The Drone Is Disposable. The Loop That Builds It Isn't.

Replicator bought drones. Ukraine built a loop. Why the design-manufacture-deploy-observe cycle, not the airframe, is the asset that compounds, for strike drones and interceptors alike.

Adamant · Riyadh

In August 2023, Deputy Defense Secretary Kathleen Hicks stood up the Pentagon's Replicator initiative with a blunt diagnosis: mass is "the PRC's biggest advantage," and the department needed to answer it by fielding attritable autonomous systems "at a scale of multiple thousands, in multiple domains, within the next 18-to-24 months" ([1]). Congress backed the ambition with roughly $1 billion: $500 million pulled into the FY2024 appropriations process and $500 million in the FY2025 budget ([2]). Three years on, one industry analysis that modeled Replicator's own disclosed budget and quantity assumptions puts the average unit cost of a Replicator system somewhere between $84,200 and $210,500, a directional third-party estimate, not an official Pentagon figure, but the only public math anyone has run ([3]). Compare that to what Ukraine is actually paying for the FPV drones it now manufactures at a rate exceeding eight million a year: $300 to $400 each ([4]).

That gap, two to three orders of magnitude, isn't a story about money. The US has more of it than anyone. It's a story about what kind of system you're actually building. Replicator bought drones. Ukraine built a loop.

A drone is a snapshot. The loop is the asset.

A drone airframe is a fixed, depreciating physical object the moment it leaves the factory. It flies into a specific electronic-warfare environment against a specific set of countermeasures, and by the time the next batch ships, that environment has already moved on. What doesn't depreciate, what compounds, is the system that observed the failure, fed it back into the design, and shipped the fix. That's the oldest idea in maneuver warfare, formalized by US Air Force Colonel John Boyd as the OODA loop (observe, orient, decide, act), and the side that closes it faster wins, regardless of whose individual platform is "better" ([5]).

Ukraine's defense-industrial base has turned that theory into a production metric. A US Army officer who spent seven months embedded with Ukrainian units, the Ministry of Defence, and manufacturers describes drone makers there updating software and iterating hardware designs three to four times a year, with field modifications from brigade innovation teams folded into the next production run "within days or weeks." Manufacturers now put QR codes directly on delivered drones, linking soldiers to support teams "around the clock" so problems get resolved immediately instead of waiting for the next acquisition cycle ([6]). Compare that to the pace it's replacing: major US defense programs typically take twelve years from requirement to fielding, and the Pentagon's own Adaptive Acquisition Framework, built specifically to go faster, still targets two to five years, which the same article argues is still too slow for a front line that iterates in days ([6]).

The output of that loop isn't a better drone. It's a drone-producing organism that's structurally incapable of staying obsolete for long. Ivan Pavlenko, chief of the Ukrainian armed forces' electronic warfare directorate, put the operational tempo at roughly 9,000 drones deployed across the front every day as of October 2025 ([6]). No single airframe survives that environment for long. The manufacturing-and-feedback system does.

The production curve nobody procured

Ukraine's own numbers show what compounding actually looks like. FPV drone output rose from roughly 3,000–5,000 units in all of 2022 to 800,000 in 2023 to 2.2 million in 2024, with 2025 estimated at roughly 3–4 million depending on the source ([4]; [7]). By early 2026, Ukraine's National Security and Defense Council put annual FPV manufacturing capacity above 8 million units, spread across more than 160 companies ranging from large factories to small workshops ([8]). Deputy Defense Minister Serhii Boiev told a NATO conference in January 2026 that Ukraine is targeting 7 million total unmanned systems built in 2026, roughly 70 times the roughly 100,000 combat drones the United States produces annually, according to Bloomberg ([7]).

None of that scale was procured. It was grown, iteration by iteration, inside a domestic manufacturing base that treats every drone lost in combat as a data point rather than a write-off. That's the mechanism this piece is actually about: a recursive loop in which design, manufacture, deployment, and battlefield observation feed the next design cycle without ever breaking the chain to wait for a new contract.

What "good drones" actually buy you

Consider what that loop is capable of once it matures. In June 2025, the Ukrainian operation known as Spiderweb used 117 FPV drones, costing roughly $117,000 in total, to damage or destroy more than 40 Russian aircraft across five airbases, assets estimated at over $7 billion ([4]). That isn't a story about one clever strike. It's what a manufacturing base that has already iterated through millions of units, absorbed thousands of field failures, and shortened its feedback loop to days can produce almost as a side effect.

Now contrast that with the logic of buying "the best available" system off a procurement catalogue. Iran's Shahed-136 hits its target less than 10% of the time, mediocre by any traditional metric, but at roughly $35,000 apiece, Russia can fire mass salvos daily and simply outlast the defense on cost, having launched over 14,700 one-way attack drones between September 2022 and December 2024 alone ([9]). Ukraine's own military intelligence, as reported by Militarnyi [10], put Russia's Geran-4/5 output at roughly 3,000 drones a month as of August 2026. The lesson generalizes uncomfortably well: a mediocre weapon inside a scalable, self-correcting production loop beats a superior weapon that has to be procured one contract at a time, because the loop, not the individual unit, is what actually compounds.

3D-printed drones are the literal version of the argument

Firestorm Labs, a San Diego startup, has built the most literal version of this thesis into hardware. Its xCell units are semi-automated, shipping-container-sized microfactories (20 or 40 feet, deployable anywhere, running largely off-grid) that 3D-print a complete Tempest drone in under 24 hours, with each container capable of roughly 50 units a month ([11]). Firestorm describes the setup as "a semi-automated, expeditionary manufacturing cell that can be operated with limited human-in-the-loop engagement" ([11]). Backers including Lockheed Martin put in $12 million early; the US Air Force followed with a $100 million contract in January 2025 to integrate the platform into AFWERX programs at Eglin Air Force Base ([12]; [11]). Firestorm's own claim is that its modular, open-architecture drone costs about one-fifth the production value and takes about one-tenth the build time of a comparable fixed-wing system ([12]).

The point isn't that additive manufacturing is cheap. It's that a 3D-printed drone can be redesigned in CAD and reprinted at the next production run without retooling a factory floor, closing the same design-build-deploy-observe loop that Ukraine runs with injection-molded frames and soldering irons. Whichever fabrication method you use, the loop is the mechanism. The printer is just a low-friction actuator for it.

The actually unlimited part is software

Here's where the thesis sharpens: the physical drone is always going to be the limited half of the equation. It has a battery, a frame, a motor, a bill of materials, and a factory floor with a finite number of machines. What isn't limited in the same way is the software layer that designs it, flies it, and learns from it, and the companies treating that as the actual product are the ones building moats rather than backlogs.

Anduril's answer is Arsenal-1, a 5-million-square-foot "hyperscale" factory outside Columbus, Ohio, backed by $1.5 billion in Series F funding when it was announced in August 2024 and designed to build tens of thousands of autonomous systems a year ([13]). The point of the facility, according to Chief Strategy Officer Chris Brose, isn't the floor space; it's unifying every product's design and manufacturing data under one software system, Anduril's Lattice, so the company produces everything on "as common a manufacturing platform as possible" ([13]). Production of Anduril's Fury autonomous fighter began there in March 2026 ([14]). US Air Force acquisition executive Col. Timothy Helfrich has set Fury's per-unit cost target at under $30 million, roughly a third of an F-35A's price ([15]). Anduril's own marketing for Fury doesn't call it a fighter jet; it calls it "model-driven, field-tested," built for "affordable mass" ([16]).

The same logic is why Anduril and Palantir launched a joint push in December 2024 to move "exabytes" of sensor, vehicle, and weapons data from the tactical edge into a usable training pipeline, arguing that most of that data currently just evaporates instead of improving the next model or the next design ([17]). It's part of why the US Army handed Anduril a 10-year, up to $20 billion enterprise contract in March 2026, then in June 2026 made it lead integrator for the data layer meant to let every Army system (drones, air-defense platforms, vehicles) feed a common software backbone ([18]). And it's why Shield AI built its Hivemind autonomy stack to be platform-agnostic in the first place: the same software flying its V-BAT reconnaissance drone was integrated onto the unrelated Destinus Hornet airframe in a two-month campaign ([19]). Applied Intuition frames the resulting shift bluntly: autonomy is "not a hardware feature locked into specific platforms, but rather a software capability" redeployable across fleets, the pattern the industry now calls "affordable mass" ([20]).

None of these companies are selling a drone as the product. They're selling the loop, and the drone is just today's output of it.

The same math applies to interceptors

Everything above applies with equal force to the defensive side of the ledger, which is where Adamant operates. A Wild Hornets Sting interceptor costs roughly $2,500; a US-made Patriot interceptor performing the same function costs more than $3 million, a cost ratio north of 1,000-to-1 ([4]). Lockheed Martin built approximately 600 of its most advanced PAC-3 interceptors last year. Ukraine's armed forces used roughly 700 of them in just four winter months of 2025–2026 ([4]). Read those two numbers together and the conclusion is unavoidable: the constraint was never willingness to spend on interceptors. It's whether an interceptor comes out of a recursive, software-defined production loop or out of a fixed annual contract that can't be re-optimized between one Shahed salvo and the next. A single two-person Ukrainian crew shot down 23 Russian Shaheds in one engagement in March 2026 using Sting interceptors ([4]), a result procurement alone cannot produce, because procurement doesn't learn.

The region's moat isn't localization. It's the loop.

The Gulf has already made the correct diagnosis on paper. Saudi Arabia's General Authority for Military Industries reports that domestic localization of military spending reached 24.89% by the end of 2024, working toward a Vision 2030 target of more than 50% by 2030 ([21]; [22]). But localization, building the hardware domestically, only answers half of what this piece is about. A factory that assembles someone else's design under license is still, in the sense this piece argues against, buying "good drones." It's a snapshot, not a loop. The harder and more valuable target is owning the design-manufacture-deploy-observe cycle end to end, so every unit fielded in the region, interceptor or strike drone, makes the next one better without waiting for a new program of record.

That's the argument we keep making internally at Adamant, and it's why we've built OWL around the data layer rather than around a single hardware SKU: sustainable scalability, not one better missile, is what actually compounds. Procuring good drones is a starting inventory. Building the loop that keeps refining them, and the interceptors built to stop them, is the moat.

A note on sources. Production and cost figures in the drone and interceptor space vary meaningfully between outlets, and we've tried to flag that rather than average it away. Ukraine's 2025 total drone output is cited as roughly 3 million by Just Security [4] and roughly 4 million by Euromaidan Press [7]: both plausible, neither definitive, and we've kept both in view rather than picking the more dramatic one. The $84,200–$210,500 Replicator unit-cost range is a third-party model built on public budget and quantity disclosures, not an official Pentagon figure, and [3] is explicit that it's directional. The 3,000-Geran-drones-a-month figure traces to Ukrainian military intelligence via [10], not to independently verified production data. Where a number could plausibly be wrong in either direction, we've said so rather than let the citation do work the source itself doesn't support.