Why the Price of AI Is Rising Instead of Falling, and What That Does to Workforce Plans
J.P. Morgan Global Research estimates that DRAM prices will rise more than 400% between the start of 2024 and the end of 2026. Bloomberg reported in July that spot prices had already climbed close to 700% over the preceding twelve months. In the first quarter of this year alone, prices rose 80 to 90% over the quarter before.
That matters to anyone planning a workforce, because most of the plans we see rest on an unstated assumption. AI gets cheaper every year, chips improve, prices fall, and the expensive people doing knowledge work eventually become optional. It is a fair assumption. It is how computing has worked for most of a lifetime. The evidence right now points the other way.
Three things landed within a few days of each other this month, from people who were not talking to one another. One tracks semiconductors, one tracks company finances, one teaches.
The raw ingredient got expensive
The memory numbers above have a straightforward cause. Factories that made ordinary DRAM were switched over to the specialised stacked memory that sits beside AI accelerators in data centres, because that product earns far more per wafer. Data centres now consume something like 70% of all memory produced worldwide.
This is no longer read as a spike that resolves next quarter. SK Hynix, one of the three companies that dominate this market, has said the shortage could run past 2030.
The memory in your next laptop and the memory feeding AI models come off the same production lines, and AI outbid you. The physical inputs are getting scarcer and dearer, and the industry expects that to hold for years.
The economics got worse as the leader got bigger
The standard answer to rising input costs is scale. Grow large enough and unit costs come down.
The numbers coming out of the industry’s flagship show something else. According to Wall Street Journal reporting on OpenAI’s finances, quarterly losses grew by roughly $3 billion between the first and second quarter of this year, reaching about $12.3 billion. Revenue over the same period rose by about $1 billion, from roughly $5.7 billion to $6.7 billion, which is growth of around 18%.
Look at the shape rather than the size. Revenue is growing and losses are growing considerably faster, so the gap widens as the company gets bigger. That is the reverse of what scale is supposed to do, and it is happening while the company prepares a listing reportedly aimed at a valuation above a trillion dollars.
A caveat belongs here. These are private financials known through reporting, and published estimates of the full-year loss vary quite a bit. We are not forecasting an outcome for any particular company.
The narrower point is harder to argue with. I have not found a demonstration that this technology gets cheaper to deliver as it gets bigger. Until someone does, falling AI costs are a bet rather than a forecast.
The people building with it report the opposite of replacement
The third signal came from someone using these tools rather than selling or financing them.
Dr Philippa Hardman, a learning researcher, built an AI system to write individual feedback for about fifty learners on a course she runs. Roughly 200 pieces of feedback over four weeks. It worked, and learners rated the feedback among the best they had received.
Her account of why it worked is the useful part. The AI was the cheap half. The expensive half was getting twenty-five years of her own judgement, decisions she had been making automatically without ever putting them into words, into rules a system could follow. That extraction took considerably longer than the build. Her conclusion afterwards was that AI raises the value of deep domain expertise.
She also offers a distinction we think is the most useful idea to come out of any of this. We tend to split work into human parts and mechanical parts, assuming we already know which is which. She found she had guessed wrong in both directions. Writing feedback that felt personal turned out to be highly mechanisable, while working out what a particular learner needed did not. The dividing line runs between what you can decide in advance and what you can only decide with the specific case in front of you.
What to do with this
Three things follow if you run an organisation.
Stop modelling AI as a falling cost. If your three-year plan assumes the price per unit of AI work declines, that assumption has no support at the moment in either the hardware market or the financials of the largest provider. Model it flat, or model a range. Ask your vendors what happens to your pricing if their compute costs rise, and pay attention to which ones will not answer.
Treat documented expertise as the scarce resource rather than tools. Every organisation can buy the same models. What they cannot buy is a clear written account of how their own best people make decisions, and that account is the input these systems run on. Very few organisations have it written down. If you do one thing this quarter, capture how your strongest performers make their calls, before they retire, resign, or stay too busy to explain.
Test which parts of a job are specifiable rather than assuming. People who have built these systems report that intuition about what can be automated is wrong surprisingly often, in both directions. The way to find out is to try writing the rules down and see where you fail.
Where this leaves us
None of this says AI is overrated. The capability is real, and the feedback system described above worked at a scale a human could not sustain alone.
The narrower claim is the useful one. The belief that AI will become cheap enough to make expertise disposable looks weaker this month than it did last month. Costs are rising, the economics have not been proven at scale, and the practitioners closest to the work keep reporting that their expertise became more valuable, because it turned out to be the thing the machine needed most.
Plan for expensive AI and valuable people. On the current evidence that is the more defensible bet.