What the Democratic Sweep Means for AI: Every Data-Center Ballot Measure Counted, Why Texas Is the Fear, What Slows, What Could Actually Collapse, What Washington Can and Can't Do, and Why a Slower Buildout Is Probably Good for AI

This piece is what the results change, and what they don't, written the morning after from the primary record: county results pages for every measure, ERCOT's own queue filings, Abbott's and Hinojosa's own words, the PJM market monitor, the hyperscalers' third-quarter releases, and the analyst notes and posts we could verify. Two corrections to the version circulating on X come first, because they matter: nothing like '41 of 41 initiatives failed' happened, and Texas was already pausing before anyone voted. Then the argument we actually want to make, labeled as ours: a slower buildout is probably good for AI, because the thing being slowed was running ahead of the turbines, the grid, the revenue, and the public's consent, and each of those was going to stop it anyway, later and harder.

The short version

• What flipped: seven Senate seats for a 54–44 majority, at least 18 House seats, four to five governorships (Texas, Ohio, Georgia, Iowa, likely Nevada), and five to seven state chambers. Democrats hold a majority of governorships for the first time since 2010. Kansas's governor, Texas's House, and about 75 US House seats were uncalled Wednesday morning.
• The ballot board, corrected: Ballotpedia counted 41 local data-center measures; the Data Centers on the Ballot tracker counts 45. Of the 40 with results, 39 went the anti-data-center way, pooled at 72% of 234,341 votes. All 13 measures that would have approved or invited data centers lost. The one pro-development result was Harwood, North Dakota, where a review requirement for an already-approved campus failed 105–325. Five were uncounted.
• Why Texas: it holds about 27% of the planned US pipeline (Wells Fargo; the dcmap tracker agrees) and nearly 30% of the capacity due in 2028 (SemiAnalysis). ERCOT's queue of large-load requests went from 220 GW in December to 474 GW in June, 90% data centers, against a record peak of 91 GW. Abbott froze grid approvals in August and air and water permits in September. Hinojosa inherits the pause and wants it written into law.
• What slows: permits for anything not yet permitted, in Texas, Ohio and Georgia first. Jefferies and GLJ put about 13 GW of Texas projects in the exposed lane; Altimeter's Brad Gerstner guessed 10–15 GW of the 43 GW expected nationally in 2027 won't arrive on time; Wedbush's Dan Ives estimates 10% of the 800–1,000 planned US sites 'will not happen now.'
• What could break: not the hyperscalers, whose 2026 capital spending is near $800 billion and continues; the leveraged layer underneath them. CoreWeave's 2031 bond yielded 15.8% three weeks ago. Fermi America's 17-gigawatt Amarillo plan needed a lender's waiver in October. The IMF counts $1.4 trillion of circular hyperscaler-to-startup financing. A pause that strands a lease those borrowers were counting on is where 'slowdown' becomes 'collapse.'
• What Washington can do: hearings, subpoenas, mandatory ratepayer protections attached to must-pass bills, and burying federal preemption of state AI laws. What it can't: a national moratorium, a federal AI statute before 2029 (a veto and a 60-vote Senate), or anything about ERCOT, which sits outside federal jurisdiction. Texas decides Texas, in a session that opens January 12.
• Our thesis: this is probably good for AI. The constraint on American AI was never ambition. It was gas turbines that ship in 2031, a grid short of its reliability target, a $4.2 trillion gap between what the data centers cost and what AI earns, and a public that now opposes the buildings three to one. A build that respects those limits produces models and products that have to be efficient, priced honestly, and wanted. The counter-case, China, is in the text.

What actually happened

The scale is not in dispute. Ballotpedia's overnight tally put Democrats at a 231–141 House majority with 14 Republican incumbents defeated and none of their own, a 54–44 Senate, four governorships flipped, five state chambers flipped and none lost. NBC's estimate ran higher, 'at least 258' House seats. CBS projected the House at 10:15pm Eastern and the Senate at 11:35pm; the Texas Senate race was called for Talarico at 10:55pm and the governorship for Hinojosa at 11:10pm, with Abbott conceding that 'mathematically, we are not going to be able to win.' Hinojosa is the first Democrat to win the Texas governorship since Ann Richards in 1990 and the first Latina to hold it; Talarico is the first Democrat Texas has sent to the Senate since 1988. Democrats also took the lieutenant governorship, the attorney general's office, and a seat on the Railroad Commission, which regulates oil and gas, for the first time since 1990.

'Sweep' is the word everyone is using and it is fair for Congress and broadly fair for the states, with the caveats that Florida kept a Republican governor, that the Texas Senate remains 19–12 Republican, and that control of the Texas House was still open Wednesday, with Democrats needing 14 flips, six called and about a dozen more leading. That last detail matters more for data centers than any federal result, and we come back to it.

The money tells you how central the issue had become. Time's count: Democrats spent more than $100 million on ads tying Republicans to Big Tech and data centers; AdImpact tracked $173 million of data-center advertising, almost all of it negative; Musk was the cycle's largest donor at $446 million, with roughly $190 million of it in Texas by Semafor's count, backing Ken Paxton. Data Center Watch's pre-election survey of 950 candidates found 66% had taken a position on data centers, and that 32% of Democrats versus 3% of Republicans backed bans or moratoriums. Amundi's Aidan Yao, quoted by Reuters, called the night 'more like a referendum on AI than a referendum on Trump'. The exit poll gives that a number.

The ballot measures, measure by measure, and the '41 of 41' myth

The claim going around is that 41 out of 41 data-center initiatives failed. No one who counted says that, and it gets the story backwards. The 41 is Ballotpedia's pre-election count of local measures on the November 3 ballot in ten states: 28 where a yes vote would restrict data centers and 13 where a yes would approve or invite them. Most of the 'initiatives' were the restrictions, and they passed. The Data Centers on the Ballot tracker, which pulls from county results pages, had 45 measures with 40 counted at 6am Eastern: 39 leaning against data centers, one for. Semafor's J.D. Capelouto, whose live thread was the night's most-shared tally at 742,000 views, counted about 40 localities across 13 states. Jim Bianco's version, '29 to 1,' and Heise's '37 of 41' are the same board read at different hours.

What the board shows is not that 41 pro-AI initiatives lost. It is that a template, Conserve Ohio's 25-megawatt charter ban, worked everywhere it was tried, and that the rare measures asking voters to say yes to a data center lost by three or four to one. Some specifics, all from county results pages unless noted:

• Trenton, Ohio passed a ban on data centers over 25 MW 3,541 to 903, about 80%, blocking the annexation of roughly 600 acres for an 18-building Amazon campus. The petition's leader, Barry Blankenship: 'We stood together and we beat a city.'
• Sunbury, Ohio banned data centers over 25 MW 2,809 to 763, on a site Amazon had planned a $2 billion campus for. Pataskala (74%), Granville (82%), Urbana (77%, with Thor Equities already suing the city), Hubbard (77%), Defiance (65%), Upper Sandusky (77%), and Conneaut (65%) passed the same bans. Ohio had at least 14 municipalities pass restrictions.
• Ashville, Ohio voted 82% to overturn the village's development agreement with EdgeConneX for a campus that is already under construction. Wilmington, Ohio rejected all four rezonings for a data-center site at about 80% each. Piqua, Ohio recalled its mayor and two commissioners over a data-center deal, each by more than 70%, and passed a ban.
• Emporia, Kansas passed a ten-year ban on 'high-impact' data centers 4,346 to 2,891, ending a 1,000-acre gigawatt-scale proposal. El Dorado (71%) and Ottawa (72%) did the same.
• Bloomington, Illinois backed a moratorium 21,851 to 6,316, the largest vote on the board. Four Madison County townships rejected advisory questions on permitting data centers, Chouteau Township by 83%. Joliet Township voted 79% against a campus the city council had already approved 8–1.
• Boulder City, Nevada refused 70–30 to open 14,000 acres to data centers, the only measure on the board that would have created new capacity. Edgewater, Florida wrote a ban into its charter at 77%; Alhambra, California at 82%; Janesville, Wisconsin now requires a public vote for any project over $450 million at the old GM site, where an 800 MW campus was planned.
• Harwood, North Dakota, the exception: a charter amendment requiring independent review before approving industrial projects over 100 MW failed 105–325. Applied Digital's $3 billion campus there was already approved.

Add the six local measures decided earlier in 2026, all restrictive, and Ballotpedia's full-year count is 50 local data-center votes in 13 states with one pro-development result. That is the number to use. The trend behind it: Data Center Watch's tally of projects blocked or delayed by local opposition went from $64 billion in its first report to $156 billion for all of 2025, then $130 billion in the first quarter of 2026 alone and $68 billion in the second, with opposition groups growing from 142 to 843. SemiAnalysis mapped more than 300 enacted local moratoriums in September and made a point worth holding onto: about 20 GW sat inside restricted boundaries, but only about 2.3 GW was 'genuinely delayed,' because developers route around towns. Tuesday changed the arithmetic of routing around, because the next three governors of the states they were routing to ran on stopping them.

Why everyone is pointing at Texas

Three reasons, in order of size. First, concentration: Texas has more data centers under development than any state, having passed Virginia this year; JLL counts 26 GW existing and under construction against Virginia's 13. Wells Fargo puts Texas at 27% of planned US data-center gigawatts; SemiAnalysis says close to 30% of the US capacity due in 2028, with 'a few months of lost time in that market' able to 'move the national number.' That is why Wells Fargo's Ohsung Kwon called a Democratic win in the governor's race 'a left-tail event to the AI trade,' why JPMorgan's Ilan Benhamou called it 'the only thing that matters,' and why Dan Ives called it the 'biggest worry' of the night. Stargate's Abilene site, Vantage's 1.4 GW Frontier campus, Meta's $10 billion El Paso build, Amazon's 7.65 GW gas-powered GW Ranch in Pecos County, Google's up-to-$40 billion Texas plan: the flagship projects of the buildout are disproportionately here.

Second, the grid is an island with a queue that stopped making sense. ERCOT's interconnection requests from large loads went from more than 220 GW in December to about 410 GW by the end of February, 438 GW in May, and 474 GW by June, about 90% of it data centers, from more than 1,800 projects. Texas's all-time peak demand, set on July 22, is 91 GW. ERCOT's ties to neighboring grids total about 1.2 GW, so unlike Virginia or Ohio, Texas cannot import its way out of a mistake; every gigawatt of real new load is built, paid for and backed up inside the state. ERCOT's own chief executive has said more than half the queue 'hasn't begun any process at all.' BloombergNEF's forecast of data-center load that actually materializes by 2030 is 17.2 GW.

Third, and this is the part the market reaction skipped: the pause was already on. Abbott, a Republican who signed the state's large-load law in 2025, ordered an audit of every queued data center on August 3, writing that 474 GW was 'more than five times Texas' record peak' and that 'unprecedented load growth could endanger the reliability and stability of the Texas electric grid.' On September 21 he directed the environmental agency to issue no permits sought by data-center projects until the audit was done, and said he would seek repeal of the industry's tax incentives in the 2027 session. Hinojosa called that an election-year stunt and said he 'has the power to do it today.' Her platform, restated Wednesday, is a moratorium on new permits until the Legislature acts, making data centers pay sales tax on their electricity (an exemption she says costs more than $1 billion a year; the Tribune's figure is $3.2 billion over two years), voiding water contracts that deny residents water, and new PUC appointees. Her words: 'These hyperscale data centers are having hyperscale impacts on our communities, and we're just left unprotected.'

So the Texas question is not whether the state slows down. It did in August. It is whether the slowdown is a six-month audit that ends with stricter fees and a pay-your-own-way rule, which is where Jefferies thinks the 2027 session lands ('more likely to increase development costs and requirements than materially restrict the Texas pipeline'), or a multi-year freeze on anything without a permit. GLJ Research's Gordon Johnson, quoted by Bloomberg, drew the lanes we use below: 'Projects with power, permits and steel keep going'; projects with power but no environmental permit slip 12 to 18 months; projects with neither 'probably go to Louisiana, Oklahoma or nowhere.' Bitcoin-miner-turned-host Cipher Mining fell 17% Wednesday on a Cantor note saying the result 'materially increases the likelihood that Cipher's next data-center lease will be signed outside Texas.' That is the mechanism in one sentence: the marginal gigawatt moves, and the marginal borrower who needed it in Texas is the one at risk.

Voters in the state were ahead of both parties. The University of Houston's Hobby School found in late October that 79% of likely Texas voters wanted limits on data centers and 99% wanted data centers to pay for grid expansion. The Texas Politics Project found 57% opposed a data center in their own community, including 60% of rural voters and 43% of Republicans. Agriculture Commissioner Sid Miller, a Republican, blamed Abbott on election night for ignoring Texans on data centers: 'Republicans quit listening to the people.'

What slows, and what could actually collapse

Start with what does not change. Wood Mackenzie's third-quarter count has about 80 GW of US data-center capacity under construction, nearly double the spring figure, with more than 40% of ERCOT's development capacity already being built. Every moratorium on the board, Hinojosa's included, is written for new permits; her campaign was explicit that projects under construction continue. The hyperscalers' spending does not pause either: Amazon guided $220 billion for 2026 and 'meaningfully more' for 2027; Alphabet $195–205 billion, rising 'significantly' next year; Microsoft about $170 billion for the calendar year; Meta $135–145 billion; Oracle $90–95 billion gross. Fortune's sum is close to $800 billion this year and about a trillion a year through 2030. Bernstein's note the day before the vote called the political fears 'overblown' while conceding 244 GW, 56% of the pipeline it tracks, was 'exposed to political risk.'

What slows is the next permit. Jefferies and GLJ both put about 13 GW of announced Texas capacity in the exposed lane. SemiAnalysis's pre-election scenario, 'pro-moratorium wins everywhere,' which is roughly what happened, exposes 'roughly 30 GW of planned capacity scheduled for 2028' out of 62 GW of US net additions that year, 'most of which hasn't started construction.' On the All-In election livestream, Altimeter's Brad Gerstner said the consensus for 2027 was 43 GW coming online and that 'with this blue wave I think it will slow it down even more. I think we can see 10-15GW of the expected build not come online', while calling the Texas fear 'overblown' in the same breath. Dan Ives: 'We estimate 10% of datacenters will not happen now' of the 800 to 1,000 planned. Those three estimates, from a short seller's skeptic, a bull and a cheerleader, land in the same place: a fifth to a third of 2027–28 capacity slips, and a tenth never happens.

The collapse scenario is specific and it is not about Microsoft. It is about the layer that borrowed against gigawatts it had not built. CoreWeave's 9.75% notes due 2031 yielded 15.76% on October 14, the day the stock fell 10%; it has raised $14.2 billion this year in loans secured by its GPUs, with five-year maturities against customer contracts of about three. Fermi America, the Amarillo developer that pitched the 'world's largest' campus at 11 then 17 GW, lost its anchor tenant in December, its chief executive in April, and on October 16 needed MUFG to waive a covenant requiring a 400 MW lease by November 10 to avoid accelerated repayment. Poolside's 2 GW Fort Stockton project lost its CoreWeave anchor in the spring. Oracle and OpenAI dropped a 600 MW expansion at Abilene in March after financing talks dragged. Oracle itself carries $125 billion of debt and ran negative free cash flow last quarter; Alphabet's and Meta's free cash flow also went negative in the third quarter. And the IMF's October stability report counted about $1.4 trillion of 'circular' deals, hyperscaler money flowing as equity and debt into weaker AI firms, with 'wrong-way risk' if those firms fail. None of those facts was created on Tuesday. What Tuesday created is time: a project that must slip 12 to 18 months is a delay for Amazon and a default for a borrower whose loan assumed the lease.

What Washington can and can't do

The incoming House leadership has said what it wants. Hakeem Jeffries, who will be Speaker, told USA Today in October that AI is a 'safety crisis, not a hoax', that 'we need guardrails that have to be put into place immediately,' and, on a select committee with subpoena power, that 'we haven't ruled anything in, and we haven't ruled anything out.' On moratoriums he was careful: 'That's a state-by-state decision.' Frank Pallone, in line to chair Energy and Commerce, told E&E in July that 'fixing this data center crisis is a top priority of mine' and that if guardrails fail, 'the possibility of a moratorium out there is certainly real'; his September letter to FERC blamed data centers for 'nearly $30 billion' in PJM capacity costs. Chuck Schumer called the House's 417–3 Ratepayer Protection Act 'toothless' because its standards are optional, and promised 'strong, enforceable consumer protections' and, 'one of the very first things,' restoring the clean-energy tax credits. Greg Casar wants AI executives to 'testify under oath.' Robert Garcia, the incoming Oversight chair, posted two words: 'Lawyer up.'

What they cannot do is pass an AI statute over a veto, and ppc.land's day-after read, that 'federal privacy or AI preemption is unlikely before 2029', is the consensus. That cuts both ways. Industry's main federal ask, preemption of state AI laws, is dead for this Congress: the Senate killed the 2025 moratorium 99–1, the June draft bill never moved, and Trump's December executive order creating an 'AI Litigation Task Force' has yet to see a court preempt a single state law. The Sanders and Ocasio-Cortez bill to halt every data center over 20 MW is equally dead; even Goldman's strategists, who expect a Democratic Congress to curb tax incentives, call a national moratorium unlikely. The real federal action is in the plumbing: FERC accepted PJM's large-load registry in October but rejected curtailment-first rules, with Commissioner Rosner writing that 'homes, hospitals, and schools should never be curtailed before data centers.' Expect the ratepayer fight to attach itself to must-pass bills, and expect hearings to produce the documents, on water, power and emissions, that companies have declined to publish.

The electricity bill is the reason this was a voting issue and not a tech-press issue. PJM's capacity auction, the price generators are paid to be available across the 13 states from Virginia to Illinois, cleared at $28.92 per megawatt-day for 2024/25, $269.92 the next year, and then at the cap three times: $329.17, $333.44 and $325. The market monitor attributes 65.5% of the 2027/28 auction's $16.4 billion to data-center load, and $23.1 billion across three auctions. Capacity is only a tenth to a fifth of a bill, and FactCheck.org's October review found the studies messier than the ads (MIT's estimate of data centers' effect on retail rates to date is 2.7%; Columbia's Douglas Arent: 'the politics has certainly grown much larger than the data'). But Ohio residential bills are up more than 20% in two years, the NRSC's own August memo called data centers 'the anchor hanging around Husted's neck,' and Sherrod Brown won by five points. In Texas, where residential prices rose a modest 3% year on year, the issue still polled at 79%. Voters were not doing regression analysis. They were looking at a 474 GW queue and a bill.

Our thesis: this is probably good for AI, because it was moving too fast

Nobody prominent made this case on election night, so we will. Investors called the result a 'left-tail event' and a 'seismic shift'; the industry's loudest voice proposed orbit. The one near-exception was Katie Koch, chief executive of the bond manager TCW, who said companies 'do have to have the social license to operate' and 'need to define what the social utility is going to be of AI and data centers to Americans.' We think she is right, and that the slowdown voters just ordered removes more risk from AI than it adds. Five reasons, each with a number.

The build was ahead of the money. Bain's September estimate is that AI needs $6 trillion a year of revenue by 2031 to justify the data centers being built, against a forecast of $1.8 trillion, a $4.2 trillion gap. Alphabet, Meta and Oracle all reported negative free cash flow last quarter while guiding capital spending higher. Ken Griffin says AI is 'almost certainly' in a bubble; RSM's Joe Brusuelas says it 'looks and feels and smells like' one. A build constrained by permits grows at the rate the revenue can carry, which is the only rate that does not end in the fragile layer defaulting. Ed Yardeni put it in the terms a bull can accept: the financial, physical and political constraints are 'healthy for the durability of the AI boom.'

The build was ahead of the hardware. A gas turbine ordered from GE Vernova today ships in 2031; its backlog and reservations reached 122 GW last quarter. Combined-cycle lead times went from three and a half years to about five. ERCOT's generation queue has 77 GW of gas in it and every one of the 20 projects that changed its date this year moved it later. In other words, the 474 GW was never going to be powered on the timeline the queue implied, with or without an election. A political pause on permits for the unpowered tail costs the industry almost nothing in real delivered gigawatts; it costs a great deal in paper gigawatts, which is what the stock prices were pricing.

The build was ahead of the grid. PJM is 6,831 MW short of its reliability requirement and its chief executive says demand is growing 'faster than electricity supply.' ERCOT's own planners built a rule this year that counts a project only once it has signed an interconnection agreement, precisely because the queue had become fiction. Grid operators were already triaging; Tuesday gave the triage a mandate. A buildout paced to transmission and generation is the one that does not get curtailed in a heat wave, which is the outcome that would actually turn the public against AI rather than against its buildings.

The build was ahead of consent. Seventy-six percent of voters would oppose a data center near them; USC's survey found 71% do not trust Big Tech 'at all' on data centers and 78% think government does too little about them. That is not a communications problem. It is what happens when an industry asks communities to host its heaviest infrastructure while offering tax abatements to itself and higher bills to them. Every winning measure on the board is a town saying it wants a vote, a bond, a noise limit, or a water cap. Those are the terms of a social license, and an industry that negotiates them builds things that last. One that litigates them, as Thor Equities is doing against Urbana, builds a decade of referendums.

And slower is better for the products. This is the part that matters to readers of a site that ranks AI apps. The models that improved most this year did it on efficiency, not raw scale: Qwen3.8-27B matches what frontier models did eighteen months ago on an 18GB download; Gemma 4's 26B model activates 3.8 billion parameters per token; the best coding agents are winning on tool use and harnesses, not on parameter count. Compute scarcity pushes labs toward distillation, caching, smarter routing, and smaller models that run on a Mac mini, and it pushes prices toward honesty: the $20 subscription that loses money on heavy users is a product of abundant capital, not abundant intelligence. Anthropic's own S-1, filed a week before the vote, lists 'failure to secure access to adequate electrical power' and organized local opposition among its risk factors. A company that must plan around those risks plans better. The AI products of 2027 will be shaped by what the labs can serve, and a constraint that forces 'serve it well' over 'serve more of it' is one we would vote for.

The counter-case deserves its full weight. China's grid added 253 GW of generating capacity in the first nine months of 2026; the US pipeline's entire under-construction total is 80 GW. Speaker Johnson's line was 'we will lose the race to China'; Trump called opposition to data centers 'a sick conspiracy'; Musk's version is that 'the limiting factor for AI deployment is fundamentally electrical power' and China has it. All true as far as it goes. The answer is that the United States constraint was never the willingness to build. It was turbines, transformers, interconnection studies, and a public that pays retail for wholesale decisions. Stanford's AI Index counts 5,427 AI data centers in the US to China's 449; Rystad puts Chinese data-center utilization at 20 to 30%. Overbuilding faster than you can power or fill is not a lead. If the race is decided by electricity, the policies that win it are the ones that add generation and transmission, which is what the ratepayer and clean-energy fights in the new Congress are about, not the ones that approve a 474 GW queue against a 91 GW grid.

What it means for the AI apps you actually use

• Prices and limits, not features. The hyperscalers' capex continues, so the frontier models keep shipping. The squeeze lands on the compute the labs rent from the leveraged layer: Microsoft already blamed a capex shortfall on 'third-party capacity deliveries we expected in Q1 were delayed.' Expect tighter usage caps and fewer 'unlimited' tiers before you see fewer models. Our chatbots and code editors rankings score usage limits and price transparency for this reason.
• Agents are the first casualty of scarce compute. Always-on agents burn tokens all day. If serving capacity slips 10 to 15 GW in 2027, the labs ration the products with the worst tokens-per-dollar first. Watch the weekly pools and the 'fair use' clauses.
• Local and open models get a tailwind. Every constraint above argues for inference you own. The open-weight field is good enough now that a $899 Mac runs a model that would have been frontier in 2025; our guide has the machines and the commands.
• Texas-hosted products carry Texas risk. Grok and SpaceX's AI stack, Oracle's cloud, OpenAI's Stargate capacity, CoreWeave's Texas sites: anything that planned to scale on Texas gigawatts in 2027–28 now has a permitting variable and a January legislative session in its roadmap.
• The Anthropic IPO prices in regulation. The S-1's risk factors on power and 'organized and coordinated' opposition were written for exactly this week. Read our explainer on what the filing says.

What people are saying, verified

Posts below were fetched and checked on Wednesday; metrics are as of fetch time. Several industry leaders, including Sam Altman, Dario Amodei, Jensen Huang, Satya Nadella, Mark Zuckerberg, Marc Andreessen and David Sacks, had not posted about the results by midday.

Posted at 1:12am Eastern, 5.5 million views and 140,000 likes by midday. Musk's second post, 'Guess my election-interfering space lasers didn't work this time,' had 8.5 million views. Reuters' analysis quoted the first line as the industry's reaction of record.

The Semafor reporter's live tally, 742,000 views, is where most of the overnight numbers came from. His entries match the county pages we checked.

Adam Tooze quoted Bianco in his newsletter: 'AI data centers were on this ballot and they suffered a devastating reversal.' Two corrections to Bianco's framing: the count was higher than 30, and most of the measures were binding charter amendments, not advisory questions.

The politicians' own words, from victory speeches: Talarico, 'We must get big money out of our politics, so no billionaire or trillionaire can buy our elections'; Sherrod Brown, 'Ohio is not for sale'; Jon Ossoff, that 'the titans of Silicon Valley dig bunkers for themselves' while warning that their AI 'risks human extinction.' Abdul El-Sayed, who won Michigan's Senate seat, had said it bluntly in July: 'People really effing hate data centers.'

What we could not verify

• Final counts. Texas House control, the Kansas governorship, Alaska's and Mississippi's Senate seats, four Maine towns, Morgan County, Illinois, and Oregon, Ohio were uncalled or uncounted at publication. Every ballot-measure figure is an unofficial county count.
• Any closing price. Market moves are overnight and premarket per CNBC and Reuters; no November 4 close existed when we published. We found no November 4 figures for Vistra, Constellation, Talen or GE Vernova.
• Any post-election statement from PJM, FERC, the Energy Department, ERCOT, a utility, or the Data Center Coalition. None existed by midday.
• The exact text of Anthropic's S-1 risk factors, quoted here as reported by Heatmap and CNBC; the filing exceeded what our tools can read.
• The dcmap.us state shares. The tracker does not publish its method; we use it only for the Texas share, which Wells Fargo and SemiAnalysis independently put at 27–30%.
• Hinojosa's '$600 a year' bill claim and the precise date the PUC adopted its large-load rule (September 18 or 20).
• Piqua, Ohio's charter amendment. WOSU and the Dayton Daily News report it passed with 67%; the tracker lists it as moved to a February special election. The three recalls are not in dispute.

What to watch, with dates

• November 10: the lease deadline Fermi's lender waived; the first sign of whether the Amarillo mega-site survives a Texas pause.
• December 10: ERCOT and the PUC's data-center audit report to the governor, the document Hinojosa will govern from.
• December 2026: PJM's next steps after FERC's partial rejection; any Senate attempt to attach ratepayer rules to year-end bills.
• January 3, 2027: the new Congress; whether Jeffries creates a select committee on AI with subpoena power.
• January 12, 2027: the Texas Legislature convenes with a Republican Senate and a House that may or may not have flipped. The moratorium, the tax exemption, and the water rules live or die here.
• January 2027: Hinojosa's inauguration and her first directives on the TCEQ permit halt; three PUC seats open to appointment.
• February 2, 2027: Piqua and Sidney, Ohio special elections, and the first test of whether the Conserve Ohio template keeps winning after the wave.
• Anthropic's IPO window: the first big AI listing to price with these risk factors live.