Drone Swarms Explained: How Cheap and Numerous Win

In early February 2026, more than 22,000 drones rose over the Chinese city of Hefei and painted three-dimensional cities and the curved horse-head walls of traditional Hui architecture across the night sky. A single computer commanded all of them. Two months later, on the other side of the world, the Pentagon quietly asked Congress for nearly $55 billion to build autonomous weapons, a roughly 24,000 percent increase over the previous year’s budget for the same office.

Those two events, separated by a few weeks and an ocean, capture the strange double life of the drone swarm. In one world it is choreography and spectacle, the most photogenic technology on Earth. In the other it is becoming the central organizing principle of modern warfare. Understanding how a swarm actually works, and where the marketing ends and the engineering begins, is now essential to understanding both.

From bird flocks to battlefield doctrine

The intellectual roots of swarming lie not in weapons labs but in biology. For decades, researchers tried to reverse-engineer the eerie coordination of starling murmurations, ant colonies, and beehives, where thousands of individuals produce coherent group behavior with no central authority issuing orders. At the University of Pennsylvania’s GRASP Lab, Vijay Kumar became one of the field’s defining figures, showing that quadrotors could form ad hoc teams and fly in tight formation by following simple local rules of alignment, cohesion, and separation rather than executing a single master flight plan.

The leap from theory to demonstration came in stages. In 2015, the Naval Postgraduate School flew a cohesive swarm of 50 autonomous aircraft, validating the multi-agent algorithms and, crucially, suggesting that the architecture which coordinated fifty units could coordinate far more. Around the same time, the concept’s darker potential surfaced in Syria and Iraq, where groups including ISIS jury-rigged commercial drones to drop munitions in loose coordination. These were not mathematically autonomous swarms; they were cheap drones flown by people. But they proved a point that would echo for the next decade: an under-resourced actor could deliver explosives through the air for almost nothing.

The U.S. Defense Advanced Research Projects Agency tried to get ahead of the curve with OFFSET, the OFFensive Swarm-Enabled Tactics program, which ran from 2017 to 2021 and aimed to give small infantry units swarms of 250 or more air and ground robots for urban combat. Working with integrators including Raytheon and Northrop Grumman, OFFSET ran six major field experiments and pioneered interfaces that let soldiers direct swarms using augmented reality, virtual reality, and tablet sketching.

Then theory met a real war. Russia’s full-scale invasion of Ukraine turned out to be the most consequential laboratory for uncrewed warfare in history. Expensive Western loitering munitions struggled against sophisticated Russian electronic warfare, while Ukraine pivoted to mass-produced first-person-view (FPV) drones assembled in workshops for a few hundred dollars apiece. The lesson was brutal and clear: in a contested electromagnetic environment, volume and expendability can matter more than the survivability of any single, exquisite platform. Military planners gave the idea a name, “attritable mass,” and it has reshaped procurement on three continents.

The Pentagon’s expensive course correction

America’s institutional response has been turbulent. In August 2023, then Deputy Defense Secretary Kathleen Hicks launched the Replicator initiative, promising to field thousands of attritable autonomous systems within 18 to 24 months to offset China’s numerical advantages. By late 2025, Replicator had become a cautionary tale. Congressional critics complained of secrecy, slow progress, and the fact that only hundreds, not thousands, of systems had reached troops. The problems were less about technology than bureaucracy: the Pentagon’s acquisition machinery is built to buy aircraft carriers on twenty-year timelines, not to iterate on what is essentially a flying smartphone.

The Pentagon’s answer was to fold Replicator’s mandate into a new organization established quietly in late 2025: the Defense Autonomous Warfare Group, inevitably nicknamed DAWG. It received a modest $225.9 million in fiscal 2026. Then, in the FY2027 budget request released in April 2026, the Department asked for roughly $54.6 billion for it, an increase of around 24,000 percent, with about $1 billion in the base budget and the rest routed through a more flexible reconciliation mechanism. That single line accounts for nearly 15 percent of the proposed defense reconciliation package, a sum that, as commentators have repeatedly pointed out, rivals the scale of an entire military service. (A widely circulated comparison holds that it exceeds the Marine Corps’ $52.8 billion budget request, though the Corps’ full FY2027 request has been reported considerably higher, so the comparison depends on which budget line you use.)

The figure is genuinely historic, and it has drawn historic skepticism. In an op-ed for The Hill pointedly titled “The Pentagon could be about to make a $55 billion mistake,” retired General David Petraeus, former CIA director and commander in Iraq and Afghanistan, and tech entrepreneur Isaac Flanagan called the request “the largest single commitment to autonomous warfare in history.” Their point was not celebration but warning. They argued that in the early Iraq and Afghanistan years, the military poured money into Predator drones while neglecting the people, doctrine, and training needed to use them, and nearly repeated the mistake. They recommended Congress dedicate at least 5 percent of the funding to doctrine, training, and force design, lest the investment produce “not a decisive advantage, but a very expensive inventory.” It is worth flagging this because the quote is often stripped of that context and recycled as pure hype. The author of the line is, in fact, the program’s most prominent friendly critic.

The two worlds of swarming

The defining fact about drone swarms today is that the word means two almost completely different things depending on the environment.

The commercial world operates in permissive conditions: clean radio spectrum, reliable data links, and uninterrupted satellite signals. Light shows like EHang’s record-breaker are the showcase, but the bigger story is industrial: beyond-visual-line-of-sight logistics, agricultural mapping, and infrastructure inspection. Here, precision is everything and electronic interference is not part of the threat model.

The military world operates in contested conditions: jammed signals, spoofed GPS, destroyed local infrastructure, and an electromagnetic spectrum saturated with hostile activity. The premium is not on precision but on surviving long enough to matter when nothing works. This single environmental difference cascades into entirely different hardware, software, and doctrine.

It also explains a dirty secret of the industry: most of the “swarms” the public sees are not swarms at all. A drone light show is closer to a symphony orchestra than a flock of birds. Each drone is a musician following sheet music, precise three-dimensional waypoints, generated in advance with 3D animation tools like Maya or Houdini, all conducted by a central computer. The drones do not negotiate with one another. If one drifts off course because of wind or a fault, it does not consult its neighbors; it triggers a failsafe and lands. Even the terrifying barrages of Russian Shahed drones are, for the most part, large numbers of individual aircraft on preset paths rather than collaborating agents.

A genuine military combat swarm is meant to behave more like a jazz ensemble. There is no central conductor and no fixed score. Each drone understands the basic rules, encoded in algorithms, and constantly listens and reacts to the others, improvising a coherent maneuver around obstacles, jamming, or enemy fire without anyone telling it exactly what to do next. That distinction, orchestra versus ensemble, is the whole game.

How a real swarm works

Strip away the marketing and a true swarm rests on three engineering pillars: networking, decision-making, and positioning.

Networking. Drones need to talk to each other in mid-air while moving fast, changing formation constantly, losing line of sight, and running on tight power budgets. The solution is a Flying Ad-Hoc Network (FANET), in which every drone is simultaneously an endpoint and a router. If the lead drone is too far from a trailing one for a direct link, the drones in between relay the data packets along the chain. Smart routing protocols tag different data by importance (video feeds need bandwidth, basic flight telemetry needs ultra-low latency) and dynamically reroute so critical flight data is never dropped in favor of a secondary sensor stream. Advanced architectures are designed to switch modes automatically: using cellular infrastructure when it’s available and falling back to lean, leader-follower coordination when the swarm flies into an area with no connectivity at all. This is the same coordination problem that a C2 platform solves for a fleet of multirotors, only with no human in the loop and a hostile actor trying to sever every link.

Decision-making. A central control computer is a single point of failure (if it dies, the swarm falls out of the sky) so swarms distribute their computation. One influential approach, Particle Swarm Optimization, is borrowed directly from flocking behavior: each drone is treated as a “particle” moving through a virtual search space, continuously adjusting its path based on its own best-known position and the swarm’s best-known position. This lets the collective fluidly route around new obstacles without a commander calculating each drone’s avoidance vector. The underlying math is brutal (coordinating routes and trajectories across hundreds of moving nodes is what computer scientists call an NP-hard problem) so engineers lean on optimization techniques and hybrid path-planners, often pairing the shortest-path A* algorithm with genetic algorithms that rapidly simulate thousands of formation permutations, to find near-optimal, collision-free solutions fast enough to be useful.

Positioning. This is where the two worlds split most sharply. Commercial swarms achieve their jaw-dropping precision with Real-Time Kinematic (RTK) GPS. Ordinary GPS drifts by several meters as signals bend through the ionosphere, fatal in a densely packed light show. RTK fixes this with a surveyed ground base station that measures the local distortion and broadcasts a correction to every drone in real time, delivering sub-centimeter accuracy. It is brilliant, and it is useless the moment an adversary jams the satellites or the correction link.

Military swarms therefore cannot depend on GPS at all. The cutting edge relies on edge AI and visual-inertial odometry: downward-facing cameras and onboard inertial sensors that let a drone calculate its own speed, altitude, and position by analyzing the optical flow of the terrain sliding beneath it, never pinging a satellite. Because the processing happens locally on the drone’s own computer, the swarm can navigate and even strike targets completely independent of the electromagnetic spectrum. This is the heart of what defense insiders call the “compute war”: the shift from radio-frequency dominance, where the best transmitter wins, to onboard intelligence, where the best silicon and algorithms win.

What’s actually fielded right now

For the United States, the clearest example of the new paradigm is LUCAS, the Low-Cost Uncrewed Combat Attack System, a roughly $35,000 one-way attack drone reverse-engineered from Iran’s Shahed-136 and deployed by U.S. Central Command. In May 2026, the Pentagon selected Shield AI, the autonomy company co-founded by former Navy SEAL Brandon Tseng, to integrate its Hivemind software onto LUCAS as the drone’s “AI pilot.” Hivemind is designed to let a group of drones sense, decide, and act together, rerouting around losses, avoiding obstacles, and adapting in real time, under the supervision of a single human operator who maintains the common operating picture across the whole formation.

It’s important to be precise about maturity here, because coverage often blurs it. LUCAS itself has already been used in combat, reportedly during operations against Iran. But the swarming capability is not yet operational: Shield AI and the Pentagon plan an operational demonstration in the second half of 2026, in which one operator will command a swarm of ten or more LUCAS drones. As Tseng frames the logic, affordable mass is valuable only if it can be coordinated, “mass without coordination is limited in value,” and Hivemind is meant to be the autonomy layer that turns a crowd of cheap drones into an intelligent team. That coordination layer is exactly what tends to break down as fleets scale. Whether it delivers at scale, against real electronic warfare, is exactly what the fall demonstration is supposed to prove.

DARPA, meanwhile, is pushing further out with AMASS, Autonomous Multi-Domain Adaptive Swarms-of-Swarms, an effort to command thousands of autonomous systems across air, sea, and land using a common language, specifically to degrade an adversary’s anti-access/area-denial defenses at the theater level.

Across the Pacific, China’s People’s Liberation Army is advancing aggressively under a doctrine it calls “intelligentization.” In late March 2026, the PLA publicly demonstrated its Atlas drone swarm system, built around a Swarm-2 ground vehicle that carries 48 fixed-wing drones and a command vehicle that can control up to 96 at once, launching one every three seconds, so a full swarm is airborne in roughly five minutes. Built by the state electronics giant CETC, Atlas is explicitly designed to saturate and overwhelm layered air defenses. Analysts read it as tailored for a Taiwan scenario: flood the airspace, force defenders to burn expensive interceptors, and clear the way for missiles and a possible amphibious assault.

The numbers that define the moment

A handful of statistics capture the economics driving all of this.

The starkest is the cost exchange. A Ukrainian FPV drone can be assembled for roughly $400 to $500. A U.S. Switchblade 600 loitering munition runs from about $70,000 to over $100,000. When a $100,000 weapon can be neutralized by a $500 jammer, the math favors whoever is poorer and more numerous, the precise inversion that “attritable mass” is built to exploit.

The battlefield impact shows up in loss figures, though these deserve a careful caveat. Ukrainian government and military sources have credited FPV and short-range tactical drones with somewhere between 60 and 80 percent of Russian equipment losses, with various reports landing at 60 percent, 60 to 70 percent, or higher. These numbers are real and widely cited, but they come largely from Ukrainian official data, the range is wide, and independent analysts note that a large share of FPV drones never reach their targets, particularly armored vehicles. The honest summary: drones now account for a dominant share of battlefield losses, but the exact percentage is contested and should be read as a trend, not a precise measurement.

On the defensive side, the most encouraging data point for Western planners came in October 2025, when the U.S. Navy tested AeroVironment’s containerized LOCUST laser weapon on the deck of the carrier USS George H.W. Bush, the first directed-energy weapon ever fired from a U.S. aircraft carrier. Powered from the ship’s grid, set up in a day, and operable by sailors with under an hour of training, it reportedly destroyed every target presented: a 100 percent kill rate against multiple drones. (Its power is generally reported in the 20-kilowatt class, demonstrated up to 26 kilowatts, though some accounts cite a higher figure.) The appeal is economic: a laser’s “magazine” is effectively bottomless and costs pennies per shot, which is exactly the counter you want against a swarm of thousand-dollar drones you can’t afford to meet with million-dollar missiles.

And the ceiling on benign coordination keeps rising. EHang’s subsidiary kept 22,580 drones airborne from a single computer over Hefei on February 3, 2026, a Guinness World Record that shattered the previous mark of 15,847 set only months earlier. It is proof that, in an environment free of interference, the scale of centralized commercial coordination is, for practical purposes, limited mainly by ambition.

The arguments that won’t go away

None of this is settled, and the debates are substantive.

The first is cost versus capability. Critics of high-end Western procurement point to Ukraine and argue the U.S. is preparing for a clean, high-tech war while ignoring the messy, attritional reality, and should buy $400 FPVs by the million. Defenders counter that basement-built drones lack the range, payload, and encrypted links needed for deep strikes against a peer like China, and warn against over-learning the lessons of one regional conflict. Both are partly right, which is why the actual answer is likely “both kinds of systems,” in proportions nobody has settled.

The second is autonomy versus ethics. As jamming pushes intelligence onto the drone itself, machines increasingly identify targets and execute terminal attacks with no human in the loop. Arms-control advocates and AI-safety researchers warn that lethal autonomous weapons combined with swarming are inherently destabilizing, prone to misidentification, accidental escalation, and violations of humanitarian law. Military leaders respond that there is no alternative: if an adversary severs the command link, terminal autonomy is the only way the mission completes, and refusing it simply guarantees defeat. This is perhaps the deepest unresolved tension in the entire field.

The third is bureaucracy versus speed. Replicator’s stumble showed that the binding constraint is often institutional, not technical. Reformers, including senior Army leaders championing the “Transformation in Contact” initiative, which pushes commercial and military drones down to the squad level, argue that buying drones on aerospace timelines fields obsolete gear on day one. Traditionalists counter that “rapid fielding” can bypass safety, cybersecurity, and testing protocols, fragment supply chains, and complicate training. The DAWG budget is, in effect, a giant bet that the reformers are right.

The fourth is the quietest but most fundamental: what even counts as a swarm. Purists insist on genuine inter-agent communication and collaborative decision-making. Pragmatists shrug: even a hundred dumb, uncoordinated drones force an enemy to light up radars and deplete a Patriot battery. In attritional warfare, the distinction between true collaborative autonomy and mere synchronized volume can be, as one line of argument goes, functionally irrelevant, at least until the defender runs out of interceptors.

Where this is heading

Over the next one to five years, several trajectories look clear. Swarms will break the single-domain barrier: DARPA’s AMASS points toward a near future where an autonomous boat detects a coastal radar, cues an aerial swarm to jam it, and lets ground robots advance under that cover, all coordinated by software rather than a human at every node. Manufacturing will continue migrating from radio dominance to the compute war, with onboard GPUs and neural processors handling vision and inertial navigation so platforms emit nothing for an enemy to detect or jam until the terminal phase.

Defense will tilt decisively toward directed energy and high-power microwaves, because firing a multimillion-dollar interceptor at a thousand-dollar drone is a fast road to bankruptcy; expect systems like LOCUST to scale from tens of kilowatts toward hundreds. And at the speculative edge, planners are already imagining fusions of swarms with hypersonic delivery, a glide vehicle that disperses a cloud of micro-drones to hunt mobile missile launchers, scenarios that would reach into the logic of nuclear deterrence itself.

The throughline connecting the light show over Hefei and the budget request in Washington is the same insight, arrived at from opposite directions: coordination at scale is now a software problem, and whoever masters the software, not the airframe, owns the sky. The commercial world has proven the ceiling is astonishingly high when conditions are perfect. The military world is betting everything on making it work when conditions are anything but. The next few years of demonstrations, from a fall 2026 LUCAS swarm to whatever China shows next, will reveal how much of the promise survives contact with a real adversary.

---

Sources

• Pentagon officials broadly detail $55 billion drone plan under DAWG (Breaking Defense)
• The Pentagon could be about to make a $55 billion mistake (Petraeus & Flanagan, The Hill)
• Pentagon selects Shield AI to plug swarm software into LUCAS drone (DefenseScoop)
• Shield AI to demonstrate AI-equipped one-way attack munitions later this year (FlightGlobal)
• U.S. Military’s Shahed-136 Kamikaze Drone Clone Is Getting Hivemind Swarming Capability (The War Zone)
• DOD touts ‘successful transition’ for Replicator initiative, but questions linger (DefenseScoop)
• We don’t have infantry: Ukraine’s war machine evolves into machine war (Defense News)
• A First Point View: Examining Ukraine’s Drone Industry (Georgetown Security Studies Review)
• DARPA OFFSET program (DARPA)
• EHang Sets New Guinness World Record with 22,580 UAVs (EHang)