Summary

The investable bottleneck is not military demand for technology but the Pentagon’s ability to stop funding its legacy stack. Alex Miller calls the DoD black box an intentional product of decades-old rules followed by “duct tape and bubble gum”; the Army receives 21% of the defense budget, has flexibility over less than half of that, and must predict needs three to five years ahead. Programs of record then create “a perverse incentive to just maintain and create inertia.”
Defense technology may have its strongest commercialization window in a lifetime, but innovation activity remains tiny. Justin Fanelli sees unprecedented alignment between national-security demand and companies willing to serve it: “Only superior military technology can credibly deter war.” He calls the next six to 12 months the best chance he has seen to move technology not yet under government contract into full-scale production—while noting DIU and similar activity still represent less than 1% of the budget.
The emerging procurement model is capability per dollar, with new products funded by actively killing old ones. Across 75 program executive offices—18 in the Department of the Navy—Fanelli wants “military Moneyball”: portfolios that compare outcomes rather than protect sacred programs of record. Horizon 1 vendors can displace an already-budgeted capability; Horizon 2 companies can use transition mechanisms such as APFIT; genuinely new Horizon 3 markets still require DIU, Congress, requirements writers, and “guerrilla marketing.”
Simpler requirements can turn an 18-month acquisition into three months—and expose how badly the old paperwork measures performance. Miller found counter-drone specifications whose combined availability, accuracy, and reliability allowed a system to work only 51% of the time; Fanelli encountered a 4,000-page requirements document that nobody likely read cover to cover. A few pages of capability-need statements compressed one process from 18 months to three. Fanelli also warns against cutting strong software capabilities down with “1,000 paper cuts, 10,000 paper cuts.”
The case for adoption is power-law overmatch, not a marginal 15% improvement. Examples include replacing potentially a billion-dollar infrastructure purchase with a small hyperconverged edge-compute stack, adapting a Navy electronic-warfare solution for the Army in three months for $300,000, and using high-performance compute to run 100,000 repetitions that would otherwise have taken 10 years. Fanelli’s test is blunt: the new capability must “kick the ass of whatever is there right now.”
Field adoption, not a polished prototype, is the decisive proof point. Miller wants companies working directly with Army units because “only the soldiers will actually be able to tell you, yes, this works at a time of crisis.” The Global War on Terror showed equipment could reach Afghanistan in 30 days, but those quick-reaction channels became side quests; “this is a period about scaling what actually works” through the main system.
The cultural reset is to price delay as risk and reward winning rather than procedural safety. Acquisition officials traditionally minimize waste, fraud, and abuse without asking whether a dollar spent after 10 years still carries a dollar of value. Miller would rather lose a dollar on a fast experiment; Fanelli compares the current incentive to grading a quarterback only on avoiding mistakes instead of touchdowns and wins.
Founders should preserve commercial discipline while getting unusually close to warfighters. Miller advises spending perhaps nine months near the operating community, while Fanelli advises avoiding bad requirements and taking “roof shots” toward the product’s moonshot. Miller warns against making DoD the sole source of revenue because that destroys genuine dual use. The green flag is a strong independent product brought into the field for tailoring—not a company reshaped around one government customer.

Deep dive

1. The black box is inherited policy, not an absence of demand

Miller describes his Army CTO role as roughly 60% education and 40% hands-on engineering: setting a technology north star, connecting it to missions, and helping a sprawling institution move toward it. Unlike a company, the Army answers to Congress and the public while operating under “60 years of policy and rules” implemented by risk minimalists.
Fanelli says government recognized the CTO role about 30 years after the private sector, and later than defense. His Department of the Navy role spans the Navy and Marine Corps: find commercial and edge innovation, scale it faster, and tie adoption to divestment because money remains locked into existing technology over three-year spending cycles.
Miller’s diagnosis is institutional: the Packard Commission’s acquisition guidance, the Clinger-Cohen Act’s information-technology rules, Goldwater-Nichols structures, and everything subsequently layered on top created the black box. “If you can’t admit there’s a problem, you can’t actually solve the problem.”
Fanelli details routes through the existing machinery; Miller’s pushback is that every such route remains a workaround. Program executive offices are not themselves required by statute, he argues, so anything below law should be reshaped rather than forcing founders to master “acronym soup.”

2. Portfolios can replace programs of record with continuous competition

The current scale problem begins with 75 program executive offices, including 18 in the Navy, containing hundreds of program-management offices. They purchase at scale, but identifying which office owns a need has been unnecessarily difficult.
Fanelli’s answer is “military Moneyball”: convert sacred programs of record into portfolios that buy capability per dollar. PEO Digital has begun that shift, letting managers compare alternatives and repurpose money instead of assuming a named program deserves permanent protection.
The route depends on product maturity. Horizon 1 improvements can approach the office holding the existing budget. Horizon 2 technologies can draw on Valley-of-Death funding, including APFIT; Fanelli calls SBIR more of a “door prize” and says companies still need operations-and-maintenance funding. Horizon 3 products creating an entirely new category still need DIU, Congress, and requirements writers, plus “guerrilla marketing,” to help create a market.
Miller objects to the destination implied by “program of record.” Once a company secures a 30-year budget line, “there’s a perverse incentive to just maintain and create inertia”; he wants products to keep earning adoption, much as product managers must keep improving Excel even after it becomes ubiquitous.

3. Capability statements are outperforming gold-plated requirements

Traditional requirements grow thick because writers fear anything omitted can never be bought. In one counter-unmanned-systems document, Miller found that the specified availability, accuracy, and reliability combined into a product required to work only 51% of the time—effectively accepting waste of the other 49%.
Fanelli’s extreme specimen was a 4,000-page requirements document that, he suspects, nobody read cover to cover. By using capability-need statements and top-level requirements under the software acquisition pathway, his team compressed an 18-month process into three months and only a few pages.
Rapid reuse is already working. A small hyperconverged edge-compute box deployed ashore could enable divestment of a billion-dollar purchase; when a ship captain volunteered, the Navy moved it afloat, deployed it to the Red Sea, and then placed it on additional ships.
In another case, the Navy recognized that an Army electronic-warfare request resembled its own system. The practical question became, “Can you just paint this brown and throw it on a Humvee?” The adaptation took three months and $300,000 before moving to scaled fielding.

4. Overmatch must make the legacy alternative indefensible

Fanelli defines overmatch as an orders-of-magnitude or “power law” leap because a 15% improvement rarely justifies switching costs. The Navy, for example, found roughly 1,000 identity-management solutions and designated Navy Identity Service as the enterprise system, freeing money and attention for harder problems.
Seronic’s evidence model captured what Fanelli wants: cloud and high-performance computing reportedly ran 100,000 repetitions of a test that would have taken 10 years conventionally. That produced an easier A/B decision on paper and reduced the risk required before an exercise.
Fanelli points to Transformation in Contact as a tech-informed execution model; Miller describes the Army’s approach of placing deployable technology into brigades in Europe, the Pacific, and the Middle East, then letting users determine whether it works. “Only the soldiers will actually be able to tell you, yes, this works at a time of crisis.”
During the Global War on Terror, Miller could request an urgent capability and receive working equipment in Afghanistan within 30 days. That proved rapid acquisition was legal, moral, ethical, and possible—but it remained a side channel. The present task is to fix “the main storyline” and scale those behaviors institution-wide.

5. Shared infrastructure frees each service to concentrate on unique missions

Army mission command accumulated 17 programs of record, each with its own map server, tiles, storage, and delivery stack. Even after Google Earth looked compelling around 2015–16, it took five years for people to realize they could share KML and KMZ files; in 2025, the Army was only beginning to “hack and slash” the vestigial systems and their protected budgets.
Miller’s preferred architecture is Lego-like: let the best data-serving company serve data and the best interface company build UI. Requiring one winner to deliver every layer multiplies mediocre tolerances until soldiers receive products “that no one would accept if I handed it to you in your everyday life.”
The enterprise should provide services of common concern—data sharing, API gateways, compute, identity, and reusable AI infrastructure—while each service spends its energy on genuinely distinctive missions. Fanelli similarly asks why separate services should make incompatible autonomy or artificial-intelligence decisions and discover the integration problem three years later.
After meeting major LLM providers, Miller’s takeaway was that Llama, Claude, ChatGPT, and even DeepSync were all available; differentiation came from customer data and how it was used. “This is not a technology problem. This is a culture and a process problem that we can apply technology to.”

6. Budget structure misprices time and trains officials to play not to lose

The pivotal historical example is 1968, when Minuteman III appeared by name in the federal budget rather than being funded from a broad portfolio. That made sense for a singular government-led strategic program, but thousands of directed line items now force the Army to predict technology three to five years ahead.
The Army receives 21% of the defense budget and has flexibility over less than half of its allocation, Miller says. Laboratories therefore attempt a “magical transition” from lab success to field deployment in one shot, even though software can move from concept to implementation in days or weeks.
In response to Hay’s venture-capital analogy, Fanelli cites In-Q-Tel’s ratio of 25 outside dollars for every federal dollar spent. Acquisition staff are tasked with reducing waste, fraud, and abuse rather than valuing time; Miller would rather lose a dollar on a prompt experiment than preserve it through a 10-year process.
Fanelli’s sharper analogy: measure a quarterback solely on avoiding mistakes and the performance changes completely. Latent risk-takers and “mavericks” already exist in government, but the system must reward outcomes, speed, and divestment; “We don’t want two successes a month. We want to overwhelm the system with how much better it can be.”

7. The coming year is about production, autonomous formations, and founder discipline

Fanelli sees the next six to 12 months as the best chance he has seen to move a capability from outside government contracting into full-scale production and operations-and-maintenance funding. The necessary permissions increasingly exist; the open question is how many teams will use them and generate repeatable wins.
The Army plans to move its next-generation command-and-control consortium into a program phase and field it to a division in Europe. Miller’s proof point was an M1A2 SEPv3 commander replacing dependence on a roughly 20-year-old mounted tablet with command-and-control capability on an Android phone.
Project Flytrap will apply lessons from Ukraine to counter-drone operations with the fifth core, combining unconventional sensors, compute, electromagnetic detection, acoustics, and automated recommendations. A separate autonomy consortium will assemble the toolchain from simulation through verification and validation because “it’s not one company” that will create an autonomous formation.
The founder playbook is equally explicit: reject unusable requirements, stay near warfighters, preserve the original product north star, and do not rely on DoD as the only customer. Bring data when policy blocks adoption—Miller says his team has gotten 12 policy exceptions that way—and bring data and stories that make a capability’s advantage “undeniably good,” not merely incremental.

Reward Hacking by Reasoning Models & Loss of Control Scenarios w/ Jeffrey Ladish, from FLI Podcast Focus Capital Advisers' Mordechai Yavneh on the greatest acquisition of all time (Valeura Energy)