The Debate Over xAI's Expansion in Memphis, Two Years In

Good morning!

Two years ago this month, the Greater Memphis, Tennessee Chamber of Commerce announced a deal to host a substantial new computing facility with Elon Musk’s artificial intelligence company, xAI. Dubbed "Colossus," the data center was built in an old Electrolux factory. It was brought online at record speed, and almost immediately drew questions about transparency, air quality, water use, and how a project of this scale had moved through local government so quickly.

Two years on, the story continues to expand alongside the company’s growing footprint, with a second campus—Colossus II—across the state line in Southaven, Mississippi; a contested gray water recycling plant; an ever-rising count of gas turbines; multiple lawsuits; and communities in South Memphis still pressing for straight answers.

Few people have tracked the details more closely than Neil Strebig, a reporter with The Commercial Appeal in Memphis who has covered the xAI story daily from the beginning. He’s attended community meetings and hearings; filed right-to-know requests; parsed the differing interpretations of the Clean Air Act by the EPA, the Shelby County Health Department and the Mississippi Department of Environmental Quality; counted turbines; and spent time with residents living alongside the facilities. The result is a command of a level of detail that few can match.

In a podcast discussion, Strebig brings us up to speed on the latest developments — including a newly updated lawsuit citing unpermitted turbinesin Southaven, the implications of the SpaceX IPO and the impending IPOs of other AI firms, and the stalled water recycling plant Memphis leaders had counted on. And, he reflects on what it has been like to chase facts as the story spread across two states and a thicket of jurisdictions.

This is the first of three episodes we’ll release over the next few days on the situation in Memphis and Southaven, and what it tells about the future of the AI infrastructure boom in the United States and the political debates and power struggles surrounding it. Listen to the discussion here.

Speaking of data centers… state public utility regulators are making decisions that will shape where AI energy infrastructure gets built, who pays for it, and how quickly it comes online. Deborah Glosser, associate professor at Western Washington University, provides a primer on how those decisions are being made.

Anthropic and the US government

After the Commerce Department issued an export directive forcing Anthropic to restrict access to its Fable 5 and Mythos 5 models, questions abound about what it means for US AI governance going forward, even as leaders in foreign capitals see it as a clear sign that they need alternatives to US tech.

• Joseph Hoefer, a DC-based government relations strategist focused on AI, argues that the question hidden within the US government's move to block Anthropic's Fable and Mythos products isn't about the models or their capabilities. He writes that it's about which legal theory the US government uses to govern them and whether a precedent has now been set.
• "It's just vibes-based regulation." Alex Stamos, chief product officer at an AI security startup called Corridor, spoke to me on a podcast about the White House yanking Anthropic's Fable with no written standard, no warning, and no procedural path back. He compares it to grounding a jet without telling Boeing why. It's sending the wrong message abroad, he says.
• Who gets to assess frontier AI, on what terms, and who finds out the results? Merve Hickok, president of the Center for AI and Digital Policy (CAIDP), warns that Trump’s new executive order on AI buries those answers inside the NSA.

AI Now Institute series on EU tech sovereignty

The Anthropic story makes debates over tech sovereignty ever more urgent. In partnership with the AI Now Institute, a new series examines the dependencies embedded in Europe's application-layer AI market — and why they make sovereignty harder to achieve than policymakers assume.

• European AI startups at the application layer build on US models, pay US clouds, and risk getting ‘sherlocked’ by the same companies supplying their core technology, write AI Now Institute’s Frederike Kaltheuner and Leevi Saari. “The question is not whether Europe can and should build applications on top of existing models — it is rather to note that applications alone do not solve the question of technological sovereignty that Europe so desperately seeks,” they write.
• Code can be forked overnight; fabrication plants can’t. Javier Serrano, deputy group leader of the Accelerator Controls Electronics and Mechatronics group at CERN, Nicholas Gates, senior policy advisor at OpenForum Europe, and Johan Linåker, an Empirical Software Engineering researcher at RISE Research Institutes of Sweden and Lund University, argue Europe’s tech sovereignty push stops at software—and that real independence means opening up hardware too, not just the code that runs on it.

Special series: Agency, trust, and governance in the age of AI

This spring, the Berkman Klein Center for Internet & Society at Harvard University hosted an essay contest in collaboration with Tech Policy Press focused on the ways in which AI systems intersect with power and the public interest. We published the first posts in the series this week:

• AI companies exploit legal gray areas to cause familiar harms through new mechanisms while evading existing legal protections, writes Amaia Aguilar, an MBA candidate at Harvard Business School and a product manager at Apple. This post is part of a series of student essays produced in collaboration with the Berkman Klein Center for Internet & Society at Harvard University.
• Don't just regulate the data—regulate what a model can infer, writes Ikenna Ogbogu, an undergraduate at Harvard studying Computer Science and Economics. He makes the case for capability-rooted AI governance and enforceable audits as part of a series of student essays in collaboration with the Berkman Klein Center for Internet & Society at Harvard University.
• The biggest AI risk isn't rogue machines—it's "cognitive surrender,” writes Evan Liu, a recent Harvard University graduate. As part of a series of student essays produced in collaboration with the Berkman Klein Center for Internet & Society at Harvard University, Liu argues AI tempts us to outsource the struggle of learning itself. “In previous eras of technological disruption, competence still mattered because there were other avenues in which humans could pivot to and still contribute to society,” he writes.
• AI didn't kill design—it exposed what design was always doing, argues Zach Deocadiz, a designer, researcher, and facilitator. He says AI makes more visible that the design process has long served company goals over users.

How do we give people more input into AI governance?

• “What if AI happened with people, not to them?” Kris Rose, who leads efforts to govern enterprise-wide AI transformation at IBM, argues that public participation is the answer to the two crises of confidence surrounding AI: public anxiety about the technology and collapsing trust in the institutions governing it.
• Right now, a handful of private actors decide the values baked into the most powerful AI models. Kevin Frazier, senior fellow at the Abundance Institute, argues that's unsustainable. His fix: a "People's AI Constitution Council" made up of citizens to choose the constitutions frontier models can follow. “Future democratic governance must be tailored to the parts of the AI stack and development process that are still amenable to deliberations and judgements that cannot transpire in seconds,” he writes.

More on AI and its implications

• As tech companies embrace AI — forcing workers to use it, using it to surveil them, or justify firing them because of it — more tech workers are organizing to fight back as they feel they are losing influence over decisions that affect their jobs, reports Tech Policy Press fellow Varsha Bansal.
• Last week, the European Parliament voted on AI Act simplification rules, including a ban on AI systems that generate nonconsensual intimate images. National regulators warn, however, that many member states still lack the authority needed to enforce the measures, Joana Soares, a Portuguese freelance journalist, reports.
• Indefinite data retention is not a feature; it is a structural vulnerability, argues Koa DeMarzo, a recent graduate of the George Washington University. The solution is straightforward: platforms should sunset inactive data, either voluntarily by changing default settings or through policymaker action.
• Surveillance pricing is often treated as a novel problem requiring new solutions. But Canada already has precedents that prohibit unfair, discriminatory pricing and protect the public, writes Emily Osborne of the Canadian Shield Institute. Now it just needs to apply them. “We must take the clear and principled stance that using personal information to set individualized prices is unfair discrimination and regulate it accordingly,” she writes.
• AI tools are commoditizing. What matters now is who can access your data—your email, finances, digital life. Chris Riley, executive director of the Data Transfer Initiative, makes the case for trust infrastructure to govern the ‘agentic’ ecosystem before the floodwaters rise.

Must reads on platform accountability

• Just hours after Brazil expanded platform liability rules, 26 legislative decrees were filed to block them — most from Bolsonaro's party, many word-for-word identical. Tech Policy Press fellow Tatiana Dias on the coordinated counteroffensive.
• The search data-sharing regime taking shape in the EU will rise or fall on how much data Google must share, how often, with what privacy safeguards, and through what process rivals actually receive it, write the Knight-Georgetown Institute’s Alissa Cooper and Zander Arnao.

I wish you the best for the week ahead!

-Justin