AI Aluminum
On June 17th, the U.S. Air Force handed Anduril Industries a contract for its FQ-44 autonomous combat drone, making it the first new entrant to win a U.S. fighter aircraft program since the 1970s. The Silicon Valley startup beat Lockheed Martin, Northrop Grumman, and Boeing – companies that between them have defined American air power for generations – to the target. Anduril, a defense technology firm founded in 2017, did not win on relationship or on legacy: it won on AI-native architecture.
The shift this represents is psychological as much as commercial. Defense procurement is arguably the most bureaucratic, relationship-driven, clearance-protected industry on earth. If AI-native vertical integration can break this barrier, anything is now open for re-negotiation.
The Great Compression Part Two: The Intelligence Trap
The compression of the human intermediary layer across the economy – the subject of this series’ opening piece – raises a question that is both philosophical and financial: if the old middlemen are disappearing, and if the AI models replacing them are themselves becoming commodities, where does value go?
What made the Air Force announcement structurally significant was a detail that went largely unnoticed: the service deliberately separated the drone hardware from its AI software, specifying that the intelligence layer could be upgraded or replaced independently of the platform. By doing so, the Air Force drew a line that the market is still catching up to – and then immediately complicated it. The aircraft is the delivery vehicle. The intelligence tier is what matters, except that intelligence is also commoditizing fast.
The commoditization signal had already arrived earlier this month, when Google cut the price of its AI plan by nearly 40% overnight. OpenAI is reportedly considering steep token-price cuts as competition with Anthropic intensifies. The models themselves are beginning to resemble a capital-intensive utility more than a premium software business. As intelligence becomes cheaper, the investment question shifts to what models cannot easily access: proprietary data, regulated workflows, institutional trust, and the systems that turn AI output into real-world action.
The answer, in the most durable cases, is proprietary domain data combined with deep sector integration. Anduril is not a defense company that adopted technology: it’s a technology company that chose defense as its vertical. Palmer Luckey, who sold Oculus to Meta when he was just 21, founded Anduril alongside veterans of Palantir with a specific thesis: Silicon Valley had abandoned defense, leaving a widening gap between what the military needed and what the traditional primes could deliver.
Where Lockheed and Boeing run bid-led organizations optimized for cost-plus contracting cycles measured in decades, Anduril built a product-first company that moves at software speed, focused on cheap, autonomous, attritable systems designed to be deployed and lost without catastrophic cost. The competitive edge that results has nothing to do with which model runs underneath it. It is purpose-built architecture, mission-specific design, and the kind of deep operational embedding that no generalist technology company can shortcut and no traditional prime can easily imitate.
Palantir built the same competitive edge a decade earlier, at the intelligence level. Their forward-deployed engineers embedded themselves inside classified environments, building proprietary data structures around defense and intelligence that competitors cannot access, let alone replicate. The model is almost beside the point. What matters is the institutional trust above it and the data structure underneath it.
The same logic plays out in banking, and the psyche shift there is equally striking. JPMorgan Chase is not an obvious candidate for AI leadership – a 150-year-old Wall Street institution steeped in regulatory obligation and institutional conservatism. Yet it has become arguably the most digitally aggressive major bank outside the fintech world, spending north of $17 billion annually on technology and deploying AI across trading, risk, legal document review, and client services. JPMorgan’s AI advantage over any fintech competitor is not compute – it is 150 years of proprietary transaction data, credit history, and market intelligence, combined with the institutional will to deploy it at scale. Many large banks sit on comparable reserves; few have built the machine to turn them into a competitive weapon. The model commoditizes; the data does not – but only in the hands of someone with the commitment to exploit it.
The pattern across defense, banking, and every sector where this is playing out is consistent: the prize migrates to whoever owns the scarce position that generic models cannot substitute for. Right now the prize sits with domain data and deep sector integration. What’s forming above it is agentic orchestration – systems that coordinate networks of specialized AI agents across high-stakes workflows: routing battlefield targeting decisions, flagging fraud across millions of simultaneous transactions, managing the exception-handling that no single model can resolve alone. Palantir’s AIP platform is the most mature example of this emerging tier, and it is no coincidence that the same company that mastered domain-specific data is now positioning for that orchestration tier. Salesforce’s Agentforce is building toward the same position from the enterprise side. The race for this trophy is not yet decided, but the companies that already own those domain data advantages are the natural favorites to own the control plane above them.
The stack, in other words, keeps moving upward. Value migrates to the next bottleneck, then the next. And below all of it – the models, the sectors, the orchestration layer – something has to hold the weight. Every control plane needs a floor. What that floor looks like, who owns it, and why it may be the most durable investment thesis of the AI era is the subject of the next piece.