2027: the year of autonomous drone swarm warfare

The Widening Grey Zone of Autonomous Killing, Part 2

Berend Watchus

38 min read

·

5 days ago

--

Listen

Share

Author: Berend Watchus · Independent AI & Cybersecurity Researcher · November 1, 2026

Press enter or click to view image in full size

Berend Watchus · Independent AI & Cybersecurity Researcher · November 2026

A note on method: I am not a military, weapons or drone expert. I am an independent researcher who has written about twenty articles on drones and autonomy, built entirely on open-source material: news reporting, official statements, think-tank analyses, academic papers and public video. Where claims rest on a single source or could not be verified, I say so. The value of this piece lies in connecting what is publicly visible, not in insider knowledge.

Part 1 in brief

In Part 1 (1 August 2026) I argued that the familiar split between human-controlled and fully autonomous drones misses where the technology actually sits. Many munitions are supervised by a human until the link is jammed, spoofed or broken, and then finish the strike on their own. A human starts the chain; a machine closes it.

Only the deploying side can know which mode a weapon was in, and no forensic standard exists to reconstruct it afterwards. That fractured attribution can turn into practical deniability. Cheap onboard AI, spreading electronic warfare and one-way tactical incentives are all widening the zone.

Part 2 corrects one piece of that argument, extends several others, and checks them against what happened between August and the end of October 2026. The short version: everything Part 1 described is accelerating, and new forces have joined it. Production is outrunning pilots. Quantity is beating size. Onboard intelligence is becoming nearly free. The swarm is becoming one distributed mind. And the bodies that physical AI needs are being perfected on stage.

The binary was wrong at both ends

Part 1 said that jamming a manual weapon makes it fail safe. For an armed drone, that holds only if safe means not aimed. A jammed drone carrying a warhead does not vanish. It comes down somewhere, and with an impact fuze it may well go off where it lands. There are three failure modes, not two.

Fail safe: the drone returns, lands, or is rendered inert. No one decides an effect, because there is none.

Fail dumb: the drone falls where physics takes it. The explosion is unaimed, not avoided, and no one decides where it lands.

Fail lethal: the onboard system completes the attack on a target it selected. A machine decides, in the final seconds.

Fail dumb is in some ways the hardest case for accountability. Terminal autonomy at least leaves a system that “decided” something, with software and logs that could in principle be examined. A jammed drone falling on a house leaves no decision at all. The operator aimed elsewhere, the jammer belonged to the other side, and the impact point was chance. On 30 September a Geran-type jet drone crossed into Moldova and crashed and exploded near a road in the Orhei district: a small illustration of how unaimed endings travel.

Jamming is therefore not purely protective. In dense areas it can turn an aimed strike into a random one.

Platforms and one-way munitions are different problems

A gunship or armed reusable drone separates the platform from the weapon. Every shot needs its own decision, human or machine. Cut the link and, absent an autonomous engagement mode, nothing fires. Here fail safe really is safe.

A loitering munition is the weapon. The explosive commitment is made at launch; everything afterwards only steers where it ends. Losing the link cannot stop it, only blur it. It behaves less like a gunship than like an artillery shell: once fired, it will detonate somewhere.

That moves the accountability question. For platforms it is “who decided to fire?” For one-way munitions the decisive act is the launch, and responsibility should attach to it, including for foreseeable outcomes. Launching into an environment where jamming is expected makes a stray impact a foreseeable result of the launch, not an accident authored by the enemy’s jammer. International humanitarian law already reasons this way about artillery, through precautions in attack and the ban on indiscriminate attacks. That framework may fit loitering munitions better than the trigger-centred framing the autonomy debate usually borrows.

The honest alibi

“Sorry, it fell, we lost control” is not a lie. Most of the time it will be true. That is precisely its power: because the true version is common, it also covers the cases where it is not true. A genuine accident and a deliberate strike dressed as one can leave identical wreckage and identical statements.

It can also become policy without any individual lie. A force that routinely launches into heavy jamming over populated areas knows some share of its drones will land at random. Each “oops” is honest. The pattern is a choice.

Late-2026 reporting adds a second complication: link-loss that is chosen rather than imposed. In a case described below, the drones reportedly cut their own link before searching for targets. Chosen link-loss creates a clearer human decision to trace, but once “cut the link and go autonomous” becomes a tactic, an outsider can no longer tell a jamming fallback from an intended autonomous attack.

Between peace and war

Attribution does not only fail between human and machine decisions. It also fails between peace and war. In autumn 2025, drones flew over Belgium’s Kleine-Brogel air base on consecutive nights, along with airports, the port of Antwerp and the Doel nuclear plant. Belgium’s defence minister described small drones probing Belgian radio frequencies first, followed by larger ones, with operators switching frequencies once the jammers were found. It was a modern version of the Cold War “ferret” flights that mapped Soviet radars, carried out with cheap drones over a NATO country in peacetime. No drone was downed, so no one could be proven responsible. Such operations may even be contracted out to criminal networks to keep that deniability.

I analysed these incidents in November 2025. A year later, the pattern is the same: the attribution gap from Part 1 now extends to European airspace in peacetime, where the question is no longer only who decided, but whether anyone can prove a state was involved at all.

What changed since August

A reported fully autonomous strike. In mid-August, Small Wars Journal reported a claim by a Ukrainian drone developer that a swarm of ten drones severed its communications link, searched for targets with onboard AI and attacked with no human approving the final engagement. If confirmed, it would be the first publicly reported fully autonomous lethal use against people. It rests on a single industry source and has not been independently verified, which is the attribution problem of Part 1 playing out in public.

Swarms as routine. On 1 October, Militarnyi reported that Ukrainian forces had been using AI-coordinated swarms on the battlefield for most of the past year, with a 100-drone test said to be coming. Ukrainian swarms already divide roles: some drones act as decoys, others conduct surveillance, others attack.

Scale on paper. Kyiv Post reported in August that Ukraine had drawn up a plan for up to 1,000 autonomous drones against Moscow’s airports, with drones diving on recognised targets without operator confirmation. It was reportedly presented to the president; there is no public sign it was carried out.

The other side closes the navigation gap. From late September, Ukrainian monitors reported a Geran-5 variant (“K5”) navigating by matching camera images of the ground against stored terrain data, a cruise-missile technique that keeps flying when satellite navigation is jammed. The designation and system are not officially confirmed. If accurate, it extends Part 1’s Shahed lineage: jamming no longer reliably produces fail dumb; it increasingly produces fail-and-continue.

Adaptation at the speed of software. According to a French parliamentary analysis cited by the Canadian Press in October, opposing forces in Ukraine need on average about six weeks to counter a new battlefield technology. Hardware is simple; software and communications change so fast that what works today may fail within weeks. Soldiers modify drones at the front with code and 3D printers.

Governance inches forward. In September the CCW Group of Governmental Experts adopted its final report by consensus, including elements of an instrument and the first agreed working characterisation of lethal autonomous weapon systems. In August the UN Secretary-General and the ICRC President asked states to launch negotiations on a new protocol at the CCW Review Conference this November. In October the ICRC told the UN First Committee that states should prohibit unpredictable systems and those designed or used to target people. The 2026 treaty deadline has been missed, but a negotiating mandate is now on the table.

Production outruns pilots

On 11 October, Ukraine’s defence ministry reported that its procurement agency had contracted 5.5 times as many drones as in the same period of 2025, with deliveries up almost 3.7 times. Drone output can nearly quadruple in a year. Trained operators cannot.

Once production outpaces the people who fly the drones, the only way to use them all is for fewer people to control more of them. Forbes reported in August that a single US soldier can now operate 21 small aerial drones at once, plus four unmanned ground vehicles, in breaching exercises. Ukraine reported more than 22,000 frontline missions by ground robots in three months, and in April it took an enemy position using only ground robots and explosive drones, with no Ukrainian casualties.

One operator per drone becomes one operator per group, then per swarm, with more of the final phase left to the machine. That is a fourth force widening the grey zone, alongside jamming, cheap onboard AI and one-way tactical incentives.

Quantity over size

In 1942 Germany besieged Sevastopol with Schwerer Gustav, an 80 cm railway gun weighing about 1,350 tonnes. It took weeks to set up, needed a crew of well over a thousand and fired a handful of shells a day. It was the high point of one idea of military power: one enormous, sophisticated weapon.

The 2026 battlefield runs on the opposite idea. Ukraine’s presidential adviser on electronic warfare, Serhiy “Flash” Beskrestnov, said in March that Russia was losing its expensive reconnaissance drones (Orlan at roughly $87,000–$120,000, ZALA at about $70,000, Supercam at around $400,000 per system) and was replacing them with cheap Molniya drones costing about $300–$800 each. Each Molniya carries a camera, a mesh modem and an AI module. Even a drone that costs less than a phone now flies with onboard intelligence.

In October he took the argument further. Rather than one expensive reconnaissance drone, he said, it would be more effective to field fifty cheap ones: far inferior individually, but giving more coverage for longer. The same logic applies across the front. Cheap interceptor drones defend against Shaheds because expensive missiles cannot be traded for cheap drones forever. Instead of one powerful jammer covering a whole sector, the front is covered in dozens of cheap ones, so losing one costs almost nothing. Progress remains the engine of war, he concluded, but new weapons must be cheap, quick to produce and scalable.

The saying that quantity has a quality all its own is often attributed to Stalin, though its origin is disputed. The 2026 version is sharper: quantity now carries its own intelligence.

Mass also changes how weapons are made. Ukrainian manufacturers split production across many sites, move often and go underground to survive constant attack. In October the Croatian drone company ORQA described building an underground component factory in Ukraine with the Ukrainian maker General Cherry, and said Western industry has lessons to learn from it. Dispersed production is less efficient but far harder to destroy. It is the industrial counterpart of the swarm: no single target whose loss stops the whole.

The same technology also saves lives. In October, Ukraine’s 475th Separate Assault Regiment used SkyFall’s StarGate, a six-rotor drone with an enclosed capsule, to evacuate two badly wounded soldiers from no man’s land about 5 km from the Russian border, where neither vehicles nor helicopters could reach them. The drone uses autonomous navigation to resist Russian electronic warfare, and each flight took less than 15 minutes. The regiment now plans a dedicated medical aviation section. The autonomy that lets a drone strike when the link is jammed also lets it carry a wounded soldier home. The technology is the same; the decision is what differs.

The brain gets cheap

In October 2025 I described two efficiency gains that together reshape small drones. The first was an event-driven, spiking-neural-network architecture that would cut a drone’s obstacle-avoidance power from about 4.8 watts to 75 milliwatts, the “observe and orient” half of the OODA loop. Independent hardware has since shown that order of magnitude is real: TU Delft flew a drone on fully neuromorphic vision and control, with the network adding only 7–12 milliwatts, and ETH Zurich’s ColibriUAV ran a closed-loop neuromorphic pipeline on about 47 milliwatts. The second was my Law of Optimized Complexity framework, which proposed a 16.4× gain in decision complexity per unit of energy, the “decide and act” half.

The point is not obstacle avoidance alone. It is the entire onboard AI: target recognition, navigation without satellite signals, swarm coordination and the decision to strike. That AI is already flying. Part 1 described the Shahed-136 MS with an Nvidia Jetson for autonomous terminal targeting, and this autumn’s Geran-5 variant reportedly navigates by matching camera images of the terrain.

Today that intelligence still costs watts, weight and heat. On a small drone, every gram of computer needs more lift, which needs a bigger battery, which adds weight. On a waiting drone, perched beside a road in ambush, the computer becomes the main power draw. Jamming makes this worse: if the link holds, the heavy thinking can stay on the ground; once it is cut, the intelligence has to fly on board.

When seeing and deciding cost milliwatts instead of watts, the last economic barrier to onboard autonomy falls. Even the smallest drone can afford to think for itself, wait in ambush for long periods and fly with almost no thermal or electronic signature. When perception and decision run at machine speed on board, the human operator becomes the slowest step in the loop. The decision to strike then moves from the operator to the drone itself, and the grey zone stops being an edge case.

One mind, two hundred bodies

Nor does each drone need to carry its own complete brain. Spread across 200 drones, the intelligence becomes a distributed system: scouts see for the whole swarm, roles divide the computing load, and losing individual drones does not stop the swarm thinking. Jamming the long link to a human operator only pushes the decision further into the swarm, whose short internal links are far harder to cut.

Then the decision to strike no longer sits in any single machine, let alone a person. One drone spots, another ranks, a third strikes. The attribution problem from Part 1 does not just fracture here. It dissolves into the network.

In December 2025, writing about the PrivLLMSwarm paper (arXiv:2512.06747), I described the end of “dumb drones”: swarms in which every drone runs its own observe-orient-decide-act loop, coordinated over secure links without a central controller, with AI reasoning performed on encrypted data so that no single drone or eavesdropper sees the whole picture. That privacy protects the swarm, but it also closes the last forensic door. A downed drone from such a swarm holds neither the full data nor the decision. There is nothing in the wreckage to reconstruct.

The remedy I proposed then still stands: mandatory audit trails, human oversight for high-consequence actions and a ban on autonomous lethal targeting without human control. LLM-driven swarms reason in language, and that reasoning can be logged. A decision log, kept like an aircraft’s black box, is the one tool that could turn a swarm’s distributed decision back into an accountable one.

The last brake: a sense of wrong

In July 2026 I asked three leading language models, Gemini, Claude and Grok, to draw a simple L-shaped hallway from my own measured plan. All three failed completely. They invented pillars in open floor, swelled a cupboard until it blocked the front door, and could not be corrected even when shown photographs of the real space. Their deeper flaw was that they had no sense of wrong. A human who sketched a door 30 cm wide would laugh before the pencil left the paper; the models labelled it neatly and defended it. Language models excel at strategic analysis and fail at tactical, physical reality.

For now, that split is one of the last brakes on autonomous warfare. The systems that reason strategically cannot reliably act in the physical world, and the systems that act in the physical world, such as a drone matching terrain or recognising a vehicle, are narrow specialists without broader understanding.

That brake is being engineered away. In May 2026, NVIDIA released Cosmos 3, an open “world foundation model for physical AI” combining reasoning, world simulation and action for robots, with an edge version for on-device use announced. Robotics researchers note that such models can still be visually plausible but physically wrong, the same missing sense of wrong my hallway exposed. But they treat it as a problem to solve, not a permanent limit. The data to solve it is being gathered from the ground up. Consumer drones with 360-degree cameras now turn a few minutes of flight into centimetre-resolution, photorealistic 3D models of real places, and since February 2026 an open standard lets those models move between platforms. Such models are not just maps. They are training grounds in which a machine can learn what a real hallway, street or building actually looks like.

The body is arriving faster than the mind. In May 2025, in Hangzhou, humanoid robots fought the first robot boxing match, throwing punches, hooks and spinning kicks and getting back up after falls, still steered by human operators. Nine months later, on China’s Spring Festival Gala in February 2026, dozens of humanoids performed martial-arts routines with weapons, which their maker says were fully autonomous, along with three-metre somersaults and wall climbing. In April 2026, a humanoid ran a half marathon in Beijing seven minutes faster than the human world record. In September 2026 a video of a six-foot humanoid sparring with a TikTok influencer in an octagon went viral as the “first-ever” human-versus-robot MMA fight; how much of it was staging or remote control remains unclear, and that uncertainty is itself part of the story. One manufacturer alone expects to ship up to 20,000 humanoids this year. These are entertainment and showcase machines, not weapons. But balance, speed, falling and recovering, and coordinated combat movement are exactly the physical competences that language models lack, and they are being mastered on stage, in public, at production scale.

The real frontier is not either capability on its own, but their integration. Vision-language-action models already put language reasoning, vision and physical action into one system; Cosmos 3 pairs a reasoning component with a world simulator; Google DeepMind’s Gemini Robotics builds embodied reasoning on top of a language model. The end point is a single unit with a strategic mind and a tactical body, or a layered version of it, with a strategic reasoner in the swarm leader and physically competent bodies in every drone, sharing one picture of the battlefield. The pieces I have written about separately (a verbal decision loop, neuromorphic perception that costs milliwatts, a self-model that knows its own state, intelligence spread across a swarm) are the layers of exactly that unit.

When it comes together, the operator’s last advantage, being the one who knows that a wall cannot float, disappears. My earlier work on synthetic self-models suggests that a sense of wrong comes from embodiment, from a body that meets consequences. An integrated machine may acquire it. Whether that sense of wrong is built to include more than physics, so that it recognises a civilian, a hospital or a surrender as something not to be struck, is a design decision being made now, mostly without public debate.

Faster engines, shorter decisions

As drones move from propellers to jets, the time available for any human decision shrinks. Russia’s jet-powered Gerans reach about 370 km/h (Geran-3) and up to roughly 500 km/h (Geran-4), against 180–200 km/h for the propeller versions, at about $80,000 per unit. IEEE Spectrum notes that at these speeds small manoeuvres cause misses and human reaction time becomes a limit, pushing defenders toward AI-guided interceptors. The speed race on both sides feeds the autonomy race.

The next efficiency frontier is the small turbofan. GE Aerospace and Kratos are developing compact, low-cost turbofans for cruise missiles and expendable drone “wingmen,” while the jet Gerans still rely on cheap turbojets. How far these small engines can be pushed depends partly on effects most readers never hear about: blade-row timing, wake recovery and flow control, in which small, well-timed inputs shape large airflows. I explored these mechanisms in a 2025 article on turbofan efficiency, assembled with an AI from public research in a single afternoon. That speed of synthesis is itself part of the story this article tells.

What still limits drone warfare

For all the acceleration described here, drone warfare still runs into hard limits. They fall into five groups.

Physics. Every drone is bound by its energy: battery or fuel sets its range, flight time and how long it can wait on standby. Speed costs energy, and flying at top speed drains it fastest. Payload competes with fuel, so every extra kilogram of explosive shortens the flight.

Waiting is becoming free. One physical limit is quietly disappearing: the energy to wait. As I described in August, the “nuclear battery” headlines promising drones that fly forever were never physically possible; a coin-sized betavoltaic generator delivers about 100 microwatts, a trickle rather than a tank. The real progress is ordinary chemistry. Lithium-thionyl chloride cells of the kind used in utility meters and remote sensors are rated for 15 to 40 years in low-drain use, and industry already pairs them with a small capacitor that stores the trickle and releases it in a short radio burst. Combined with event-driven sensing that costs milliwatts, a drone or sensor can lie dormant for months or years and wake only when something passes. The limit that remains is the energy to move, not the energy to wait.

A landmine that can fly. If the energy for waiting is separated from the energy for flight, a device could lie dormant for days or years and still wake with full strength. Anti-personnel mines are banned for most states precisely because they wait indefinitely and cannot tell a soldier from a child years later, and Verdun’s farmers still dig up shells more than a century on. A long-dormant autonomous drone would stretch Part 1’s handover from human to machine across time as well as distance: deployed by a human in one war, it might act in a different situation, on a moved front line, perhaps after the war itself has ended. At that point no one decided the strike at all.

A spy gadget with a long fuse. In 1945 Soviet schoolchildren gave the US ambassador in Moscow a carved Great Seal hiding a listening device with no battery at all; it slept until a radio beam woke it, and it went undiscovered for seven years. A dormant drone merges the two old ideas, the mine that waits and the bug that listens, into one cheap device that can watch for years, then act, and no one outside can tell which it is doing.

Underground, the link always breaks. Rock and concrete block radio and satellite signals, so a drone or robot in a tunnel cannot be steered from outside the way it can in open air; it must either trail a cable or act on its own. Militaries already send small drones and ground robots into tunnel networks rather than risk soldiers, and the same sheltered, rarely inspected spaces under cities, from sewers to metro lines, are where a dormant device could wait longest unseen. Even there, though, the surface can betray what happens below: radar satellites using interferometry measure ground movement down to millimetres, and peer-reviewed studies have used it to map the subsidence caused by tunnelling under cities.

Guidance. What a drone carries ranges from a simple bomb to a smart munition to the drone itself as a kamikaze. How it finds its target depends on its link. Radio control is limited by range and by jamming, so the farther and harder the fight, the more the drone must rely on its own autonomy. This is where the physical limit turns into the grey zone: when the link breaks, autonomy takes over.

Industry. Drones need money, factories, components and people. Production in Ukraine and Russia is measured in millions a year and thousands a month, and factories are dispersing and going underground. The weaker links are often the supply chain, with motors, batteries, chips and cameras from a handful of suppliers, and the trained operators and engineers who keep the systems adapting.

Will. No production line runs without the political decision to keep producing and fighting. This is the limit that decided Verdun, and it remains the one no technology removes.

Mind. Finally, a drone is limited by its intelligence: what it can perceive, understand and decide, and, in a swarm, how well many drones think together. This is the limit falling fastest, and the one that matters most for accountability.

Two outside limits act on all five. The enemy’s defences (jamming, interceptors, guns, microwaves and lasers) turn every advantage into a race. And weather (wind, rain, icing, fog and darkness) still grounds or blinds small drones more than any adversary.

Most of these limits are variables. Funding rises and falls with budgets. Production rises and falls with factories, components and strikes on them. Political will rises and falls with elections, casualties and public mood. Defences gain and lose ground in a six-week cycle. But one thing only moves in one direction: research and development. Knowledge does not go backwards. A design proven at the front stays proven; a model trained on a million flights stays trained; a technique published or captured stays known. Even if a war ends and budgets shrink, the know-how remains, ready to be scaled again. That is the ratchet underneath every other limit, and the reason the grey zone widens even when the fighting pauses.

Boast what deters, hide what works

The public picture of autonomous warfare is shaped by what governments and companies choose to show. They advertise numbers, spectacle and cheapness: production figures, swarm sizes, kung-fu humanoids, interceptors that cost a few thousand dollars. They also advertise reassurance: a human stays in the loop. What they hide follows an older rule. Britain held back its radar-jamming chaff until July 1943, partly for fear the enemy would copy it. The U-2 was presented as a camera plane, while much of its payload was electronic eavesdropping equipment. The F-117 stealth fighter flew for years before it was acknowledged.

The same logic now applies to the capabilities that matter most: the frequencies and anti-jamming tricks that keep drones flying, what their sensors can hear and recognise, the battlefield data that trains the next generation of autonomy, and above all how much autonomy is really switched on in combat. That last fact is the most sensitive of all, because it is legally explosive and decisive at once. It means the visible state of autonomous warfare is very likely behind the real one, and that the commitment hardest to verify, the human in the loop, is the one we are most often asked to take on trust.

The same silence now reaches orbit. The Finnish radar-satellite company ICEYE, which raised over a billion euros in 2026 and has supplied Ukraine since 2022, documents automatic target recognition that classifies aircraft, vessels and vehicles in a radar image in under 15 seconds, and in April 2026 its deployable intelligence cell worked inside a French infantry brigade’s sensor-to-shooter loop during the ORION exercise. As I noted in August, everyone repeats “AI-assisted, human-in-the-loop,” but no one publishes how long the human review actually takes, or how often the system misclassifies in real operations. Those are exactly the numbers that matter when the output cues a strike, and exactly the numbers that stay unpublished. Nor does the commercial framing mention that, as legal scholars at West Point’s Lieber Institute have noted, satellites feeding a targeting loop can themselves become lawful military targets.

The Reaper mirror

The targeted-killing programmes of the 2010s had the opposite profile. A few expensive MQ-9 Reapers flew deliberate strikes after long surveillance and legal review. Nobody doubted who decided; the deniability came from secrecy, classified rules such as signature strikes, and disputes over how civilian deaths were counted.

Today’s grey zone flips that. Cheap mass replaces scarce precision, and the deniability comes not from a hidden decision but from the possibility that no single decision exists. Then: we know who decided, but it is classified. Now: it may be unclear that anyone decided at all.

The Reaper era also shows what worked. Outside groups such as the Bureau of Investigative Journalism and New America built strike databases from local reporting and open sources. That pressure helped produce the 2016 US executive order requiring public reporting of civilian casualties from strikes outside declared war zones, a requirement revoked in 2019.

This points to a realistic answer to the gap Part 1 left open. The missing forensic standard may never come from reconstructing one drone’s final seconds. It is more likely to come from patterns: impact mapping over months, stray rates compared across fronts and jamming conditions, and repeated “accidents” near the same kind of target. One stray drone in a market is unprovable. Fifty strays clustering on the same category of site is evidence of policy.

The tools for that already exist in orbit. Commercial radar constellations see through cloud and darkness, and their archives record what changed on the ground, building by building, whether or not anyone asked them to look at the time. Reinsurers use the same data to quantify flood damage within hours. Pointed at a front line, it can document destruction and patterns of strikes after the fact. As I argued in May, there is no historical operational security against a system that analyses the past continuously: the archive already holds years of change, recorded before anyone thought to look. And the sensing layer is accelerating on the same timeline as the drones. Radar that filters out everything stationary and keeps only what moves, so-called ground moving target indication, already runs on commercial data, and the US Space Force and National Reconnaissance Office are reportedly building a dedicated constellation for it, aiming at full capability around 2028. The uncomfortable part is that the archive built for accountability is the same archive built for targeting. Which use dominates is, again, a human decision.

Nothing new about numbers

It is tempting to see drone swarms, millions of drones a year and a front saturated with jammers as something unprecedented. History says otherwise. Trench warfare, mass and signals intelligence are all more than a century old.

Mass is not new. At Verdun, the German bombardment opened on 21 February 1916 with more than 1,200 guns along a front of about 80 km. On the first day alone, the Germans fired roughly a million shells at the French army. In one small wood, the Bois des Caures, an area of about 1,300 by 800 metres, an estimated 80,000 rounds fell. Exploding shells caused around 60% of all battlefield casualties in the First World War. The battle lasted ten months, and both sides together lost around a million men killed and wounded.

Trenches are not new. Ukraine’s front has rebuilt the Western Front’s logic: a no man’s land of several kilometres up to around 30 km deep, watched constantly from the air, where anything that moves is spotted and struck. It is not covered by machine guns but by drones, and the result is the same: movement becomes nearly impossible, and soldiers dig.

Signals intelligence is not new. In August 1914, at Tannenberg, Russian commanders sent orders by radio without encryption, and German listeners read them. In 1917, British codebreakers decrypted the Zimmermann telegram, which helped bring the United States into the war. By the Second World War, signals had become industrial: at least 35,616 Enigma cipher machines were delivered to the Wehrmacht by a conservative count (the often-quoted figure is around 100,000), and by the end of the war about 12,000 German Army personnel worked in radio interception alone.

The air war was already an electronic war. From 1940, Germany built an integrated air defence network, the Kammhuber Line, reaching from occupied France into Germany. It divided the sky into boxes, each combining radars, searchlights, flak guns and night fighters guided by ground controllers watching a projected map, while radio listeners tracked Allied bombers as they formed up over England. Germany produced thousands of Würzburg radars and about 1,500 of the larger Würzburg-Riese. Flak was radar-directed and enormous in scale: it accounted for close to half of all American aircraft lost in Europe.

Jamming is not new either. The Allies answered with electronic warfare. A secret 800-person laboratory at Harvard, the Radio Research Lab, flew aircraft packed with receivers over Germany to map its radars, then built jammers to blind them. The cheapest countermeasure was the most effective: strips of aluminised paper cut to match the radar’s wavelength, called Window by the British and chaff by the Americans. First used over Hamburg in July 1943, it swamped German screens with false echoes. Searchlights wandered, flak fired blindly and night fighters could not find the bomber stream. The RAF lost 12 of 791 bombers that first night.

And the numbers were vast. According to the entrepreneur and historian of military technology Steve Blank, the Western Allies had some 28,000 combat aircraft over Europe at their peak and lost about 40,000 aircraft destroyed or damaged beyond repair, while the Soviet Union lost around 46,000 in the East. About 160,000 American and British airmen were killed, wounded or captured. Germany defended itself with some 15,000 flak guns and thousands of radars. The Allies answered with more than 30,000 jammers, 24,000 of them built against the Würzburg alone. Over Hamburg in July 1943, crews threw out 46,000 packets of chaff by hand, each holding about 2,000 metal strips.

Today’s numbers continue that line. Russia reportedly produces about 3,000 jet-powered Gerans a month. Ukraine plans millions of drones a year. The front is covered with dozens of cheap jammers rather than a few large ones. And every drone is both a radio emitter and a target for interception, which makes electronic warfare as pervasive as artillery was in 1916.

What is new is not the volume. Verdun’s million shells a day were each aimed by a human crew. Enigma’s tens of thousands of machines encoded messages written by humans. In 1943 every one of those 40,000 lost aircraft carried a crew, and every bomb was released by a person. Numbers have always killed in war. The grey zone is about who, or what, decides.

The 1914 lesson

In 1914 most political and military leaders expected a short war. The weapons that would decide it were already in their hands: the machine gun, quick-firing artillery, barbed wire and the railways that could feed men and shells to a front faster than any army could advance. A few saw it coming. The Polish banker Jan Bloch predicted as early as 1898 that modern firepower would produce stalemate, entrenchment and slaughter. He was largely ignored.

What followed was not ignorance, but momentum. By 1916 everyone knew what modern weapons did. At Verdun, and on the first day of the Somme, when the British army suffered some 57,000 casualties, leaders who understood the cost kept committing men, inside a war whose momentum made stopping feel impossible.

Europe is now planning as if attrition could return. In 2026 Britain began locating the tens of thousands of former soldiers who still carry a legal liability to be recalled. Germany began registering its young men again and revived a long-dormant rule requiring them to seek the army’s approval before long stays abroad, deemed granted for now. France renamed its citizenship day “Mobilisation Day.” The Netherlands is acquiring casualty trains, rolling hospitals planned, according to press reporting, around 1,000 to 1,500 wounded a day. As I described in September, almost none of this is new: it is dormant Cold War machinery switched back on, with the threshold for emergency powers moved earlier, into the uncertain space between peace and war.

I am not predicting a third world war, and I am not claiming the death toll will double or triple. My point is narrower. Looking at the pace of change described here, 2027 to 2030 could raise warfighter casualty levels wherever both sides stay committed, and more so if more nations become involved. The technology is not the hardest part to see. What is hard to imagine is how fast it compounds, just as it was in 1914.

Three wars at once

2027 to 2030 will not be a rerun of either world war. It may be something stranger: the trenches of 1916, the electronic war of 1943 and a technology curve that compounds faster than doctrine can follow, fused into one battlefield.

The First World War layer is already visible: a front frozen by firepower, a no man’s land watched from the air, mass rather than manoeuvre. The Second World War layer runs on top of it: integrated sensor networks, signals intelligence, jamming and counter-jamming, and whole economies geared to mass production. The third layer is new: acceleration and autonomy. Countermeasures age in about six weeks, AI speeds up the design of the next weapon, and the decision to strike is moving from people to machines. In 1916 and in 1943, every shell and every bomb was released by a human hand. That is the line now being crossed.

The capacity to go beyond world-war scale is not new. Since Hiroshima, on 6 August 1945, humanity has known that a single aircraft with a single bomb can destroy a city. That capacity has been held back for eighty years by deterrence, because it is concentrated in a few hands, visible and unusable without catastrophe.

Drone mass is the opposite kind of capacity. It is cheap, distributed, usable every day and sits below the threshold that deterrence was built to stop. Small, cheap, tactical devices can now reach strategic effect. On 1 June 2025, in Operation Spiderweb, Ukraine smuggled first-person-view drones deep into Russia inside trucks and launched them at airfields hosting strategic bombers. Ukraine claimed 41 aircraft hit at four bases; US officials estimated up to 20 struck and about 10 destroyed. Even the lower figure means a handful of consumer-grade drones damaged part of a nuclear-capable bomber fleet without coming anywhere near the threshold of nuclear deterrence.

And drone mass scales easily. Ukraine aims to raise production towards twenty million drones a year. Russia reportedly builds about 3,000 jet-powered Gerans a month. The Pentagon has launched a programme to buy hundreds of thousands of attack drones. Production lines that already run at this rate can be turned up far faster than any army of 1914 could mobilise.

Sadly but predictably, the maker phase of this technology is happening at the front rather than at home or in small businesses. Soldiers write code, 3D-print parts and rebuild drones in weeks. History suggests the civilian benefits will follow later, as the radar and electronic warfare labs of the Second World War later seeded Silicon Valley. For now, war is the workshop.

The flow also runs in the other direction, and it is really a circle. In October 2025, writing about Sikorsky’s U-Hawk, a Black Hawk with its cockpit removed, flown by autonomy software and managed from a commercial tablet, able to carry quivers of small drones, I described how military technology completes a full loop. GPS, inertial sensors and packet-switched networks began as defence research, became cheap mass-market technology inside every tablet and phone, and now return to the military in that civilian form, bringing the weaknesses of the consumer world with them.

The loop now spins faster than ever. In 2016 and 2017, ISIS turned off-the-shelf quadcopters into grenade-droppers over Mosul, at times about one drone a day. Ukraine’s “Baba Yaga” heavy night bombers partly grew out of an agricultural octocopter built to spray crops. Today’s FPV attack drones descend from hobby racing drones. The same airframe can spray a field or bomb a trench, and the same autonomy that steers a strike can carry a wounded soldier home. Uranium could be controlled. Farm equipment, hobby electronics, tablets and 3D printers cannot, and that is what makes the governance problem of 2027 so different from the nuclear age.

The ceiling, in other words, is no longer set by industry. It is set by politics, and by how committed both sides remain. In 1916 the leaders at Verdun knew the cost and kept going, carried by a war’s momentum. In 2027 the same momentum would drive a system that can grow faster, fight longer and decide more of its own targets than anything before it. That is the real lesson of three wars at once: not that catastrophe is certain, but that the brakes are now almost entirely human, at the very moment humans are being taken out of the loop.

Accelerating returns

The futurist Ray Kurzweil, long a director of engineering at Google, built his forecasts on what he calls the law of accelerating returns: each generation of technology helps build the next, so progress compounds. He has predicted human-level AI around 2029. The singularity he describes is not a fact. But the stack of signs is growing: increasingly capable AI, drone swarms, the Internet of Things, big data, 3D printing at the front line, humanoids that run faster than world-record holders, and capabilities that never reach the civilian world at all.

One public example shows how much may sit beyond public view. In April 2026 Anthropic announced Project Glasswing, giving a small group of organisations access to an unreleased model, Claude Mythos Preview, that the company said it would not make generally available because of its cybersecurity capabilities. According to Anthropic, it found thousands of serious, previously unknown vulnerabilities across major operating systems and browsers. If a commercial lab now holds back a model for that reason, it is reasonable to assume that defence establishments hold capabilities we do not see.

We are only in 2026. It could very well be that by 2028 such systems routinely produce solutions that multiply force, cut cost and time, and increase speed and lethality, and that they help design the next generation of the very drones described here.

Open

Part 1 asked who pulls the trigger when the link breaks. Part 2 adds harder versions: what happens when there is no trigger at all, only a launch and a fall; what happens when the decision belongs to a swarm rather than a drone; and what happens when the swarm’s mind and body are finally one machine.

The answers have to start at the launch, they have to be built from patterns rather than single incidents, and they will increasingly depend on whether autonomous systems are required to keep a record of their own reasoning. Every trend in this piece points the same way for 2027: more drones than pilots, quantity over size, swarms moving from tests to routine use, navigation that survives jamming, onboard intelligence that costs almost nothing, and research that never goes backwards. If that holds, 2027 is the year swarm warfare stops being a headline and becomes the default.

Whether this month’s CCW Review Conference produces a mandate that reaches the launch, the pattern or the log will show whether governance is moving at the speed of the hardware. On current evidence, it is not.

Sources

1. Watchus, B., The Widening Grey Zone of Autonomous Killing (Part 1), OSINT Team, 1 Aug 2026 — https://medium.com/@BerendWatchusIndependent/the-widening-grey-zone-of-autonomous-killing-0f4b4149633e
2. United24 Media, Geran-5 Gets Kh-101-Like Guidance (incl. Moldova crash, 30 Sep), Oct 2026 — https://united24media.com/war-in-ukraine/geran-5-gets-kh-101-like-guidance-as-russia-pushes-its-jet-drones-toward-full-autonomy-23108
3. Watchus, B., NATO Counter-Drone Operations in Belgium: Deep Technical Analysis for Cyber Security Professionals, System Weakness, 12 Nov 2025 — https://systemweakness.com/nato-counter-drone-operations-in-belgium-deep-technical-analysis-for-cyber-security-a8528e214e73
4. Small Wars Journal, Fully Autonomous Drones Reportedly Kill in Ukraine, 17 Aug 2026 — https://smallwarsjournal.com/2026/08/17/fully-autonomous-drones-reportedly-kill-in-ukraine/
5. Militarnyi, Ukraine Uses Autonomous Drone Swarm Against Russian Forces, 1 Oct 2026 — https://militarnyi.com/en/news/ukraine-uses-self-deciding-drone-swarm-against-russian-forces/
6. Kyiv Post, Ukraine Planned 1,000-Drone Swarm to Paralyze Moscow Airports, 21 Aug 2026 — https://www.kyivpost.com/post/82817
7. Militarnyi, Russian Forces Begin Using New Geran-5 Version With Autonomous Optical Navigation, Sep 2026 — https://militarnyi.com/en/news/russian-forces-begin-using-new-geran-5-version-with-autonomous-optical-navigation/
8. The Canadian Press via CP24, Drone warfare is evolving faster than armies can adapt, experts warn, 25 Oct 2026 — https://www.cp24.com/news/canada/2026/10/25/drone-warfare-is-evolving-faster-than-armies-can-adapt-experts-warn/
9. ICRC, Statement to the UN General Assembly’s First Committee, 14 Oct 2026 — https://www.icrc.org/en/statement/statement-un-general-assemblys-first-committee
10. New Zealand MFAT, UNGA First Committee 2026: Cluster 4 Joint Statement on LAWS, 22 Oct 2026 — https://www.mfat.govt.nz/en/media-and-resources/united-nations-general-assembly-first-committee-2026-cluster-4-joint-statement-on-lethal-autonomous-weapon-systems
11. UN News, UN chief, Red Cross renew call for rules on lethal autonomous weapons, Aug 2026 — https://news.un.org/en/story/2026/08/1168196
12. Kyiv Independent, Ukraine’s drone production jumps in 2026, says defense ministry, 11 Oct 2026 — https://kyivindependent.com/ukraines-drone-production-jumps-370-in-2026-says-defense-ministry/
13. Forbes, How Drone Swarm Tactics Are Becoming The New Norm For U.S. Soldiers, 25 Aug 2026 — https://www.forbes.com/sites/zitaballingerfletcher/2026/08/25/how-drone-swarm-tactics-are-becoming-the-new-norm-for-us-soldiers/
14. DefenseScoop, 101st Airborne unit tested attack drones, UGV, AI in recent exercise, 25 Jun 2026 — https://defensescoop.com/2026/06/25/101st-airborne-unit-tested-attack-drones-ugv-ai-in-recent-exercise/
15. LIGA.net, Russia attacks the front with cheap drones because the Defense Forces effectively destroy expensive ones (Flash), 5 Mar 2026 — https://news.liga.net/en/politics/news/russia-attacks-the-front-with-cheap-drones-because-the-defense-forces-effectively-destroy-expensive-ones-flash
16. Defense Tech News (Preston Stewart), video episode on quantity versus quality and the StarGate evacuation, Oct 2026
17. Business Insider via ChinaPulse, Ukraine: NATO drone maker says underground factories are a lesson for West, 24 Oct 2026 — https://www.chinapulse.com/data-news/2026/10/24/ukraine-nato-drone-maker-says-underground-factories-are-a-lesson-for-west/
18. Kyiv Post, Ukraine’s General Cherry and Croatian drone manufacturer ORQA to build underground component factory — https://www.kyivpost.com/post/73493
19. Associated Press via Click2Houston, A wounded Ukrainian soldier waited 2 days to be rescued, then a drone flew him to safety, 26 Oct 2026 — https://www.click2houston.com/tech/2026/10/26/a-wounded-ukrainian-soldier-waited-2-days-to-be-rescued-then-a-drone-flew-him-to-safety/
20. Watchus, B., EXCLUSIVE: The Autonomous Knowledge Accelerator: How a Commercial LLM Generated 200× Power and 16.4× Cognitive Speed, System Weakness, 31 Oct 2025 — https://systemweakness.com/exclusive-the-autonomous-knowledge-accelerator-how-a-commercial-llm-generated-200-power-and-16-4-e772eceea3db
21. Watchus, B., ED-HPSNN research paper, Oct 2025 — https://archive.org/details/ed-hpsnn-research-paper-updated-ready-for-publish
22. Watchus, B., The Law of Optimized Complexity, 2025 — https://archive.org/details/looc-paper-watchus-2025-final-with-funding
23. Paredes-Vallés et al., Fully neuromorphic vision and control for autonomous drone flight, Science Robotics, 2024 — https://research.tudelft.nl/en/publications/fully-neuromorphic-vision-and-control-for-autonomous-drone-flight/
24. Bian et al., ColibriUAV: An Ultra-Fast, Energy-Efficient Neuromorphic Edge Processing UAV-Platform, 2023 — https://arxiv.org/pdf/2305.18371
25. Ayana & Qiming, PrivLLMSwarm: Privacy-Preserving LLM-Driven UAV Swarms for Secure IoT Surveillance, arXiv:2512.06747 — https://arxiv.org/html/2512.06747v1
26. Watchus, B., The Need for a Thinking, Private Brain & The Death of Dumb Drones, OSINT Team, 10 Dec 2025 — https://osintteam.blog/the-need-for-a-thinking-private-brain-the-death-of-dumb-drones-05a568532c50
27. Watchus, B., Why LLMs Are a Liability in Tactical Contexts: The L-Shaped Hallway Paradox, OSINT Team, 5 Jul 2026 — https://osintteam.blog/why-llms-are-a-liability-in-tactical-contexts-the-l-shaped-hallway-paradox-8c98e4a9b434
28. NVIDIA, NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI, 31 May 2026 — https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai
29. Karcini et al., Robots Need More Than VLAs & World Models, arXiv:2606.06556, 4 Jun 2026 — https://arxiv.org/html/2606.06556v1
30. Wikipedia, Gemini Robotics — https://en.wikipedia.org/wiki/Gemini_Robotics
31. Interesting Engineering, Robot boxing in China (Mecha Fighting Series, Hangzhou, May 2025) — https://interestingengineering.com/innovation/robot-boxing-china-humanoids-wins
32. South China Morning Post, Kung fu, somersaults and scale: Unitree eyes 20,000-robot output in 2026 after gala — https://scmp.com/tech/big-tech/article/3343825/kung-fu-somersaults-and-scale-unitree-eyes-20000-robot-output-2026-after-gala
33. iTWire, The machines are running, fighting, and herding pigs: 2026 is the year robots got real, 23 Apr 2026 — https://www.itwire.com/your-it-news/mobility/the-machines-are-running,-fighting,-and-herding-pigs-2026-is-the-year-robots-got-real
34. Futurism, Behold: Video of “First-Ever” Human Versus Robot MMA Fight, 21 Sep 2026 — https://futurism.com/robots-and-machines/first-ever-human-versus-robot-mma-fight
35. Watchus, B., Trend update: “First-Ever” Human Versus Robot MMA Fight, 29 Sep 2026 — https://medium.com/@BerendWatchusIndependent/trend-update-first-ever-human-versus-robot-mma-fight-44219b473537
36. IEEE Spectrum, Ukraine Can’t Stop Russia’s Jet-Powered Shahed Drone, 30 Sep 2026 — https://spectrum.ieee.org/jet-powered-shahed-drone
37. The Defense News, GE Aerospace and Kratos Team Up to Build Affordable, Scalable Turbofans for Drone Wingmen — https://www.thedefensenews.com/GE-Aerospace-and-Kratos-Team-Up-to-Build-Affordable-Scalable-Turbofans-for-Drone-Wingmen/
38. Watchus, B., How an LLM and One Curious Non-Engineer Eliminated the Last 25% Mystery of Turbofan Thrust in One Afternoon, System Weakness, 26 Nov 2025 — https://systemweakness.com/how-an-llm-and-one-curious-non-engineer-eliminated-the-last-25-mystery-of-turbofan-thrust-in-one-30a6036c806e
39. Wikipedia, Chaff (countermeasure) — https://en.wikipedia.org/wiki/Chaff_(countermeasure)
40. Osprey Publishing, 100th anniversary of the Battle of Verdun — https://ospreypublishing.com/uk/osprey-blog/2016/100th-anniversary-of-the-battle-of-verdun
41. National WWI Museum and Memorial, Artillery — https://www.theworldwar.org/learn/about-wwi/artillery
42. History Hit, Why Was the Battle of Verdun so Significant? — https://www.historyhit.com/1916-battle-verdun/
43. Crypto Cellar, Enigma Development and Production History — https://cryptocellar.org/enigma/e-prod-history/index.html
44. Wikipedia, German radio intelligence operations during World War II — https://en.wikipedia.org/wiki/German_radio_intelligence_operations_during_World_War_II
45. Wikipedia, Würzburg radar — https://en.wikipedia.org/wiki/W%C3%BCrzburg_radar
46. Osprey Publishing, German Flak Defences vs Allied Heavy Bombers — https://ospreypublishing.com/blog/German_Flak_Defences_vs_Allied_Heavy_Bombers
47. Steve Blank, The Secret History of Silicon Valley (resources and bibliography) — https://steveblank.com/secret-history
48. Steve Blank, Hidden in Plain Sight: The Secret History of Silicon Valley (slides, rev. Feb 2018) — https://stewart.sdsu.edu/2026_SteveBlank-SecretsOfSiliconValley_Master%20Presentation.pdf
49. Computer History Museum, The Secret History of Silicon Valley (event) — https://computerhistory.org/events/secret-history-silicon-valley/
50. Militarnyi, Russia Produces Around 3,000 Geran-4/5 Kamikaze Drones per Month — https://militarnyi.com/en/news/russia-produces-around-3-000-geran-4-5-kamikaze-drones-per-month/
51. Watchus, B., UK and Western Europe Turn to (Continental) Civilian and Veteran War Preparation, 2026, OSINT Team, 10 Sep 2026 — https://medium.com/the-first-digit/uk-and-western-europe-turn-to-continental-civilian-and-veteran-war-preparation-2026-the-25a39146c538
52. Kyiv Independent, Reuters: Ukraine struck fewer Russian bombers during Operation Spiderweb than estimated, US officials claim — https://kyivindependent.com/reuters-ukraine-struck-fewer-russian-bombers-during-operation-spiderweb-than-estimated-us-officials-claim/
53. United24 Media, Ukraine targets annual production of 20 million drones under new EU partnership — https://united24media.com/war-in-ukraine/ukraine-targets-annual-production-of-20-million-drones-under-new-eu-partnership-20805
54. The War Zone, Pentagon Launches $1B Program To Rapidly Buy Hundreds Of Thousands Of Kamikaze Drones — https://www.twz.com/air/pentagon-launches-1b-program-to-rapidly-buy-hundreds-of-thousands-of-kamikaze-drones
55. Watchus, B., HAWK TUAH: Why This Drone Mothership (U-Hawk) Matters Beyond the Battlefield, System Weakness, 22 Oct 2025 — https://systemweakness.com/hawk-tuah-why-this-drone-mothership-u-hawk-matters-beyond-the-battlefield-6c83754a7be6
56. The Aviationist, Sikorsky Unveils the S-70UAS U-Hawk, First Fully Autonomous Black Hawk Helicopter, 13 Oct 2025 — https://theaviationist.com/2025/10/13/u-hawk-unveiled/
57. Wikipedia, Baba Yaga (aircraft) — https://en.wikipedia.org/wiki/Baba_Yaga_(aircraft)
58. Defense One, The Drones of ISIS, 12 Jan 2017 — https://www.defenseone.com/technology/2017/01/drones-isis/134542/
59. VentureBeat, Anthropic says its most powerful AI cyber model is too dangerous to release, Apr 2026 — https://venturebeat.com/data/anthropic-says-its-most-powerful-ai-cyber-model-is-too-dangerous-to-release
60. Watchus, B., The “Nuclear Battery” Hype, Decoded for People Who Track Capabilities, OSINT Team, 22 Aug 2026 — https://medium.com/the-first-digit/the-nuclear-battery-hype-decoded-for-people-who-track-capabilities-7b573e00d56b
61. Nuclear batteries: Potential, challenges and the future, The Innovation Energy, 2025 — https://www.the-innovation.org/article/doi/10.59717/j.xinn-energy.2025.100079
62. Wikipedia, Ottawa Treaty — https://en.wikipedia.org/wiki/Ottawa_Treaty
63. Wikipedia, Iron harvest — https://en.wikipedia.org/wiki/Iron_harvest
64. Wikipedia, The Thing (listening device) — https://en.wikipedia.org/wiki/The_Thing_(listening_device)
65. Watchus, B., I Flagged ICEYE in May. Now Raised a Billion Euros: Here’s What Its Satellites Can Actually Do, OSINT Team, 26 Aug 2026 — https://medium.com/the-first-digit/i-flagged-iceye-in-may-now-raised-a-billion-euros-heres-what-its-satellites-can-actually-do-9ce66516812e
66. Watchus, B., The Satellite+AI Stack: Why a Forestry Article Is the Most Important OSINT Read of 2026, OSINT Team, 31 May 2026 — https://osintteam.blog/the-satellite-ai-stack-why-a-forestry-article-is-the-most-important-osint-read-of-2026-8e901c3a60e1
67. Lahssini, K., Mapping Forest Height with Lasers from Space, The New Climate, 25 May 2026 — https://medium.com/the-new-climate/mapping-forest-height-with-lasers-from-space-28760f4de37c
68. ICEYE, Space-Based Tactical ISR System — https://www.iceye.com/missions-isr-system
69. Aerotime, ICEYE brings satellite intelligence to French Army brigade at ORION 2026, 7 May 2026 — https://www.aerotime.aero/articles/iceye-satellite-intelligence-brigade-orion-2026
70. Tunnel-induced land subsidence assessment in a densely populated residential area using Sentinel-1 PS-InSAR, Discover Geoscience, 2024 — https://link.springer.com/article/10.1007/s44288-024-00086-2
71. Lieber Institute, West Point, “The Eye in Space”: ICEYE’s SAR Satellites and the Law of War, Apr 2025 — https://lieber.westpoint.edu/eye-space-iceyes-sar-satellites-law-of-war/

Further reading

1. CSIS, The Art of Swarming, 14 Oct 2026 — https://www.csis.org/analysis/art-swarming
2. British Army, Strength in swarms: shaping the future of allied drone warfare — https://www.army.mod.uk/news/strength-in-swarms-shaping-the-future-of-allied-drone-warfare/
3. European Parliamentary Research Service, Hybrid warfare, 1 Oct 2026 — https://eprs.europarl.europa.eu/EPRS_BRI(2026)797001
4. European Parliamentary Research Service, Drones and new systems of warfare: Adapting the EU to today’s security challenges — https://www.europarl.europa.eu/thinktank/en/document/EPRS_ATA(2026)782599
5. Small Wars Journal, The Miniaturization of Military Robots, 23 Oct 2026 — https://smallwarsjournal.com/2026/10/23/the-miniaturization-of-military-robots/
6. Small Wars Journal, The New Kalashnikov: Governing the Poor Man’s Autonomous Drone, 20 Oct 2026 — https://smallwarsjournal.com/2026/10/20/autonomous-drone-proliferation-governance/
7. Small Wars Journal, The Coming Wave of High-Powered Microwaves, 30 Sep 2026 — https://smallwarsjournal.com/2026/09/30/the-coming-wave-of-high-powered-microwaves/
8. Small Wars Journal, Army looks to small UGVs as Ukraine war reshapes battlefield robotics, 19 Jun 2026 — https://smallwarsjournal.com/2026/06/19/army-small-ugvs-battlefield-robotics/
9. CNA, PRC Concepts for UAV Swarms in Future Warfare, 2025 — https://www.cna.org/analyses/2025/07/prc-concepts-for-uav-swarms-in-future-warfare
10. Raj et al., Recent Developments and Applications of Drone Swarm, Sensors 26(10), 2026 — https://pmc.ncbi.nlm.nih.gov/articles/PMC13210634/
11. Kallenborn, Z., InfoSwarms: Drone Swarms and Information Warfare, Parameters 52(2), 2022 — https://press.armywarcollege.edu/parameters/vol52/iss2/13/
12. Lviv Herald, The Flock Goes to War: Drone Swarms and the Next Revolution in Warfare, 5 Sep 2026 — https://www.lvivherald.com/post/the-flock-goes-to-war-drone-swarms-and-the-next-revolution-in-warfare
13. Oxford Changing Character of War Centre, Cagri Sagiroglu publishes “The Ghost in the Machine”, 14 Oct 2026 — https://www.ccw.ox.ac.uk/news/2026/10/14/cagri-sagiroglu-publishes-the-ghost-in-the-machine
14. Africa Center for Strategic Studies, Lessons from Africa’s Drone Wars — https://africacenter.org/publication/asb49-lessons-africa-drone-wars/
15. Foreign Affairs (Mara Karlin), The Fog of Tech, 20 Oct 2026 — https://www.foreignaffairs.com/united-states/fog-tech-artificial-intelligence-mara-karlin-ai
16. Foreign Affairs, Back in the Trenches: Why New Technology Hasn’t Revolutionized Warfare in Ukraine — https://www.foreignaffairs.com/ukraine/back-trenches-technology-warfare
17. Foreign Affairs (Michael Horowitz), Battles of Precise Mass — https://www.foreignaffairs.com/world/battles-precise-mass-technology-war-horowitz
18. Jamestown Foundation, AI Lessons From Russia’s War in Ukraine — https://jamestown.org/ai-lessons-from-russias-war-in-ukraine/
19. Stacker, The 5 biggest changes making drones harder to defeat, 13 Oct 2026 — https://stacker.com/stories/aviation-space/5-biggest-changes-making-drones-harder-defeat
20. Army Times, 101st soldiers use drones to drop grappling hooks, breach razor wire, 26 Jun 2026 — https://www.armytimes.com/news/your-army/2026/06/26/101st-soldiers-use-drones-to-drop-grappling-hooks-breach-razor-wire/
21. Military Times, US soldiers train eyes and ears for drone swarms, 15 May 2026 — https://www.militarytimes.com/news/your-military/2026/05/15/no-sound-of-silence-us-soldiers-train-eyes-and-ears-for-drone-swarms/
22. The Defense Post, British Army drone swarm, 18 Sep 2026 — https://thedefensepost.com/2026/09/18/british-army-drone-swarm/
23. Secure UAV Swarms in Low-Altitude Wireless Networks: Challenges and Solutions, arXiv:2605.26876 — https://arxiv.org/pdf/2605.26876
24. DefenseScoop, DoD Drone Dominance program orders and deliveries, 5 Mar 2026 — https://defensescoop.com/2026/03/05/dod-drone-dominance-program-orders-deliveries-military-units/
25. DefenseScoop, Hegseth’s Drone Dominance Program, gauntlets and RFI, 2 Dec 2025 — https://defensescoop.com/2025/12/02/hegseth-drone-dominance-program-ddp-gauntlets-website-rfi/
26. CNAS, Drones, Missiles, and the U.S. Munitions Shortage, 18 Sep 2026 — https://www.cnas.org/publications/reports/drones-missiles-and-the-u-s-munitions-shortage
27. Council on Foreign Relations, Out of Ammo, Aug 2026 — https://assets.cfr.org/images/Out-of-Ammo_August26/Out-of-Ammo_August26.pdf
28. RUSI, The Evolution of Russia’s Precision Strike: Adopting Mass Precision — https://www.rusi.org/explore-our-research/publications/research-papers/evolution-russias-precision-strike-adopting-mass-precision
29. Hudson Institute (Nadia Schadlow), Three Ds: Mass, Decentralization, Democratization, Deliverability — https://www.hudson.org/three-ds-mass-decentralization-democratization-deliverability-nadia-schadlow
30. National Defense University, A New Arsenal for a New Era: Balancing Affordable Mass with Exquisite Capabilities, 29 Aug 2025 — https://es.ndu.edu/Archive/Media/Publications/Article-View/Article/4406686/a-new-arsenal-for-a-new-era-balancing-affordable-mass-with-exquisite-capabiliti/
31. NATO JAPCC, The Relevance of Quantity in Modern Conflict — https://www.japcc.org/essays/the-relevance-of-quantity-in-modern-conflict/
32. Militarnyi, Production of the Geran-4/5 Jet Drones Has Surpassed That of the Geran-2 — https://militarnyi.com/en/news/production-of-the-geran-4-5-jet-drones-has-surpassed-that-of-the-geran-2/
33. United24 Media, How did Russia go from testing jet drones to producing 3,000 a month — https://united24media.com/war-in-ukraine/how-did-russia-go-from-testing-jet-drones-to-producing-3000-a-month-22160
34. Militarnyi, Ukraine Plans to Produce Over 7 Million Drones in 2026 — https://militarnyi.com/en/news/ukraine-plans-to-produce-over-7-million-drones-in-2026/
35. Inside Unmanned Systems, Counter-UAS: The Price of the Shot — https://insideunmannedsystems.com/counter-uas-the-price-of-the-shot/
36. Ukrainska Pravda, Cheap but dangerous: how Russia’s plywood Molniya drone has become a big problem, 15 Jan 2026 — https://www.pravda.com.ua/eng/articles/2026/01/15/8016293/
37. Ukrainska Pravda, Russia plans to increase share of jet-powered Shahed drones to 50%, 14 May 2026 — https://www.pravda.com.ua/eng/news/2026/05/14/8034678/
38. Al Jazeera, Humanoid robots race, tumble at China’s first robot Olympics, Aug 2025 — https://www.aljazeera.com/gallery/2025/8/18/humanoid-robots-race-tumble-at-chinas-first-robot-olympics
39. Dellibarda Varela et al., Sensorimotor Features of Self-Awareness in Multimodal Large Language Models, arXiv:2505.19237 — https://arxiv.org/html/2505.19237v1
40. MiGFlug, How Does a Counter-Drone System Work? Detect, Track, Defeat — https://migflug.com/afterburner/?p=29923216
41. Price, A., Instruments of Darkness: The History of Electronic Warfare (book)
42. Jones, R.V., Most Secret War (US title: The Wizard War) (book)