Fermi names nuclear veteran Lee McIntire as CEO
Two stories from the AI buildout's two most contested frontiers landed Wednesday: the power side and the data side. Fermi, the company behind the nuclear-powered Project Matador campus in Texas, ended its three-month CEO search by hiring a veteran of the nuclear industry, and publishers gained a new weapon in the fight against AI scrapers — a font that poisons training data while staying perfectly readable to humans.
Fermi, the AI data center developer behind the nuclear-powered Project Matador campus near Amarillo, Texas, has named Lee McIntire as CEO, ending a three-month search that began when the company fired co-founder Toby Neugebauer in April.
McIntire, an independent board member since September 2025, previously ran CH2M Hill and TerraPower, the nuclear startup founded by Bill Gates, and brings four decades of large-project experience to a company co-founded by former Energy Secretary Rick Perry. Chairman Marius Haas framed the hire as a shift in mission: "Fermi's next chapter is a construction and power delivery story," he said, and the job now is to "build on schedule, on budget, and safely."
The timing points to momentum. On Monday Fermi signed TensorWave as the first binding customer lease at Project Matador, and the stock is up nearly 23 percent since then. The appointment caps a turbulent stretch — Neugebauer and CFO Miles Everson were pushed out in April in a restructuring the company branded "Fermi 2.0," complete with a new Dallas headquarters. For the AI-nuclear sector, whose bottleneck has shifted from raising capital to actually delivering gigawatts, putting someone who has built nuclear projects in charge is the clearest sign yet that Fermi intends to be a construction company, not a story.
A new font called ShieldFont promises publishers a readable page for humans and a subtly scrambled one for AI scrapers — the same page, two different texts.
Designers Isaque Seneda and Gabriel Abrucio built it on ligatures, the typographic feature usually used to pair letters, to swap content words for decoys of the same part of speech: "horse" becomes "potato." Humans see the correctly rendered page; bots that pull raw HTML get a version where roughly a quarter of all words, and nearly half of content words, have been replaced. In tests across six scraper pipelines, more than 90 percent of pages that would have passed a quality filter were rejected after the swap, and pages that slipped through carried about 20 percent "training-time garbage": real English, correctly spelled, asserting nothing true. "Dropped means they did not get your work. Kept means they got something wrong," the authors write.
The approach has limits — screen readers, translation tools, and search engines all get tripped up, and a scraper that renders pages and runs OCR would defeat it, at five to thirteen times the cost. That cost math is the point: ShieldFont aims to make unauthorized scraping less useful and more expensive, in service of what its creators call a basic principle — "being discoverable does not mean consenting to AI training." It is the scraper arms race's latest escalation, and whether or not it scales, it is clever enough to be worth watching.
What to watch: whether Fermi's next filings show Project Matador construction starts, and whether any publisher of scale actually adopts ShieldFont.
The scraper arms race keeps escalating — how far should publishers go to keep their work out of training data? Tell us in the comments.
