Needed: A congressionally supported, Rickover plan for sovereign AI
by Mike Hewitt · The Washington TimesOPINION:
By the mid-1950s, the U.S. Navy had put the world’s first nuclear-powered submarine to sea. The reactor was supposed to take a decade. Hyman Rickover did it faster, and then he ran his program accident-free for the rest of the century.
Rickover did not invent physics. He invented a way for the federal government to develop and build a complex, strategic program.
Rickover’s program had three features that the rest of the Pentagon did not and still does not: single-point accountability that owned the technology end to end, a long-tenure leader who outlasted political cycles, and an industrial base disciplined to one standard.
He ran it with dual hats, the Navy and the Atomic Energy Commission, and he owned design, fuel, training, operations and safety as one integrated system.
Seven decades later, the program he built has logged more than 6,200 reactor years of safe operation. No other federal program comes close to the Nuclear Navy.
Yet Rickover’s real lesson was never about reactors. It was about what becomes possible when authority, accountability and time horizon are aligned inside one office structured to deliver.
We are facing a problem with exactly that shape. Artificial intelligence and the sovereign compute infrastructure that runs it are becoming strategic capabilities on par with nuclear propulsion.
Sovereign compute means nuclear-powered AI facilities operating on a dedicated, resilient optical backbone, islanded from the commercial grid, resilient against disruption, and operating under government authority to run our most critical workloads. It does not mean isolation from the commercial domain or government ownership of every component of the supply chain.
Advertisement Advertisement
That authority is the same kind Rickover held over every reactor the Navy ever put to sea. The data centers, the power, the networks, the silicon supply chain — none is a commercial convenience. Each is a platform for national power.
The Defense Department does not buy compute the way Rickover bought reactors. It buys it the way it buys office supplies: fragmented, at the lowest bid, year by year, riding on a commercial market never designed for mission assurance.
That architecture is fundamentally flawed. The authority structure, acquisition pathway, infrastructure, certification regime and long-term demand signal are designed and stand ready for a congressionally supported sovereign AI industrial base.
We need a single accountable engineer, dual-hatted across the relevant authorities. We need a network of dedicated infrastructure yards that produce sovereign compute the way Newport News and Electric Boat produce hulls. We need long-term capacity purchase agreements that give the industry a demand signal worth investing in.
One standard. One accountable office. One generational time horizon.
Advertisement Advertisement
The nuclear race was built on the knowledge that Russia was pacing us and China would not be far behind. The AI race is no different: China has been focused on sovereign AI while downplaying its usefulness.
This is a Sputnik moment; a Kennedy moment. Those events galvanized the country around the strategic importance of the nuclear and space domains and produced the national assets that defined American power for generations. The AI domain demands the same response.
The requirements for AI across our military and intelligence industrial bases represent the largest, fastest and most dynamic capability transition since the Manhattan Project. Whether it is the Golden Dome, AI-enabled carrier strike groups or cognitive warfare preventing conflict, AI will be the nuclear power plant of the 21st century.
A class of capability, such as the Nimitz-class aircraft carrier and the Virginia-class submarines: purpose-built national assets, sovereign by design, accountable to one standard and capitalized at a scale that only a long-term government demand signal makes possible.
Advertisement Advertisement
Nuclear-powered AI requires applying Rickover’s first principles: a confident supply chain, an innovation hub that paces the threat and the technology, and the authority and accountability to deliver.
I have spent my career on both halves of this problem. I came up in the Navy and later worked alongside the first generation of officers who stood up cyber and information operations as a warfighting domain. That work taught me — mostly the hard way — what happens when you try to bolt a fast-moving capability onto a slow-moving acquisition system. The cost shows up later, in capabilities we do not have when we need them. We do not have time to relearn that lesson on AI.
Every transformational technology in this country required someone with the authority, standard and stubbornness to build through the resistance.
The next one is sovereign compute. The Rickover first principles still work. We just need the resolve to apply them.
Advertisement Advertisement
• Rear Adm. Mike Hewitt retired in 2014 after 31 years of service in the U.S. Navy. He is co-founder and CEO of IP3.