Poll check: 44% still think AI is a bubble (249 votes). Interesting.

One reason is that people keep framing the Pentagon angle like it’s something new. It’s not. “Defense uses AI” has been true for years. The DoD has been pushing AI since at least Project Maven (2017), then the JAIC (2018), and today it sits under the CDAO. Defense R&D has funded “AI-ish” work for decades (DARPA timeline, expert-systems era, etc.). So the headline is not “AI reached the military.”

What is new is that frontier LLM capability is now good enough that it’s being deployed into classified environments — and that the policy fight is happening in public, between the leading labs, with real consequences. That’s a different level of adoption, and a different level of strategic importance.

There’s also an important nuance in the “who competes with whom” story: Palantir is not a frontier model lab like OpenAI or Anthropic — it’s the workflow / orchestration layer where models (including Claude) get deployed into real defense systems (Gotham/Foundry/AIP, Maven Smart System). In practice, Palantir is usually a distribution and integration channel, not a direct competitor to the labs. That’s why partnerships like Palantir + Anthropic exist — and why NATO adopting Palantir’s Maven capability is a big tell for Europe’s current posture.And yes, it raises a fair question: if the control point sits in the deployment layer, why did the “red lines” debate suddenly become a model-provider issue now — and not earlier, when platforms like Palantir were already deeply embedded?

Zoom out: China is explicitly accelerating military modernization including military AI (DoD’s annual report is blunt about this).

Europe has initiatives and debate — and yes, there’s Mistral — but at the “platform stack” level Europe is still heavily dependent on US tech and US distribution. And in defense, that dependency is now visible in procurement: Palantir-style deployment and orchestration systems are becoming the default.

AI bubble? The real question is: who owns the workflow and the control point when frontier models move from demos → enterprise systems → classified operations.

We mapped the broader market side of this shift (dispersion, repricing, agent layer, control-point migration) here: studioalpha.substack.co…

Sources (plain URLs, copy/paste ready) Source: war.gov/News/News-Stori… Source: war.gov/News/Article/17… Source: csiac.dtic.mil/articles… Source: openai.com/index/our-ag… Source: investors.palantir.com/… Source: shape.nato.int/news-rel… Source: media.defense.gov/2025/…