The cost of AI stopped being abstract on May 31. Not in one place — in three at once. The price of running a model, the price of the memory it runs on, and the price of the electricity it draws all got repriced and pushed onto the person using the model, on the same day, by four actors who never spoke to each other. A fifth story, published the same day, argued that the value all that cost is supposed to buy may be the hardest thing in economic history to measure. The meter started running before anyone finished building the dashboard.
Here is the day, laid flat. GitHub moves Copilot to token-usage billing effective June 1, ending flat-rate subscriptions and triggering what developers described as "massive cost increases." Memory chip makers are leveraging newfound pricing power to lock customers into multi-year contracts, restructuring around a physical shortage that runs years deep. US energy regulators are readying a June proposal to speed data-center connections to regional power grids. And SoftBank pledged up to €75 billion to build 3.1 gigawatts of AI computing capacity in France by 2031.
None of these is a story by itself. A billing change, a procurement trend, a regulatory docket, a capex pledge — each would scroll past on an ordinary day. The pattern is the story. Four different inputs to the same product, all being repriced in the same direction in the same week, by actors with no reason to coordinate. When that many uncoordinated parties move together, the explanation isn't a coincidence of news. It's a structural force moving underneath all of them.
The force is the end of the subsidy
Name it plainly: AI demand has outrun the era of subsidized inputs. For three years, every layer of the stack ran below cost on purpose. Inference was priced to win developers, not to clear. Memory was abundant and cheap. Electricity was a line item nobody at a model company thought about. The subsidy was never a feature of the technology — it was a feature of a market with more capacity than demand. May 31 is the day the subsidy stopped being abstract at every layer at once, because demand finally caught the capacity.
Watch how the same mechanism shows up in each story. The Copilot developer used to pay a flat $10 — a number that hid how much compute their requests actually burned. Now they pay per token, which means they pay the real cost, which a discussion of the change summarized bluntly: AI demand is such that it no longer makes sense to subsidize tokens. The memory buyer used to take spot prices for granted; now they sign a multi-year lock because the maker has the leverage to demand one. The ratepayer in a grid region near a new data center used to pay for their own consumption; now the regulator is fast-tracking the hookup that will load the grid. In every case the cost that used to be absorbed somewhere upstream is being pushed downstream to whoever touches the model.
A subsidy doesn't end when someone decides to end it. It ends when demand catches the capacity that made it free.
This is not a one-day spike, and the prior year of reporting bears it out. Contract prices for NAND chips were up more than 600% since September 2025, with DRAM up nearly 400% — a memory crunch that is precisely why a chipmaker now has the leverage to demand a multi-year contract. The shortage spread outward through the whole physical supply chain: by late May, surging demand was straining supplies of lasers, optical fiber, and connectors, not just the chips. The price of renting the compute moved the same way; model providers started raising prices 30% or more on coding plans. May 31 is not the start of this. It is the day three separate threads of it surfaced together, which is what made the force visible.
The cost is loud. The value is dark.
And then, on the same day, SemiAnalysis published the other half of the picture — the half that turns a procurement story into something stranger. It argued that "Dark Output," the economic value AI generates that is currently invisible to national statistics, may be one of the hardest measurement problems in history. Hold that against everything above. The cost of AI is loud: it arrives on a developer's invoice, in a memory supplier's contract, on a ratepayer's bill, in a regulator's docket. The value is dark: it does not show up in GDP, in productivity figures, or — for most firms — on the income statement.
This is the asymmetry that governs the moment, and it is not speculative. An MIT report found that 95% of enterprise GenAI pilots delivered little to no financial impact. An IBM survey of 2,000 CEOs found that just 25% of AI initiatives delivered the expected ROI — three out of four came up short. The cost is measurable to the cent. The return is, for most buyers, still a rumor.
On May 31 the cost side got three new instruments — a meter, a contract, a grid rule. The value side got an essay explaining why no instrument exists.
However: the J-curve is real
The strongest objection arrived the same day, and it is not weak. A piece argued that AI adoption is following the J-curve of general-purpose technology — like the electrification of American factories, which took years of capital outlay before any productivity gain showed up in the data. On this reading, the bears measuring "no ROI" are using instruments that simply can't see the output yet. The cost comes first because it always comes first; the value is dark because it is early, not because it is absent. Factory electricity looked like a cost sink for two decades before it rewired the entire economy.
This is the right counter-argument, and it does not break the pattern — it sharpens it. Concede the whole J-curve. Assume the value is real and merely lagging. The asymmetry still holds, and it is the entire point: the cost is legible now, and the value will not be legible for years. That gap is not a measurement error to be corrected. It is the structural condition of the next phase. The bull and the bear are not disagreeing about whether AI works. They are disagreeing about a dashboard that neither of them can read — while the meter, which both of them can read, keeps climbing.
The therefore
When the meter runs faster than the dashboard, the people paying the bill demand proof of return before proof is possible. That is the force that governs what comes next — not the capability of the models, but the widening distance between a cost that is legible to the cent and a value that is dark for years. Every actor on May 31 was responding to the cost side, because the cost side is the only side with numbers. The developer cancels the subscription, the regulator writes the rule, the sovereign wealth fund commits the €75 billion — all reading the same loud meter, none of them able to read the dashboard that would tell them whether it was worth it.
The subsidy era hid this. It let everyone build, adopt, and deploy without confronting the price, on the implicit promise that the value would arrive before the bill did. On May 31 the bill arrived first — at three layers, in one day — and the value stayed dark. The question for the next phase is not whether AI delivers. It is who keeps paying a meter they can read for an output they can't, and for how long. The reckoning was never going to be technological. It was always going to be the moment the meter got ahead of the dashboard, and everyone could see the gap.