Silicon Sovereigns

Walk into any lab today and you can feel the air crackling with the same energy that once surrounded the first particle accelerators. In a recent interview for Los Alamos National Laboratory’s house magazine, Dr. Jason Pruet, who directs the Lab’s new National Security AI Office, explains why. For most of his career Pruet thought of artificial intelligence as a clever instrument, something to slot in beside a spectrometer or a laser. Over the past five years that view collapsed. Transformer models, the architecture revealed by a small team at Google in 2017, kept revealing new capabilities the larger they became, until the line between “tool” and “colleague” started to blur. Pruet now argues that AI is forcing a wholesale rewrite of our approaches to scientific discovery, and that the United States needs a public commitment to computing power on the scale of the post‑war push that built national laboratories. No university, he points out, can afford the exascale machines required to run the latest models, so the federal government must once again underwrite the core infrastructure of discovery.

Pruet’s analogy only makes sense if you recall what happened after 1945. Vannevar Bush, the engineer who orchestrated the research side of the Manhattan Project, convinced Washington that fundamental science was a public good. The result was a half‑century of federal investment in telescopes, reactors, mainframes and payload fairings—capital projects too large for any campus ledger. Pruet argues that AI now occupies the same position: it is the next general‑purpose technology, but its frontiers lie behind GPU walls that cost hundreds of millions of dollars to build.

A few weeks after the Los Alamos interview appeared, Congress heard from the other end of Pennsylvania Avenue. Dr. Douglas Matty, recently appointed as the Pentagon’s Chief Digital and Artificial Intelligence Officer, laid out how the Department of Defense intends to master the same terrain. Matty told lawmakers that the military has organized its AI strategy around three verbs: enable, speed and scale, a sequence that starts with standards, moves through talent and hardware, and ends with global deployment across every classification level. He described a contracting platform called Tradewinds that can move a machine‑learning startup from first pitch to field test in weeks rather than the year‑long cycles that once defined defense acquisition. He also acknowledged that none of this works without serious metal, and so the Pentagon is standing up secure, high‑performance GPU clusters that war‑fighters can reach without waiting in a commercial cloud queue.

Read the two expert reports together and a paradox emerges. On one axis knowledge is flattening. Nowadays a motivated layperson can develop research agendas and outputs that rival those produced by experts. Such cognitive systems, paired with the appropriate knowledge base, can and do facilitate discovery. On the other axis the costs of the cognitive machines are rising. Access to computed inference expands just as infrastructural access to compute contracts. Pruet worries that this asymmetry could ring‑fence the next age of discovery inside a handful of data centers; Matty’s testimony shows how quickly a determined institution can decide to build its own fence.

The Pentagon’s decision is pragmatic: it cannot allow tomorrow’s C2 (command and control) software to depend on cloud resources it does not own. For the civilian science ecosystem the same logic points in the opposite direction. If the frontiers of AI become private estates, the combinatorial explosion of ideas that fuels real progress will slow. The cure is the one Pruet prescribes: treat compute the way the twentieth century treated accelerators and telescopes. Build national GPU parks, allocate time the way astronomers book nights on a shared mirror, and let the real competition return to the layer of insight rather than the layer of electricity.

Both the Los Alamos scientist and the Pentagon technocrat are sounding the same alarm from different watchtowers. Artificial intelligence is no longer a clever add‑on; it is the next infrastructure, as basic to modern power as highways or electrical grids. Either the public rebuilds Vannevar Bush’s pact for the silicon age, or the most egalitarian idea engine humans have ever invented will be switched on only by those who already own the generators.