Skip to main content

The Year Intelligence Gets Cheap

The Last Normal Year - Week 3

By Rick Aman
on

The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use. Moore's law changed the world at 2x every 18 months; this is unbelievably stronger. — Sam Altman, "Three Observations," February 2025

What a Thought Used to Cost

Let me put the AI cost concept into plain terms, because the whole argument of this week hides inside it. AI sells intelligence by the token, a fragment of a word. A million tokens is about 750,000 words, roughly five long novels in length, reasoned through, with questions answered. Price that cost per unit honestly, and you can watch the ground move under every organization in this series.

As the founding president of College of Eastern Idaho, analysis of this level was a budget line, and a serious one. A program demand study cost five figures and took a semester, and by the time the report landed, the labor market it described had already shifted. Consultants billed by the day. Boards rationed their questions because answers were expensive, and presidents guessed more than any of us admitted at the time.

This week I asked a machine for the first draft of an environmental scan of eastern Idaho aimed at a community college demographic, and it arrived before my coffee cooled and it cost less than the coffee.

This concept of declining costs of compute is the second force of the Last Normal Year, and it is the one leaders feel last, because it arrives dressed as good news. Intelligence is getting cheap. Clarity is not. Everything still ahead in this series, the delegation of work, the drawing of lines, the premium on judgment, runs on this curve. So let us look at the curve itself.

The Collapse of Compute, in Three Numbers

First number. Stanford's AI Index, the most careful annual count of this field, tracked the cost of querying a model at the performance level of the systems everyone met in late 2022. In November 2022, that work cost $20.00 per million tokens. By October 2024, it cost $0.07. The price of a fixed unit of machine intelligence fell more than 280 times in twenty-three months.

Hold that against the computing curve our generation planned by. Moore's law doubled computing power about every eighteen months, and it rebuilt the modern world at that pace. To fall 280-fold at Moore's pace would have taken roughly twelve years. This took less than two.

Second number. Epoch AI, which tracks the cheapest way to buy a fixed level of capability, found prices falling between 9 and 900 times per year depending on the task, with a median near 50 times per year. Performance at the level of GPT-4, the breakthrough system of 2023, fell from $37.50 per million tokens to $0.18 between March 2023 and February 2025. Read that as a board member. A capability that three years ago only the best-funded labs and companies could afford to use at scale now rents for pennies. Other trackers, the venture firm Andreessen Horowitz among them, put the sustained rate at roughly ten times cheaper per year for the same capability. So far, Altman's 10x rule has been the conservative estimate, not the aggressive one.

The third number may matter most because it helps explain the first two. In 2022, achieving a score above 60 percent on a widely used AI knowledge and reasoning test required an enormous AI model with 540 billion parameters, the internal values a model uses to process information and generate answers. The test, known as MMLU (Massive Multitask Language Understanding), measures how well AI answers questions across dozens of subjects, from history and science to mathematics and law. At the time, a model of that size required the computing power of a data center.

By 2024, a model of 3.8 billion parameters cleared the same bar. Intelligence is not only getting cheaper. It is getting smaller, which means it is getting local: onto the laptop, the phone, and soon the device on the student's desk and the trustee's kitchen table, no procurement process required and no permission asked.

Translate the seven cents into institutional terms. At that price, a college could have a machine read its entire catalog, every syllabus it offers, and ten years of board minutes, then summarize all of it, for less than what one hour of a consultant's time used to cost. The constraint on institutional self-knowledge is no longer the money. It is whether anyone thinks to ask, and whether anyone checks what comes back.

Every era gets one input whose price collapses: steam, electricity, computing, cellular airtime. Each time, the institutions that won were not simply the ones that bought the input. They were the ones that redesigned around it. Add these three numbers together and the conclusion is not merely that AI became affordable. It is that the economics beneath every knowledge institution have inverted. The scarce input around which, for a century, we built our budgets, our hierarchies, and an entire consulting industry is on its way to free.

What Scarcity Becomes

When a resource gets cheap, the value does not vanish. It moves. For a hundred years, our institutions were organized, reasonably, around the scarcity of analysis. Staff offices, research shops, and consulting firms existed because thinking at scale was expensive, and the executive's craft was deciding which few questions deserved the cost.

The rationing is now obsolete. The craft is not; it has been promoted. When answers are abundant and nearly free, the bottleneck of leadership moves upstream, to three things that are not getting cheaper. The first is the question. Peter Drucker warned that the important and difficult job is never to find the right answers; it is to find the right question. In a world of seven-cent answers, the quality of leadership increasingly depends on the quality of inquiry.

The second is verification. Cheap answers arrive fluent, confident, and occasionally wrong, and they arrive by the thousand. Someone has to check them against the record, and checking well is slow, disciplined human work. The third is judgment: deciding what the answers mean for this institution, these students, this mission. AI can make an answer cheap. It cannot make the Why cheap, and the Why still belongs to the board. And notice: none of the three, question, verification, or judgment, can be delegated without losing the thing itself. A board that outsources its questions has outsourced its leadership; however cheap the answers get.

There is a budgeting trap inside the good news, and boards with collaboration of CEOs should name it before their CFOs do. When the price of a unit falls, total use explodes. Economists call it Jevons paradox. Enterprise spending on LLM services rose from $3.5 billion to $8.4 billion in six months during 2025, even as the unit price was collapsing. Do not budget for savings from cheap intelligence. Budget for reallocation, because every department will find ten new uses for a resource that costs seven cents, and they will be right about most of them.

One caution, because clarity includes honesty about the curve itself. Epoch AI notes that the fastest price declines may not persist, and the frontier of capability keeps getting more expensive even as yesterday's frontier gets cheap. Institutions will live on both curves at once. Knowing which curve a given decision stands on is part of the new clarity.

Three Questions for the Boardroom

So put three questions on the agenda, in the boardroom and in the cabinet meeting alike.

First question. If analysis costs nothing, what work do we stop doing? Every institution carries reports nobody reads and studies commissioned to postpone a decision. Cheap intelligence exposes them, because neither defense holds anymore: the reports were excused by the cost of thinking, and the studies by the safety of delay.

Second question. Where are we still rationing thought out of habit? Policies, approval chains, and planning cycles built in the age of expensive answers now ration a resource that costs seven cents. Some of that rationing still protects the mission. Much of it is simply the past, governing the present.

Third question. Which of our past decisions were really purchases of scarce expertise, and what are they worth now? Contracts, positions, and partnerships bought for access to analysis deserve a fresh look, not because they were wrong, but because the scarcity they answered is ending.

The governing question of this series still sits above all three: does this move us toward our Preferred Future? The institutions that thrive after normal will not be the ones with the most intelligence. Intelligence is about to be free, and free things differentiate no one. They will be the institutions whose leaders stayed clear: clear about purpose, clear about their questions, and clear about the decisions that remain stubbornly, valuably human.

Next week: what happens when change arrives in every sector at the same time.

At Aman and Associates, we work with governing boards, CEOs, and executive teams to strengthen organizational focus and align resources with mission and impact. In a focused two-hour Zoom Futuring Session for your board or executive team, we help leaders step beyond traditional planning to define their Preferred Future and examine what should continue, what should change, and what may have completed its useful life.

Rick Aman, PhD | Aman and Associates

rick@rickaman.com | rickaman.com/articles