In January 2025, OpenAI, Oracle, and SoftBank announced Stargate, a joint venture to build up to $500 billion of AI data center capacity in the United States, with the explicit backing of the federal government. A government-adjacent venture reserving compute capacity years in advance does not read like ordinary technology purchasing. It reads like the way nations have historically managed strategic resources — oil reserves, steel capacity — where supply, not just price, is the thing being managed.
The underlying logic is that advanced AI capability now depends on a narrow, physically constrained input: chips manufactured by a small number of fabs, largely built on equipment only a handful of firms make. That concentration makes compute a chokepoint in a way software rarely is, since software can be copied freely while advanced semiconductor manufacturing cannot be replicated quickly even by wealthy states that want to.
Washington's export controls on NVIDIA's most advanced chips, tightened repeatedly since 2022 and extended further in 2025, are the clearest expression of this on the restriction side. An adversary with limited access to NVIDIA's top-tier hardware is presumed to have a harder time developing frontier AI capability, so controlling the flow of chips becomes a lever with effects similar to historical restrictions on dual-use technology. Whether this actually slows capability development, or mostly slows the legitimate commercial sector while determined state actors find workarounds, is contested and difficult to verify from outside classified channels.
Treating compute as strategic inventory changes incentives for everyone downstream. Firms that can secure guaranteed allocations, the way Stargate's partners have, gain a planning advantage independent of technical merit — they can commit to training runs a competitor without secured access cannot reliably promise. This concentrates advantage around access to hardware supply chains as much as around research talent.
The domestic version shows up as subsidized data center construction and national compute clusters — the US CHIPS Act's manufacturing incentives, the EU's EuroHPC initiative, and Gulf states like the UAE negotiating direct chip-supply agreements with Washington. This resembles earlier eras of industrial policy around semiconductors and energy, where governments accepted that a strategically important input would not be efficiently allocated by markets alone.
The risk in this framing is treating compute scarcity as permanent when it may be transitional. Manufacturing capacity expands, more efficient chip architectures reduce the compute needed per unit of capability, and today's chokepoint could loosen faster than policy built around its permanence anticipates.
What is clear is that compute stopped being a purely commercial procurement decision the moment Stargate needed a president's endorsement to make sense to its investors. Whether that produces a more secure and more broadly capable AI ecosystem, or mostly a new set of gatekeepers controlling who gets to build at the frontier, is being decided now, mostly in rooms without public minutes.