The US military needs a commercial backup plan for an AI chip crisis

As Washington and Beijing prepare for a high-stakes meeting next month, policymakers are once again debating tariffs, rare earth minerals and advanced technology exports. But a less obvious issue could also shape the outcome of U.S.-China AI competition: AI chips for classified systems.

In a conflict, the ability to retrain AI systems rapidly as conditions change could determine how quickly militaries adapt on the battlefield. Updating modern AI models requires enormous computing capacity — often thousands of specialized chips known as graphics processing units, or GPUs.  That reality raises an urgent question: What happens if the U.S. military suddenly needs far more classified AI chips than it currently has available? 

There are two basic approaches to solving this problem. The military currently relies on data centers that are physically walled off from commercial networks — facilities typically located on military bases, staffed exclusively by U.S. citizens, and subject to strict government access controls.

The first approach is to double down on that model and stockpile these classified data centers with enough chips to handle a surge in demand. Programs such as Stratus are already physically installing GPUs inside classified facilities. But this approach has real limits. It risks spending enormous sums on infrastructure that sits idle between crises, while competing with commercial buyers for chips already in record demand. More fundamentally, there is no guarantee it would be enough — especially against an adversary like China.

The second approach is more creative, and potentially more powerful. Rather than build redundant infrastructure for classified workloads, the military could, under predefined emergency conditions, temporarily run classified workloads on GPUs in commercial data centers using software-based security controls rather than physical separation. These sophisticated software controls govern who can see what and ensure nothing sensitive is left behind once the work is done.

Google Cloud already employs this kind of software-based security model to host sensitive but unclassified government information — proving that strong access controls can protect government data without the expense of entirely separate physical infrastructure. This approach would leverage billions of dollars in private investment while avoiding the need for the government to permanently build and maintain equivalent infrastructure.

Furthermore, the scale of that private investment in computer infrastructure is striking — and largely untapped for national security purposes. An estimated $320 billion was poured into AI hardware in 2025 alone. Yet today’s cybersecurity compliance requirements mandate that classified workloads run only in dedicated, government-controlled environments.

Unlike classified data centers, which are staffed exclusively by U.S. citizens and military physical security, commercial data centers are not subject to the same access controls, meaning foreign nationals may work in those facilities with less physical security. As a result, the vast commercial AI infrastructure that powers everyday applications remains largely off-limits for classified military operations.

Critics will rightly note that software controls are not foolproof. But in a fast-moving crisis, the trade-off may be worth it. Faster access to commercial GPU capacity could mean the difference between retraining a critical AI model in days versus weeks. And as American technology companies continue to innovate — developing more secure chips, more intelligent software controls and better tools for ensuring classified data is sanitized from commercial systems — the security gap between physical and software separation will continue to narrow. 

Policymakers can take several steps now to hedge against the risk of a future military AI chip crisis. First, they should resist the urge to stockpile chips in classified data centers when private capital is already building the world’s most powerful AI infrastructure. Government spending should leverage that investment, not duplicate it. Flooding the market for chips that aren’t immediately necessary risks driving up costs and undermining commercial development.

Second, they should fund research that makes it possible to run classified workloads on these commercial systems — ensuring that sensitive data can be processed without ever being permanently stored outside government control.

Finally, they should prioritize establishment of contracting mechanisms that treat commercial AI data centers as a mobilizable national resource, much like the Civil Reserve Air Fleet, which allows the military to rapidly leverage commercial aircraft during emergencies without owning the entire fleet. 

The U.S. holds a rare advantage: the world’s most dynamic AI industry, the most advanced chip designs and a technology ecosystem that is already the envy of every competitor. The question is whether the government will build the frameworks to harness that advantage before a crisis forces the issue.

The cost of getting this right is modest. The cost of getting it wrong — a military caught flat-footed, unable to adapt its AI systems fast enough — is not.

Maj. Katherine L. Carroll is a U.S. Space Force fellow at Georgetown University’s Center for Security and Emerging Technology. 

More Opinions - Technology News

GOP patience with Iran operation growing thinner as Trump escalates tensions

Trump’s gone too far: Invoke the 25th Amendment before it’s too late

Greene raises 25th Amendment after new Trump Iran post: ‘Evil and madness’

US soldier trying to halt deportation of wife detained on military base

Tucker Carlson: Trump expletive-filled Easter post ‘vile on every level’

Why so many Americans now sympathize with the villain

Pentagon cancels Tuesday press briefing with Caine, Hegseth

Jeffries calls on Republicans to ‘stop the madness’ after latest Trump Iran ...

Rahm Emanuel makes case for change to a divided Democratic Party

GOP rep: Trump 'not talking about obliterating innocent people' with new Iran ...

Live updates: As Iran’s deadline nears, Trump cautions ‘a whole ...

Freedom Caucus calls for full DHS funding in GOP-only bill, rejecting ...

Iran calls for human chains around power plants ahead of Trump deadline

Supreme Court clears way for dismissal in case of ex-Cincinnati City Council ...

Trump on Iran: ‘Whole civilization will die tonight’ unless ‘something ...

GOP Sen. Johnson: I do not want to see the US blow up Iran infrastructure

America is heading for a recession — and it may be the worst yet

Former NATO commander on Iran rejecting ceasefire proposal: ‘They still have ...

The Hill Podcasts – Morning Report

© The Hill

visit website