The Swarm Problem.

A latticework reading of Y Combinator's call for counter-swarm founders — which mental models predict this crisis, which ones fail it, and what new ones the battlefield demands.

▶ Play video

Y Combinator — YC is funding the counter-swarm stack

When the cost curve flips, everything follows.

The drone swarm is not a military curiosity. It is a cost curve that crossed. For most of weapons history, the defender had the advantage: walls beat battering rams, fortresses beat cannons, Patriot missiles beat Scud missiles. The asymmetry favored those who already owned territory. That changed the moment a $500 FPV drone could threaten a $3 million Patriot round — and the calculus collapsed entirely when a thousand of them could fly at once.

Garry Tan's YC pitch is short enough to watch twice in the time it takes to read this essay, and it is worth doing exactly that. It is not a product specification. It is an argument about where the frontier of asymmetry has moved, and why the existing defense stack is the wrong shape for the new problem. The existing stack was built for single targets; swarms are a distributed system problem.

The latticework reading: what this pitch is really describing is a regime change — a moment when the rules of an entire domain reorganize because a single cost has crossed a threshold. Those moments have a predictable structure, and the mental models that describe them are both reinforced and contradicted by what Tan is proposing.

Old models, sharper edges.

Asymmetric cost dynamics is the engine of the entire pitch. When attacker unit cost falls below defender unit cost, you get a reversal — not a disadvantage but a structural inversion. Tan's "$500 drone vs. $3 million Patriot" is not an anecdote; it is a 6,000× ratio, and the ratio is widening as autonomy pushes drone costs toward commodity electronics. The model predicts this inversion would eventually produce what we are seeing: attackers freely spending; defenders rationing responses.

The pitch also gives a sharp new instance of systems thinking. Tan describes the current counter-drone stack as radars, cameras, jammers, interceptors, and humans with binoculars — all heterogeneous, all non-communicating. This is a textbook failure mode: high redundancy in components, no integration at the system level. A well-designed system centralizes the picture, not the hardware. The winning architecture is the one that makes every sensor and every response act as one.

Finally, inversion — Munger's practice of asking what guarantees failure — runs cleanly through the pitch. Work backward from defeat: you lose if your cost-per-kill is higher than the attacker's cost-per-drone, if your systems don't share state, if your non-kinetic options don't exist, and if your defenses are optimized for the last war (individual drones) rather than the next one (swarms with onboard autonomy). Every positive proposal in the pitch is the negation of a failure mode.

Models that do not survive intact.

Deterrence theory — the Cold War orthodoxy that mutually assured destruction keeps adversaries from acting — assumes rational actors with symmetric cost structures. Swarm warfare breaks both assumptions. A non-state actor willing to spend $500 is not deterred by a response that costs $3 million; the economic asymmetry makes escalation cheap for the attacker and expensive for the defender. Classical deterrence was calibrated for symmetric superpowers; it has no purchase on this regime.

Specialization — generally a source of efficiency — becomes a liability in integrated defense. The existing stack of specialists (radar operators, jammer operators, interceptor crews) creates inter-unit latency that a coordinated swarm can exploit. The winning defense platform is the one that fuses all inputs and outputs in real time, under a single decision layer. Specialization within the system survives; specialization of the system does not.

The most subtle contradiction is to kinetic primacy — the assumption that every defense problem has a hardware solution. Tan explicitly calls for non-kinetic defenses: aerosols that foul rotors, streamers that entangle swarms, and — most consequentially — attacks on the autonomy stack itself. When radio jamming is obsolete, the target is the AI decision layer. The conflict moves from physical space to software space. Defense engineers who think only in munitions are playing the wrong game.

New entries for the latticework.

The most portable new model is cost-floor advantage: in any adversarial contest with repeated engagements, the side whose marginal unit cost is lower will eventually dominate by attrition, regardless of who wins individual rounds. This generalizes well beyond drones — to cybersecurity (where one successful breach can cost less than the defenses it penetrates), to legal disputes (where the party with lower per-hour costs can outlast the other), and to any marketplace where volume is the decisive variable.

The Cloudflare analogy delivers a second new model: distributed defense architecture. Cloudflare doesn't stop individual packets; it routes, absorbs, and reroutes traffic in real time across a global network. The same topology — decentralized nodes, shared state, resilience through redundancy — is what the counter-swarm stack needs. A defense platform built like Raytheon (large, hierarchical, slow to respond) loses to a swarm. A defense platform built like Cloudflare (distributed, state-sharing, algorithmically coordinated) can absorb and respond faster than any centralized system. The Cloudflare model is the right shape for the problem.

Finally, autonomy attack surface: as swarms become radio-jam-resistant and onboard-autonomous, the attack surface shifts from the radio link to the AI decision layer. You cannot jam software in the classical sense, but you can deceive it — with spoofed sensor data, adversarial signals, or environmental conditions the model was not trained on. The new frontier of counter-swarm R&D is adversarial machine learning applied to edge hardware running in the field. This is a new discipline, barely named, with enormous strategic value and almost no deployed solutions.

When to reach for which.

The latticework, after the swarm.

The pitch is short, urgent, and designed to recruit. But underneath the founder-facing language is a rigorous structural argument: a cost curve has crossed, the existing defense stack has the wrong topology, and the winning company will look more like a software firm than a defense contractor. That is a strong claim, and the mental models described here either confirm it or explain why it might fail.

The winning companies will look more like Cloudflare than Raytheon. — Garry Tan, Y Combinator

The enduring contribution of this pitch may be precisely that reframe — not the specific technologies proposed, but the recognition that physical security problems are increasingly distributed computing problems. Watch closely whenever a physical domain starts to look like a software domain, because the models that work in one do not automatically transfer to the other — and the gap is where both breakthroughs and catastrophic failures live.