The Midterms Didn't Reject AI. They Priced the Moat. That Is Good News, Even for the Billionaires.

On Tuesday, voters in ten states looked at 41 data center measures and said no to 40 of them. The same night they handed Democrats the House and the Senate, flipping seven Senate seats, including Texas. By Thursday, AI stocks were selling off and the word "bubble" was back in every headline.

Most of the AI world read that week as a loss. I read it as the best thing that has happened to AI products in two years. I'm going to make that case, including the part where it's good for the people who just lost money on it, and I'm going to concede every hard fact along the way, because a case that skips the hard facts is a pep talk.

The short version

• The 2026 midterms were not a vote against AI.
• They were a vote against the bill for AI: who pays for the power, what a town gets for hosting the building, and whether a state can say no. For two years the moat in AI was compute. Tuesday priced the moat.
• That moves the scarce resource from GPUs to proof of value, which is a product problem, and the jobs and the money follow it to where the customers are.
• The industry gets its terms two years before 2028, with a presidential veto capping the downside. Communities and states got a bargaining position they didn't have on Monday.
• The wealth part is bargaining power today and dollars later, and I'll say exactly which dollars I expect to move and when.

What the vote was, and wasn't

Start with what the ballots actually said. CNBC's count: 41 data center measures in ten states, and voters backed restrictions or rejected pro-development measures in every contest with results on Wednesday except Harwood, North Dakota. Ohio alone had 21. In Wilmington, four rezonings tied to data center projects lost roughly 80 to 20. In Wells, Maine, more than 80% approved a retroactive 180-day moratorium.

Now what the ballots didn't say. The AP Voter Poll found about a quarter of voters called AI and data centers crucial to their vote, behind cost of living and corruption. NBC's exit poll, asking a broader question, got 56% calling it a crucial or major factor. And about six in ten said AI will do more to hurt society than help it. But when CNN asked about a data center in the respondent's own community, opposition ran roughly three to one.

Read those together and the picture is specific. People are not against models. They're against a building that raises their electric bill, drinks their water, and leaves 300 jobs behind. Meta's $10 billion, four-million-square-foot campus in northeast Louisiana expects 300 to 500 permanent jobs after construction. That is the deal voters priced on Tuesday, and they priced it low.

The governors matter more than the Congress here. Gina Hinojosa unseated Greg Abbott in Texas, and Abbott had already turned: in June he told regulators to make data centers bear the cost of their own grid infrastructure, and in August he ordered a stop to new grid connections pending audits. Texas, Ohio, and Georgia now have governors publicly skeptical of data centers, and by one analyst's count those three states host nearly half the US AI data center buildout.

The moat got priced

For two years, the AI strategy conversation had one variable. How many GPUs can you get behind how much power, how fast. Everything else was downstream. I wrote earlier this year that the product budget is a compute budget now, and I meant it as a warning about where the money had gone.

Tuesday changed the variable. Compute now comes with a price the public sets: pay for your own power, leave something behind, and accept that a town can say no. Nomura's note the next morning said projects planned for 2028 face greater exposure to permitting delays and regulatory uncertainty. BMI warned a sharp slowdown in construction could produce a greater-than-expected decline in overall investment.

The market did the arithmetic in one session. Microsoft and Amazon each rose more than 6% on Wednesday. CoreWeave fell about 10%. JPMorgan called it a widening cost-of-capital gap between hyperscalers and neoclouds. CoreWeave, Nebius, and Iren had already fallen at least 30% in October as credit tightened.

So the moat didn't disappear. It got a price, and capacity already standing got more valuable, not less. That is the first thing the "AI lost" read gets wrong.

Why the billionaires should be relieved

This is the part I expect pushback on, so I'll be direct.

OpenAI's Chris Lehane has said AI will be the defining issue of the 2028 presidential election. Take him at his word. The industry had two paths to 2028. Path one: keep the preemption fight alive, keep building on the ratepayer's tab, and meet the public's anger for the first time in a presidential year, with no veto in the way and a movement that has spent two years practicing on 41 ballot measures. Path two: find out the terms now.

Tuesday chose path two for them. The Democrats' majority is real, and the bill list is already drafted: Reuters reports plans for bills on electricity costs tied to data centers, independent testing of advanced systems, and liability. But President Trump holds the veto, and Democrats would need a veto-proof majority for anything sweeping. Ed Mills at Raymond James made the sharper point: the probability of legislation is higher than people anticipate, through must-pass bills. That is a negotiation, not a wipeout. Every term gets tested against a president who wants the buildout and a public that wants the bill paid. Terms that survive that are terms the industry can live with in 2028.

And watch how fast the industry moved once the price was visible. On November 5, two days after the vote, OpenAI published a National Community Compact and a Compute Campus Policy Blueprint. Its own framing: how it builds these campuses is one of the first tests of whether the benefits of AI will be broadly shared, and it starts with who pays for power. In Pike County, Ohio, that meant a $6 million childcare commitment with SB Energy. Meta is reportedly building a bipartisan team. The president said Friday he would work with states that want to "become RICH" by embracing AI.

Two years of lobbying produced none of that. Forty ballot measures produced it in forty-eight hours. If you hold AI equity, that speed is the good news. The thing your valuation needs is adoption revenue, and adoption revenue needs communities that don't hate you.

The money the industry spent on the election is the cautionary half. Leading the Future, backed by Andreessen Horowitz, Greg Brockman, and Joe Lonsdale, started with about $100 million and ended with just over $20 million left. More than 86% of its backed candidates won or were leading, but the biggest general election bet, $4.8 million behind Darline Graham in South Carolina, lost to Annie Andrews. Molly White's Tech Influence Watch counted $79 million from AI industry super PACs across 68 House and Senate races. The lesson isn't that the money was wasted. It's that you can't buy your way past a utility bill.

Where the jobs actually come from

The jobs argument for data centers was always thin, and Tuesday called it. Three hundred permanent jobs for $10 billion is not an employment program.

The jobs are in adoption. Here's the mechanism. When compute was the moat, the question a buyer asked was "can I get access." When compute comes with a public price, the question becomes "did this deployment pay for its power." Nobody answers that question in a data center. It gets answered inside a hospital, a utility, an insurer, a bank, one workflow at a time, by someone who can take a model into a customer's process and make it stay there.

That work already has a name and a hiring curve. In the FDE series I cited Lightcast's count: more than 7,300 US forward deployed engineer postings in the first nine months of this year, against fewer than 200 in all of 2023. The FDE is the Product Builder wearing a visitor badge, the seat where product, engineering, and customer proximity collapse into one person. Every public software company that put FDEs in its earnings script this quarter did so because the customer, not the GPU, is now where value is decided.

And look at where those jobs sit. A compute campus concentrates a few hundred jobs on a farm site in a county that voted against it. Adoption work sits where the customers are, which is everywhere there is a payroll. Seventy-five percent of data centers under construction are in Republican-leaning states; the companies that have to make AI land are in all fifty. That is a better geographic distribution of AI employment than the one the compute boom was producing, and it happens without a single bill passing.

I've lived the gap between the two. At Commure, several health systems bought our platform to build their own applications on it. That was the pitch, and they paid for it. Then, for about two years, the applications didn't get built. Not for lack of money or infrastructure, both were in place. Nobody inside a hospital wakes up wanting to build software on someone else's platform. So we built the applications ourselves, put them in front of the clinicians, and that is when adoption of the platform took off. Same platform, same customers, same compute. The thing that was missing for two years was the work of making it land, and the moment we treated that as our job instead of the customer's, the asset we had already sold started producing value. That is the whole argument of this post in one company. The infrastructure is the cheap part to buy and the easy part to blame. The jobs, and the returns, are in the two years nobody owned.

The wealth part, honestly

I promised to say which dollars move and when, so here is the honest version.

What moved on Tuesday is bargaining power, not money. State preemption, the industry's top legislative goal for two years, is dead for the length of this Congress. Communities proved forty times that they can say no, which means every future siting conversation starts from "what do we get" instead of "please." States like Texas have already shown the template: make the large load pay its own interconnection and generation costs. That shifts cost off the ratepayer and onto the company, which is a transfer, and it needs no federal law.

What hasn't moved yet is the worker side. Mark Kelly's AI Horizon Fund is the serious proposal on the table: a federal trust fund financed by taxes on the largest AI and digital advertising companies, seeded with $30 billion to be repaid, paying for retraining, infrastructure, and unemployment benefits at 75% of lost wages for 26 weeks. It has zero cosponsors. It will not pass this Congress over a veto. I include it because it is the shape of the thing, and because the midterms made it a live question instead of a white paper.

So I will not tell you the midterms redistributed AI wealth. They redistributed the right to negotiate. In my experience that is the step that matters, because every dollar that moves later moves through a negotiation someone had to be allowed into first. Before Tuesday, a town negotiating with a hyperscaler had a zoning board and a prayer. Now it has forty precedents and a governor who ran on its side.

The wing that wants a pause

The case I've made assumes the new majority behaves like Zoe Lofgren, who is seeking the Science Committee chair and says Democrats will focus on responsible regulation while fostering innovation. Not everyone in the caucus wants that. Yassamin Ansari called on day one for a pause on advanced AI development and a ban on artificial superintelligence. Abdul El-Sayed, the new senator from Michigan, campaigned on a cabinet-level AI agency and partial public ownership of AI companies. Bernie Sanders wants a national data center moratorium. Greg Casar wants AI CEOs under oath.

I take that wing seriously, and I still don't think it sets policy in the next two years, for a reason that cuts both ways. The veto that protects the industry from a pause also protects the pause wing from having to own the result of one. Neither side can pass its maximal version. What passes is what both a Democratic Congress and a Republican president can sign, and that is a list that looks like ratepayer protection, independent testing, and liability rules. Boring, specific, and exactly the kind of thing a product team can build to.

The risk I do see is a crisis. Mills said it plainly: a crisis could build enough support to overcome presidential opposition. A grid failure traced to a data center, a model that hurts someone at scale. If that happens, the pause wing gets its moment, and the industry will wish it had negotiated harder in 2027. Which is one more reason to negotiate now.

Five predictions, dated, under my name

These are mine, not facts. Hold me to them.

1. By the end of 2027, at least ten states have large-load rules on the books that make data centers pay their own interconnection and generation costs, and no hyperscaler exits the US over them. The Texas template spreads because it is the one both parties can sign.

2. Hyperscaler capital spending still grows year over year in 2027. The buildout slows in 2028-dated projects, where Nomura says the exposure is, not in 2027 spend. Anyone calling the 2026 midterms the top of AI capex will be wrong on the calendar.

3. Lightcast's count of US forward deployed engineer postings doubles again in 2027. Adoption becomes the hiring story of AI, and deployment roles outgrow research roles in postings.

4. At least one debt-funded neocloud is absorbed by a hyperscaler or restructures its debt by the end of 2027. The K-shaped trade that opened on Wednesday closes with a transaction.

5. An industry-funded worker fund, the Horizon Fund or a successor, is in both parties' 2028 platforms. The Republican version will call it something else and route it through states. The mechanism, AI companies paying into a fund for displaced workers and grid costs, will be the same.

If I'm wrong on three of five, I'll write that post too.

One thing to try this week

Take your AI roadmap and sort it by cost per watt. Not by customer value, not by strategic priority. By how much compute each item burns to deliver its outcome. Then draw a line where a customer could reasonably ask "did this pay for its power" and you couldn't answer.

Everything below the line just became political, because somewhere a ratepayer is about to ask the same question with a ballot. Everything above it is where the jobs, the pricing power, and the next two years of product work live. If the whole roadmap is below the line, you have a compute strategy, not a product strategy, and Tuesday was the week that stopped being the same thing.

Part of the running argument on SaaS to AI Business Models: when the unit of value stops being the seat and the cost of delivering an outcome gets a public price, the margin structure moves, and so does the org that has to defend it.

Sources: CNBC, voters rebuff data centers in every ballot measure except one, November 4, 2026. Scientific American, the data center boom was on Tuesday ballots 41 times, it won once. TIME, how AI backlash shaped the 2026 midterms, November 4, 2026. CNN, how Democrats used anger over data centers to flip red states, November 7, 2026. AP Voter Poll analysis via KSAT, the midterm verdict on Trump's economy, November 5, 2026. CNBC, how the midterms could fuel the K-shaped AI trade, November 5, 2026. Council on Foreign Relations, the midterms rattled AI stocks. Euronews, data centre backlash helps Democrats win in Texas, Georgia and Ohio, November 4, 2026. Foreign Policy, the post-midterms AI reckoning, November 5, 2026. Reuters via US News, Democrats planning flood of AI bills, November 6, 2026. The Hill, Democrats eye Big Tech after midterm wins. CNBC, top Democrats call to rein in AI buildout, November 5, 2026. Fast Company, the midterms weren't the rebuke of AI money many expected. OpenAI, setting a new standard for responsible AI infrastructure, November 5, 2026. Sen. Mark Kelly, AI for America roadmap. Roll Call, tech taxes would drive Kelly's labor-focused AI Horizon Fund, September 25, 2026. CBS News, data center boom puts Republicans in a bind, with Cleanview Research's construction counts. Fortune, AI data centers as political flashpoints, with Meta's Louisiana jobs figure, October 22, 2025. Lightcast FDE posting counts as cited in Buying From an FDE Company.