The Drone Foundry

On the morning of June 1, 2025, at a fuel depot outside Olenya, in Russia’s Murmansk Oblast, the roof panel of a parked cargo truck retracted on a remote signal. Inside the steel container, racked in shallow trays, more than a hundred quadcopter drones rose under their own power, the way a wasp’s nest does when poked. Within minutes the first of them was over the runway at Olenya Air Base. Within hours, at four more air bases scattered across five Russian time zones, identical trucks had opened and identical drones had flown. By the time Russian air-defense units understood what was happening, the Ukrainian Security Service operation called Spiderweb had reportedly hit forty-one aircraft, including Tu-95MS and Tu-22M3 strategic bombers and at least one A-50 airborne early-warning plane. Ukrainian officials put the dollar value of the lost airframes at roughly seven billion. The drones themselves had cost a few hundred dollars each.

The trucks had been driven into Russia weeks earlier through ordinary logistics channels. Their drivers, civilians hired through brokers, did not know what they were hauling. The launch was timed by SBU operators sitting in offices in Kyiv. The drones, once airborne, ran on commercial radio links and consumer-grade autopilots. Several were configured to navigate to their targets using onboard machine vision after losing the radio link, a fail-safe that worked in at least some cases. The most dramatic single attack on strategic aviation in the war had been executed by an apparatus that could be purchased, in its component form, from electronics distributors who would also sell the parts to a high-school robotics team.

The cost ratio was the part that mattered. By the spring of 2025, Ukrainian and Western analysts had converged on a number that earlier in the war would have been treated as propaganda. A typical first-person-view attack drone, the small fixed-rotor airframe that Ukrainian and Russian forces by then used to kill armor at the contact line, cost between three hundred and five hundred dollars assembled. The vehicles those drones were destroying cost three to ten million. Russian armor losses through early 2025 ran into the thousands. Independent analyses estimated that small drones, mostly FPVs, accounted for roughly sixty-five percent of confirmed Russian tank losses, displacing artillery as the leading killer of armor in a war whose first year had been defined by it. For Russia’s T-90M, the modernized variant the army had hoped would be its conventional answer to Western tanks, FPV strikes were credited with about half of all combat losses.

A weapon costing one part in ten thousand of its target, in any other era of warfare, would have been called an anomaly. By 2025 it was the dominant economic fact of the war. It rested entirely on a particular industry and a particular silicon stack.

The industry had not existed before the invasion. In February 2022, Ukraine had perhaps ten companies designing or assembling unmanned aerial vehicles, most of them building small numbers of reconnaissance airframes for the territorial defense forces. By late 2025, depending on the count, there were between five hundred and a thousand. The Ministry of Digital Transformation, working with the General Staff and the Ministry of Defence, had launched a state platform called Brave1 in April 2023 to channel grant money, technical accreditation, and frontline feedback into the new ecosystem. By the second half of 2025, Brave1 had pushed out two hundred and forty grants and the state budget was earmarking roughly seventy-five million dollars a year for defense-tech awards across drones, electronic warfare, counter-Shahed turrets, mother-ship platforms, and a Ukrainian analogue of the DJI Mavic that operators could buy without depending on a Chinese export license. Monthly drone production climbed from twenty thousand units in 2024 to two hundred thousand a month by mid-2025. The full-year projection by November 2025 was around four million drones, with one Kyiv-government estimate placing the figure at over two and a half million for the first eleven months. Bloomberg’s reporting that month judged that Ukraine was building more drones than every other NATO member combined.

The shape of that industry, looked at from above, did not resemble any defense industrial base the United States had built since the Second World War. There were no walled campuses. There were several hundred small workshops, often in residential buildings or converted warehouses, producing batches of a few thousand drones a month. Many of the firms were three or five engineers with a CNC mill and a soldering bench. Capital costs were measured in tens of thousands of dollars rather than hundreds of millions. The bottleneck was not floor space or financing but components, and the components were almost entirely off-the-shelf.

The flight controller on a Ukrainian FPV in 2025 was, in roughly nine cases out of ten, an STMicroelectronics STM32 microcontroller, the same ARM Cortex-M-based part that hobbyists had been using to build quadcopters in their kitchens since the mid-2010s. Wholesale, the chips cost a few dollars apiece. The companion radio module was usually an Espressif ESP32 or a clone of one, made for the consumer Internet-of-things market by a fabless Shanghai company that had begun life selling Wi-Fi modules to smart-lightbulb manufacturers. The video transmitter was an analog 5.8-gigahertz module of the type sold by the thousand on AliExpress. The camera was a consumer charge-coupled device meant for action cameras. The motors were brushless outrunners of a design used in drone-racing leagues. The batteries were lithium polymer cells from the same Chinese plants that supplied the e-scooter and electric-bicycle markets. Total component cost, even at small-batch retail prices through wartime supply chains, was under two hundred dollars. The remainder of the unit price was assembly, testing, and the warhead.

The Russian side was running on the same kind of inventory, with some politically inconvenient overlap. Teardowns of Iranian-designed Shahed-136 loitering munitions, which Russia produced under license as the Geran-2 from a plant in the Alabuga Special Economic Zone in Tatarstan, repeatedly found Western chips. Reports through 2023 and 2024 from Ukrainian intelligence, the Royal United Services Institute, and the open-source community at Aeronaut and others catalogued Texas Instruments TMS320 series digital signal controllers, NXP microcontrollers, Analog Devices and Onsemi analog parts, Infineon power components, and GPS modules from Hemisphere GNSS. By late 2024 some teardowns had identified an Nvidia Jetson Orin module, the embedded edge-AI computer Nvidia sold for robotics and autonomous-vehicle work, sitting in the nose of upgraded Shahed variants with onboard camera-based terminal guidance. Through 2025, as Western firms tightened compliance and the sanctions teams in Brussels and Washington blocked more routing channels, Russian engineers had begun substituting domestic-friendly parts where they could. The Xilinx Kintex-7 FPGA that had handled signal processing on the eight-channel Nasir anti-jam GPS module in earlier Shaheds was replaced, in 2025 builds, with an integrated circuit from Beijing Microelectronics Technology. STM32 microcontrollers were being swapped out for GigaDevice GD32 parts, a near-pin-compatible Chinese family that had built its market by being a drop-in for the European original. The substitutions worked. They also told a story about the war that the public discussion of EUV scanners and three-nanometer fabs did not. The decisive silicon on the battlefield was not at the leading edge. It was at the mature node, and it was widely available.

What separated the two sides, then, was not the chips. It was what surrounded them.

A Ukrainian FPV drone in 2025 was less a fixed product than an open platform. The standard practice at Brave1-funded firms, and at the larger independent workshops like Vyriy Drone and Skyeton, was to publish the airframe schematics, document the autopilot in a way that allowed third-party plug-ins, and release software updates over the air on a cycle measured in days. Operators at the contact line would record what worked and what failed in a given week and post the results to encrypted channels that fed back to the manufacturers. A new Russian electronic-warfare frequency or a new tactic for masking infantry against thermal cameras could be observed, characterized, and answered with a software push to deployed drones within seven to fourteen days. The Pentagon’s procurement system, structured around five-year programs of record, could not match this cadence by an order of magnitude. The Russian system, organized around centralized design bureaus and ministry approvals, could not match it at all.

This rapid iteration was the most important thing the war had produced, more important than any single weapon. It had a name in the Ukrainian defense-tech vocabulary borrowed from American software practice. The teams called it the feedback loop. By 2025 most serious participants in the Ukrainian drone industry had organized their entire operation around shortening it. The Defence Innovation Cluster Bravinkov, an irregular gathering of engineers from forty-odd firms that met outside Kyiv every six weeks, treated the loop as the central design problem of the war, ahead of payload weight or flight time. A British review of Ukrainian unmanned-systems doctrine published in late 2025 made the point starkly: the side that had achieved the shortest path from frontline observation to factory-floor design change had achieved, for the first time in modern warfare, an advantage that compounded weekly.

Autonomy was the natural next step. By the middle of 2025 most of the credible Ukrainian drone designers were adding optical terminal guidance, the ability for a drone to recognize and track its assigned target across the final hundred meters of its flight without operator input. The motivation was electronic warfare. Russian jammers had become dense enough at the front, by 2024, that a meaningful share of FPV sorties were being lost to severed radio links before reaching their targets. If the drone could be released outside the jamming bubble and complete the strike on its own using a small onboard neural network running on a low-power inference chip, the kill chain became immune to jamming in its terminal phase. A CSIS analysis of Ukrainian deployments through the second half of 2025 found that mission success rose from roughly fifteen percent for fully manual sorties to roughly sixty percent for sorties that used AI-assisted terminal guidance. A few firms, most prominently Saker, fielded reconnaissance drones with onboard target classification trained on Russian armor, infantry, and electronic-warfare equipment. The Saker Scout was openly described by its designers as a system that identified targets but did not initiate the strike, the operator’s finger still on the trigger; analysts watching the deployment cycle judged that the human-in-the-loop constraint was likely to weaken under combat pressure within a year.

Counter-drone work converged on the same architecture from the opposite direction. By the summer of 2025 a Ukrainian firm had deployed a system called Sky Sentinel, an AI-controlled heavy machine-gun turret with a commercial-camera sensor stack and an onboard inference chip that detected, tracked, and engaged airborne targets automatically. The prototypes recorded their first interceptions of Shahed drones in May. Cost per turret, including the gun, sensors, mount, and compute, ran around a hundred and fifty thousand dollars, less than a single high-end surface-to-air missile and far less than the loitering munitions it shot down. Production was set to ramp to a few dozen units a month before the end of the year. Conventional FPV interceptors, smaller drones tuned to chase and ram incoming Shaheds, were being produced by Ukrainian firms at a pace of nine hundred to fifteen hundred a day by early 2026, with kill rates in the sixty-to-eighty percent range under combat conditions.

The naval war ran in parallel. The Ukrainian Magura-V5, an unmanned surface vessel built from a fiberglass hull a little under six meters long and packed with explosives, electronics, and Starlink terminals, became operational in 2023. By the summer of 2024, Magura V5s and the related Sea Baby boats had crippled the Russian Black Sea Fleet by sinking or damaging eighteen warships, including the Tarantul-class corvette Ivanovets and the Ropucha-class landing ship Tsezar Kunikov, in attacks that forced Moscow to relocate most of its operational hulls from Sevastopol to Novorossiysk on the Russian mainland. Each Magura cost something between two hundred and fifty and three hundred thousand dollars. The ships they sank cost between fifty and three hundred million. In May 2025 a Magura-7, the second-generation variant, became the first naval drone in history to shoot down a manned combat aircraft, downing two Russian Su-30 fighters over the Black Sea with AIM-9 Sidewinder missiles launched from a modified attack boat. The achievement was not registered in any naval treatise’s category of expected outcomes.

In Washington, the Pentagon had been running its own version of this argument for a decade, and was discovering by 2025 how much harder it was to execute. Replicator, the program Kathleen Hicks had announced in August 2023 with the promise of fielding “multiple thousands” of attritable autonomous systems within eighteen to twenty-four months, reached its target date of August 2025 with what officials acknowledged was a count in the hundreds rather than the thousands. The exact figure remained classified, but defense reporters who had covered the program closely judged that the gap was a function of the same acquisition system the program had been created to bypass. Contracts had gone disproportionately to firms new to defense work, many of which were still moving prototypes through environmental qualification when the deadline arrived. By the end of 2025 the program had been folded into a new organization called the Defense Autonomous Weapons Group, the DAWG, with a budget line in the fiscal 2027 request that placed it as the single largest year-over-year increase across the Pentagon. The Hicks bet had been right in its direction and behind on its arithmetic.

The American firms growing into the gap looked, from the outside, more like the Ukrainian model than the traditional contractor base. Anduril, the company Palmer Luckey had founded in 2017 with a hundred million dollars in venture funding after his exit from Facebook, had grown by mid-2026 into a Costa Mesa-headquartered manufacturer with a five-million-square-foot factory under construction in Ohio, a thirty-one-billion-dollar private valuation, and Lattice, the AI command-and-control software platform that had become the de facto user interface for several Army and Marine Corps autonomy programs. Its Bolt vertical-takeoff loitering munition, its Pulsar electronic-warfare jamming suite, and its Roadrunner counter-drone interceptor were all available, at least in concept, in months rather than years. The first-generation Anduril systems sent to Ukraine in 2023 and 2024 had received mixed reviews from Ukrainian operators in the field, with several units quietly stopping use after performance complaints; that experience was filtering back into the second-generation designs by 2026 in a way Anduril’s leadership openly described as the company having to earn the right to build for the Ukrainian style of war.

The deeper question that the Spiderweb attack and its predecessors raised was the one Bob Work had not quite been ready to articulate at the Reagan Library in 2015. The Third Offset had argued that American military power would be vindicated by combining cutting-edge silicon with human-machine teaming. In one sense the argument was right. The fabs in Hsinchu and the GPUs they produced did sit at the foundation of every algorithm that mattered, and every modern military’s path to AI-enabled autonomy ran through that foundation. In another sense the argument had missed the war that was actually being fought. The drones that destroyed Russian armor and bombers and warships, every day, in every weather, did not run on three-nanometer parts. They ran on twenty-eight-nanometer microcontrollers, on twenty-eight-nanometer Wi-Fi radios, on forty-five-nanometer power-management chips, on whichever commodity silicon happened to be in stock at the JLCPCB warehouse in Shenzhen the week the batch was assembled. The leading edge mattered for training the models. It barely mattered for fielding them. What had become decisive at the contact line was not who could fabricate the smallest transistor but who could rewrite the firmware fastest.

That asymmetry inverted, in some uncomfortable ways, the bet Perry had made in 1977. Perry had wagered that attaching American military advantage to American silicon was a safe move because the United States made all the silicon that mattered. The bet had held for three decades and then collapsed slowly, as the leading edge moved to Taiwan and the country’s industrial capacity migrated to East Asia. Work had updated the wager in 2015 for a world in which the silicon was no longer American, but it was at least confined to a small number of leading-edge fabs the United States could reach with policy. The Ukrainian war had introduced a third proposition. The silicon that won the war did not have to be leading-edge at all. It had to be available, cheap, easy to integrate, easy to update, and produced in volumes high enough that losing tens of thousands of units a month was an operational matter rather than a strategic one. By 2026 most of that silicon was Chinese. The rest was Western parts that the sanctions machinery had not yet sealed off from gray-market routing.

For Beijing the implications cut both ways. Chinese-made commodity microcontrollers, image sensors, motors, and batteries had become the supply chain of choice for every party in the conflict, including the European NATO members rushing to build their own drone manufacturing in the second half of 2025. China’s export controls on civilian drones and drone parts, which had taken effect on September 1, 2024, had been written to give Beijing leverage over the inputs the war required, but in practice the parts kept moving. The Chinese export of complete drones to Ukraine had largely stopped. The export of the components Ukraine assembled into drones had not. For Washington, the leverage worked in the opposite direction. The October 2022 and October 2023 export controls had been designed against a leading-edge target. They had little to say about the twenty-eight-nanometer mainstream, which would have been a category error to attempt to control even if anyone had wanted to. The chip war as it had been fought in policy memos was a war against a thin slice at the top of the cone. The chip war as it had been fought in Ukraine was a war about the whole base of the cone.

What that meant for the next decade of American policy was not yet legible in any single document. The CHIPS Act’s leading-edge fab subsidies were sized to the top of the cone. The Brave1 program in Kyiv had been sized to the base. Both had been the right answers to different questions about which kind of chip war the country thought it was fighting. By the spring of 2026 it had become clear that any serious American answer would have to address both, and that the answer would have to be willing to take seriously the manufacturers who had spent the previous decade being treated as a footnote: the firms that made microcontrollers, the firms that made power management chips, the firms that made motors and sensors and batteries, the firms whose products sat in every drone over Ukraine and almost none of which had been mentioned in any major American chip-policy speech. The next offset, if it came, would have to be assembled out of the parts the last one had not bothered to count.