What are the relevant sources of power here? Capital and talent ultimately. The capital required for frontier models is small compared to defense budgets so this is about talent. How to, one way or another, get it to work for national interests including military purposes. This is reminiscent of the Manhattan Project, where talented folks each had to decide what they would and would not do, and for whom.
Fair, I exaggerated. DoD doesnt need to match hyperscalers as a group, they just need their own. MSFT capex ~$120B is nontrivial for DoD but not out of the question. That said, it's easier for DoD to contract with compliant hyperscalers and/or lease instead of buy the datacenter capacity. If lenders put up the capex, which seems doable, then the DoD opex would be more like let's say $20B/year... forgive the handwaving but it all seems doable.
Backing up, my bigger point is that the talent gap seems far more daunting than this capital / budgetary gap.
What are the relevant sources of power here? Capital and talent ultimately. The capital required for frontier models is small compared to defense budgets so this is about talent. How to, one way or another, get it to work for national interests including military purposes. This is reminiscent of the Manhattan Project, where talented folks each had to decide what they would and would not do, and for whom.
Are the AI budgets small? The hyperscalers are spending $600B on capex this year. The DoD budget is, what, $850B?
Fair, I exaggerated. DoD doesnt need to match hyperscalers as a group, they just need their own. MSFT capex ~$120B is nontrivial for DoD but not out of the question. That said, it's easier for DoD to contract with compliant hyperscalers and/or lease instead of buy the datacenter capacity. If lenders put up the capex, which seems doable, then the DoD opex would be more like let's say $20B/year... forgive the handwaving but it all seems doable.
Backing up, my bigger point is that the talent gap seems far more daunting than this capital / budgetary gap.