Source: raw/How_a_Data_Center_Slowdown_Benefits_Big_Tech.md (The AI Daily Brief, Nathaniel Whittemore, https://www.youtube.com/watch?v=ssEWGHjKHTM); raw/Is_the_AI_Boom_Over_Let_s_Talk_About_It..md (Nate B Jones, https://www.youtube.com/watch?v=illf65HD_t4); raw/Anthropic_s_IPO_in_Question_Midterms_Hit_AI_Nasdaq_Drops_5%_and_RSI_Accelerates_309.md (Moonshots ep. 309, Salim Ismail, Dave Blundin and Alex Wissner-Gross, https://www.youtube.com/watch?v=WF-ADiC7Khs); raw/Big_Tech_s_3T_AI_Bet_Claude_Writes_90%_of_Anthropic_s_Code_AIs_Editing_AI_Brains_EP_308.md (Moonshots ep. 308, https://www.youtube.com/watch?v=x09f_BuJsoY). All fetched 2026-11-05 to 2026-11-09. Platform: YouTube.
Local votes against data centers on November 3, 2026 made it harder and costlier to build new AI capacity in the US, above all in Texas. Three shows reached the same counter-intuitive reading within days: a slower build-out hurts startups and neoclouds that need new sites, but it can help hyperscalers and anyone who already owns working capacity, because demand keeps rising while supply grows more slowly. This article records that argument, the evidence each show gives, and what it means for buyers of compute. The electoral results themselves are covered elsewhere; this article covers only the data-center consequences. Every figure is a host relaying a report or a poll, not primary data.
Key Takeaways
- The votes were lopsided. NLW counts 43 of 44 moratorium-type data-center votes passing. Nate B Jones (NBJ) cites unofficial results in Edgewater, Florida, where more than three quarters of voters backed a ban on new data centers, and calls those margins typical. Moonshots relays a Goldman Sachs estimate that 45% of planned US data-center build-out could face restrictions.
- Use and opposition coexist. NBJ: 65% of US adults had used an AI chatbot in the previous week (Edison Research/SSRS, May, up from 52% in February), while 72% opposed a new data center in their community (ABC News/Washington Post/Ipsos, October).
- Demand has not slowed. Google said on its latest earnings call that usage through its model APIs rose 40% quarter on quarter and is more than 4× a year earlier, and that it has started buying outside compute because Google Cloud is supply-constrained (NBJ).
- The bottlenecks predate the election. Google, Meta, Amazon and Microsoft each came in under their Q3 capex targets, citing construction delays. Microsoft attributed a $5B shortfall to a rented facility that was not ready, and Oracle warned in September it might declare force majeure on a New Mexico project with power-connection problems (NLW).
- Who wins: operators with cash and working data centers. Fewer new facilities means higher utilisation and pricing for existing ones, and distressed projects can be bought cheaply (NBJ, NLW, Matt Farley).
- Who loses: neoclouds and other new builders that borrowed against delivery dates. A delay pushes out revenue while debt payments and depreciation continue (NBJ). Satrini Research (as captioned) flags “takeout risk” and “completion defaults” when completion dates slip past debt maturities.
- The glut risk may fall. Derek Thompson’s argument, endorsed by NBJ: slower supply gives demand time to catch up, which lowers the chance of a 2027 overbuild.
- Markets read it as a split trade on November 4: Nvidia −3% and CoreWeave −10%, while Microsoft rose 6.7%, Amazon 6.1% and Google 3.3%, and the Magnificent Seven added about $1T (NLW). A broader crash followed the next day; see AI Bubble Signals.
What changed for builders
- Texas is the centre of it. The outgoing governor tried “an 11th hour switch” away from data centers. The incoming governor, Gina Hinojosa, is expected to be less accommodating, and Texas has been “the epicenter of the data center build out” (NLW). NBJ adds that Texas was already restricting permits before the election.
- The terms on offer. An X account relayed by NLW summarises Hinojosa’s plan as no subsidies, union labour, no NDAs, and guarantees that household power and water bills will not rise. Texas’s incoming senator (name garbled in the captions) asked that data-center companies pay for their own energy, recycle their own water and get community consent before building.
- Satrini Research’s view (via NLW): “The End of the Trade as We Know It” says AI capex is past its peak and that new data centers will cost more. Builders face larger municipal concessions, clean-energy requirements, ratepayer protections and slower permitting, on top of interconnect, labour, turbine and input-cost bottlenecks. It does not expect the incoming governors of Texas, Ohio or Georgia to stop the build-out outright. Its key line: “perhaps the bottlenecks become the excuse some people need to quietly trim their own projections without causing a rush for the exits.”
- NLW’s own forecast: data centers keep getting built, but on new terms: no NDAs, full process transparency, builders paying the full cost of infrastructure including subsidising other users’ electricity, quality-of-life and environmental review, and direct payments or community investment. He expects builders to accept all of it.
- An analyst’s view: JP Powers (RWA Wealth Partners) calls the fears “a little bit overblown”. He expects more onus on states and higher costs, but little effect on existing projects.
- Moonshots’ reading: Dave Blundin does not expect approved projects already under construction to stop, given sunk costs and jobs. He reads the market as pricing in slower power generation and slower construction of projects that were planned but not yet zoned.
Why a slowdown can favour incumbents
NBJ lays out the mechanism step by step:
- A neocloud borrows, commits to equipment and signs a customer who expects capacity by a set date.
- Approval or power hook-up slips. Revenue moves out or disappears, while loan payments keep arriving and the equipment keeps depreciating.
- It may have to rent capacity elsewhere to keep the customer, which eats its margin. “The danger in this game is running out of money before you can serve demand.”
- The project does not die with the company. A buyer with cash and existing operations picks up the equipment, contracts or half-built site “at a competitive price the original investors wouldn’t have accepted”.
- Operators with working data centers face fewer competing facilities, which supports “higher utilization, better pricing, lower depreciation rates, and stronger negotiating power”.
- Smaller AI companies end up renting more capacity from the large companies they compete with.
Supporting commentary relayed by NLW:
- Matt Farley (market analyst): “slower AI capex is bullish, not bearish. Harder to build data centers mean GPU prices flatten, compute spot rents climb, capex falls, margins expand, FCF flips faster.” He asks: “Can the constraint now be the moat?”
- Skanda Amarnath (Employ America): his group chat discussed “whether the hyperscalers would welcome political blocking of data center construction as a capital discipline coordination mechanism.”
- The Hill: hyperscalers are in talks with environmental groups on a joint framework for data-center development.
- Jessica Lessin (The Information; captioned “Lesson”): hyperscalers up while chips and neoclouds fell “suggests that the AI trade doesn’t play just one way”, which she reads as good news for anyone worried that weakness in some companies would pull the whole market down.
Demand keeps rising
- Google: API usage up 40% quarter on quarter and more than 4× year on year; buying outside compute because of supply limits (NBJ, from the earnings call).
- OpenAI’s CFO: “We are not slowing down data center investment and infrastructure investment” (Salim Ismail on Moonshots, relaying remarks made hours before recording). OpenAI has also announced a community compact to cover energy costs.
- Other moves listed on Moonshots (Ismail): a $17B Google data center in Wyoming; SpaceX building 3.6 GW on Earth and targeting 13 GW by the end of 2027; Meta warning of construction challenges; Microsoft taking “some oversight” of some neocloud providers.
- Big tech’s long-term commitments: Moonshots ep. 308 puts Alphabet, Meta, Microsoft, Amazon and Oracle at more than $3T in combined lease, building, power and equipment commitments. Details and the Q3 capex figures are in AI Bubble Signals.
Where new capacity goes instead
- Specific sites, not states (NBJ). What matters is deliverable power, water, a supportive community and a developer with good local relationships. Being able to open on time is worth more than cheap land. Communities that can offer a credible site gain bargaining power over infrastructure costs, union jobs and local investment. If displaced projects pile into the same few welcoming places, those places will face pressure too.
- States that want it. Blundin cites Tennessee, where xAI put its Colossus cluster, and Ismail cites Virginia, as places that have welcomed data centers and seen economic uplift.
- Outside the US. NBJ names Canada and Mexico for proximity, and India, Southeast Asia and parts of South America for stable land and power. He hears more interest from developers outside the EU than inside it, because building in the EU is not seen as faster. He mentions the EU gigafactory initiative and a new UK agreement setting expectations for developers. Blundin argues Canada has far more power per capita than anywhere else in North America, and that a Canadian project started now would come online around the next US election.
- Orbit. NBJ thinks the election raises the odds of orbital data centers. xAI and Musk have said they are pursuing it, and Google has a Project Suncatcher prototype in orbit. Radiation, heat, depreciation and maintenance are still unsolved. He moves his own estimate for orbital compute at scale from about 10 years to “maybe six to eight”. Wissner-Gross says the political uncertainty pushes compute toward space. Blundin’s limit: best case, space is “a couple percent of all compute in two years”.
- The cost of keeping it out (NBJ): “If you keep a data center out of the community, it doesn’t keep the technology out. It just means you’re renting the technology and you don’t own it.” Construction and operations jobs follow the investment abroad, and the service is rented back to US customers.
Compute access for smaller buyers
Moonshots hosts describe the shortage from the buyer’s side. These are first-hand anecdotes from investors with stakes in the companies named.
- Blundin could not buy an NVL72 (Nvidia’s 72-GPU GB300 rack) at any price. He signed a $15M three-year lease, and the machine comes online on February 1. He says he got it only through an inside contact at Ornn, a GPU-futures company he and Wissner-Gross have financial interests in.
- One of Ismail’s pilot clients applied to a cloud provider and was told it was not big enough; the provider was focusing on very large customers.
- Blundin claims Google, Anthropic and xAI have reserved most US compute and chip supply for years, and calls for a “right to compute” so startups are not shut out. He relays Musk’s estimate of about 15 GW of compute sitting idle because it cannot find a liquid-cooled data center to plug into.
- A host on ep. 308 notes that H100 prices are rising although the chips are about three years old.
- Wissner-Gross: if supply is restricted while demand rises, the price of compute should rise, which signals where scarce capital should go.
Added 2026-11-11: more estimates, and the supply trend
Sources: The Artificial Intelligence Show ep. 250 (raw/Ep._250_-_Midterm_Voters_Reject_Data_Centers_AI_Speeds_Up_Anthropic_s_Research_Nvidia_s_Spy_Probe.md, Roetzer reading Axios, the Washington Post and the Wall Street Journal); The Neuron Daily, November 10 (raw/newsletter-theneurondaily-com-47e5bd68d8.md); The Epoch Brief, November 11 (raw/newsletter-epoch-ai-fa4c3ebf04.md), checked against ai-research/epoch-compute-capacity-doubling-9-months-2026-11-11.md.
- More estimates of what is at risk. Axios: “nearly two-thirds of 2028 data center capacity could face delays”. Brad Gerstner: 25-35% of next year’s US build-out. Goldman’s 45% above is a third figure; they may measure different horizons. ^[ambiguous] The Washington Post puts a quarter of planned new data centers in Texas.
- Timing (Roetzer, opinion). Commitments run about two years ahead, so projects already broken ground in 2025-2026 continue and the first real capacity effect lands on what would have come online in 2028-2029. Sequoia’s Pat Grady expects “elongate[d] timelines”; a construction-industry economist quoted by the Wall Street Journal expects the boom to “really soften next year”.
- The price pass-through, as Roetzer predicts it: less inference supply means higher prices or usage limits for ChatGPT, Claude and Gemini, especially on compute-heavy video, image, world-model and agent features.
- OpenAI’s terms. OpenAI published a “national community compact” and a “compute campus policy blueprint”, renaming training data centers “compute campuses” and promising local compacts, paying its own energy and infrastructure costs, and impact reporting. Its Pike County, Ohio example: 2,500 long-term operating jobs averaging $80,000 a year and 26,000 construction jobs.
- Building where it is welcome. SpaceX sought a tax abatement on the first phases of a roughly $24B gas power plant and supercomputing project at Starbase, Texas (The Neuron). Musk, after the vote: “data centers in space are looking pretty appealing right now.”
- Global supply is still growing fast. Epoch estimates new AI computing capacity sold worldwide grows about 2.5× a year (doubling every 9.1 months), down from about 6× in 2023, with deceleration since 2024 not statistically significant. US permitting limits where that supply can be installed, not how much is made. ^[inferred] Detail: Epoch AI, November 2026.
Try It
- Budget for higher compute prices through 2027. Every source here expects scarcity, not a glut. Lock in capacity or rate limits you depend on, and keep a second provider configured.
- Ask whether your provider’s capacity is operating or still under construction. NBJ’s point: a building under construction and an operating facility are different things until the power is on. A delayed site can strand your launch date.
- Check the counterparty if you buy from a neocloud. Its exposure to delayed sites and debt is now the main risk. See the counterparty checklist in AI Bubble Signals.
- Do not read construction headlines as adoption forecasts. NBJ: “Your employer is going to keep buying AI services and they will figure out how to get AI from somewhere.”
- Watch what companies do with their next planned project. NBJ’s tell: are they waiting, moving, paying more for a deliverable site, or buying capacity from someone else? That says more about commitment to AI than a press release.
- If you advise local businesses or communities, treat a credible, deliverable site as a bargaining asset. NLW and NBJ both expect builders to accept tougher terms in exchange for certainty.
Update (2026-11-11): two more capacity signals (AI Daily Brief, relaying Bloomberg)
- “Microsoft declined to renew the contract” for a compute rental arrangement with Oracle that had been in place since 2023.
- Google, “who had been a prospective tenant, has now bought out the project and plans to invest 17 billion into the development” (Crusoe’s project).
- Relays only; Bloomberg’s articles were not fetched. Chip supply and export context: the Nvidia-Megaspeed article.
Open Questions
- Goldman’s 45% figure reached the wiki through one host on Moonshots, with no report title or date. Moonshots also cites “56% opposition to big tech and data centers” without naming the poll.
- Name spellings from auto-captions: “Satrini” (also “Sutrini” and “Satriani” in the same episode) is kept as captioned; it may be Citrini Research. “Jessica Lesson” is taken to be Jessica Lessin. The incoming Texas senator’s name is garbled. Ornn is spelled from captions (“Orin”, “ORE”).
- Musk’s 15 GW of idle compute is Blundin’s relay of a claim Musk “has been saying for a while”. No date or source is given.
- Whether capex trims are demand-driven or delay-driven cannot be told from the outside. Satrini’s line suggests some may be dressed as politics. This matters for the “hyperscaler walks away from capacity” trigger in AI Bubble Signals.
- Project Suncatcher is NBJ’s description (“a prototype in orbit”); no launch date or specification was given.
- SpaceX’s 13 GW target is Ismail’s figure. It roughly matches Musk’s earlier “2.5–3 GW by end-2026 plus 10 GW in 2027” recorded in the bubble-signals article, but the two were not given as the same number.
Related
- AI Bubble Signals — the November sell-off, the $3T commitments and NLW’s escalation triggers
- The Compute Economics of the AI Buildout — supply-side numbers and neocloud credit stress
- Will Financing Bottleneck AI Compute? — how data-center debt is structured
- How Americans View AI — the polling behind the local opposition
- Anthropic’s Public S-1 — the filing lists data-center opposition as a risk factor
- Enterprise AI Spend, November 2026 — the demand side
- Cost & Intelligence Levers — routing and cost controls if compute prices rise