What Trump's AI Manhattan Project Means for the AI Industry
On Monday, the President signed an executive order laying out his vision for a “Manhattan Project” for AI. The plan calls for the national labs—the Department of Energy-funded research centers—to work together with AI companies to train AI models based on government science data. For instance, it tasks the Secretary of Energy with assessing the computing resources that “industry partners” could provide to the effort.
That collaboration, along with others in the order, would build on existing relationships between AI companies and their counterparts in universities and national labs.
For example, earlier this year, system administrators from xAI sought out advice on how to set up data centers from Dr. Dan Stanzione, who runs the Texas Advanced Computing Cluster at the University of Texas, Austin. They talked about “high speed file systems to support AI workloads,” Stanzione said, meaning how multiple servers can efficiently read the same files at the same time.
According to Stanzione, those xAI employees worked on the Memphis data center that the company established last September and used to train Grok 4, a model that caught xAI up to the frontier of AI companies, at least at the time.
“Everybody who runs big systems, we all talk, because we all have the same problems,” he said. TACC’s former employees work at Amazon, Nvidia, Google and Microsoft, so they tend to discuss best practices with TACC.
The executive order also calls for internships to train students to perform AI research for science. That provision also builds on existing government programs to equip the next generation of workers with AI experience. For example, TACC has been providing computing capacity to the National AI Research Resource, a government pilot program to support AI researchers and small companies.
“We need a lot of AI literate workers,” said Stanzione, who was on the task force that helped set up the program. Unless you get a job inside one of the leading AI companies, “where do you learn those skills?” he asked. “How do you get good at large language models?” Student researchers use TACC for AI projects such as fine-tuning language models to excel at chemistry, he said.
NAIRR has not yet received permanent funding, but Trump’s AI action plan gave a shoutout to the project, as did the House Bipartisan AI Task Force in its report last year. A bill to make NAIRR permanent has been introduced in the House.
Expanding the pipeline of AI researchers is surely a boon to AI startups, but a more immediate benefit to startups would be getting access to the scientific AI models that result from the mission. That could free those startups from training models themselves, saving them a fortune on Nvidia chips.
Academic data centers have managed to get a hold of a surprising number of those chips. TACC, for instance, will install about 4,000 of Nvidia’s latest generation of AI chips, the Blackwell, in its next facility. TACC negotiated a price for the chips in 2020, before ChatGPT, and “Nvidia has stuck by the deal, to their credit,” Stanzione said.
The total price of the computing capacity for the new data center, including the Blackwells and thousands of Vera chips, along with the networking equipment and cooling for them, came out to $150 million, Stanzione said. With current Nvidia prices, he estimated the Blackwells alone would likely cost a couple hundred million dollars.