The Productivity Mirage Hiding Behind America's AI Buildout

On 25 June 2026 at 16:15 UTC, the US Bureau of Economic Analysis published its third estimate of first-quarter GDP and quietly improved the number to a 2.1% annualised rate — a sharp upward revision that Washington treated as vindication. Two other data points landed the same week and told a less flattering story: a forecast that US electricity prices will rise roughly 6% annually as AI data centres strain the grid, and a new study finding that LLM-based trading strategies mostly failed to outperform a simple buy-and-hold approach over twenty years of backtesting. Read together, the three numbers describe an economy in which the macro headline is being propped up by capital expenditure on artificial intelligence whose productivity payoff, at the operational level, remains unproven.

The thesis is uncomfortable but defensible: the AI buildout is currently a cost centre disguised as a growth engine. Headline GDP captures the brute spending on chips, land, turbines and transformers. It does not yet capture a productivity dividend at the firm or household level that justifies the strain on power systems or the rerouting of capital away from other sectors. The market's exuberance is rational only if one assumes that the spend translates into output that the data has not yet shown.

The macro headline is real, and so is what it hides

A 2.1% Q1 print is not a recession number. It is, however, a number that owes a meaningful share of its composition to a narrow band of investment: hyperscale data centres, accelerator chips, and the power infrastructure to feed them. When capital expenditure on a single category rises fast enough to bend quarterly GDP, the macro line and the microeconomic reality can diverge sharply. The Q1 revision rewards the spending; it says nothing about whether the systems being built are producing returns commensurate with the resources being poured into them.

That distinction matters for monetary policy, for fiscal forecasting, and for the politics of who pays the bill. GDP is a flow; the AI buildout is a balance-sheet commitment that runs for a decade. The 2.1% revision celebrates the front end of that commitment while leaving the back end in shadow.

The 6% electricity number is the tell

The forecast of 6% annual increases in US electricity prices is the most politically combustible of the three figures, and the most economically honest. Data centres are industrial loads with unusual characteristics: they run around the clock, they tolerate almost no interruption, and their appetite for power scales with model complexity rather than with revenue. When industrial buyers of that scale move into a regional grid, the marginal cost of generation, transmission and balancing reserves rises for everyone else.

This is not an ideological claim about AI. It is the standard incidence of a large new baseload buyer entering a regulated market. The 6% figure functions as a quiet transfer: hyperscale operators and their cloud customers capture the upside of the compute buildout, while households and small businesses absorb the rate increases required to finance the grid expansion. The headline GDP number credits the spend. The electricity forecast tells readers who actually pays for it.

The trading-study result punctures the productivity narrative

The third data point is the most damaging to the case that AI is already transforming the productive economy. A new study found that LLM-based trading strategies mostly failed to outperform a simple buy-and-hold strategy over a twenty-year backtest. That is a narrow result — it concerns a single application in a single market — but it lands at the centre of the productivity story. Equity trading is exactly the kind of information-rich, fast-moving environment where machine learning was supposed to deliver a durable edge. If large language models cannot beat the cheapest passive strategy across two decades, the case that they are quietly rewriting the productivity frontier across the broader economy needs to be made more carefully than its boosters have made it so far.

The honest reading is not that AI is useless. It is that the gap between the macro story (record capex, rising GDP, soaring valuations) and the micro story (no clear edge in the most quant-friendly market in the world) is wide, and that the gap is being financed by ratepayers and by capital allocators who have decided to wait for the productivity story to materialise rather than demand evidence of it.

Counterpoint, and what the framing gets right

There is a serious defence of the current trajectory. Compute demand may be creating the conditions for breakthroughs that the next twenty-year backtest will capture. The productivity gains, on this read, are latent in the infrastructure itself and will appear as the cost of inference falls and as new use cases compound. Electricity prices, the argument runs, are rising because the grid was underbuilt for a generation and is finally catching up to demand; the alternative was stagnation.

That defence holds in the long run. It does not hold yet. The sources do not specify which sectors will absorb the productivity dividend, how quickly inference costs will fall, or how rate increases will be distributed across income groups. Until those questions have firmer answers, the 2.1% revision is a description of effort, not of payoff.

Stakes

If the AI buildout delivers the productivity its boosters promise, the 6% electricity rises and the underwhelming trading results will look like the costs of a successful industrial transition. If it does not, the same numbers will read as a transfer of wealth from households to infrastructure operators and a misallocation of capital at scale. The window for telling those stories apart is not infinite, but it has not closed yet.

This piece ran in the opinion column. Monexus frames the AI buildout as a productivity bet whose macro numbers are visible and whose payoffs are not, rather than as either miracle or bubble.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://x.com/polymarket/status/
• https://x.com/polymarket/status/
• https://x.com/polymarket/status/