The State of the Edge Operator 2026

Something fundamental changed in how military capability gets built. For seventy years, battlefield technology flowed in one direction: from defense primes, through program offices, down to the operator. The operator was the last to touch the tool and the last to shape it.

Ukraine broke that model. Soldiers procured drones off retail shelves, volunteers shipped artillery software that reached 90 to 95 percent of artillery units without ever being an official program, and commercial-component aircraft became the war's primary kill mechanism. At the same time, the Pentagon and NATO formalized what the battlefield already proved: the January 2026 Department of War AI Strategy directs investment in AI compute "from datacenters to the edge" and makes commercial model parity a primary procurement criterion.

This report assembles the evidence. It draws on primary doctrine, congressional research, frontline fieldwork from RUSI, CSIS, the Belfer Center, and IFRI, and verified market data. Every factual claim is cited and dated.

The conclusion we draw is not that bottom-up innovation replaces enterprise command and control. Ukraine itself pairs unit-level commercial procurement with Palantir-class targeting synthesis. The conclusion is that a layer is missing. Enterprise C2 platforms see the theater. The operator holds a supercomputer. Almost nothing connects the two when the network goes dark. That layer, operator-owned AI running locally on consumer-class hardware, is where we built EdgeLance, and this report explains why we believe it is the most consequential open opportunity in defense technology today. The thesis runs through every section that follows, and it fits in five words: the operator is the platform.

The most common drone of the war in Ukraine has not been a defense program of record. It has been the Chinese-made consumer DJI Mavic, available for roughly $2,000, lost by the thousands per month and tolerable to lose precisely because replacement cost is trivial.¹ Before the full-scale invasion, Ukrainian units were already repurposing commercial "wedding" camera drones for combat reconnaissance.² By 2024, Ukrainian FPV production exceeded 2 million units, of which only about 1 million had been contracted by the government as of November 2024. The rest moved through volunteer networks and direct unit purchases.³

The frontline technology stack is recognizably consumer-grade. As Zhytomyr-born tech executive and frontline supplier Vsevolod Liscovich described it: a Starlink terminal with a generator or a battery, a drone "you could buy on Best Buy," and WhatsApp or Signal for communications.¹

These systems are not a sideshow. RUSI fieldwork published in February 2025 found that tactical UAVs, mostly FPVs built largely from commercial components, accounted for 60 to 70 percent of damaged and destroyed Russian systems, even though 60 to 80 percent of FPVs fail to reach their targets.⁴ Cheap, attritable, commercially sourced systems became the dominant kill mechanism of a peer war.

Capital has noticed

Defense tech had its best funding year ever in 2025: $49.1 billion in VC deal value by PitchBook's count, up roughly 80 percent from $27.2 billion in 2024. CB Insights, using a narrower equity definition, recorded $17.9 billion, more than double 2024, with US equity funding nearly tripling to $14.2 billion.⁵ "Ukraine demonstrated drone and autonomous system effectiveness in real combat, fundamentally shifting how VCs view defense investments," CB Insights' head of insights Jason Saltzman noted.⁵ The record did not hold long. By May 2026, Crunchbase counted $14.6 billion already invested in its military and national security categories, blowing past its full-year 2025 record of $9.6 billion.⁶

The signal for capability planners is direct: the consumer electronics supply chain, with its volumes, price points, and iteration speed, is now a battlefield input. The question is no longer whether commercial hardware belongs in the fight. It is who puts intelligence on it first.

On January 12, 2026, the Department of War released its Artificial Intelligence Strategy, a Secretary-level directive that reads like a commercial technology adoption mandate. It commits the Department to "invest substantial resources in the expansion of our access to AI compute infrastructure, from datacenters to the edge," and directs the Chief Digital and AI Office to establish a cadence with AI vendors "that enables the latest models to be deployed within 30 days of public release," declaring that this "shall be a primary procurement criterion."¹

The strategy stands up seven Pace-Setting Projects. One of them, Swarm Forge, pairs elite operator units directly with commercial technology innovators to test AI-enabled combat concepts such as drone swarms, with validated packages transitioning to operational units in 90 days or less.² The operating model the Pentagon is institutionalizing is precisely the one Ukraine improvised: put commercial technology in operators' hands, iterate at unit level, field fast.

NATO moved earlier. The Alliance's revised AI strategy of July 2024 states that "it is vital for NATO to use these technologies, where applicable, as soon as possible," explicitly including frontier and foundation models. It also concedes the hard part: the need to overcome "technical and governance issues when civilian-market dual-use solutions are applied in military context."³ That admission defines an alliance-level capability gap that commercial edge AI is positioned to close.

Leadership and money are aligned

The Defense Innovation Unit's mandate, stated jointly by then-DIU Director Doug Beck and then-CDAO Radha Plumb in January 2025, is to bring commercial AI and autonomy into the Department "with the focus, speed, and scale required for strategic effect." The same leaders wrote that AI tools must be exercised "every single day from the command level down to the tip of the spear."⁴ In one DoD exercise, DIU and the Navy cut the time to update AI models on mine-hunting unmanned underwater vehicles from six months to a few days.⁴

The budget follows. The FY2026 request leans into AI and autonomy across the services: $789 million for Air Force collaborative combat aircraft, $147 million for tactical autonomy, $486 million for Army basic AI research, and $582 million for data labeling and model deployment under Alpha-1.⁵ Within the Department's $66 billion IT budget, each service posted double-digit AI growth.⁶

Consider two artillery systems. Dzvin was Ukraine's official, top-down battle management program. It "fell victim to bureaucratic hurdles and corruption," in CSIS's words. Kropyva was a volunteer-developed artillery app with no official status. It reached 90 to 95 percent of Ukrainian artillery units.¹ The volunteer-built Delta situational awareness system became the de facto standard the same way, and was only formally institutionalized into the armed forces in 2023, after adoption was already won.¹

CSIS identifies the mechanism: "One of the most significant accelerators of military AI development in Ukraine has been direct communication between engineers and military personnel."¹ Nearly 700 frontline units hold direct commercial purchase authority, and IFRI's February 2026 study of the war's technology lessons concluded that "tactical innovation consistently emerged at the unit level, often faster than formal procurement could respond."² The Belfer Center put it plainly in 2023: Ukraine innovates "bottom-up with soldiers procuring off the shelf technology."³

The Replicator benchmark

The US ran the inverse experiment. Replicator, unveiled August 28, 2023 and led by DIU, was the Pentagon's flagship top-down effort to field thousands of low-cost attritable autonomous systems by August 2025. It was explicitly grounded in Ukraine lessons, with observers estimating Ukraine expends up to 10,000 low-cost systems per month, and framed by Deputy Secretary Hicks as the answer to China's advantage in mass.⁴ By the deadline, per a former defense official cited by the Congressional Research Service and confirmed by the program's former DIU director, the Department had fielded only hundreds.⁴

The honest read is not that top-down fails and bottom-up wins. Critics correctly note that Ukraine's model produced fragmentation and scaling problems, and Kyiv is now consolidating data standards around Delta.⁵ The honest read is about clock speed: when the operator owns procurement, software, and feedback, iteration runs in weeks. When capability must flow down through a program office, it runs in years. A force that wants both scale and speed has to build for both, which is the subject of Section 09.

On June 1, 2025, Ukraine's SBU launched 117 FPV drones from trucks parked inside Russia against five strategic bomber bases, including Belaya in Irkutsk, roughly 4,300 kilometers from Ukraine. President Zelensky said the operation, Spiderweb, had been in preparation for a year and a half, and that one staging site sat "directly next to FSB headquarters in one of their regions."¹ Kyiv claimed 41 aircraft hit, a third of Russia's cruise missile bomber fleet, and some $7 billion in damage. Independent assessments are lower and still extraordinary: NATO assessed 10 to 13 aircraft completely destroyed, and Janes confirmed four destroyed Tu-95MS bombers at Belaya alone via satellite radar imagery.² The exchange rate is the point. CSIS put it at $600 to $1,000 FPV drones destroying $250 million bombers that are no longer in production. The losses are not expensive. They are irreplaceable.³

Twelve days later, Israel opened its war against Iran the same way. The Mossad had established an attack drone base inside Iran, with systems smuggled in via "ships, trucks, and suitcases," and activated it overnight on June 13 against missile launchers and air defenses near Tehran, clearing the corridor for the manned strike packages that followed.⁴ Two intelligence services, two theaters, one playbook: the decisive air weapon was not flown into contested airspace. It was carried in through the front door, in pieces, on commercial hardware.

The supply chain cuts both ways. In September 2024, thousands of Hezbollah pagers, sourced through a Hungarian shell company licensing a Taiwanese brand, detonated across Lebanon in an operation Israel later confirmed.⁵ Every force now has to assume its hardware and software can arrive compromised. Provenance, signed delivery, and cryptographic verification of what runs on a device are no longer IT hygiene. They are force protection.

The defender's broken cost curve

The same asymmetry is bankrupting the defensive side of the ledger. In the Red Sea, the US Navy fired roughly 220 Standard-series missiles at Houthi drones and missiles by January 2025, and the Navy Secretary put the cost of Mideast engagements at nearly $1 billion as early as April 2024. The interceptors run $2.1 million for an SM-2 and around $4 million for an SM-6. The Houthi drones they kill cost $2,000 to $20,000.⁶

Read together, Spiderweb, Rising Lion, and the Red Sea say something larger than any one operation. Offense at the tactical edge has been democratized by the commercial drone supply chain. Defense has not. Whoever closes that gap, cheap, distributed, intelligent detection and response at the perimeter, owns the next decade of force protection. The Outlook section returns to this.

Strip away the doctrine and the funding charts, and the person at the point of contact faces three compounding problems: the connectivity is not reliable, the infrastructure behind it is not theirs, and the kit they carry generates a record they do not control.

The link you depend on belongs to someone else

On July 24, 2025, Starlink suffered a global outage from a failure of its own core network software. Ukraine's Unmanned Systems Forces commander Major Robert Brovdi posted from the front: "Starlink is down across the entire front." Combat missions ran without video feeds; reconnaissance fell back to strike drones.¹ It was not the first time. Outages dented Ukraine's 2023 counteroffensive, and in 2022, when Ukraine made an emergency request to extend coverage toward Sevastopol for an attack on the Russian fleet, the company's owner refused, judging the escalation risk himself. The geofence held because one civilian decided it should.² The dependence problem has reached US testing too: an August 2025 Starlink outage left two dozen US Navy unmanned surface vessels idle off the California coast, and the Navy's safety report concluded that "Starlink reliance exposed limitations under multiple-vehicle load." Europe drew its conclusion at budget scale, activating the first elements of its €10.6 billion IRIS2 constellation in January 2026 to reduce dependence on a single US commercial provider.³

The cloud fails on quiet Tuesdays

The cloud behind the link is no sturdier. On October 20, 2025, a latent race condition in AWS's DNS automation took down us-east-1 for roughly 15 hours, degrading more than a thousand downstream services. The July 2024 CrowdStrike update crashed some 8.5 million Windows machines, cancelled over 5,000 flights, and cost Fortune 500 companies an estimated $5.4 billion. Cloudflare's November 2025 outage took X, ChatGPT, and UK banks offline; Azure's October 2025 Front Door failure ran more than eight hours.⁴ Every one of these happened in peacetime, in friendly territory, with nobody jamming anything. An architecture that fails on a quiet Tuesday in Virginia has no claim to reliability under electronic attack in a contested theater.

The kit collects on its wearer

The third pain is quieter. The operator's own equipment builds a record about the operator, held elsewhere, usable later. US Naval Institute Proceedings reported in 2023 that sailors and Marines are wary of command wearable programs, specifically whether punitive action could be taken based on the data.⁵ The Pentagon warned troops off consumer DNA kits in 2019 because the data "could expose personal and genetic information" and create "increased risk to the joint force and mission."⁶ The evidence concern is not hypothetical: a Royal Marine's own helmet camera footage convicted him at court-martial in 2013, and the body-camera literature finds such footage is more commonly used to prosecute than to protect the wearer.⁷ Add the Strava record from Section 08 and the pattern is complete: sensors that operators do not control become liabilities that operators carry. The predictable result is distrust of the very tools that could keep them alive.

What operator-first actually means

These three pains define the bottom-up design brief, and they are EdgeLance's founding assumptions rather than edge cases. No cloud in the mission loop, so there is nothing in Virginia to fail. No third party in the kill chain, so no vendor, carrier, or constellation owner can geofence the mission; the operator's own devices are the infrastructure. And data minimization under operator control: collection is governed by signed mission policy, surveillance and physiological data inherit classification-driven expiry, and cryptographic burn means the record of the mission ends when the mission authority says it ends, not when a vendor's retention policy does. The operator is not the product, not the data source, and not the weak point. The operator is the customer.

RUSI's frontline fieldwork describes the operating environment in unsparing terms: "navigational jamming is ubiquitous throughout the combat area," command-frequency jamming is widespread, and drone mass observation has made the battlefield transparent within 3 kilometers of the line of contact.¹ Any capability that depends on a continuous link, to a satellite, a cloud region, or a command server, degrades exactly when it matters most.

The battlefield's answer has been to move intelligence onto the munition and the device. Ukrainian FPVs have been upgraded with autonomous terminal guidance, on-board AI that flies the final attack phase after jamming severs the link, and with fiber-optic spools that RUSI describes as rendering them "impervious to electronic disruption."¹ Upgrade kits adding last-meter autonomy under heavy jamming have been reported at price points around $100.² Notably, AI in this war functions mainly as an enabler of human-led processes, concentrated in narrow computer vision tasks: terminal guidance, route correction, target tracking.³ The pattern is consistent: small, specialized, local models doing one job well, on whatever compute is on hand.

Consumer silicon caught up

The constraint that once made this impossible, model quality at edge-deployable scale, is dissolving. EdgeRunner 20B, a fine-tune of OpenAI's open-weight gpt-oss-20b trained on 1.6 million curated military records, matched or exceeded GPT-5 on military-specific test sets spanning combat arms, combat medicine, and cyber operations, with 95 percent-plus statistical significance and two narrow exceptions.⁴ The caveat matters: it is a vendor self-benchmark on author-created test sets, not yet independently replicated. The strategic point survives the caveat. A 20-billion-parameter open-weight model runs on operator-class hardware, high-end laptops today, premium mobile silicon next. Domain-tuned local models no longer concede capability to the cloud on the tasks operators actually need.

The market is pricing this in. The military edge computing market was valued at $3.42 billion in 2025 and is projected to reach $10.32 billion by 2034, a 13.9 percent compound annual growth rate.⁵

Put the two trends together, an electromagnetic environment that punishes connectivity and silicon that no longer requires it, and the design requirement for operator-level systems writes itself: assume disconnection, run inference locally, sync when the link returns.

For over a decade, tactical mobility meant Android. The Army's Nett Warrior program selected a chest-mounted Samsung Galaxy Note II running an NSA-approved Android build as its end-user device back in 2013, and the pattern held.¹ The TAK ecosystem reflects it today: TAK.gov notes that ATAK testing runs "across a range of Samsung devices due to device fielding primacy," while iTAK offers only "a subset of ATAK capabilities" for iOS.² The force itself lives on the other platform. iPhone holds roughly 58 to 60 percent of US mobile OS usage and 62 percent of US smartphone shipments, and among American teens, the cohort the services recruit from, 87 percent own an iPhone.³ The institution buys Android. The people who fill its ranks carry iPhones, and every recruiting class widens the gap.

In February 2026, the institutional barrier cracked. iOS 26 and iPadOS 26 were listed on the NATO Information Assurance Product Catalogue, approving iPhone and iPad to handle classified NATO information up to NATO RESTRICTED, following what Germany's BSI described as exhaustive technical assessment, and requiring no third-party add-ons: the certification covers the devices' built-in security, native apps included.⁴ For the first time, the hardware operators already prefer is cleared, at alliance level, for restricted mission data as shipped.

And the silicon arrived at the same moment

Four months after the NATO listing, Apple shipped the inference leap to match. The third generation of Apple Foundation Models includes AFM 3 Core Advanced, a 20-billion-parameter sparse on-device model that activates just 1 to 4 billion parameters at a time, natively multimodal, alongside the 3-billion-parameter AFM 3 Core.⁵ The Foundation Models framework now gives third-party developers direct API access to these on-device models, including a Spotlight-powered search tool "for implementing fully local Retrieval-Augmented Generation" and image input at any size, with watchOS 27 extending the intelligence stack to the wrist.⁶ Frontier-adjacent inference, fully offline, on the phone in the operator's pocket, is no longer a research demo. It is a platform API.

Read Sections 02 and 06 alongside this one and the door is plainly open. Doctrine demands commercial AI at the edge. The EW environment demands that it run locally. NATO has now certified the consumer device fleet operators already carry, the data custody model favors it, and the device vendor has shipped 20B-class local models with developer APIs. What remains is mission software built natively for that stack. EdgeLance has been built for commercial devices from the start, with command nodes, field nodes, and readiness wearables on the hardware operators already carry, precisely because this convergence was coming. The certification and the silicon did not create EdgeLance's architecture. They caught up with it.

For a force whose location and communications metadata are targeting data, where the data lives is operational security. Apple's stated architecture keeps processing on the device, and where cloud assist is used, Private Cloud Compute is designed to be stateless and non-targetable, with its software published to a tamper-evident transparency log; these are vendor design claims, but unusually verifiable ones.¹ The independent evidence runs the same direction. Trinity College Dublin researchers measured a Pixel handset sending roughly 1MB of telemetry to Google at startup against roughly 42KB from an iPhone to Apple, about 20 times more, with both platforms transmitting "despite the user explicitly opting out"; follow-up work found Google Messages and Dialer sending message hashes, call times, and phone numbers to Google with no opt-out.² The custody difference has already mattered in practice: Google's server-side location history database made it the one-stop shop for geofence warrants, at one point roughly a quarter of all US warrants Google received, until Google moved location history on-device in December 2023 specifically to end them. Apple, holding no equivalent server-side store, reported receiving such warrants and having no responsive data to give.³ A platform that hoards less exhaust is simply a smaller attack surface for an adversary, or a subpoena.

Operators are already sensor platforms. Nobody is managing them.

The fitness tracker has become a recurring counterintelligence incident. In March 2026, a sailor aboard the French carrier Charles de Gaulle logged a deck run on Strava and exposed the ship's position northwest of Cyprus during its Iran-bound deployment; Le Monde verified the leak against satellite imagery.⁴ It was the latest in a long pattern: the 2018 Strava global heatmap that revealed US bases in Syria and Afghanistan and lit up the flight deck of HMS Queen Elizabeth, Le Monde's 2024 "StravaLeaks" investigation tracking the bodyguards of Biden, Macron, and Putin, its 2025 follow-up identifying French ballistic missile submarine crews, and the 2022 FakeReporter operation that defeated Strava's privacy settings to expose personnel on Israeli bases.⁵ The Pentagon prohibited geolocation features in operational areas back in August 2018. The leaks kept coming anyway, because the devices kept coming.

They kept coming because the data is genuinely valuable, and the services know it. The Department has fielded Oura rings by the tens of thousands since 2019 by the vendor's count; the DIU and DTRA RATE program pairs a Garmin watch with an Oura ring across 165 biomarkers to predict illness up to 48 hours before symptoms; the Navy selected wearables for watchstander fatigue monitoring and signed WHOOP separately. When the Defense Health Agency tried to consolidate with a $96.1 million wearable award, the contract drew protests and was cancelled in March 2025.⁶ The picture is a force that issues biometric sensors with one hand and bans their data exhaust with the other, while the data itself rests in half a dozen consumer clouds. Physiological and biometric data is mission data. It needs mission-grade management: synced device to device, encrypted in motion and at rest, expiring with the mission, and never resident in a commercial fitness platform an adversary can scrape.

This is where the ecosystem argument completes itself. Apple's platform already ships the substrate for exactly that management model: Health data is end-to-end encrypted in iCloud by default, keychain material syncs end-to-end encrypted with hardware-security-module escrow, and Continuity moves state between an operator's devices directly over Bluetooth within one account, per Apple's platform security documentation.⁷ These are vendor claims, but they describe a sync fabric where encryption is the default rather than an add-on. EdgeLance builds the mission layer on top: Apple Watch readiness and duress nodes feed the operator's own mesh rather than a fitness cloud, biometric presence anchors authentication, and the same classification-driven expiry and cryptographic burn that governs mission data governs the operator's physiology. The wearable stops being an OPSEC liability and becomes a managed sensor.

In February 2026, reports that Anthropic's Claude models had been used in the operation to capture Venezuelan President Nicolás Maduro brought a simmering conflict to a boil. Anthropic held a Department of War contract worth up to $200 million, signed in July 2025 alongside OpenAI and Google, but the Department's January 2026 AI strategy required contracts to permit "any lawful use" of AI within 180 days. Anthropic refused two exceptions: mass domestic surveillance of Americans and fully autonomous weapons. Pentagon chief spokesman Sean Parnell announced the relationship was under review, stating that "our nation requires that our partners be willing to help our warfighters win in any fight."¹

The standoff escalated fast. Secretary Hegseth directed the Department to designate Anthropic a supply chain risk; Anthropic publicly refused to move, declaring that "no amount of intimidation or punishment from the Department of War will change our position," and took the designation to court.² The contract was terminated. In March 2026 a federal judge enjoined retaliation against the company on First Amendment grounds, while the D.C. Circuit in April left the supply-chain-risk designation in effect for covered systems.³ OpenAI, meanwhile, signed its own DoW agreement on different terms, with contract language barring intentional use for domestic surveillance of US persons.⁴

The lesson for capability planners is not about who was right. It is that a mission capability built on a rented frontier model inherits the vendor's policy surface, and the political relationship between that vendor and the government, as operational dependencies. Doctrine already implies the answer: the 30-day model parity mandate of Section 02 is only achievable if models are swappable. Model flexibility, the ability to change models and vendors without rearchitecting the mission stack, is now a readiness requirement.

The economics point the same direction

Even where policy permits, the bill is climbing. At the top tier, prices are rising again: Claude Fable 5 lists at $10 per million input tokens and $50 per million output tokens, double the prior Opus generation, and its new tokenizer can consume up to 35 percent more tokens for the same text.⁵ Unit prices below the frontier have fallen nearly 80 percent year over year, yet consumption has grown faster still, more than 100x over two years by one infrastructure estimate, and AI now accounts for 18 percent of cloud spend at AI-forward enterprises, up from 4 percent in 2023. Agentic workloads add failure-mode exposure: a single three-hour recursive agent loop generates roughly $3,700 in unplanned compute, and $37,000 per incident at ten concurrent agents.⁶

EdgeLance's mission cost model puts tactical numbers on this. A cloud-dependent intelligence node runs roughly $20 per hour all-in: about $10 per hour in API calls for continuous frame analysis, transcription, and threat queries, plus roughly $10 per hour in supporting infrastructure and licensing. A 20-node mission running 12-hour days burns $4,800 per day, and that is at mid-tier model pricing; routing the same workload through the newest frontier tiers such as Fable 5 pushes the API half of the curve substantially higher.⁷ Local inference on operator-owned silicon carries zero marginal token cost, which is why EdgeLance's compute policy defaults to local-first routing under an explicit per-hour cost ceiling.

It would be easy to read this report as an argument against enterprise platforms. Ukraine's own practice refutes that reading. Alongside unit-level commercial procurement, Ukraine employs Palantir's AI software, which "synthesizes knowledge of friendly force weaponry and ranges, and then makes recommendations for targeting solutions," per the Belfer Center, and in January 2026 expanded its Palantir relationship to train AI on sensitive military data.¹ Theater-level synthesis is a genuinely hard problem that enterprise platforms solve well.

What Ukraine demonstrates is an architecture, not a vendor preference. IFRI's conclusion: "Rather than a single monolithic battle management system, Ukraine demonstrates the value of open, adaptable architectures that tolerate partial failure, function under degraded connectivity, and integrate new tools rapidly."² Combat effectiveness has depended on software as an integration layer between bottom-up sensing and top-down synthesis.

The white space at the edge

The US is buying the top of this stack. It is still missing the bottom. The Army's NGC2 modernization, with Anduril and Palantir on the team, drew an internal Army memo citing "deep flaws" in 2025, and Modern War Institute analysts warn that new C2 tools risk overwhelming company commanders at the tactical edge with data they cannot process under fire.³ Pushing more enterprise software down the echelons is not the same as giving the operator intelligence that works when the network does not.

The complementarity thesis, supported by every evidence stream in this report, runs as follows. Top-down platforms aggregate, deconflict, and allocate at theater scale, and require connectivity and data centers to do it. Bottom-up edge AI senses, decides, and acts at the point of contact, and must assume disconnection. Each is weak precisely where the other is strong. The force that wires them together, local inference feeding enterprise synthesis when links permit, enterprise context cached forward to the operator before links fail, gets compounding returns from both.

NATO has already conceded the gap on the commercial side, and the DoW AI Strategy has already mandated the cadence on the model side. The remaining question is execution: who delivers operator-owned AI that plugs into the enterprise stack instead of competing with it.

EdgeLance is a working implementation of the operator-level layer this report describes: local AI command and control for denied, degraded, intermittent, and limited (DDIL) environments, running entirely on hardware an operator can carry, with zero cloud dependency.

Intelligence on the operator's own silicon

An EdgeLance command node runs eight-plus AI models simultaneously on a single commercial laptop: a large language model for scene understanding, medical reference, and query answering, alongside specialized vision, speech, and identification models for detection, transcription, OCR, and face matching. An inference broker distributes work across a multi-device compute mesh that auto-detects each node's tier, from 8GB phones to 128GB workstations, and a capacity policy keeps heavy inference within each device's envelope. The models are open-weight and swappable, with no per-token meter and no vendor usage policy in the kill chain, and the compute policy engine enforces an explicit per-hour cost ceiling whenever cloud routing is permitted at all. The fleet spans macOS command nodes, iOS and Android field nodes, Windows service nodes, and Apple Watch readiness nodes.

Built for the jamming-saturated battlefield

Section 06's environment is EdgeLance's design assumption. The mesh is DDIL-aware: encrypted envelopes, CRDT-based state sync, multi-hop relay, and bandwidth-adaptive routing, with a LoRa bridge for 220-byte signed bursts when everything else is down. Events queue locally during disconnection and burst out automatically when a link recovers. Nothing is lost, and nothing requires a cloud.

Plugged into the enterprise stack, not against it

Per the complementarity thesis, EdgeLance speaks the standards the top-down world already uses: a bidirectional TAK bridge (CoT over UDP multicast and TCP) with a native ATAK plugin, and Zarf plus UDS packaging for airgapped Kubernetes deployment, the same delivery pattern used across the DoD software ecosystem. Bottom-up sensing feeds top-down synthesis whenever links permit.

Operator control as a first principle

Autonomy is bounded by design. Thirty-one mission agents operate under signed policies; six low-risk actions are auto-authorized, and everything else requires operator approval. Mission data carries classification-driven auto-expiry and cryptographic burn: data is encrypted under an ephemeral key, and destroying the key makes it irrecoverable. Security is anchored to operator presence through multi-factor chains (NFC, facial, voice, and device posture), never to network connectivity, because a security model that depends on a heartbeat fails exactly when the EW environment gets serious.

The operator already carries the hardware. Doctrine already demands the capability. The battlefield already proved the model. EdgeLance is the layer that connects them.

On the night of September 9, 2025, some 19 to 23 Russian drones crossed into Polish airspace. NATO aircraft shot several down, Poland invoked Article 4, and within three days the Alliance stood up Eastern Sentry, a standing air defense activity on the eastern flank.¹ NATO Secretary General Mark Rutte drew the lesson in cost terms: "it is the drone technology which is really changing the way we conduct these wars. You see now $400 Ukraine drones taking out multi-million dollar... Russian tanks."² The threat is not confined to the front. In late 2024, unidentified drones loitered for days over RAF Lakenheath, Mildenhall, and Feltwell, and earlier over Langley Air Force Base, with attribution still unresolved.³ The Atlantic Council's prescription names the gap directly: NATO "needs a 'hellscape' defense at 'Replicator' speed."⁴

The money has already moved

Washington has now priced this future. The FY2027 request, released in April 2026, asks for $1.5 trillion for the Department of War, the highest single-year level since World War II and roughly 38 percent above FY2026 in real terms.⁵ Where the growth lands matters more than the topline. Drone and counter-drone funding triples to roughly $75 billion. A new Drone and Autonomous Warfare Group account jumps from about $200 million to $54.6 billion, the largest year-over-year boost of any program in the request. Army counter-small-UAS procurement nearly triples. AI and autonomy draw $58.5 billion.⁶ Secretary Hegseth's framing at the rollout and in testimony tracks this report's argument almost line for line: "Drones are so central to the future of warfare... we have to be able to both make the exquisite ones better than anybody else, and also the attritable swarm, and then the ability to defeat them ahead of our adversary, learning from battlefields like Ukraine." A sub-unified command for autonomous warfare was announced as forthcoming.⁷

The allies are moving on the same vector. At The Hague in June 2025, NATO committed to 5 percent of GDP by 2035, split as 3.5 percent for core defense and up to 1.5 percent for defense-related spending that the declaration explicitly directs toward protecting critical infrastructure, defending networks, and unleashing innovation. Secretary General Rutte called it "a quantum leap that is ambitious, historic and fundamental."⁸ The money is already visible in the trend line: NATO Europe and Canada went from $359 billion in 2021 to more than $574 billion in 2025, a 20 percent real jump in the final year alone, with EU-27 spending hitting a record €343 billion in 2024 and an estimated €381 billion in 2025.⁹ More telling than the totals is the plumbing being built for AI and edge technology specifically: a Rapid Adoption Action Plan mandating that new technology reach the field within 24 months of an identified need, the €1 billion NATO Innovation Fund as the world's only multi-sovereign venture fund, a DIANA accelerator cohort of 150 innovators for 2026, the EU's €150 billion SAFE instrument oversubscribed within months, the UK committing £4 billion to autonomous systems, and Germany signing multibillion-euro loitering munition frameworks.¹⁰

Section 04 showed why the current answer fails: $2 million interceptors against $2,000 drones is not a defense, it is a depletion schedule. Base and FOB defense against small drones and ground intrusion is, at bottom, an edge AI problem. Detection has to be continuous, local, and cheap. Continuous means every camera, microphone, and RF sensor on the perimeter watched all the time, which no human staffing model and no per-frame cloud API bill survives. Local means it keeps working when the link is jammed or cut, which Section 06 established as the default condition. Cheap means it runs on hardware the unit already owns.

This is the defensive half of the EdgeLance stack, running on the same nodes that support offensive tasking. Perimeter cameras under continuous AI watch with zone alerts and camera tasking among the system's auto-authorized agent actions. Local vision models handling object, person, and face detection with NFC-backed IFF to separate friendlies from unknowns. Audio transcription and speaker challenge at checkpoints. A detected threat triggering drone reposition for a closer look, with the operator approving anything beyond it. Alerts propagating across the mesh to the watch on a sentry's wrist, with no cloud round trip in the loop.¹¹ Surveillance data collected at the perimeter inherits the same classification-driven expiry and cryptographic burn as every other mission artifact, so defending the base does not mean building a surveillance archive an adversary can steal.

The future of warfare visible from June 2026 is symmetrical. Cheap, intelligent, distributed systems attack, and only cheap, intelligent, distributed systems can affordably defend. Both run on consumer-class hardware, local models, and operator control. The force that fields that stack first, at the edge where the fight actually happens, sets the terms in both directions. This report opened with a missing layer. It closes with the thesis that fills it: the operator is the platform.