Product1 publisher3 min readPublished
Heatmap finds three-quarters of Americans oppose new data centers where they live
The opposition numbers land on product teams, because an AI feature that fires without being asked bills compute for every user and adds load in towns already organizing against the next build.
The Product Desk · Product desk
What happened
- A Heatmap survey published last week found three-quarters of Americans oppose the construction of new data centers in their own communities.
- Activists have organized more than 130 protests across dozens of states, and data center construction has entered the campaign talk for the coming US midterms.
- The International Energy Agency puts a single query to an AI assistant such as ChatGPT at up to 10 times the electricity of a traditional Google search.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost The per-interaction compute bill sits inside the product's gross margin every month, and no amount of caching work recovers it, because each user gets a different computation.
- constraint Once siting is a midterm talking point, permit timelines are set in county meetings, and a signed power contract does not buy a roadmap past them.
- decision If efficiency gains get absorbed by demand, the lever a product team still controls is which features are allowed to recompute for every user and which run once for all of them.
- contradiction Lightbits pitches better utilisation of GPUs already installed while its own senior vice president says demand keeps climbing, so buyers should treat efficiency claims as something to add to a capacity plan.
Someone shipped a thread summary that generates the moment a message opens, whether anyone scrolls far enough to read it or not. Under the hosting economics that call was made in, an unread summary cost nothing. Andrew Marshall, vice president of developer relations at Yugabyte Inc., described why that no longer holds. "Traditional software has very low marginal cost because computation happens primarily on the user's device or cheaper multitenant infrastructure," he said. "AI inference incurs a real computational cost for every interaction." [10]
"Conventional software runs one code path for up to millions of users," Marshall said. "An AI application runs a different computation for each one. That's what makes it worth paying for, and it's what defeats the caching and code reuse that give software its margins." [11]
Three-quarters of the $750 billion Fitch Group expects the top five hyperscalers to spend on construction this year is $562.5 billion of AI-directed building in twelve months [1][16], or $112.5 billion each on an even split [17]. The 4.4 percent of national electricity that US data centers drew in 2023 is projected to triple by 2028 [5], which puts them near 13 percent [18]. SiliconAngle reports one large facility using up to 5 million gallons of water a day and equates that to a city of 50,000 people [8], so about 100 gallons per resident per day [19].
The opposition side of the record is softer than the spending side. The protest count and the Heatmap survey measure what residents say and what organizers can turn out [2][3], and the report does not name a project that was cancelled or delayed.
Efficiency work will not settle it either. The report invokes Jevons Paradox, the observation that per-unit savings get eaten by rising demand until total consumption goes up [14]. Ramesh Chettuvetty, senior vice president of AI product and business at Lightbits Labs Ltd., said "demand is still skyrocketing for more and more capacity" [15].
Two questions sort an AI feature list, and they can be answered in an afternoon: whether the compute runs once for many users or separately for every user every time, and whether a user asked for the output. The features that recompute per user and that nobody requested are the first cut, because they carry real electricity behind a trigger the team wrote. The scale of that electricity is not trivial per unit: SiliconAngle reports a five-second generated video consuming as much power as a household microwave running nonstop for over an hour [13].
For an auto-firing feature, invocation counts measure the trigger. The figures that survive a Friday review are how many users returned to the feature in week two and how long the first useful output took.
What to watch
- A county or state denying a hyperscaler permit outright would put a number on the opposition's effect on delivered capacity.
- Fitch's next capex estimate, and whether the AI share holds at three-quarters of hyperscaler construction spend.
- Whether Heatmap repeats the survey after the midterms and opposition stays at three-quarters.