Today’s midterm elections have huge implications for the way forward for AI

• AI is on the ballot.
• Anthropic discloses its Q3 results and even more massive spending plans.
• Investors push Anthropic to lower valuation target as its IPO math looks tough.
• Palantir reports record revenue growth.
• Report: U.S. intelligence believes Chinese spy targeted Nvidia CEO Jensen Huang.
• A new way to ‘program’ AI models.
• AI lab employees offer chilling video testimonials on the tech’s big risks.

It’s election day in the U.S., of course, and in many places AI is on the ballot.

Voters in dozens of states and local municipalities from California to Maine are deciding on measures that could impact the construction of hundreds of billions of dollars’ worth of data center projects. Bloomberg has a great breakdown of the measures by type and state. Overall, there are at least 41 ballot initiatives involving data centers, according to election tracking site Ballotpedia. My native Ohio has the most pending, with 21 measures in total, including 11 that would either ban construction or force a pause in further data center developments.

Among the most closely watched municipalities is Piqua, a 20,000-person city in west central Ohio, where Meta is planning to spend $1 billion to construct a data center with a capacity of up to 1.4 gigawatts. Meta promises the project will bring 1,000 construction jobs to the area, but local opposition is running high. Voters there will be deciding today whether to recall the mayor and two city commissioners who have supported the project. There’s also a $2 billion AWS data center project on the ballot in Sunbury, Ohio, as well as a $10 billion AWS project on the ballot in Adams County, Ohio.

With a majority of Americans being more fearful than hopeful about AI, according to a recent Pew Research Center poll, many data center projects could be in trouble. In the second quarter of 2026 alone, at least 45 data center projects worth some $68 billion were disrupted due to local opposition, according to a report from Data Center Watch, a research project run by the company 10a Labs. These trends could spell trouble for the leading AI companies, which are increasingly compute constrained, as well as for the neocloud companies that have sprung up to provide AI compute. (For more on the risks facing the neoclouds, see my colleague Beatrice Nolan’s essay in last Thursday’s Eye on AI.)

Many AI industry leaders are belatedly saying data center developers made a mistake in trying to rush many of these projects to market without doing enough to ensure political buy-in from the communities in which they are located. OpenAI has recently been trying to convince people to support data center projects by rebranding them as “compute campuses”—although it’s not clear that alone will do much.

When Bea and I recently sat down with Nvidia CEO Jensen Huang in London, he told us, “before we bring in massive amounts of investment into a small community, I think we have to go first and let the community know what we’re planning to do. Partner with the community, invest in the community, and if the community really doesn’t want us to be there, find another community.” He said that he thought the industry now realizes this and has changed its approach. “Everybody is much more mindful about it,” he said. “I’m hoping we turn the tide on [data centers].” But this belated “hearts and minds” strategy from the AI industry and its data center partners may be too little, too late.

All of this certainly heightens tensions around Anthropic’s upcoming IPO. In an updated prospectus made public on the SEC website yesterday, the company disclosed even more infrastructure and compute commitments than it had in its initial S-1 filing, with the figure now totaling some $895 billion. The company also revealed that it had to spend $5 billion in the third quarter to source short-term computing capacity to meet demand, saying that the added spending was one of the major reasons its adjusted operating margin fell to 15% from 19% in the prior quarter. Anthropic wants to lock in data center capacity on a longer-term basis, and even build out its own data centers through its new “Anthropic Direct” project, to lower some of these costs. But there’s a debate about whether the data center capacity Anthropic and its rivals such as OpenAI need will actually be available on the time scales they require, due to problems getting data centers built and connected to the grid. In fact, some think that Anthropic may be committing to buy more capacity from neoclouds than it actually needs precisely because it is assuming some portion of it won’t ever get delivered.

A vibe shift on Capitol Hill—but what will make it into law?

If the midterms result in a Democratic-controlled Congress, as seems possible, even likely if the polls prove correct, then AI policy could shift dramatically. The Democrats are far more likely to try to enact some kind of federal AI regulation. Before the election, the bipartisan Frontier Act looked like the bill that had the best odds of passing. But a new Congress will rejuggle that calculus and could open the path to other bills.

Of course, any legislation that Congress does pass faces a likely veto from President Trump, who has been steadfast in his opposition to AI regulation. That means it will still require a bipartisan bill with very broad support to have any hope of garnering the two-thirds majority in both houses needed to override a veto. This probably still favors a bill that looks like the Frontier Act, but you never know.

Another thing we may see in a Democratic-controlled Congress: a new wave of investigations into the leading AI companies and the safety and security protocols they’re implementing. Already Sen. Elizabeth Warren (D-Mass.) has written to Anthropic CEO Dario Amodei to demand answers to questions about the risks the company disclosed in its IPO prospectus. Expect much more of this if the Democrats retake the Senate.

Will AI play a role in trying to manipulate the vote?

Another thing to watch for in the midterms will be what role, if any, AI plays in attempts at election interference, political misinformation, and manipulation. For years, experts have been concerned about deepfakes and other AI-generated content being used in this way. Increasingly, AI is being used to create political advertisements, as well as satirical social media memes, that can influence, or at least reinforce, people’s perceptions of candidates. But in most cases, it is clear that this political communication is AI-generated and the effort is aimed more at persuasion than deception. So far, in U.S. elections, true deepfakes have not proved decisive.

The AI-focused publication The Transformer ran a piece in which it asked Anthropic’s Claude model how it would try to manipulate the midterm elections and it came up with an 11-page dossier on how to run an AI-powered influence campaign aimed at suppressing voter turnout. Interestingly, one tactic it did not pick up on, but which experts increasingly warn about, is poisoning the data that AI chatbots use to answer questions. This can take the form of trying to shift how an AI chatbot will answer questions about a candidate, their background and positions. But more insidiously, it can include efforts to mislead people about the mechanics of voting in their state, such as whether they can register to vote online.

Also concerning are attempts to suppress the vote in certain districts with misinformation about problems at specific polling stations. And any attempts by AI-powered hackers to influence the election would present even more problems. OpenAI and Anthropic both announced that they have stood up war rooms to try to guard against bad actors using their models for political dirty tricks today. We’ll see what they report.

With that, go out and vote! And here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

Before we get to the news, just a reminder to check out the latest episode of our vodcast, Fortune AI Weekly. This week, Beatrice Nolan and I dissect Anthropic’s S-1, discuss a disturbing new lawsuit against OpenAI, and new research from Anthropic on how AI can impact professionals’ judgment. You can check out the vod here on YouTube.

Also, if you want to find out what’s really happening in AI and learn more about how you can use this transformative technology to supercharge your business, join me and my colleagues for Fortune Brainstorm AI San Francisco. We have an amazing line up of speakers for you, including OpenAI President Greg Brockman, Microsoft AI CEO Mustafa Suleyman, AI luminaries Fei-Fei Li, currently cofounder and CEO of World Labs, and Andrew Ng, currently cofounder and CEO of LearnVector, Palo Alto Networks CEO Nikesh Arora, LinkedIN cofounder and Greylock Partner Reid Hoffman, and many many more. The conference is taking place December 7-8. You can apply to attend here.

FORTUNE ON AI

Anthropic’s growth slowed and costs jumped in the third quarter, updated IPO filing shows—by Beatrice Nolan

Anthropic’s third quarter results mean IPO math and history are not on its side, analysts say—by Wen Shao

Who pays for AI’s $2.6 trillion build-out? Look to corporate budgets—by Sheryl Estrada

Palantir posts ‘highest-ever’ revenue growth rate of 96% as investors’ patience rests on a hair trigger—by Amanda Gerut

AI IN THE NEWS

Wall Street pushes Anthropic toward a lower valuation. That’s according to a story in the Wall Street Journal, that reports institutional investors are telling Anthropic and its investment bankers that they only have appetite for an Anthropic IPO if it values that company in the range of $1.5 trillion to $1.6 trillion, down from earlier talk of nearly $2 trillion. The push for a lower valuation comes amid a tech-stock selloff, rising interest rates, concerns about excessive AI exposure, data-center backlash, and uncertainty over returns on AI spending. Anthropic, which could begin its roadshow next week and have its stock market debut the week before Thanksgiving, reported third quarter revenue of $17 billion, up from $11.6 billion in the second quarter, while also disclosing a staggering $895.4 billion in long-term compute purchase and lease commitments. Investors are watching the offering as a key test of appetite for AI stocks, with a successful Anthropic debut potentially opening the door to a wave of other IPOs early next year.

OpenAI slashes funding target and valuation in latest round. The AI company has cut its latest fundraising target from $30 billion at a $1.4 trillion valuation to about $20 billion at a $1.25 trillion valuation, as a sharp tech sell-off and mounting concerns over AI spending have dampened investor enthusiasm, the Financial Times reports. Some investors have reduced their allocations, although a UAE consortium including state-backed MGX is expected to be the largest participant, alongside firms including Founders Fund, Coatue, Sequoia, Andreessen Horowitz and SoftBank, the newspaper said. Meanwhile, the Wall Street Journal reported that OpenAI has so far only managed to secure about $11 billion of the funding it is seeking.

U.S. intelligence agencies say Nvidia CEO Jensen Huang targeted by Chinese spy. That’s according to a story in the New York Times that quotes unnamed sources saying that U.S. intelligence agencies have concluded that Alice Huang, an executive associated with Singapore-based data-center company Megaspeed, was a Chinese intelligence asset who was providing information to Beijing while maintaining a “close relationship” with Nvidia CEO Jensen Huang. President Trump was briefed on the findings roughly a year ago but chose not to act, amid concerns about disrupting Nvidia’s chip sales, the stock market, and U.S.-China trade relations, the paper said. It also said intelligence agencies had chosen not to provide an official warning, called a “defensive intelligence briefing,” to Jensen Huang about Alice Huang. U.S. officials already suspected Megaspeed of helping Chinese companies circumvent restrictions on advanced AI chips, with investigators examining whether some Nvidia chips acquired through Megaspeed’s Southeast Asian operations were being illegally moved into China. The episode raises broader questions about the Trump administration’s approach to AI export controls, particularly as Nvidia’s Huang has successfully lobbied for looser restrictions on the company’s China business while developing a close relationship with the president.

OpenAI looks to make data centers less burdensome on power grids. As part of an effort to convince state power regulators as well as local communities to support data center projects, OpenAI said it was launching a new “Grid Protectors Program.” The initiative will seek to make its future AI data centers more flexible in their electricity consumption, reducing demand during periods of grid strain to help protect power supplies for homes and businesses. The program will begin with a planned data center in Effingham County, Georgia, where an agreement with Georgia Power provides for up to 1 gigawatt of flexible demand reduction—equivalent to the average electricity consumption of roughly 700,000 households. OpenAI says it will cover the full electricity and infrastructure costs associated with its facilities, while also working with utilities to use demand management, on-site power generation and battery storage to improve grid reliability and potentially lower costs for other customers. You can read OpenAI’s blog on the effort here.

U.K. chip startup Fractile raises $774 million funding round. London-based AI chip startup Fractile has raised £577 million (about $774 million) in Series C funding, according to a filing with British business registry Companies House, trade publication Tech Funding News reports. The new funding comes just five months after Fractile raised $220 million in its Series B. The company is developing inference chips that keep data closer to processing transistors and claims its architecture can dramatically accelerate AI workloads. Anthropic has reportedly agreed to an initial $250 million chip order, with deployment expected in 2027.

EYE ON AI RESEARCH

AI can edit the weights of another model, no fine-tuning needed. That’s the conclusion of a research paper from a group of PhD. students at MIT’s Computer Science & AI Lab (CSAIL) that is getting a lot of attention this week.

Normally, if you want to tune an AI model to perform particularly well at a specific task, you have to do some kind of additional training. Usually this is in the form of supervised fine-tuning, where the model is shown at least dozens and often thousands of examples, until it eventually excels at the task. (And remember, this is after the base model has already been trained on millions or billions of examples to perform a task such as predicting the next word in a sequence.) In other cases, models can be tuned through reinforcement learning, where they learn by trial and error to take actions that will maximize a reward that the person conducting the training specifies. Either way, the learning takes place through “gradient descent,” the process by which a model gradually adjusts the weights of its neural network in order to minimize the amount of error on a given task.

But what the CSAIL team has discovered is that a large frontier model—in this case, OpenAI’s GPT-6 Astra—can, without any specialized training, read and interpret the weights of a small, open-weight model. Just by looking at the weights, the researchers say, Astra knows what the model was trained to do. And it can then edit these weights directly to improve the small model’s capabilities markedly—with no need to further train the small model at all. CSAL calls this method “weight space programming,” because it points the way towards being able to create neural networks that are “born” being able to excel at certain tasks, rather than having to learn these skills through training on thousands or millions of examples.

The team demonstrated this potential by getting Astra to directly write the weights for a small game-playing model and found that it was better than a model that had been trained with a commonly-used reinforcement learning algorithm for 100 million steps. It also found that Astra could directly read the weights of Alibaba’s small Qwen3.5-0.8B model and diagnose the source of a defect that model has when it’s doing certain mathematical functions (the model gets stuck in loops). By editing just 0.016% of the Qwen model’s weights, Astra is able to fix this looping defect and also boost the Qwen model’s score on a math benchmark from 7.9% to 48.1%.

This is a big deal for three reasons:

•First, it could dramatically lower the cost for enterprises that want to train or fine-tune their own models for specific tasks.

•Second, it could be a big breakthrough in interpretability—our ability to understand the internal “thought” processes of large language models. It might be possible to simply ask one model to read and interpret the weights of another model and explain the model’s thinking to us in natural language. If true, this would be a lot simpler and more effective than other kinds of interpretability techniques. And that’s critical because interpretability is one of the ways we are hoping to keep AI agents under human control.

•Lastly, while this interpretability finding would be a huge win for AI safety, I think the research also has potentially worrying implications for the safety of open weight AI models. It seems it would be easy for a bad actor with no technical skills to use a frontier AI model to edit out any guardrails or alignment measures that the developers of open weight models try to bake into their models. It was always true that a machine learning expert might have been able to tamper with these baked-in guardrails. But the CSAIL research would seem to open the possibility up to non-technical users too.

You can read the paper here.

AI CALENDAR

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Nov. 18-19: Slush, Helsinki, Finland. 

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

What do people working inside leading AI companies really think about AI’s risks? That is the premise of a new film series called “frominside.ai” in which current and former employees of OpenAI, Anthropic, and Google DeepMind provide on-camera testimonials about what they see as the dangers of the AI technology they are building or once helped to build. A few of the interviews are with former AI company employees who have emerged as prominent whistleblowers on AI safety, such as former OpenAI researcher Daniel Kokotajilo. But more interesting are the five testimonials from current company employees who are much less well known. In particular, the testimonial of Mary Phuong, a Google DeepMind researcher who works on AI safety and alignment, has gone viral in AI circles because of how visibly emotional Phuong gets when the filmmakers ask her if she has any message she wishes to convey to her family or the people she grew up with about AI. 

Phuong says that although her family sees her as a workaholic and obsessed with her career, which she says is “partly true,” she then adds, tears in her eyes and her voice choked with emotion, “I want them to know that I am partly doing this for them. And for all of us.”

The film series was produced by Palisade Research, a nonprofit AI safety research organization in Berkeley, California, with clear ties to Effective Altruism and the more “doomer” wing of the AI research community. But, bearing that in mind, the testimonials are still fascinating, as a historical record if nothing else. You can disagree with those who think AI poses an existential risk, you can think they are deluded, but watching these video testimonials it is impossible to doubt the sincerity of these AI company employees.