PHOENIX — The construction site looked like any other in the sprawling desert outskirts of Phoenix: chain-link fencing, stacks of drywall, the low hum of generators. But by nightfall in December 2024, what had vanished from one such site was not copper wiring or heavy equipment. It was 100 NVIDIA H100 graphics processing units, each retailing for $30,000, altogether worth $3 million.
The Phoenix Police Department arrested three individuals in connection with that theft, but the bigger picture is far from resolved. The U.S. Department of Energy’s March 2025 report documented that 500 H100 GPUs were reported stolen from data center construction sites across the country in 2024, with a total value of $15 million.
The geography of the thefts is stark: Phoenix, Arizona, lost 200 GPUs. Dallas, Texas, lost 150. These are not smash-and-grab jobs.
They are surgical strikes on the single most sought-after piece of hardware in the artificial intelligence boom. The H100 is the engine training the largest large language models.
NVIDIA shipped 2 million H100 units globally in 2024, according to the company’s annual report, but supply constraints remain severe. Demand far outstrips supply. That scarcity has created a black market where stolen GPUs are resold to cryptocurrency miners or to overseas AI labs willing to pay without asking questions about provenance.
The U.S. Department of Commerce’s Bureau of Industry and Security issued a warning in January 2025 that stolen GPUs could be diverted to countries subject to export controls, including China. The FBI’s Cyber Division is investigating the thefts as part of a broader organized crime pattern. The implication is clear: these are not isolated heists by opportunistic thieves.
They appear coordinated, targeting construction sites where security is still being installed and inventory tracking is loose.
A $30,000 Piece of the AI Puzzle
The H100 GPU is not a consumer product. It is a specialized data center component, roughly the size of a large book, designed to run the matrix math that powers generative AI. Its $30,000 retail price reflects the immense computational capacity packed into each unit.
Training a frontier-level large language model requires clusters of thousands of these chips running for weeks or months. That concentration of value in a portable object has made the H100 a target unlike any previous generation of hardware.
The 500 GPUs stolen in 2024 represent only a small fraction of the 2 million units NVIDIA shipped that year, but the thefts are concentrated in a way that disrupts specific projects. A data center construction site losing 100 or 200 GPUs faces a months-long delay while replacements are sourced and security is overhauled. The Phoenix arrest in December 2024 — three individuals, 100 GPUs, $3 million — was a rare public win for law enforcement.
The Phoenix Police Department has not disclosed whether the suspects were part of a larger network. The FBI’s involvement suggests the investigation is looking beyond the local fences.
Export Controls and the Black Market
The Bureau of Industry and Security warning from January 2025 adds a national security dimension. The United States maintains strict export controls on advanced AI chips to prevent them from reaching China and other adversaries. The H100 is covered by those controls.
If stolen GPUs are flowing to overseas AI labs in restricted countries, the thefts are not just property crime — they are a potential breach of export law. The FBI’s Cyber Division is investigating the pattern as organized crime.
The bureau has not released details of its investigation, but the characterization suggests a structure: planners, financiers, logistics handlers, and buyers. The thefts from construction sites in Phoenix and Dallas fit a playbook: casing the site, bypassing minimal security, removing the GPUs before inventory is completed, and moving them through resale channels before the theft is even discovered. By nightfall on any given day, a stolen H100 could be across the border or on a cargo flight.
The resale market operates on speed and silence. Cryptocurrency miners, who need massive GPU arrays but face supply constraints of their own, are one known destination.
Overseas AI labs, particularly those in countries under export restrictions, are another.
What Comes Next
The 2024 theft numbers are a snapshot of a problem that the industry is still learning to counter. Data center construction sites are now installing real-time tracking for high-value components. NVIDIA has not publicly changed its shipping or distribution practices in response to the thefts.
The company’s 2024 annual report, which noted the 2 million units shipped, did not address theft or security directly. The FBI investigation continues.
The Bureau of Industry and Security is monitoring export channels. The Phoenix and Dallas police departments have the most direct experience with the thefts, but the pattern suggests the next heist could be anywhere a new data center is rising from the desert floor.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “NVIDIA H100 GPU Thefts: $30K Chips Stolen from AI Data Centers”, “description”: “PHOENIX — The construction site looked like any other in the sprawling desert outskirts of Phoenix: chain-link fencing, stacks of drywall, the low hum of…”, “datePublished”: “2026-07-08T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Isabella Moore”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}}
