How is the defense tech landscape being reshaped by AI, drones, and the shift from legacy contractors to startups (Anduril, Palantir, Shield AI)
Why the Military Is Buying From Silicon Valley, and Why That's More Complicated Than It Sounds
Based on analysis of a 132-node, 471-edge knowledge graph covering AI, drones, defense startups, and the restructuring of the global defense industry.
The Short Version
For most of the last century, building weapons was the job of a small number of huge companies — Lockheed, Raytheon, Boeing — that had decades-long contracts, massive factories, and deep government relationships. That model is being challenged by software companies and startups that think warfare is becoming more like running an app than assembling a jet. This analysis maps out how that shift is happening, what’s holding it back, and where the contradictions are buried.
The Two Big Labels on Everything
Imagine you’re sorting a huge pile of news stories into folders. Two folders end up with almost everything in them: one labeled “war is becoming software” and one labeled “AI is speeding up how fast you can find and hit a target.” In this graph, those two ideas — called “Software-Defined Defense Paradigm Shift” and “AI Kill Chain Compression” — connect to more than 35 other nodes each.
But here’s the non-obvious part: they’re not really causes of anything. They’re more like category labels. Almost every other node in the graph connects to them by saying “this validates the label” or “this is an example of the label.” The actual cause-and-effect relationships are happening in the middle of the graph, between smaller, more specific nodes. The two big hubs tell you what theme everything fits under — they don’t explain why any of it is happening.
Ukraine as the Graph’s Only Lab
Science requires experiments. If you want to know whether a drug works, you need people who took it and people who didn’t. This graph has a problem: almost all of its real-world proof comes from one place — Ukraine.
The “Ukraine Defense Tech Laboratory Effect” node connects to at least 12 other nodes as their proof source. Fiber-optic drone cables that can’t be jammed? Proven in Ukraine. Cheap drones destroying expensive tanks? Ukraine. Faster targeting with AI? Ukraine data. A new European defense startup called Helsing? Validated by Ukraine contracts.
This isn’t a criticism of Ukraine’s importance — it genuinely is where most modern drone and AI warfare has been tested under real conditions. But the graph encodes a structural fragility: a large portion of the theoretical claims rest on evidence from a single ongoing conflict. If that conflict ends, or turns out to be exceptional in ways we don’t yet understand, a lot of the nodes that say “proven” would need to be reclassified as “plausible but not yet demonstrated elsewhere.”
The Battery Problem at the Center of the Drone Revolution
Here is one of the less obvious findings. The whole idea of “cheap drones that can be lost without catastrophe” — the doctrine called attritable warfare — depends on a $400 drone being able to destroy a $4.5 million tank. That math only works if the drone is actually cheap to build.
What makes it cheap? Batteries. Specifically, a battery chemistry called lithium iron phosphate, or LFP. And who dominates global LFP manufacturing? China.
So the graph contains a loop that looks like this: the US wants to build lots of cheap autonomous drones to compete with China; those drones are cheap because of battery chemistry that China controls; therefore the cost structure that makes the strategy viable depends on the supplier the strategy is designed to counter. The drone cost revolution and the supply chain trap are not separate problems — they’re wired together through battery chemistry.
Four Self-Reinforcing Spirals (And One That Cuts the Other Way)
The graph contains several feedback loops — situations where A causes B, B causes C, and C causes A again, spinning faster over time.
Loop one is about the defense industry itself. The shift toward software-based defense creates conditions where a few large software platforms (from companies like Anduril or Palantir) start absorbing contracts. That consolidation makes the software paradigm stronger, which attracts more contracts, which consolidates further. Nothing in the graph acts as a brake on this process.
Loop two runs through Ukraine. Combat generates new drone innovations. New drone innovations drive the market for counter-drone systems. Counter-drone systems generate more combat data. More combat data feeds back into new innovations. This loop has no stop condition in the graph — it just keeps running.
Loop three is happening in Europe. Trade tensions with the US are causing European countries to fund their own AI defense companies (notably Helsing, a German startup). European investment in Helsing operationalizes European strategic independence. That independence makes European countries more willing to buy European, which further separates their procurement from American suppliers, which reinforces the original trade tension. Poland is specifically encoded as a satellite beneficiary of this loop.
Loop four runs in reverse — it constrains rather than accelerates. The AI-powered targeting that the US military is developing requires advanced chips that only Taiwan’s TSMC can manufacture at scale. But TSMC is located on an island that the US military AI would theoretically be used to defend. The capability depends on the supply source, and the supply source depends on the capability. It’s not a growth spiral — it’s a ceiling. The more dependent US military AI becomes on TSMC chips, the more the vulnerability of TSMC’s location becomes a strategic weakness in the very systems that are supposed to address that vulnerability.
The Governance Window Is Closing, and Nobody in the Graph Is Opening It
There are ongoing international discussions about whether autonomous weapons — systems that can select and engage targets without human approval — should be regulated or banned before they proliferate widely. The graph has a node called “LAWS Governance Pre-Proliferation Window,” which represents the period during which such regulation might still be possible.
Four separate things in the graph are closing or undermining that window: the demonstrated use of AI targeting systems in operational contexts, the split between AI companies that refuse military contracts and those that don’t, the irreconcilable disagreement between AI safety advocates and military users, and the gray market spread of commercial chips that can run military AI in countries that shouldn’t have access to it.
The non-obvious finding: not a single edge in the graph represents anything that strengthens the governance window. No treaty, no technical constraint, no international agreement points toward reopening it. The graph doesn’t encode governance as a solved problem, or even as a solvable one in its current structure.
Some Things That Aren’t Obviously Connected
A few relationships in the graph deserve attention because they’re not intuitive:
Budget cuts increase platform lock-in. The government program associated with efficiency-driven spending cuts (called DOGE in the graph) is encoded as amplifying the advantages of existing software platform holders. When you cut budgets and need to consolidate contracts, the vendors who already have the integrations in place win. The graph treats austerity and concentration as directionally linked rather than opposed.
One AI company refusing military contracts strengthens a rival political network. When Anthropic declined to build tools for certain military applications, the graph shows this increasing the relative concentration of frontier AI access among a different set of actors — specifically the network around Peter Thiel, Palmer Luckey, and Marc Andreessen. This is not about intent. It’s structural: if the most capable AI model available withdraws from a market, the remaining capable models gain leverage in that market. The graph encodes the consequence independent of the motivation.
Scale AI is both feeding and competing with its biggest customer. Scale AI provides the data infrastructure and training pipelines that make Palantir’s Maven system more capable. Scale AI also has its own military planning product that competes directly with Maven. The graph shows both relationships as simultaneously true, but does not encode which one is dominant.
The Six Questions the Graph Can’t Answer
The analysis ends with six specific testable hypotheses — places where the graph makes a structural prediction that could be checked against real-world evidence. Can new manufacturing at scale relieve the munitions shortage? Will directed energy replace electronic warfare as the primary way to stop drones? Will Europe’s buy-European procurement rules actually cause a measurable split from US defense suppliers? Can China run effective military AI without the chips that US export controls are designed to block?
These aren’t rhetorical questions. They’re places where the structure of the graph makes a specific, falsifiable prediction. The graph is built in a way that, if you tracked the right metrics over the next three years, you could confirm or disconfirm each one.
Bottom Line
The defense industry is going through a restructuring driven by software, AI, and cheap autonomous systems — but the graph reveals that several things assumed to be straightforward are actually tangled:
- The cost revolution in drone warfare depends on Chinese manufacturing dominance in battery chemistry.
- Almost all real-world proof of concept comes from a single conflict theater.
- The same chips enabling US military AI are produced in a location that US military AI is supposed to defend.
- Efficiency-driven budget cuts are concentrating market power in existing platform holders, not distributing it.
- The window for international regulation is being closed by multiple independent forces simultaneously, with no counterforce encoded anywhere in the graph.
- European strategic independence and US-allied integration are happening at the same time, in the same countries, with no resolution of the contradiction currently represented.
The graph is not a prediction. It is a map of structural relationships as they currently exist. What it shows most clearly is that the defense technology transition is not a clean story of innovation replacing incumbency — it is a system full of dependencies, loops, and contradictions that make the outcome significantly less certain than the dominant narrative suggests.