I first learned how to program on a Soviet-era clone of a PDP-11 in Mumbai in 1985. We spent two weeks in our second semester learning the nuances of how higher-level code written in Fortran consumed scarce computing resources.
We had to build a Fortran program to generate the Fibonacci series of numbers and had to do it four different ways. One that optimized for CPU cyckes the other that optimized for memory, one that optimized for I/O (punched cards and line printers), and one that balanced all of the above.
We didn’t have the luxury of sloppy logic. Every instruction and CPU cycle had a price. Every memory allocation was a tradeoff that required us to spend hours sharpening an algorithm so it could sip, not gulp, from the scarce pool of resources. That discipline was a virtue built out of necessity, and it made us better engineers.
Forty years later, I watch executives wave generative AI and now agentic AI at every problem in their organization the way you’d wave a magic wand, and I keep thinking about that machine in Mumbai.
Over the past 4 years, AI has been the most aggressively subsidized technology I have seen in my career. Hardware (GPUs), data centers, frontier models, chatbot applications, every layer of the stack has been priced well below its true economic cost. My instinct, and I’ve been saying this in public forums for the past 3 years at every opportunity I get. I believe that the real cost of what people have been consuming is at least ten times what they’ve been paying. Venture capital and market-share ambition have been quietly footing the bill so the rest of us could believe that intelligence was now a commodity, available by the token, at prices that defied physics (and economics).
This is Garrett Hardin’s tragedy of the commons, recast for the AI era. When a resource feels free, no individual user has any incentive to protect it. The rational move is to graze freely, because if you don’t, someone else will. Multiply that across millions of developers, product managers, and AI-curious executives, and you get exactly what we have now: a pasture stripped bare, dressed up as innovation.
The causal chain that nobody is naming out loud.
Subsidized costs created a false sense of abundance. False abundance produced poor usage choices, because there was no economic signal to encourage discipline. Poor usage choices, combined with shallow understanding of what AI actually does, produced bad architectures, because nobody was forced to design for cost, scale, latency, or reliability when the bill was being absorbed somewhere upstream. Bad architectures, combined with rushed implementation and unclear business intent, produced failed programs. The ones that demoed beautifully and then quietly died on the way to production. The real cost of AI was never the token price, whatever the Valley would have you believe. It was the compounding waste of every bad decision made in an environment that punished none of them, inside a system actively encouraged by the players who profit when consumption goes up.
Let me give you the image I use with clients to make this real. Imagine you own a fine-dining restaurant, and you've been handed a master chef, trained for decades, capable of dishes your team could not approximate in months of trying. The Food Network is picking up her tab, so you pay almost nothing for her time. And what do you do with her? You don't redesign your menu. You don't rethink your operation. You don't even ask what she's actually capable of. You put her in the kitchen and have her chop vegetables. All day. Every day. You hand her a knife and a pile of onions, congratulate yourself on how much faster the prep work is going, and tell the world your restaurant employs a master chef.
That is what most enterprise AI looks like today. A generational capability, applied to trivial tasks, inside processes that were never reimagined. The subsidies didn’t just distort prices, they distorted thinking. Teams stopped asking what AI was for and started measuring how much of it they were using. Tokens consumed became a proxy for progress. Architectures got lazier because they could. Why design for retrieval, caching, model selection, or workflow integration when you can throw the biggest model at the largest prompt and let the invoice be someone else’s problem?
It’s time to pay the piper.
The stories are everywhere now. AWS and Microsoft are pulling back on the code-generation capabilities they rolled out to their developers. Google, Anthropic, and OpenAI are imposing serious usage limits on the tiers people actually use. The frontier model providers, under pressure from public markets that have stopped clapping for revenue and started asking about profit, are quietly resetting the terms. And there is a deeper economic shift underneath all of it.
Rapidly changing technologies and surging compute demand have wreaked havoc with the amortization schedules that let organizations spread large capital spending across multiple years. The hardware you (or your Cloud Provider) bought to last five years is obsolete in twelve months. The VC-funded subsidies and the valuation premium that AI companies have been enjoying are unwinding at the same time. And the organizations that built their AI strategy on the assumption of permanent abundance are about to discover that their economics only worked when someone else was paying. This is not a crisis. It is the correction to what Alan Greenspan would have called irrational exuberance. And like most corrections, it will be brutal for the unprepared and a significant advantage for the disciplined.
Where do we go from here?
So how do you build this advantage of discipline, and what does it look like in practice?
It starts with AI fluency, and it has to run at every level of the organization.
At the executive level, fluency starts with knowing when AI is the wrong tool. Most “Faster, Better, Cheaper” problems don’t need AI. They need RPA, plain automation, or a process redesign that should have happened a decade ago. Using a frontier model to do work that a deterministic script could handle is using a sledgehammer to put in a picture hanger. AI earns its keep when it drives top-line growth, opens up work that was previously impossible, or augments human judgment in ways no rules engine ever could. That is where the economics work. Everywhere else, AI is an expensive hammer looking for a nail.
Once you know where AI actually belongs, the next layer of executive fluency is understanding what it truly costs to put it there. Not the line item on the cloud bill, but the full economic picture. The compute. The integration. The human supervision. The rework every time a model changes underneath you. The regulatory exposure. The operational risk when things go wrong. A CEO who cannot think about AI costs the way she thinks about labor costs or capital costs is going to make decisions that look great in the slide deck and brutal in the P&L two years from now.
At the practitioner level, fluency means knowing how AI actually works so you know what it should be used for. Language fluency is not just vocabulary. It is the feel for register, idiom, and where meaning collapses when you push a phrase out of its native context. A practitioner who understands context windows, retrieval, model strengths, and the difference between probabilistic generation and deterministic execution does not reach for a frontier model the way an executive reaches for a buzzword. She knows when an LLM is the right call, when a skill or agent should handle it, and when plain code is the cleaner answer. She knows that throwing language at a problem that needs logic is not innovation, it is mismatch. The engineer who calls a frontier model for what a regex could solve isn’t being innovative. She’s being expensive.
At every level, fluency means understanding the true potential and limitations of these systems. Not as autocomplete on steroids, but as cognitive partners in work that no human team could do alone. That is a different posture than the one most organizations have adopted. It requires people who can think with AI, not just type at it.
Good Architecture is Timeless
Fluency without architecture is still waste. The principles of good architecture are timeless, and they haven’t changed because the technology did.
Business architecture has to come first. The operating model has to be redesigned around what human-AI teams can actually do together, not retrofitted with copilots dropped on top of broken processes. Underneath that, technical architecture has to be architected to adapt as the business changes, designed to evolve as AI capabilities shift faster than anyone can predict, and engineered to operate at the cost, latency, and reliability that production work actually demands. Those three are not slogans. They are “North Stars” that should guide your AI implementation. If your architecture fails any one of them, you have a demo, not a system.
The organizations that come through this correction well won’t be the ones with the biggest models or the largest token budgets. They will be the ones who built fluency into their people, discipline into their architecture, and judgment into every decision about what AI is actually for.
ROI from AI doesn’t come from the technology. It never has. It comes from the discipline you bring to it. That was true in Mumbai in 1985, when every cycle counted. It is true now, when the cycles only felt free.
The piper is at the door. The leaders who prepared for this moment didn't prepare because they saw the subsidies ending. They prepared because they understood, all along, that good architecture is timeless and false abundance is borrowed time.