Back in my April column, I explored how the hidden expense of artificial intelligence poses a serious threat to its profitable long-term commercial viability. Since then, a series of major news stories from the tech sector have strikingly supported this very point, revealing impacts at a dramatic scale.
The AI industry is shifting at such a blinding pace that it’s hard to follow. Just a short time ago, tech firms—and even other industries—were pushing employees hard to adopt AI, mandating its integration into workflows whether there was a genuine need or desire for it or not.
Looking Back with Perfect Clarity
It seems predictable in hindsight: when you link someone’s job security to using a particular tool more, many people will indeed start using it more. This sparked trends like “tokenmaxxing,” internal token usage rankings at companies such as Amazon, and staggering quarterly AI costs revealed by Uber (and other firms that preferred to stay anonymous). I’m honestly puzzled why these organizations were shocked by these outcomes. Still, the result has been a sharp reversal in corporate directives, both because the spending is unsustainable long-term and because AI adoption hasn’t delivered the transformational business results they’d hoped for.
Perhaps leadership assumed AI would trigger some kind of productivity miracle—but if so, they clearly hadn’t done enough research. Experts in the field and journalists covering tech had already cautioned that AI is simply a tool, powerful when used wisely and wasteful when misapplied.
I’ve drawn this analogy before, but picture a construction company that just invented electric drills—tools capable of dramatically speeding up building work. The wrong move would be to buy every drill possible until supplies ran out and prices soared, then force workers to use drills for every single task, tracking how many minutes each person used one. The result? Buildings riddled like Swiss cheese with pointless holes, enormous spending on equipment and power, and barely any meaningful progress—much like what we’re seeing with AI in tech today.
Budgets Have Limits
At last, reality is hitting hard—and fast. While some companies continue investing heavily, major players are now realizing the cost-benefit math simply doesn’t add up. So they’re pivoting. Yet as I warned in April, correcting course won’t be straightforward. Some firms are telling staff to use AI only where it genuinely adds value, abandoning the “tokenmaxxing” mindset to curb costs without losing potential benefits.
What they’re missing is that forecasting token expenses and pinpointing when AI truly helps is far more unpredictable than budgeting for traditional software. Let’s revisit my April piece to recall how AI usage plays out for an individual user.
“On the surface, you can limit how many tokens you send—and thus your costs—by writing short prompts and cutting extra details. But once agentic tools enter the picture, with LLMs generating prompts for other LLMs, you lose control over prompt length. More importantly, you have almost no say in how many tokens a model returns (aside from asking it to ‘be concise’). Output token count remains largely unpredictable. And keep in mind: each output token costs five times more than an input one.”
To dig deeper: every time you use AI, there’s a real chance your question won’t be answered correctly on the first try. That uncertainty compounds the problem. A developer doesn’t know (A) how many tokens a response will consume, or (B) how many attempts—possibly with tweaked prompts—will be needed for success. The true cost equals the sum of all input tokens plus all output tokens (which is unknown), multiplied by the number of tries required (also unknown). Both A and B fluctuate unpredictably based on model design, the nature of the task, inherent randomness in the system, and hidden backend factors. Then multiply all that by per-token pricing—which itself varies, as I explained in April.
So if you’re on the finance team of a tech company trying next year’s dollar budget for AI tokens, good luck. Even with past usage data or detailed productivity targets, your odds of getting the number right remain slim. But since open-ended spending isn’t an option, you’ll inevitably have to cap usage—which raises uncomfortable questions.
Real-World Consequences
What will this actually look like in practice? Will the second half of the year revert to all-manual coding after the first half leaned heavily on AI? Will every email and marketing draft be handwriting in Q3 and Q4? Will transcription and voice tools shut off once a spending threshold is reached? This puzzles me because I’ve seen firsthand how different coding with AI feels versus doing it the old way—and constantly switching between the two would be deeply disruptive.
There’s also the ripple effect on AI providers. Last October I wrote about how hyperscalers—Anthropic, OpenAI, Google, and others—are pressuring startups to embed AI features into their products, hoping to drive revenue back to investors who’ve poured billions into the sector. But as AI delivery costs rise and pay-per-use models spread, this cycle is starting to falter. If companies cut back on AI tools due to budget constraints, hyperscalers’ revenue streams will shrink. That’s the last thing OpenAI and Anthropic need as they plan high-stakes IPOs this year, both facing uncertain profitability and owing hundreds of billions to backers.
Also worth noting: Apple just unveiled its own AI ambitions at WWDC, and early reactions are favorable. Powered by Google’s Gemini, the new Siri promises strong privacy safeguards—processing on-device or in a private cloud with minimal data retention—and won’t charge users extra. If the quality holds up, this could pull everyday consumers away from ChatGPT and Claude.
Final Thoughts
Keep watching this space. Headlines about “companies stunned by AI bills” and “OpenAI and Anthropic aiming for record-breaking IPOs” may seem unrelated, but they’re really two sides of the same coin. Even if tech firms believe AI boosts productivity, their budgets aren’t infinite. And neither are consumers’—with grocery prices straining household finances and public confidence at historic lows. That forces a hard question: where exactly will the tens of billions OpenAI, Anthropic, and others expect to earn actually come from? Layer on growing public resistance to data centers and widespread skepticism toward AI itself, and hyperscalers are facing a serious reckoning.
Check out more of my writing at www.stephaniekirmer.com
