# TechWave EP161: Do Americans Hate Data Centres, or AI? Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/techwave-2026-11-09-ep161-ai/ > Notes on TechWave EP161 (9 November 2026): host Harry works through the anti-AI backlash that followed the US midterms, citing polls where 87% of voters weighed AI data centres in their vote and 76% opposed one in their own community; he traces a water figure that was inflated 1,000-fold and sets it against Google's 0.25 millilitres per Gemini prompt. Educational listening notes only — no investment advice, no price targets, and the judgements here may be wrong. Published: 2026-11-10 Locale: en Tags: TechWave, podcast-notes, data centres, AI sentiment, US midterms ![Dusk on an American suburban street, the lamps just lit, the road running away into the distance, and on the horizon the silhouettes of long data centre halls and transmission towers](/covers/techwave-2026-11-09-ep161-ai-cover.png) > The hearts of the many can raise a city wall; the mouths of the many can melt metal. > > —— *Discourses of the States*, "Discourses of Zhou, Part Two" (pre-Qin; translation mine) In TechWave EP161, published 9 November 2026, host Harry works through the anti-AI backlash that followed the US midterm elections. The polling he cites has 87% of voters weighing AI data centres in their decision, 56% calling it a key or important factor, 76% opposing a new data centre in their own community, and 57% saying AI does more harm than good. Harry traces the most widely repeated water figure back to its source: a Chilean government document wrote litres where it meant cubic metres, which made one Google data centre look like it used 1,000 times the water of the town around it, while Google's own report puts a median Gemini prompt at 0.25 millilitres. His read is that the water and electricity-bill complaints are mostly emotional, and that what actually happens next is construction slowing without stopping; everything after the first ten minutes, he says, is his own subjective view. ## First came the doom talk, then the book Harry traces the resentment backwards, and his first stop is 2023 and 2024. Those were the years Anthropic and OpenAI talked to the public at length about AI risk: how AI might destroy humanity, how many people it would put out of work. Anthropic's CEO Dario Amodei had a few lines he repeated back then — all code would be written by AI, 90% of white-collar jobs would disappear — and Harry notes he says them less often now, that both companies have pulled back. The words are already out, though, and the public's picture of AI was shaped in those two years. The second stop is a book, Karen Hao's *Empire of AI*. It says one Google data centre uses 1,000 times the water of the small Chilean city of Cerrillos. The number came from a Chilean government report that wrote litres where it meant cubic metres, which made residential water use look 1,000 times smaller than it was. Hao later acknowledged the error, apologised and corrected it — but before the correction she had finished a round of interviews, the book had reached the New York Times bestseller list, and she had been named one of Time's most influential people in AI. A second claim was circulating in the same period: that one ChatGPT prompt burns several litres of water. Harry says TechWave corrected this a year ago, because Google's report is explicit — a median Gemini prompt uses 0.25 millilitres, about five drops. He offers an arithmetic handle: finish a 500 millilitre bottle of water from a convenience store and you have spent the equivalent of 2,000 Gemini prompts, which is more than a year's use for someone who only talks to a chatbot. One toilet flush covers a decade. ## Over the same period, bills went the other way On water, Harry says newer builds have changed: closed-loop cooling keeps the same water circulating, and anything that can run on fans skips water cooling altogether. On electricity he offers no defence — a 1 GW data centre peaks at close to 1 GW. He reframes the question as what happens to a resident's bill once the data centre arrives. Then he points to Project Rainier in Indiana, the Amazon and Anthropic site that supports Claude's training and inference, now at 1 GW. The local utility's latest filing takes rates down, and expects the average household to pay about US$100 less a year from 2027. The mechanism here is the part of the episode I most want to remember. Part of your bill pays for generation; another part covers the fixed cost of the grid itself. That fixed cost used to be divided among everyone drawing power, so when a data centre draws an enormous load it absorbs a large share of it and the household share shrinks. On capacity, the hyperscalers help build generation and transmission, and absorb the grid costs their own load creates. Harry says he searched for counter-examples and found nothing clear. The one rate increase he found was xAI's Colossus 1 in Memphis, up 4%, under a schedule approved back in 2023 — the city council stated that no residential, commercial or industrial rate had moved because of the data centre. The most extravagant package on offer is Meta's Hyperion in Louisiana: a power contract projected to save other customers 2.65 billion in electricity costs over 20 years, more than a billion in local infrastructure, and a commitment to replenish 100% of the water used. Harry says if he lived there, he would want the data centre. ## Before the vote: the anger was already tied to AI So why, with terms like those on the table, are people still on the street with signs? Harry splits the opposition into three layers: the data centre's own effects (bills, water, noise); dislike of AI itself (lost jobs, hacking, the chance it eventually destroys us); and politics — Trump backs data centres, so the Democrats took the opposite side, the anger about the cost of living transferred across, and the data centre became the symbol of the rich getting richer, taking your water and your power to make money for the tech giants. Harry then opens up the premise under the second layer. The latest Microsoft and Gallup research has 55% of Americans not using AI at all, or using it once every few months; a large share of the remaining 45% are not daily users, and Harry guesses more than nine in ten only use a chatbot — ChatGPT as a better Google search, the occasional email. He cites Muse: as big as it got, 2.5 million daily active users is 0.73% of the US population. For an average American the value delivered by AI is limited, while the stories about its harms are everywhere. The line of his I remember longest: very few technologies in human history have earned a reaction this negative before any real catastrophe happened. He sets it against nuclear power — Three Mile Island and Chernobyl were an explosion and a meltdown, and the fear grew after the events; before them, nuclear had public support. He also checks whether "nothing has happened yet" holds up. The Yale Budget Lab report concludes AI has caused no meaningful unemployment: the big tech layoffs were mostly correcting years of over-hiring, with AI used as the reason. The effect shows up more in recruiting — hiring more senior people and fewer of them, while the organisation still grows. On security, the PHP vulnerability of the past fortnight was serious, yet few consumer-facing companies have taken large losses from it. Whether that changes, he puts the odds as not low. ## Election week: politics became the bottleneck first The results came in, the Democrats took both chambers, Taiwan's market fell far enough that the national stabilisation fund stepped in, and the semiconductor sector took its worst single day since the pandemic. Harry's causal chain is short: the Democrats oppose data centres, new projects stall, and data centres that never get built never need the GPU servers. His own reaction to the day was that nobody expected politics to become the bottleneck before electricity did. Does construction stop? He says no. The expensive part of a data centre is the hardware, there is no labour cost to save, and staying in the US buys speed rather than anything irreplaceable; Canada, Mexico, the Nordics and Southeast Asia all have land, the intelligence you train is the same wherever you train it, and the inference latency penalty is small. He adds a view: with the Middle East racing to build its own sovereign AI capacity, the US exporting its AI infrastructure is bad for it long term. The slowdown, though, he treats as certain. He relays the analyst base case: projects half-built with all permits in hand get finished, those still missing permits risk being stopped, and those at the planning or proposal stage mostly die. Finding new sites takes time, so the pace falls over the next few years. With compute scarcer and pricier, his reasoning runs like this: Anthropic's gross margin on model usage is 85% — of every 100 dollars of revenue, compute takes 15 — so a doubling of compute prices still leaves 70%, and they keep buying while the company as a whole starts losing money. The price is cuts: projects that burn compute without earning, and without feeding the direction the company is heading, go first, the way OpenAI cut Sora. He guesses a few of Anthropic's industry-specific products take the hit; research compute does not, because that is where the competitive edge comes from. ## What to watch next: whether the trust flywheel turns Harry thinks sentiment reverses once AI agents — AI that can operate tools in sequence and finish a task on its own — become a mainstream product. The same Microsoft and Gallup research contains a finding that holds across every country surveyed: the more often someone uses AI, the less they worry about it. The worry is concentrated among people who barely use it. Where is it stuck? He says agents can already do anything — book a restaurant, shop, plan a trip, deal with customer service — and the problem is the last mile. An agent can shortlist restaurants at your price, cuisine and location and analyse the reviews, then hit a place that only takes bookings over LINE or by phone, and stop there. The tooling is crude too: his example is the Gmail connector, which expands the whole message history into every email it sends. LINE and Shopee, which people in Taiwan use daily, have nothing to connect to. The flywheel looks like this: the more you trust it, the more tools and personal data you give it, the more it can finish, and the more you trust it. For ordinary consumers it is not turning. Harry's most optimistic call is six months, and he says plainly that nobody knows. For him it is already turning, and this is the most vivid stretch of the episode. Using OpenAI's Dot as his agent, he reckons two weeks produced more than NT$200,000 of value: six videos he handed over as raw footage and never touched again, each of which used to cost NT$2,000 to outsource; and when two sponsorship slots opened up, ten minutes of direction was enough for the agent to find candidates and build a list. He picked a dozen, asked it to write, and it read his entire history with each company, looked at their products, and sent tailored emails — by the next midday both slots were sold, for NT$140,000. Then he undercuts his own example: he is a business owner, and an ordinary consumer has no videos to edit and no ad slots to sell, with little high-value work to hand over in the first place. ## The news sounds terrifying — how much of it should I believe? What stopped me longest was that factor of 1,000, whose source was a units column filled in wrongly in a government document. I have been caught by this myself: I see an outrageous multiple, and what I retain is the multiple, not the units behind it. Numbers like that get amplified as they travel, because they are useful and because they sound like a complete story. The one step I added since: when a number arrives as a shocking multiple, find who computed it and in what units. Most of what this stops is material I wanted to believe. Harry's 0.25 millilitres is the same move run in reverse — Google's report states a median and states a unit, which makes it checkable in a way "several litres" never was. ## The technology didn't change — so why do I still want to sell? That single day in semiconductors happened, and Harry's position is that the technology's fundamentals did not move. When both are true at once it is uncomfortable to hold anything: the account is falling and there is no new bad number to point at. Harry offers two test questions, and says the long-run direction holds as long as both answers stay yes: will future AI produce far more value than it does now, and will future AI still need large amounts of compute. One midterm election, he argues, moved neither. He says he treated the fall as a chance to buy, and states his conditions up front — be able to hold for five years, and don't let a drawdown touch your life or your mood — adding explicitly that none of it is financial advice. I would add the failure conditions myself. Both questions can break in visible ways: an architecture change that cuts the compute needed for the same intelligence, or AI's value staying concentrated among a small group of heavy users and never spreading. The second is harder to spot, because it never appears in a single day's news — it looks like adoption numbers sitting still year after year. His two questions point at direction and say nothing about timing. A political bottleneck can hold for years, and you have to survive those years first. ## Further reading - TechWave EP161 (9 November 2026), hosted by Harry; the first stretch is the election and polling, the rest is his subjective read - Karen Hao, *Empire of AI*, together with her correction to the Chilean data centre water figure - Google's report on energy and water use per Gemini prompt, which gives a median and a unit - The Microsoft and Gallup cross-country research on AI adoption - The Yale Budget Lab report on AI's effect on employment ## One thing to take away An impression can set before any evidence arrives, and once set it goes looking for material. American hostility to AI grew on ground where no catastrophe has occurred; hostility to nuclear power grew after Three Mile Island and Chernobyl. The second is learning, the first is something else, and both feel equally strong from the inside. Here is something I have tried. Take something you have disliked recently — it needn't involve investing: a new policy at work, a rule at your child's school, a system the company is rolling out. Find where your impression first came from: who said it, which document the number came from, what units that document used. If you can trace it, you know what you are against; if you can't, what you are against is whoever first told you about it.