The Ghost in the Machine

Congress just passed the largest defense budget in American history. Shipyards are being asked to double submarine production. Munitions plants are running three shifts. The money is there. The demand signal is unambiguous. And yet the people who actually know how to build things — who hold in their hands and in their memories the accumulated knowledge of how to weld a hull section, route a wire harness, or cure a composite panel — are retiring at a rate the industry cannot replace. We are about to discover that appropriations do not equal production. And we are about to learn it the hard way.

But this is not really a story about defense. It is a much older story about a particular kind of knowledge — one that humans have been losing, and failing to protect, for as long as civilization has existed.

What Knowledge Looks Like Before It Disappears

Sometime around the third century AD, Roman engineers mixed a concrete that modern materials scientists still cannot fully replicate. The formula used volcanic ash from the region around Pozzuoli — a compound now called pozzolana — that reacted with seawater to form crystals that actually grew stronger over time, self-healing microscopic fractures as they spread. The Pantheon, built with this concrete, has stood for 1,900 years without reinforcement. The sea walls of Caesarea Maritima have survived millennia of wave action that would destroy anything we pour today.

When the Western Roman Empire collapsed, the knowledge of how to mix it collapsed too. Not destroyed — simply untransmitted. For over a thousand years, nobody could explain why Roman structures outlasted everything built after them. The formula wasn’t in any text. It didn’t need to be. Every engineer who worked with it knew it the way you know how to ride a bike: completely, wordlessly, and in a way that makes writing it down feel almost beside the point. We cracked the chemistry in the last decade using electron microscopy. We still can’t fully reproduce the process.

Antonio Stradivari built roughly 1,100 violins between 1644 and 1737. Modern luthiers have his original blueprints, wood samples from the same Alpine forests, access to the finest materials on earth, and centuries of obsessive reverse engineering. They still cannot produce an instrument that sounds like a Stradivarius. The leading theory — confirmed only recently through chemical analysis — is that Stradivari soaked his wood in a mineral solution to protect against woodworm infestation, a preservation technique so routine to 17th-century craftsmen that it warranted no documentation. The wood cured differently. The resonance shifted in ways that even Stradivari probably didn’t consciously understand. The knowledge died with his workshop, invisible even to the man who held it.

The Chamorro navigators of Micronesia could cross thousands of miles of open Pacific without instruments of any kind. They read the ocean itself — the swell patterns that bend around islands hundreds of miles away, the star paths that shift with the seasons, the color of water over different depths. This was not intuition. It was a rigorous, transmissible body of knowledge — but one that could only pass through years of apprenticeship on the water, navigator to student, generation to generation. Spanish colonization disrupted that chain in the 16th century. A single Micronesian elder, Mau Piailug, spent the last decades of his life attempting to reconstruct and pass on what remained before he died in 2010. The tradition survives today in fragments.

The knowledge that seems most obvious to those who hold it is exactly the knowledge that never gets written down. It doesn’t feel like knowledge. It feels like common sense.

This is what anthropologists call tribal knowledge — the accumulated, unwritten, experiential intelligence that a community builds over generations and that evaporates when the community disperses. It is not ignorance or carelessness that causes the loss. It is the simple human assumption that transmission will continue the way it always has. The apprentice will watch the master. The student will sit beside the navigator. The next engineer will work alongside this one. And then, at some point, that assumption proves wrong — and what seemed permanent reveals itself to have been, all along, extraordinarily fragile.

Parker is Retiring in April

Parker has been welding submarine hull sections for 31 years at a yard on the Gulf Coast. He knows that the particular grade of steel his crew works with requires a specific pre-heat sequence when the temperature drops below 50 degrees — a detail not in the procedure manual because when the manual was written, everyone just knew it. He knows that the interpass temperature gauge on bay seven runs three degrees low and compensates without thinking about it. He knows that the seam at frame 14 has a tolerance quirk — a consequence of a design revision made in 1997 — that every experienced welder on the floor accounts for by feel, that has caused exactly one failure in 29 years, and that a new hire would have no way of knowing to look for.

Parker is not unusual. He is the American defense industrial base. A quarter of the aerospace and defense workforce is at or beyond retirement eligibility. The Navy estimates it needs 140,000 new skilled workers over the next decade for submarine production alone. These numbers describe a capacity crisis, which is serious. But they understate the knowledge crisis, which is existential. You can train a new welder in months. You cannot train the 31 years of contextual judgment that tells Parker when the metal sounds wrong before any sensor catches it. That knowledge has never been written down because it never needed to be. It transferred body to body, floor to floor, the same way it always had.

Until now.

The standard response to this problem is automation. Buy robots. Install them. Close the gap. And that instinct is correct — as far as it goes. The problem is that it treats the workforce shortage as a capacity problem when the deeper crisis is a data problem. A robot is only as smart as its brain. Its brain is only as smart as its model. Its model is only as good as the data it was trained on. You can spend $50 million on the most advanced robotic assembly system ever built, and on day one it will still underperform Parker — because Parker’s 31 years of judgment are not in the training set. His model was built on every variance, every environmental quirk, every undocumented failure mode this specific floor has ever produced. A lab-grade dataset is not a substitute for that, at any price.

Wire Harnesses and The Half-stack Trap

Take wire harnesses. Every aircraft, submarine, armored vehicle, and missile system in the American arsenal runs on dense, custom-routed bundles of electrical wiring — no two identical, routing logic shaped by decades of design iteration and hard-won production knowledge about what fails under vibration, temperature, and stress. They are assembled almost entirely by hand, by people who have internalized that logic so thoroughly they couldn’t fully articulate it if you asked. It is exactly the kind of knowledge the Roman engineers had, the kind Stradivari’s apprentices were supposed to absorb, the kind the Chamorro navigators passed down on the water for generations.

Now imagine automating that assembly without first capturing the knowledge embedded in it. You’ve built a very expensive robot that makes harnesses the wrong way, faster.

This is the half-stack trap — and it is where most of the capital in defense manufacturing automation is currently pointed. Pure hardware plays assume the knowledge problem is someone else’s job. Pure software plays — factory analytics platforms, process monitoring tools, digital twin vendors — sit above the production environment without the authority to actually capture what’s happening inside it. The hardware becomes a commodity. The software has no signal worth training on. Neither wins. And the Parkers of the world keep retiring, taking with them something that a purchase order cannot replace.

Wire harnesses are not the only example. Fluid line assemblies, structural bonding, composite layup for airframes, precision weld fabrication, propellant loading, depot-level maintenance for next-generation platforms — in every one of these categories, the craft is everywhere and the data is nowhere. Quality is invisible until failure. Tolerance for error is zero. And the knowledge that prevents failure has never once been written down.

The Full-stack Imperative

The companies that will own this space are not selling robots or software to manufacturers. They are manufacturers — or so deeply embedded in production that the distinction barely matters. This is the only position from which you can actually solve the problem.

Here is what full-stack looks like in practice. You put engineers and instrumentation on the floor alongside the Parkers of the world — before they retire, not after. Computer vision watches the hands. Sensors read the material. Structured data capture translates judgment calls into machine-readable ground truth. You build a model on that data. You run the human and the robot side by side, using the human’s decisions to continuously refine what the machine knows.

Imagine Parker in his last two years before retirement, working beside a system designed to learn from him rather than replace him. Every compensation he makes for the temperature gauge. Every adjustment at frame 14. Every pre-heat decision on a cold morning. All of it captured, structured, fed back into a model that is, slowly, becoming capable of replicating not just the physical motion but the judgment behind it. When Parker leaves, the knowledge doesn’t leave with him. For the first time in human history, it doesn’t have to.

This is not a transitional architecture. It is the permanent architecture. Human and robot working side by side is not a phase to be managed through — it is the model. The goal is not to eliminate the craftsperson. It is to do for Parker’s 31 years of floor intuition what we failed to do for Roman concrete, for Stradivari’s mineral baths, for the Chamorro navigators’ star paths. To externalize what was always transmitted body to body, and make it durable.

What makes this possible now, when it wasn’t five years ago, is not a single breakthrough but a convergence. Multimodal data capture has become cheap and precise enough to instrument a factory floor without disrupting it. On-device inference has become fast enough to run models in real time. The models themselves have become capable enough to find signal in the messy, high-variance data that a real production environment actually generates. The technology has crossed the threshold. The bottleneck now is organizational will.

The obvious counterargument is that you don’t need to own the floor — you just need access to it. Hire Parker earlier, pay him more, have him train the next generation before he leaves. That is not a bad instinct, but it misunderstands the problem. Apprenticeship works when knowledge transfers body to body and the trainee has years to absorb it through repetition. The math no longer holds. Parker’s generation is leaving faster than apprentices can be trained, and the tacit knowledge they carry is too contextual, too environmental, and too varied to transmit through instruction alone. You cannot teach someone to hear that the metal sounds wrong. You can only build a system that learns to hear it for itself — and that system needs to sit on the floor, beside Parker, accumulating ground truth before the clock runs out.

And here is the part that makes this a genuinely durable business rather than a one-time data capture play: the moat widens with every unit produced. More production generates more data. Better data trains better models. Better models improve yield and reduce error rates. Better yield wins more contracts and justifies higher prices. More contracts mean more production. The companies that get on this flywheel early and stay on it — that resist the temptation to license out the software and walk away from the hard work of running production — will find themselves in a compounding position that a late entrant cannot close. The data gap between the incumbent and the challenger grows every day the incumbent is on the floor and the challenger is not. That is a different kind of moat than a patent or a contract. It is structural, and it deepens over time. 

That is the hard part — and it is hard deliberately. Owning production means carrying manufacturing risk. It means slower scaling than a pure software model. It means building something that looks, from the outside, more like a manufacturer than a tech company — at least for a while. The businesses willing to make that commitment will accumulate a data asset that no half-stack competitor can acquire at any price. You cannot buy your way into the training set. You cannot license your way to 31 years of floor intuition. You have to go earn it, the same way Parker did — by showing up, day after day, until the knowledge is yours.

The production operation is the data flywheel. You cannot license your way to the training set you need.

Where The Capital is Going — And Where It Isn’t

The venture numbers tell a story of extraordinary momentum with a telling gap at its center.

Defense technology attracted $27.2 billion in venture investment in 2024, then nearly doubled to $49.1 billion in 2025 — the best funding year the sector has ever recorded. 

Manufacturing-focused defense investment specifically rose from $2.6 billion across 24 deals in 2024 to $4.7 billion across 39 deals in 2025. As broader venture markets cooled, defense tech re-accelerated, with the number of firms actively investing in the sector increasing 41% in a single year.

The vast majority of that capital has gone to autonomous systems, battlefield AI, advanced computing, and drone technology. These are important categories. They are also categories where the knowledge problem is relatively tractable — software is updatable, model weights can be retrained, and the learning happens in simulation as much as in the field.

The manufacturing knowledge layer — wire harnesses, structural bonding, composite layup, precision weld fabrication — has attracted a fraction of that capital. The $4.7 billion in manufacturing-focused defense investment in 2025 sounds significant until you note that most of it went to drone factories, space infrastructure, and defense electronics. The unglamorous, craft-intensive, ITAR-sticky production processes at the core of the DIB’s capacity problem remain dramatically underfunded relative to their strategic importance.

In 2026, execution, not invention, will determine returns. Companies that convert facilities into repeatable output will disproportionately capture both capital and contract velocity. The market is beginning to understand what this thesis has argued all along — that the bottleneck is not ideas or hardware, it is the unglamorous, knowledge-intensive work of building production systems that actually work at scale. Capital is starting to follow. It has a long way to go.

What We Look For

When we evaluate companies in this space, we run a simple three-part screen.

Where We’re Putting Our Conviction

At Alumni Ventures, we’ve been building a portfolio around this thesis for several years — not as a single sector bet, but as a coherent view that the intersection of defense, advanced manufacturing, and AI-enabled knowledge capture is one of the most durable investment opportunities of the next decade. A few of the companies we back that embody different facets of this thesis:

These three represent a slice of what we’re building. We have more portfolio companies in this space that we’ll be sharing in the coming weeks, and a pipeline of companies we’re actively evaluating that push further into each of these layers — the floor, the software above it, and the ownership structure beneath it.

The Clock is Running

The money is in place. The platforms are being built. The demand is real and growing. What cannot be appropriated, contracted, or ordered into existence is the knowledge that makes production actually work — the 31-year mental model, the fingertip judgment, the institutional memory that lives in people and not in documents.

The companies that solve this will not look like defense primes, and they will not look like Silicon Valley robotics startups. They will look like something the industry hasn’t really seen: manufacturers who understand that the robot and the data are inseparable, who had the patience to go earn their training set the hard way, and who built something that — unlike Roman concrete, unlike the Stradivarius, unlike the navigators’ star paths — actually survived the moment of transmission.

Parker is retiring in April. The clock is running.

And a special thank you to Jordan at Senra, Fil at Dirac, and Garuth at Ironstead for your review and inspiration of this essay.

Referenced investments are for illustration purposes only. These investments are not intended to suggest any level of investment returns; not necessarily indicative of investments invested by any one fund or investor. No representation is intended that any result discussed is representative of the outcomes experienced by any AV Fund or investor. These investments are not available to future fund investors except potentially in certain follow-on investment options. The sample co-investors listed are some that AV has historically co-invested with and is not a predictor of future co-investors for any given portfolio company. The identity of a co-investor is not necessarily indicative of investment outcomes. There is no guarantee of who will be the co-investors. Many returns in investments result in the loss of capital invested.

This communication is from Alumni Ventures, a for-profit venture capital company that is not affiliated with or endorsed by any school. It is not personalized advice, and AV only provides advice to its client funds. This communication is neither an offer to sell, nor a solicitation of an offer to purchase, any security. Such offers are made only pursuant to the formal offering documents for the fund(s) concerned, and describe significant risks and other material information that should be carefully considered before investing. For additional information, please see here. Achievement of investment objectives, including any amount of investment return, cannot be guaranteed. Co-investors are shown for illustrative purposes only, do not reflect all organizations with which AV co-invests, and do not necessarily indicate future co-investors. Example portfolio companies shown are not available to future investors, except potentially in the case of follow-on investments. Venture capital investing involves substantial risk, including risk of loss of all capital invested. Diversification cannot prevent investment loss; it is a strategy to mitigate investment risk. This communication includes forward-looking statements, generally consisting of any statement pertaining to any issue other than historical fact, including without limitation predictions, financial projections, the anticipated results of the execution of any plan or strategy, the expectation or belief of the speaker, or other events or circumstances to exist in the future. Forward-looking statements are not representations of actual fact, depend on certain assumptions that may not be realized, and are not guaranteed to occur. Any forward-looking statements included in this communication speak only as of the date of the communication. AV and its affiliates disclaim any obligation to update, amend, or alter such forward-looking statements, whether due to subsequent events, new information, or otherwise.