Don't Blame the Democrats. The AI Build-Out Was Already Stuck.

gcwagner.com | November 2026 | G.C. Wagner

TL;DR: The Nasdaq lost $2.5 trillion in a day, and the headlines blame the Democrats' midterm win. I think that's the wrong culprit. The build-out was already slowed by the grid, by the price of memory, and by revenue growth that is still huge but no longer exponential. Politics just gave investors a headline for a doubt they already had.

The day the Nasdaq lost $2.5 trillion

On Thursday the Nasdaq fell 5.5% and lost an estimated $2.5tn. Nvidia slid more than 9%, and AMD and Micron each dropped more than 17%. The Philadelphia semiconductor index fell 12.2%, and the Nasdaq closed 11.9% below its all-time high.

The story on the tape was politics. Democrats won both houses of Congress. Texas governor-elect Gina Hinojosa pledged a moratorium on new data centres, and Senator Bernie Sanders posted "NATIONAL DATA CENTER MORATORIUM NOW".

But look at the order of events. Le Monde says the chip collapse was triggered by disappointing Intel results, which suggested demand was slowing. Investors had come to believe chips were immune to the cycle because of AI. Alphabet had already fallen about 12% in a day over its capex bill. One MarketWatch analyst, quoted by the BBC, called Thursday a "continuation of worries" about AI spending.

So politics lit the match. The fuel was already there.

The grid says no before the voters do

Here's the part most commentary skips. A data centre shell takes 18 to 24 months to build. But the wait for a grid connection now averages more than four years in primary US markets, and seven in Northern Virginia, according to JLL. In PJM, projects spend over three years just reaching an interconnection agreement, then another four waiting to come online.

Note that seven years is Northern Virginia, not the whole country. I'd rather be exact, because a sceptic will check.

The grid is not a hard stop, though. Developers are building their own power. Enverus forecasts 22.5 GW of on-site data centre demand by 2030, about 36% of expected US capacity additions. Grid delays are the main reason. So the grid doesn't kill data centres. It makes them slower and dearer, and that is bad for a story built on speed.

Is politics a real factor too? Yes, but it's the same story. Voters worry about congested grids and higher power bills. As of 4 November, 38 of 39 data centre restrictions on ballots had passed. Polls say seven in ten Americans don't want one in their community. Texas's construction pause was imposed by a Republican, outgoing governor Greg Abbott.

How big is the political risk? BMI estimates 53% of hyperscaler pipeline growth sits in states that flipped governors. But 2027 schedules are largely locked in, and a national moratorium looks unlikely. BMI also notes that "bring your own power" rules could push capex up, not down.

My view: "NIMBY" is just the grid and the power bill wearing a different coat. Anyone saying politics alone stops the build-out has missed the physics. Anyone saying politics is irrelevant has missed the ballot box.

The line went up. Then it got complicated.

The bull case was a chart. Weekly tokens on OpenRouter grew from 0.5 trillion in January 2025 to 128.4 trillion in mid-September 2026, about 250 times. Agents overtook humans as the main users on 6 February.

Since then, a plateau? Some say so, but I wouldn't bet on it. The thread claiming it rests on one week slightly below the last, and OpenRouter skews towards cost-conscious users who prefer open-weight models.

Cheap Chinese models are a real part of the story, but a specific one. They took 69% of OpenRouter tokens last month. Yet the three biggest open-weight Chinese labs are still only in the high single digits of spend. Cheap models take tokens, not dollars. That is price deflation, and deflation is awkward when you've borrowed to build for volume.

Anthropic shows the same pattern. Q3 revenue was $17.0bn, a record. But growth fell from about 144% quarter on quarter to 47%, price cuts took $600m off the quarter, and compute costs rose from 48% to 58% of revenue. Reuters also reports that growth slowed. That is still spectacular growth. It is just no longer the curve that $725bn of planned Big Tech spending was sized for.

Where is the return?

Outside software, the evidence is thin. A St. Louis Fed study of nearly 490,000 earnings calls found no measurable aggregate productivity bump after accounting for capital investment. About 95% of AI productivity comments describe future gains, not realised ones (Fortune's write-up). Only 7% of business leaders report a positive return, according to KPMG. The same piece notes that a UK Department for Business and Trade evaluation found no evidence of improved productivity.

Here is where I'll dial back my own thesis. The Dallas Fed finds productivity grew 3.7% a year in information, finance and professional services, against 1.7% elsewhere. Those sectors have 16% of hours but 40% of the gains. It's correlation, not proof, and it isn't only IT. Finance and professional services made the list. So AI seems to work where work is digital and checkable. That's a narrower market than the build-out assumes.

Do enough people pay? Consumer use has levelled off. US adult usage rose just three points, from 61% to 64%, while spending tripled to about $40bn. Estimates of who pays vary wildly, from 55% of AI users in one survey to 2.2% of households in PNC card data. Consumers are a thin base. The money is in agents and enterprise.

The memory question

I can't prove anyone cornered the market, and I won't say it. Here's what's documented.

In October 2025, OpenAI signed letters of intent with Samsung and SK Hynix targeting 900,000 DRAM wafer starts a month. Estimates put that at up to 40% of global output. Letters of intent are not contracts. Hyperscalers have also locked in supply under agreements that in some cases run for five years or longer, and J.P. Morgan estimates DRAM prices will rise more than 400% between early 2024 and the end of 2026. Meta raised its capex guidance by $10bn, citing higher hardware and memory costs.

Part of the capex growth is price, not volume. Scarcity looks like demand. When memory stocks led Thursday's fall, that is what the market was telling you. Micron fell more than 17%, and SK Hynix more than 10%.

Add the web of interlinked deals between chipmakers, labs and clouds, and you get a system where one party's spending is another's revenue. That is not fraud. It's a feedback loop, and loops run both ways.

What would prove me wrong

Capex guidance holds next quarter even as memory prices stabilise.
Enterprise revenue keeps compounding, so price cuts are offset by volume.
On-site power scales fast, so the grid stops being the constraint.
Congress does more than tweet. If it passes real limits, I was too relaxed about politics.

My takeaway

I don't think the data centre build-out is nonsense. Demand is real, and Anthropic's revenue proves it. I think the build-out was sized for a curve that is flattening, hampered by a grid that can't deliver on time, and inflated by the cost of scarce parts. Thursday's crash was politics on top of arithmetic.

If you want one test: next time someone blames the Democrats, ask how many megawatts were waiting for a grid connection in Northern Virginia.