The State of the Edge Operator 2027

In June we argued that the operator had become the platform. The most capable computer in a fight was the one in a soldier's pocket, the most lethal weapon was a cheap drone built from commercial parts, and the layer that would connect the two to enterprise command and control did not yet exist. We called that the operator gap.

A great deal has happened in one year. In 2026 the United States went to war with Iran, watched drones probe its own bomber bases, and proposed the largest single commitment to autonomous warfare in its history. In the four months since our first report it ordered tens of thousands of small attack drones through competitive Gauntlets, put sea drones into combat for the first time, announced a four-star command for autonomous warfare, and split the Army's next command and control system into a platform and a set of applications. Ukraine moved from drones flown by pilots to drones that finish the strike on their own.

This edition records what changed. As before, every factual claim is cited and dated, and we have drawn on official releases, congressional and defense press reporting, research institutions, and Ukrainian government data. Where sources disagree we say so. We also score our own 2026 calls against what happened, which a research practice should do every year.

The second half of the report is about EdgeLance. Section 12 summarizes what we built in 2026 in response to the evidence, and Section 13 sets out our 2027 roadmap. Each item on the roadmap traces to a finding in the first ten sections. The pattern of the year is that the money, the mass, and the policy all arrived, and the operator gap got wider. Closing it is the work of 2027.

The FY27 budget request asks for $54.6 billion for the Defense Autonomous Warfare Group, the office that replaced Replicator. The group received $225.9 million in fiscal 2026. Only $1 billion of the request sits in the base budget. The remaining $53.6 billion depends on a reconciliation bill, which gives the department up to five years to obligate it.¹ David Petraeus called it the largest single commitment to autonomous warfare in history.²

On September 30, Secretary Hegseth announced a four-star Autonomous Warfare Command, with a target activation date of October 1, 2027, subject to congressional action. As an interim step, the unmanned systems portfolio becomes Project Agincourt, led by the director of the Defense Innovation Unit.³ Two weeks later at AUSA, the Army said its Futures and Concepts Command would become Futures and Autonomous Systems Command, building autonomy into every echelon.⁴ At the same conference Hegseth said the department needs roughly $50 billion to take autonomous systems from thousands of units to millions.⁵

The money is real, and it is not yet appropriated

Readers should hold two facts at once. The direction of policy is unambiguous, and the money has not arrived. The government is operating under a continuing resolution through December 11, and Breaking Defense reported in October that more than 150 new programs face delay until a full-year bill passes.⁶ Analysts have also pointed out that the request is large and thin on detail, and that swarm doctrine has not caught up with the money.²

For programs and vendors the practical lesson is that 2027 will reward work that is already funded, already fielded, or easy to buy through existing vehicles. The Army's drone marketplace, which we cover in Section 02, is one example. Software that rides on platforms already being bought is another.

Capital is concentrating

Private money moved first and moved big. PitchBook counted $19.8 billion invested in defense and adjacent startups across 262 deals in the first quarter of 2026, the largest quarter in its dataset, and $35.6 billion for the first half.⁷ One round, a $5.0 billion raise by a single defense prime-in-waiting, made up about a third of the second quarter. The top ten rounds took 62 percent of the half. Autonomy was the largest theme, at about $16 billion over twelve months.⁷ By October, PitchBook data showed investors still concentrating on larger companies while early-stage dealmaking slowed.⁸

The pattern matters to anyone building for the edge. Capital and contracts are flowing to platforms. Specialized edge software will reach units by riding those platforms and the programs that buy them, not by competing with them.

The allies are paying too

At Ankara in July, NATO reported that European and Canadian defense spending rose 20 percent last year. Allies announced about $50 billion in new multinational defense-industry deals, including a drone and counter-drone marketplace, and banks committed $217 billion in additional financing.⁹ The European Union's SAFE program offers $175 billion in loans as the first stage of a plan to support up to $934 billion in defense spending.¹⁰ The money for the edge is now transatlantic.

The Drone Dominance Program is the clearest success of the year. The first Gauntlet led to orders for about 30,000 small drones. The second, scored at Fort Carson in August and posted September 15, ran 1,858 sorties across 23 platforms from 19 companies, and every long-range platform reached 15 kilometers, against 24 percent reaching 10 kilometers in the first round. The department plans to buy about 60,000 drones from the second pool and more than 200,000 across all phases by 2027. A third Gauntlet follows in about six months and a fourth six months after that.¹

Buying also got easier. The Army's UAS Marketplace, launched in March on commercial cloud infrastructure, lets units and allies compare vetted systems and place orders. At AUSA the Army reported more than $800 million in sales over about four months, and eight more countries signed on, bringing participation to 33 nations.²

The pilots are the gap

The deadline for a one-way attack drone in every Army squad was the end of fiscal 2026. In September the Army said it had fielded 10,000 drones and was fielding 30,000 more, enough for the squad level, and units had accepted 12,080 aircraft from the first Gauntlet alone. Col. Gregory Merkl, who leads the effort at Army headquarters, named the remaining problem plainly: "Where we still have some work to do, candidly, is getting everybody trained at the appropriate level."³ The Army is closing its experimental 600-person Unmanned Assault Battalion and spreading its lessons across every unit instead of building dedicated drone formations.⁴

Decentralizing is the right call and it has a cost. Every squad leader now carries a training, tasking, and accountability burden that used to sit with specialists. In February small drones were still nonexpendable property, so losing one could trigger a property-loss investigation.⁵ By September a training commander described them as "like a munition."⁴ Munitions come with their own accounting: drawn, assigned, expended, on whose authority. The attritable force needs that discipline without the paperwork, which only software can provide at this volume.

The Marine Corps is moving the same way. Its ground combat vision for 2040, published in August, puts counter-drone defenses, loitering munitions, and a dedicated drone operator in the squad, which is back to 13 Marines.⁶

The opposing force already knows how to fight this way

The clearest warning came from an exercise. During drills in Germany this year, Ukrainian drone operators playing the opposing force took out a US armored brigade.⁷ At Project Convergence in July, a drone took about ten minutes to get airborne after a long battery charge, against under a minute for a comparable Ukrainian aircraft, and soldiers flew at half of maximum endurance to save batteries and avoid losing aircraft.⁷ The gap is practice, sustainment, and the habits that come from fighting with drones every day.

The program is moving to close it. In July the Pentagon added reusable bomber drones to Drone Dominance, with about $32 million for up to 1,200 prototypes and a Gauntlet in October.⁸ Reusable aircraft turn one tasking decision into dozens per shift, which raises the stakes for every job around the airframe: training, tasking, sustainment, accountability, and review. We set out those jobs in our September paper, Running an Attritable Force.

On September 30 the Army awarded Anduril a contract worth up to $1.8 billion over five years to take Next Generation Command and Control from experimentation to fielding, with an initial $162.8 million for the common data baseline across I Corps. On October 7 it announced about $93.6 million in one-year awards to nine companies to build applications on that baseline across the six warfighting functions. NGC2 is planned to reach all of I Corps by the end of 2027.¹

This is the complementarity thesis from our 2026 report, written into a contract. The enterprise layer is a platform. The value that reaches a specific unit for a specific mission comes from applications that sit on it. The question for 2027 is what those applications do when the platform's network is under attack.

Growing pains at the far end of the network

The early evidence is mixed. In September the 25th Infantry Division tried to run the NGC2 platform during Lightning Surge 5 in Hawaii, with Lockheed Martin as integrator. "It wasn't seamless, and we learned a lot," said Brig. Gen. Shane Taylor. Data passing through the division's legacy radios did not work reliably with the new platform, the 25th used a different waveform than the 4th Infantry Division had, and tablet integration was requested and completed in the middle of the exercise. Dense jungle and light vehicles were harder than Fort Carson's open terrain and Strykers.² I Corps' commander has said the Pacific brings a "tyranny of distance and disruption" that earlier tests did not.³

None of this is a failure. It is what happens when software built on a strong network meets the network a light infantry division actually has. It is also the clearest description we have seen of the operator gap: the enterprise picture is improving fast, and the stretch between that picture and the soldier at the end of a jammed, mountain-shadowed link is where capability drops away.

The enterprise layer is consolidating

Maven Smart System became a program of record in March and passed 100,000 users by September.⁴ The FY27 request includes more than $2 billion to move joint command and control from fragmented deployments to a program of record.⁵ The enterprise market is consolidating around a small number of platforms, which is good for interoperability at the top and raises the stakes for whatever runs underneath them at the edge.

Where the software already pays

The same exercises showed what works. At Project Convergence, a battalion commander said new data-sharing tools cut the time to plan and communicate missions by two-thirds, and satellite connectivity let him reach an operations center across the base without masts.⁶ The Army has asked for nearly $4 billion this year for NGC2.⁶ The transport underneath is being bought as well: Persistent Systems received $121.5 million in mesh radio orders for the NGC2 prototype across two awards.⁷

That is the order of operations the Army chose: transport, then a data foundation, then applications. It is sound. The applications that matter most at the edge are the ones that keep working when the transport is jammed, the satellite is out of view, and the division is spread across a thousand miles of ocean. Those are not yet on contract.

In the week of March 9, security forces at Barksdale Air Force Base reported multiple waves of 12 to 15 drones over the flight line and other sensitive areas. The drones had long-range control links and resisted jamming, and a briefing suggested the operators were testing the base's response. At the time Northern Command had a single counter-drone fly-away kit for the homeland, and its commander said about a quarter of detected drones were being defeated, up from almost none a year earlier.¹

The response is now funded. Joint Interagency Task Force 401's Domestic Shield program put a three-year, $500 million vehicle in place on July 1, with a first task order of $80.5 million for multi-sensor systems at Air Force Global Strike Command sites, and a second $500 million vehicle followed in September.² Five installations were selected for a directed-energy pilot: Fort Huachuca, Fort Bliss, Naval Base Kitsap, Grand Forks, and Whiteman.³ The Army awarded Anduril a ten-year counter-drone contract worth up to $20 billion in March, and JIATF-401 chose the company's software as its command and control layer for counter-drone operations, with a first task order of about $87 million.⁴

Sensors multiply faster than pictures

The Mitchell Institute argued in July that point defense is necessary but insufficient against small drones.⁵ We agree, for a reason that has more to do with software than with effectors. A base that buys radar, RF detection, and cameras from different vendors has several pictures and one watch officer correlating them by eye. A drone that crosses from one installation's airspace into a neighbor's starts over as a new detection. Nobody sees that the same operator is probing five sites in one night.

The defense of an installation has four seams: between sensors, between the picture and the authority to act, between neighboring installations and civil partners, and between the event and the review. Every one of them is a place where a correct detection can turn into no action. We set out the requirements for closing them in our August paper, Defending the Installation.

The second wave of awards bought fusion software

The September awards show buyers learning the same lesson. JIATF-401 added a second Domestic Shield vehicle worth up to $500 million over three years, covering RF sensing, electronic countermeasures, command and control, AI, and multi-sensor integration, and an award to a firm whose core product is edge-based software for fusing many sensors.⁶ Sensor fusion is now a line item. The task force's own guide stresses passive detection that reads signal characteristics rather than content, which keeps privacy rules intact on domestic installations.⁷

Europe is building the same wall

Europe faces the same problem at continental scale. The EU's European Drone Defence Initiative aims for initial capability by the end of 2026, and Eastern Flank Watch for full function by the end of 2028.⁸ The command and control arrangements are still unclear, including who manages interceptions and how a recognized air picture will be built. The cost case is plain: when NATO jets engaged drones over Poland in September 2025 they shot down four of an estimated 23, with fighters that cost $33,000 to $42,000 an hour to fly and missiles that cost about $1 million each. An EU primer called that approach "fiscally unsustainable."⁸

On September 8, Brave1 reported that successful strikes using AI terminal guidance had increased tenfold since January. In standardized tests with the Ministry of Defence, each drone had to fly two kilometers, lock onto a moving target from more than 500 meters, and strike without pilot input, hitting at least two of three times. Six of seven companies passed. Brave1 stressed that the decision to engage the autonomous mode stays with the operator.¹ Brave1 did not publish the base number behind the tenfold figure.

The data behind those modules is also being organized. The Brave1 Dataroom, launched in January, gives more than 100 Ukrainian companies access to frontline data for training detection and interception models under security controls.² Ukraine's Delta situational awareness system reports 200,000 users and an AI assistant inside the interface, and it was used to coordinate Ukrainian activity at NATO's REPMUS maritime exercise in October.³

Units ordered more than 590,000 drones through the Brave1 Market from January to September, and Ukraine's National Security and Defense Council reports 100,000 interceptor drones produced in a year, with mission success above 60 percent and a cost per Shahed downed more than 25 times lower than a Western air defense missile.⁴ These are Ukrainian government figures and we could not independently verify them, but the direction matches every other source.

The decision moves upstream

For three years the best counter to an FPV was to break its link in the final seconds. A drone that guides itself for the last 500 meters does not need that link. The implication for command and control is direct. When a pilot flies the whole way, the pilot is the last check on the target. When a module flies the last stretch, the last human check happens earlier, at nomination, approval, assignment, and the moment the mode is engaged.

DoD Directive 3000.09 already requires appropriate levels of human judgment over the use of force. With terminal autonomy, that judgment lives in the tasking record. Armies adopting these modules need a record of who approved which target, under what constraints, with what picture, and which asset was sent, and it has to survive a jammed network and a lost aircraft.

Ukraine is standardizing, and buying like a unit

Ukraine is also doing what the United States has not yet done: standardizing the data. At NATO's REPMUS maritime exercise in October, its Defence Ministry coordinated Ukrainian activity through Delta and reported that telemetry from assets following NATO STANAG 4817 is already integrated.⁵ Procurement is moving the same way. Brave1 Market now lets units order drones built to their own specification for payload, cameras, and endurance through competitive bidding.⁶

The lesson for American planners is that Ukraine's advantage is no longer improvisation. It is a common picture, common data standards, unit-level buying, and a testing regime that turns a frontline idea into a procurement recommendation in months. Each piece can be copied. Together they are a system.

On July 12, as part of a wave of strikes on Iranian targets, three uncrewed surface vessels hit the submarine and ship maintenance facility at Bandar Abbas Naval Base. It was the first American combat use of sea drones. A month earlier a boat of the same type, operated by Task Force 59, recovered two Army aviators after their Apache went down off Oman. The boats are 24 feet long, carry up to 1,000 pounds, and range more than 1,000 nautical miles.¹

The Navy is now weighing how the campaign will speed up autonomy in mine countermeasures, where its first uncrewed surface vehicle program of record has already deployed.² The Council on Foreign Relations argues that drones and missiles have largely replaced mines as the favored tool against chokepoints, because they let the attacker choose which ship to hit.³

A sea drone at range spends most of its life out of contact. Supervision is intermittent by design, so the vessel acts on orders, constraints, and abort conditions it received before the link dropped. After a strike the explanation has to be rebuilt from data: what it was assigned, by whom, what it saw on the way, and when it last heard from a person. The maritime fight is the purest version of the problem every uncrewed system now faces, and it argues for tasking with leases and receipts and for a mission record that lives on the assets themselves.

Uncrewed vessels now sail with the fleet

On October 8 the Theodore Roosevelt Carrier Strike Group deployed with a medium uncrewed surface vessel, the first carrier strike group to do so. The acting Navy secretary announced it at AUSA on October 14, and the builder identified the vessel as its Seahawk prototype, first delivered as an unarmed surveillance and anti-submarine sensor ship.⁴ Uncrewed boats are moving from task forces and experiments into the formations that fight.

Your own drones can be taken

The risk runs both ways. In June, Ukraine lost contact with four uncrewed surface vessels about 250 kilometers off the Romanian coast. One appeared at a berth in the port of Constanța the next day and detonated after the area was evacuated. Ukraine assessed Russian electronic warfare as the most likely cause.⁵ A fleet that cannot verify where its own vessels are, and who is directing them, can be turned against its hosts. Signed tasking, verified positions, and a record of every command are force protection at sea.

On February 28, as the war with Iran began, electronic interference disrupted navigation for more than 1,100 commercial ships in Gulf waters in a single day. Lloyd's List Intelligence logged 1,735 interference events against 655 vessels in the first four days, and navigation systems placed tankers at airports and on dry land.¹ In the Baltic, thirteen European nations and Iceland warned in January about interference affecting shipping, and commercial flights over Sweden lost satellite navigation as recently as September.²

The response on the ground has been to train for it. The Army's lesson from Saber Junction, as Col. Merkl put it, is that soldiers must learn to navigate the spectrum the way they navigate terrain.³ The Marine Corps wants tools that let Marines assess their own electromagnetic emissions so the enemy cannot find them first.⁴ Ukraine has gone around the problem with fiber: its forces received 374,000 fiber-optic drones under 2025 contracts.⁵

Two consequences for software

First, any position report can be false, so a picture that accepts every report is a picture the adversary can draw on. Trackers have to validate, refuse what they cannot verify, and keep disputes visible. Second, every transmission is a signature. A squad that streams video and beacons all day has made itself easy to find. The quietest design processes data on the device and sends the result: a detection, a track, a line of transcript.

Procedure fills the gap where the picture is missing

Units are adapting their doctrine to the spectrum they expect. In February the 3rd Combat Aviation Brigade published a three-color rule for small drones sharing airspace with its helicopters in Poland: green, drones up to 500 feet when no rotary-wing traffic is planned; amber, drones laterally clear of corridors; red, every drone grounded. The brigade noted that communications in a contested fight are limited, so the Army's handbook favors procedural control over positive control.⁶ The drones that probed Barksdale in March had long-range control links and resisted jamming.⁷ Defenders and attackers alike are building for a spectrum they cannot count on.

Procedure is the right fallback. It is also expensive, because a rule written for the worst case is applied all day. Section 08 shows what that cost at Fort Irwin, and Section 10 sets out the alternative.

At Project Convergence Capstone in July, about 10,000 soldiers ran simulated large-scale combat at Fort Irwin, many with hardened commercial Android phones on their chests and battery packs on their backs. The 25th Infantry Division fielded about 2,200 of these end user devices. Soldiers used them for maps, enemy markers, photos, chat, and planning, and officers praised the real-time targeting feedback. When connectivity dropped, from simulated electronic warfare and other causes, apps slowed and maps lost detail. A platoon sergeant said his device was not working properly without a connection. Troops kept paper maps alongside the phones, and one captain said the analog copy is what he fights and lives on in degraded environments.¹

The Wall Street Journal's October account of the same exercise filled in the rest. With temperatures near 110 degrees, servers overheated, batteries died and could not be recharged, and soldiers draped ice-soaked T-shirts over failing antennas. One drone overheated and landed. Another took about ten minutes to launch after a long charging wait, against under a minute for a comparable Ukrainian aircraft. Drones lost signal behind the mountains and were flown at half their maximum endurance without autonomy or AI assistance. A five-drone swarm turned back when three Black Hawks entered its airspace. The same reporting found that new data-sharing tools cut one battalion's mission planning time by two-thirds.³ Most of what failed was the layer around the hardware: power, heat, links, launch readiness, and a shared picture of the airspace. Most of what worked was software moving information faster.

The devices have plenty of compute. On-device AI is now mainstream on commercial phones and laptops, and transcription, summarization, and assistant features increasingly run with no network at all.² The limit is software that treats the network as the place where the data lives. Lt. Gen. Michael McCurry said the exercise was informing the Army about who needs these devices and how far down they should go.¹ We think the answer depends on what the device does with no network.

The design requirement is local-first. The mission lives on the device: map, tracks, plan, messages, and approved models. Losing the link costs freshness, not function. Changes sync device to device over the team's own mesh and catch up with the wider network when it returns. That is the standard units should test against by pulling the link on purpose at the worst moment and seeing what is left on the screen.

Commercial devices are cleared; commercial habits are not

The clearance barrier for commercial phones is largely down. NATO cleared iPhone and iPad for RESTRICTED information in February 2026, and the Army is putting hardened commercial Android devices on soldiers' chests by the thousand.⁴ What has not changed is how the software on those devices is built. It is designed for a network that is always there, because that is how commercial software is built.

Project Convergence also tested augmented-reality headsets. The Army has three companies under contracts worth more than $100 million each, and a senior official at the exercise said the headsets showed deficiencies and the service was uncertain about moving forward. An Army spokesman said the program remains on track.⁵ Hardware on the soldier keeps arriving. The question for each device is the same one: what does it do when the network is gone, the battery is half drained, and it is 109 degrees?

GenAI.mil had about 1.7 million users in September, roughly 500,000 of them heavy users, out of a force of about 3 million. When the current administration took office, about 80,000 people in the department used generative AI. Users have built about 100,000 agents, and frontier models are slated to extend from the controlled unclassified network to secret, top secret, and special access networks.¹

The most important sentence from that event came from the chief digital and AI officer, Cameron Stanley: "AI is getting put into everything. That scares me, mainly because of capacity." Compute capacity, he said, is going to be difficult to overcome.¹ Every one of those 1.7 million users depends on a data center and a network path. That works at a desk. It does not work for a platoon under jamming or a ship in emissions control.

Special operators asked for the alternative at SOF Week in May: smaller models that need less compute and still understand intent, and processing close to where data is collected, because cloud tools do not reach remote areas.² In August the Army's AI Integration Center and a commercial partner released an open-weight model tuned on Army data and built to run locally in air-gapped environments.³ The edge version of enterprise AI is starting to appear. It needs a way to be staged, governed, and used offline by units that do not have a data science team.

There is also an economic argument. Every inference that runs on a device the unit already owns is one the data center does not serve. Edge AI is part of the answer to Stanley's capacity problem as well as to the jammer.

Congress is writing rules for the models

The FY27 defense bills begin to govern the models themselves. Provisions summarized by War on the Rocks would create a department-wide program for reporting, analyzing, and fixing AI incidents and vulnerabilities, allow the department to ban and remove certain AI platforms from its systems, and direct a review of policy for autonomous weapons and AI-enabled targeting and planning.⁴ Meanwhile GenAI.mil's agent designer lets the department's roughly 3 million personnel build their own assistants without writing code.⁵

Two implications follow for the edge. Units need the ability to swap a model quickly, because a model can be banned or recalled by policy. And they need a record of which model produced which answer, because incidents now have to be reported and analyzed. Both are easier when the unit, not a vendor, controls the model loadout.

In October 1983, American forces landed on Grenada and could not talk to each other. Army, Marine, and Navy units carried radios that did not interoperate. Ships in sight of the Rangers could not take their calls for fire, and Marines in the north could not raise the Rangers in the south. The best-known story has an officer placing a commercial long-distance call to Fort Bragg to get support routed to the fleet offshore; its details have never been fully confirmed.¹ The radio failures are not in dispute, and they helped drive the Goldwater-Nichols Act of 1986.² Goldwater-Nichols unified command. It could not unify equipment.

Forty years later the problem is back in a new form. The second Gauntlet alone tested 23 platforms from 19 companies.³ The Army's drone marketplace serves 33 nations.⁴ Soldiers carry commercial phones, vehicles carry commercial satellite terminals, and helicopters now launch and control drones of their own. Each product arrives with its own app, its own ground station, and its own idea of the picture.

Every seam showed up in the field this year

At Lightning Surge 5, data did not pass reliably between new command and control software and the 25th Infantry Division's radios, which ran a different waveform than the division that tested before it.⁵ At Project Convergence, a drone maker attributed launch delays in part to communications parts built by a separate company, and a five-drone swarm turned back when Black Hawks entered its airspace, because its tasking had no picture of the helicopters.⁶ The 3rd Combat Aviation Brigade's color rule exists for the same reason: without a shared picture of crewed and uncrewed traffic, aircraft are separated by procedure.⁷

Closed boxes impose a tax that grows with the square of the systems. Six systems wired to each other need fifteen adapters; twelve need sixty-six. Each adapter is a contract, a test event, and a failure point, often built in the field at the worst moment. The tactical cost is worse. A drone that cannot take a task from the unit's plan is flown by one soldier for one purpose. A sensor that reports only to its own screen adds a screen. An aircraft the drone crew cannot see stops the whole sky.

Congress and Kyiv point the same way

The House version of the FY27 defense bill would direct the department to establish a common operating system for small drones, with interoperability as the stated purpose, and require dedicated test and training corridors.⁸ Ukraine already integrates telemetry from assets that follow NATO STANAG 4817 into Delta.⁹ A common standard is coming. The open question is who will own it and whether it will work at the edge, disconnected, on the hardware units already carry.

This is the second thesis of this report. The operator is still the platform, and the platform is now dozens of devices, vehicles, and aircraft from dozens of makers. Software that binds them into one plan, one picture, one task list, and one record is the multiplier the purchases are waiting on. Our October paper, One Mission, Any Hardware, sets out the full case.

The 2026 report made ten claims in its executive read. We have scored the ones that could be tested against events since June.

The miss worth noting is speed. We expected the budget to roughly triple the relevant accounts. The request went far beyond that, and most of it now depends on a reconciliation bill and a continuing resolution. We also underestimated how quickly the US would adopt a competitive, fly-before-you-buy model for drones. The Gauntlets are the most Ukrainian thing the Pentagon has done.

The operator gap, the central claim of the 2026 report, got wider. The enterprise layer is consolidating and well funded. The number of aircraft, sensors, and devices at the edge is growing by an order of magnitude. The software that connects an approved decision to an asset at the end of a contested link is still, in most units, a soldier with a radio.

EdgeLance is mission software for the tactical edge. It gives disconnected teams one place to plan, see, decide, task, and review the mission, runs local AI on the hardware the unit already carries, and syncs across whatever link is left. It sits between enterprise command and control and the transport layer. In 2026 we focused on the parts of that job the evidence said mattered most.

Plan-to-asset tasking

An approved plan task becomes an offer to every asset that can take it, from any vendor. Assets bid on what they can do given position, payload, and endurance. The winner holds the task on a lease, and if it goes quiet the lease lapses and the task goes back on offer. Results return as receipts on the task. Approvals are signed by a role allowed to give them, and dispatch only accepts signed requests. This answers the tasking problem in Sections 02, 05, and 06.

A picture that fails closed

Our multi-domain tracker fuses radar, RF, camera, ADS-B, and vehicle reports into one set of tracks. It validates every observation, refuses backdated, foreign, and downgrading inputs, keeps contested identities visible as separate tracks, and cannot be exhausted by floods of conflicts. It was built for the environment in Sections 04 and 07.

Mission blocks

Capabilities ship as blocks that share one mission vocabulary of observations, tracks, assets, and tasks. Each block declares its inputs, outputs, and permissions, runs with its own least-privilege credentials, and is validated before deployment. A full tasking chain composes from five blocks, and the whole system installs into air-gapped Kubernetes from one signed package.

A first-party mesh

Devices sync directly with each other with multi-hop relay, partition tolerance, and signed events. A device that loses every peer keeps the full mission and its own record, and catches up when peers return. This is the local-first requirement from Section 08.

Staged mission AI

Units use any approved model, local or API. Models, documents, and memory budgets are staged in mission packs per device class and checked by a fit gate before step-off. Answers cite the unit's own sources, and an inference gateway routes each request to the best place it can run right now. This is the edge half of the AI picture in Section 09.

Review and burn

Every task, approval, receipt, and observation lands in a mission record the unit owns. The review replays the mission for the debrief, and the Coach scores radio traffic against doctrine. When the mission ends, the burn destroys the keys for sensitive data while the record of decisions survives.

Our roadmap for 2027 has one aim. EdgeLance will keep the independent record of what every machine in the fight was told, what it did, and whether it stayed inside its authority. Every item below runs disconnected, works with any vendor's hardware, and keeps the operator in authority. None of it adds a second operating picture. Units already have one, and we build to work with it.

Why assurance leads

The force is about to field autonomy faster than it can certify it. The Army's new acquisition executive for autonomy must field autonomous fires, combat vehicles, watercraft resupply, and crewed-uncrewed teaming no later than fiscal 2028.¹ In August, Army Test and Evaluation Command asked industry for continuous assurance of AI and autonomous systems that works in disconnected conditions and keeps a human override, across aviation, ground maneuver, and air defense.² Directive 3000.09 already requires that verification before an autonomous weapon is fielded.³ Every one of those fielding decisions needs evidence that the autonomy stays inside its limits.

That evidence is worth only as much as its independence. A vendor's own report on its own autonomy will not settle a review board. We do not build flight control or guidance, and we do not sell the autonomy we check. Section 05 showed the last human decision moving upstream as terminal autonomy spreads. The record of that decision, and of everything the machine did after it, is where accountability now lives. Assurance is how a commander holds it.

Why the rest follow in this order

The vehicle contract comes with assurance, because assurance needs telemetry and a unit with aircraft from a dozen makers needs one way to see their state and send them approved work. Observe-only comes first, so trust is earned before commands are enabled. The airspace measures that ground sorties today (Section 10) become the limits the assurance check holds every task to. Sustainment and plan comparison ride on the same roster.

Installation defense builds on both, on top of what the task force has already bought. JIATF-401 selected an enterprise command and control layer for counter-drone operations this year.⁴ The seams in Section 04 remain open: the site procedures commanders must now issue within 60 days of the December guidance,⁵ authority that must hold when the link to higher drops, tracks that restart at the fence line, and the record the review board needs the next morning. We close those seams. Europe is the second market. Its drone wall still has no common command layer, and one analyst called EU drone defenses "largely missing in action" in July.⁶

Training answers the gap Section 02 measured. The Army is extending its drone specialist to every applicable unit and standardizing advanced training in a five-week course at Fort Rucker.⁷ A unit that rehearses on the system it fights with, and cannot launch a configuration that differs from the rehearsal, carries its training into the fight. Every training sortie produces the same evidence an operational one does, so qualification, readiness, and assurance come from one record.

EdgeLance Data answers Sections 07 and 08 for the whole ecosystem. The layer that keeps EdgeLance working without a network should be available to other mission applications, including those built on enterprise platforms. The Defense Innovation Unit has named the failure it prevents: when a disrupted link returns, retry floods choke it.⁸ We hold the scope narrow in 2027: offline sync, a TAK bridge, and one SDK.

What we will not build

We will not build a competing command and control platform or a second operating picture. Units have those, and an independent check loses its value the moment it competes with what it checks. We will not build flight control, waveforms, or guidance modules. Those belong to the vehicle and radio makers, and the roadmap is designed to work with theirs. We will not add a path for high-consequence action without explicit authorization. And we will not require a cloud, a specific model, or a specific device. Any model, any hardware, any mission remains the rule.

Three questions will decide the year. Will the reconciliation money arrive, and on what schedule? Will the drone programs score how well aircraft fit a common tasking and accounting layer, or only how well they fly? And will the applications being built on the new enterprise platforms work at the far end of a jammed link?

Our view has not changed since last June. The operator is still the platform. What changed is the scale: hundreds of thousands of aircraft, millions of AI users, and a new command for autonomy. The layer that connects them to each other and to the operator, and keeps working when the network does not, is the most valuable thing left to build at the edge.

This report draws on official Department of War and service releases, congressional and defense trade press, research institutions, and Ukrainian government data published between June and October 2026, with earlier sources where they set the baseline. Figures are dated where they appear. Ukrainian production and procurement figures are government statements we could not independently verify. Vendor performance claims are attributed to the vendor. Sections 12 and 13 describe EdgeLance from its own engineering documentation.