The Signal — October 26, 2026
Half of Americans now say they would oppose a data center within three miles of home, and voters in battleground races are hearing about it in campaign ads eight days before the midterms. Apollo's chief economist argues that, outside the biggest tech companies, AI hasn't yet shown up in revenue, margins or productivity. And last week Google and Meta released software that lets PyTorch code run on Google's chips with minimal changes, loosening one of the ties that keep AI work on Nvidia hardware.
Half of Americans now oppose a data center near home, up from 28% in January, POLITICO poll finds
A POLITICO poll conducted by the London-based firm PF Global found that 50% of U.S. adults now oppose a data center being built within three miles of their home, the highest share since POLITICO began asking. In January, 28% opposed and 36% supported; now 21% support it and 35% are strongly opposed, according to Newsmax's write-up of the published results. The online survey of 2,019 adults ran October 16–19 and has a stated overall margin of error of plus or minus 2 points.
The poll also asked about hypothetical House candidates. For a Republican who backed a nearby data center, 41% said they would be less likely to support, or definitely wouldn't support, that candidate, against 24% who would be more likely to. For a Democrat taking the same position, the figures were 39% and 23%. Opposing a local project increased respondents' stated willingness to support candidates of both parties, especially Democrats: 47% said they would be more likely to back a Democrat who fought one, compared with 36% for a Republican. POLITICO says pro-data center candidates had a slight edge in January. Nearly half of respondents now name Republicans as the party most eager to build data centers, up from 38% in January, and 46% don't know where their own House candidate stands. The candidate questions went to subsamples of roughly 950 to 1,100 people, so their margins of error are wider.
POLITICO's analysis of AdImpact data counted 579 ads in battleground House, Senate and governor races this cycle that mention data centers; virtually all portrayed them negatively, and almost 80% were run by Democrats. Some Republicans have shifted as well: Michigan Senate candidate Mike Rogers has backed a "pause" on construction, and the National Republican Senatorial Committee warned the industry in August that the issue could sink Ohio Sen. Jon Husted. President Trump, who supports the buildout, urged developers last week to pay host communities a "dividend," writing that projects that were "highly unpopular" have "become extremely popular when a financial deal is struck." Sam Chen, a Pennsylvania Republican strategist, told POLITICO the pitch sounds fine at first because of the jobs, but that as people learn about "the water usage, the power that it might pull," it turns into "that 'not-in-my-backyard' mindset."
A separate PBS News/NPR/Marist poll of registered voters in September found 65% opposed construction in their own communities, with Republicans split almost evenly and self-identified MAGA supporters the only group in favor. The two polls used different samples and wording, so their numbers can't be compared directly, though both show broad local opposition. Both measure what people say about a hypothetical rather than how they will vote, and we couldn't check the POLITICO questionnaire ourselves because its results file blocked our download.
Sources: POLITICO · Newsmax via Yahoo News · PBS News
Apollo's chief economist says AI hasn't paid off for companies outside big tech yet
Torsten Slok, chief economist at the investment firm Apollo, wrote in an October 24 note that AI "has delivered zero financial payoff to corporate America outside of tech." He compares the seven largest tech companies with the rest of the S&P 500, which he calls the S&P 493. Revenue per worker and profit margins are rising for the first group and flat for the second, he writes, and his chart of the San Francisco Fed's productivity series shows "still no signs of AI boosting productivity growth."
Slok argues that AI spending is concentrated in a small number of firms, citing spending data from Ramp, and that "mainstream corporate America does not need the latest model and is unwilling to pay for it." He says companies are starting to treat token spending "not as a source of growth but as a cost center that weighs on margins." His deeper doubt is about what businesses need from intelligence: "AI may be a great calculator, but in business, intelligence is mostly judgment," and models don't understand "accountability, trust, culture, unwritten rules, ethical trade-offs or fairness."
The note is five charts and four paragraphs, not a study. Flat averages across 493 companies can hide firms that gained alongside firms that lost, and margins move for many reasons besides AI, including interest rates, wages and demand. The claim that companies are pulling back on token spending rests on Ramp's data as charted; the note shows no survey of the companies themselves. TechCrunch, which Apollo used to own, adds that other recent studies have also found little productivity gain so far. Slok's conclusion is that "the AI story will be decided in the S&P 493, not the Mag 7," which gives a test that can be checked as new earnings and productivity data arrive. Apollo's note carries the usual disclaimer that it isn't investment advice, and we're reporting it as evidence about AI adoption, not as a view on any stock.
Sources: Apollo · TechCrunch
Google and Meta release TorchTPU, letting PyTorch code run on Google's chips with minimal changes
At PyTorch Conference in San Jose last week, Google open-sourced TorchTPU, a backend co-developed with Meta's PyTorch team that lets code written in PyTorch run natively on Google's TPU chips. The PyTorch blog post went up October 20 and Google's announcement October 21. PyTorch could already run on TPUs through PyTorch/XLA, a 2020 bridge that the TorchTPU team says "required intrusive source-code modifications," compiled slowly and lacked PyTorch's ordinary step-by-step ("eager") mode. The new goal is that moving a model starts with changing model.to("cuda") to model.to("tpu").
The team says TorchTPU supports about 450 core PyTorch operations, has run the 5,000 most downloaded Hugging Face models out of the box, and is already used in production by some of Google Cloud's largest customers. Hugging Face Transformers, the vLLM and SGLang serving engines, and the TorchTitan training library now have TPU support built on it. Meta's director of PyTorch engineering, Jana van Greunen, said in Google's announcement: "We cannot be constrained by a single hardware architecture." InfoQ reports that Meta engineer Zongwei Zhou said in a keynote that Meta plans to move larger-scale LLM workloads onto TPUs with it. Amazon made a similar move the same week: TorchNeuron, a native PyTorch backend for AWS's Trainium chips, went into public beta October 20 with .to('neuron') as the device change.
For PyTorch users, Nvidia's CUDA has been the default path, and moving to another vendor's chips has usually meant rewriting code. Native backends lower the cost of trying alternatives without removing the work. The TorchTPU team says the one-line change gets a model running but not at peak speed, because shapes, sharding and operations may be "skewed toward a GPU architecture." Some features aren't supported yet, including FlexAttention and reinforcement-learning setups that run training and generation on the same chip. InfoQ cites The Next Web's report that TorchNeuron doesn't yet handle inputs whose shapes change, and Blair Hayes argued for Forkast that "the optimization and supply layers remain firmly under vendor control," meaning tuning, capacity and pricing still belong to each chipmaker. Google reports up to 52% model FLOPs utilization training Qwen 3 8B and near-linear scaling across as many as 8,192 of its Ironwood TPUs; those figures are Google's own and haven't been independently reproduced.
Sources: PyTorch (TorchTPU) · Google Cloud · InfoQ · PyTorch (TorchNeuron) · GitHub
On the Editor's Desk
We held an NBER working paper by David Autor and Google researchers on 133 patent lawyers. AI help improved their drafts, especially for junior lawyers, but after three months only the senior lawyers beat the control group on a review task done without AI. It is a careful randomized trial, but it was published in early September and the recent coverage adds no new findings, so we'd rather give it a proper look later. We also held Anthropic's advice to fold subagents back into a single main loop, because the evidence is Anthropic's own unnamed customer comparisons. A report that OpenAI is offering Anthropic's customers discounts of up to 90% blurs a customer's total cost savings with the discount itself, and the original report is paywalled. Allegations that influencers were asked to promote Moonshot's Kimi K3.1 Pro without disclosing a commercial relationship rest on circulating screenshots that haven't been authenticated, and we found no company response.
