types require data volumes they can't generate. A similar dynamic applies to AI operational costs. The companies with the largest engineering teams absorbed tokenmaxxing as an expensive lesson. Smaller operators don't have that margin for error. The waste is proportionally more damaging, and the correction, when it comes, is harder to absorb.

There's a parallel failure mode that operates at an entirely different scale — not in organisational budgets, but in individual cognition.

A bicycle amplifies your capability. You still steer, you still balance, you still decide where to go. A crutch bears weight you could carry yourself, and the longer you lean on it, the weaker the muscles become.

The tokenmaxxing era was the organisational expression of the crutch. Employees were incentivised to offload as much work as possible to AI — not because the output was better, but because the act of offloading was the metric. The individual version is quieter, more personal, and potentially more consequential. It's the daily habit of reaching for AI before reaching for your own thinking. The draft you ask Claude to write before you've formed a view on what it should say. The analysis you delegate before you've looked at the data yourself. The competitive research you automate before you've developed the instinct to know what matters in the results.

Each of these, in isolation, looks like efficiency. Over time, the cumulative effect is the erosion of the skills that make you good at your job. The seller who never manually analyses a search term report gradually loses the ability to spot patterns the model wasn't trained to flag. The brand manager who outsources every strategic brief slowly forgets how to construct an argument from first principles. The operator who automates all supplier evaluation eventually can't tell when the automation is wrong.

Here's where it gets genuinely fascinating — and slightly uncomfortable. If the subsidy window closes and AI becomes meaningfully more expensive, the people who maintained their own capability will adapt. The people who outsourced it will be stranded. Not because AI stopped being useful, but because they stopped being useful without it.

Six months ago, an employee who consumed the most tokens was a top performer. Today, that same employee is a cost centre. The metric flipped. But the skills you did or didn't develop during the tokenmaxxing era are already baked in. You can't retroactively un-atrophy a capability you spent six months not using.

(I realise this is the professional equivalent of "never skip leg day," but it doesn't make it less true.)

The two traps — organisational overspend and individual cognitive outsourcing — are expressions of the same underlying failure: adoption without strategy. The companies that burned through their AI budgets by April and the individuals who can't draft a competitive analysis without a chatbot both made the same mistake. They treated volume as a proxy for value.

For ecommerce operators, the framework starts with three questions. Which tasks genuinely benefit from frontier model capability, and which can be handled by cheaper alternatives? Where does human judgement remain essential — not as a sentimental preference, but as a genuine quality control mechanism? And are the systems you're building today designed to survive a pricing environment that looks materially different from the current one?

The AI tools available to us right now are more capable than they've ever been and cheaper than they're likely to remain. That combination won't last. The companies providing these tools are heading toward public markets and profitability targets that are structurally incompatible with today's pricing.

Build lean. Build intentional. And maybe — just maybe — keep doing a few things manually that you could automate, so you remember how they work when the bill arrives.

P.S. If you're wondering whether your AI workflows would survive a 2x price increase, the answer is almost certainly "not without some surgery." Now is the time to audit, not after the pricing email lands in your inbox.

P.P.S. Uber blew their annual AI budget by April. April. I've made some questionable budgeting decisions in my time, but at least I usually make it to June before the spreadsheet starts weeping.

📹 Hedra - Turn an image plus audio into talking, expressive AI avatar videos.

🤖 Synthesia - Avatar-led explainer and training videos in 140+ languages.

⚙️ Bolt - In-browser AI app builder supporting many frameworks.

🗣️ Cartesia - Ultra-low-latency text-to-speech and voice cloning for real-time apps.

😍It was great, all I needed

🙂Good, not great

🙄It sucked