An 149th Intelligence Squadron airman conducts training in a computer training lab at Mather Field, California Dec. 2, 2023. (DVIDS)

WASHINGTON — If OpenAI can actually implement its Stargate Project to build $500 billion-worth of AI infrastructure in the US, one of the major beneficiaries may be the US military.

“It depends on how much of that they devote to gov[ernment] cloud and AI cloud,” said Roy Campbell, chief strategist for the Pentagon’s High Performance Computing Modernization Program and deputy director for advanced computing in the undersecretariat for research & engineering. And if the Defense Department can get a slice of Stargate’s computing power, he told Breaking Defense, it could bypass a major bottleneck for its current high-tech ambitions.

“We really don’t have the footprint inside [DoD], the in-house capabilities, to answer all the questions, so we’re really going to have to ‘phone out’ to all these assets” that the private sector is building, Campbell told Breaking Defense on the sidelines of the Potomac Officers’ Club annual defense R&D conference.

What Stargate will build, precisely, is still a little vague. Announced in a high-profile press conference Tuesday at the White House, featuring newly elected President Trump, the Stargate Project will be a new company created to develop that “AI infrastructure” to train and operate ever-larger artificial intelligence models, including new data centers and the electrical power to run them.

Managed by OpenAI and Japan-based Softbank, Stargate will also draw funding from computing giant Oracle and UAE-based investment firm MGX. OpenAI will also provide technology to the new Stargate company, along with Arm, Microsoft, NVIDIA, and Oracle.

“We will begin deploying $100 billion immediately,” OpenAI’s launch announcement said. “The buildout is currently underway, starting in Texas, and we are evaluating potential sites across the country.”

The ambitious half-trillion project is aimed at commercial customers: Neither the White House press conference nor the OpenAI announcement mentioned national defense. But it could also help DoD solve its own AI dilemma.

What DoD Needs

For over a decade, the Pentagon has placed ever-bigger bets on artificial intelligence to achieve a wide array of goals, from streamlining procurement contracts, to analyzing intelligence data, to coordinating all the services’ disparate forces over land, sea, air, space, and cyberspace, a concept called CJADC2. But with over three million troops to pay, train, equip, and keep supplied around the world, even the nearly trillion-dollar DoD budget can’t keep up with AI’s exponentially growing hunger for computing power.

Cutting-edge Large Language Models and other generative artificial intelligence tools require so much electricity, in fact, that industry is looking to nuclear power. By one calculation, OpenAI’s flagship ChatGPT burns almost three times as much energy per query as Google Search, and it consumes enough power per year — over 226 gigawatts — to charge three million electric cars (or 186 time-traveling DeLoreans).

Meanwhile DoD is already straining to meet its own computing needs. That includes just figuring out how to implement AI by running field experiments with battle command networks, swarming drones, and other autonomous systems. At last year’s Autonomous Warrior 24 in Australia, for instance, as well as similar events stateside, “they have to bring the supercomputer out to the test site,” Campbell told the conference today. “The amount of data that has to be pulled off of all of the [test] assets and annotated, that has to be crunched overnight to be able to make sense of what happens, is phenomenal.”

RELATED: Don’t blow the budget on ChatGPT: Army CIO sounds alarm on big bills for GenAI

As more and more military AI moves from the proving ground to the battleground, Campbell continued, that kind of supercomputing power may become necessary in the warzone as well. While AI can sift inhumanly large amounts of information at superhuman speed, experts warn it can also be “brittle,” prone to breaking down or producing nonsense results — “hallucinations” — when exposed to new inputs that are sufficiently different from the data it was trained on. That means AI algorithms can’t be trained once and used indefinitely: They have to be constantly updated and retrained on new data.

“If you encounter something new in the battlefield, you may have to retrain, which then begs the question of, how much do I need in my Forward Operating Base?” Campbell told the conference. “I may only have hundreds of kilowatts of capability in that Forward Operating Base,” he said, rather than the megawatts required to train and retrain a cutting-edge AI model.

To meet future power demands, the Pentagon is looking at smaller, safer nuclear power plants to provide one to five megawatts of power in a compact or even truck-transportable package. But where long-range, high-bandwidth connections are available — and as long as a savvy adversary doesn’t  jam them — then forward-deployed forces may be better off referring their AI problems to supercomputing clusters back in the US, Campbell said.

That would include not just DoD’s own high-performance computing assets, he emphasized, but the larger capabilities in the private sector. That’s where Stargate could come in, providing computing resources beyond the Pentagon’s capability.

Of course, that assumes Stargate pans out. Within a day of OpenAI’s high-profile announcement, which featured newly inaugurated President Trump, no less a figure than Trump ally and multi-billionaire Elon Musk publicly dissed Stargate’s backers, claiming they did not having adequate cash.

“They don’t actually have the money,” he said on his social media network X.com, known before he acquired it as Twitter. “SoftBank has well under $10B secured,” Musk said of one of OpenAI’s three major partners on the project. “I have that on good authority.”

OpenAI CEO Sam Altman promptly shot back that Musk was “wrong,” while the CEO of Microsoft, Satya Nadella, said “I’m good for my $80 billion.”

But a close reading of the Stargate press conference at the White House suggests the partners did scramble to quintuple an initial $100 billion proposal in just a few weeks, under pressure from Trump.

“Mr. President, last month, I came to celebrate your winning and promised that we will invest $100 billion,” said Softbank CEO Masayoshi Son. “And you told me, ‘Oh, Masa, go for $200!’ Now I came back with $500 because, because this is the — as you said yesterday — this is the beginning of golden age of America.”