Manufacturing yield and regulatory approval now outlast software clones, making physical complexity venture capital’s newest competitive moat.
Defensibility Through Complexity
A generative coding assistant can reproduce a software competitor over a weekend, but it cannot reproduce a Federal Aviation Administration type certificate or years of manufacturing yield data. That gap is the specific reason venture capital began treating physical complexity, not clever software engineering, as the moat that lasts after 2023. The old assumption inside venture capital held that complexity was a cost, something a founder minimized wherever possible because every additional physical dependency slowed a company down relative to a pure software competitor. That assumption has inverted.
The inversion follows directly from a mechanical observation about what large language models and generative coding tools can and cannot copy. These tools excel at reproducing patterns that exist abundantly in their training data: common software architectures, standard user interface conventions, and well documented application programming interface integrations. They cannot reproduce a physical manufacturing process that has never been publicly documented in detail, a proprietary alloy composition developed through years of trial and error, or the specific calibration data a sensor company has accumulated from thousands of hours of field deployment. The moat that used to come from clever software engineering has largely dissolved into something any well resourced team can approximate. The moat that comes from physical manufacturing depth has not dissolved at all, because the knowledge required to replicate it was never written down anywhere a model could learn it from.
Consider the concrete mechanics of what makes a deep technology company hard to copy. A battery manufacturer’s cell chemistry involves a formulation that took years of iterative testing to optimize for a specific balance of energy density, thermal stability, and cycle life, and even with a full bill of materials in hand, a competitor attempting to replicate the formulation faces months of its own trial and error to match the original company’s yield rates and defect rates. A semiconductor company’s process node involves thousands of individually tuned manufacturing steps, each interacting with the others in ways that are not fully documented even inside the company that developed them, passed down instead through the tacit knowledge of the specific engineers who ran the original development runs. Neither of these forms of knowledge exists as a dataset a model can be trained on, because neither was ever fully captured in writing in the first place.
Regulatory approval adds a second, distinct layer of defensibility that operates independently of the underlying engineering difficulty. A medical device company that has obtained Food and Drug Administration clearance, an aerospace company that holds a Federal Aviation Administration type certificate, or a nuclear technology company that has navigated the Nuclear Regulatory Commission’s licensing process has cleared a barrier that took years and tens of millions of dollars to pass, and that barrier does not reset for a competitor simply because the competitor can build a functionally similar product faster. A well capitalized competitor entering a regulated physical category still has to run the same multi year approval process the incumbent already completed, which means the incumbent’s head start compounds rather than eroding, the opposite of what happens in software categories where a fast follower can close a feature gap within a single product cycle.
Specialized infrastructure creates a third form of defensibility that is easy to underestimate because it looks, from the outside, like ordinary capital expenditure. A company that has spent three years commissioning a specialized test facility, a wind tunnel, a cleanroom fabrication line, or a battery cell testing chamber calibrated to its specific chemistry, has built an asset a competitor cannot simply purchase off the shelf, because the facility’s value comes as much from the years of process refinement layered on top of the physical structure as from the structure itself. A competitor that builds an equivalent facility from scratch still has to run its own multi year process refinement cycle before the facility produces comparable yield and quality, which means the capital expenditure alone does not close the gap.
Tesla’s own manufacturing history is instructive on how long this kind of complexity takes to master even for a well funded, technically sophisticated team. The company’s Gigafactory Nevada battery cell production line took years longer to reach its targeted output and yield rates than the company’s own public projections suggested when the facility was announced, not because the underlying chemistry was a secret but because converting a laboratory process into a high volume manufacturing line involves thousands of small process adjustments that can only be discovered by actually running the line at full production volume and observing where it fails. A well capitalized competitor with access to the same published chemistry still has to run its own multi year version of that same discovery process, which is why Tesla’s manufacturing lead over competing battery producers persisted for years after its underlying cell chemistry had become broadly understood within the industry.
Legal protection for this kind of knowledge operates differently than for a software innovation, a distinction that matters directly to how a founder should structure intellectual property strategy. A software feature can often be protected, if at all, through a patent that describes a specific method, but the tacit manufacturing knowledge described throughout this analysis rarely lends itself to patent claims because it consists of thousands of small operational decisions rather than one novel, claimable method. Trade secret law, protected under the Defend Trade Secrets Act of 2016 at the federal level and under state trade secret statutes modeled on the Uniform Trade Secrets Act, is the more natural legal vehicle for protecting this category of advantage, provided the company takes the specific, documented steps, restricted facility access, employee confidentiality agreements, and careful control over which vendors and partners see which parts of the process, that trade secret law requires to maintain protection. A founder who assumes the difficulty of replication alone provides legal protection, without taking the affirmative steps trade secret law requires, may find that protection is far less enforceable than assumed once a departing employee joins a competitor.
The unexpected implication for venture fund diligence is that the traditional software test for evaluating a competitive moat, network effects, switching costs, and data advantages, now applies with less force than a more old fashioned test: how many years of accumulated, undocumented physical knowledge would a well funded competitor need to replicate this company’s manufacturing yield, regulatory position, or test infrastructure. A diligence team that only asks whether the underlying technology is impressive misses the more important question of whether the difficulty of replication is measured in weeks, because the technology is documented and reproducible, or in years, because the knowledge is tacit and embedded in a specific team’s accumulated experience.
This same test also changes how a fund should think about the risk of a well capitalized incumbent, rather than a venture backed startup, entering a category. A large technology company with abundant capital can hire away a startup’s key engineers or acquire a competing team, but it cannot instantly acquire the years of accumulated manufacturing yield data, the specific regulatory relationships built over a multi year approval process, or the tacit process knowledge embedded in a specific production line’s history. This is a meaningfully different competitive dynamic than software, where a large incumbent with abundant capital and engineering talent can often close a feature gap within two or three product cycles regardless of how much of a head start the original innovator had.
For founders, the strategic implication is that defensibility through complexity has to be built deliberately rather than assumed to emerge naturally from operating in a hard industry. A founder who treats regulatory approval as a checkbox to clear as quickly as possible, rather than as a source of lasting advantage to be documented and defended, is leaving value on the table relative to a founder who treats the accumulated regulatory relationship and the specific approval pathway navigated as a strategic asset worth protecting through careful trade secret practices and disciplined internal documentation of exactly what makes the company’s process difficult to replicate.
Anduril Industries has applied this same complexity moat argument directly to its manufacturing strategy rather than treating it as an incidental byproduct of building defense hardware. The company’s Arsenal-1 production facility in Ohio, announced as a multi billion dollar investment aimed at high volume autonomous systems manufacturing, is designed explicitly to accumulate the kind of production line knowledge that a well capitalized but manufacturing inexperienced competitor cannot simply purchase, since the facility’s value compounds with every unit produced and every defect pattern identified and corrected. A venture fund backing a company at the seed stage in this category should evaluate whether the founding team has a credible plan to build exactly this kind of compounding manufacturing knowledge, rather than assuming a strong initial prototype automatically translates into a defensible production advantage once volume increases.
Corporate development executives evaluating a partnership or acquisition in a deep technology category should specifically probe how much of the target company’s advantage rests on documented, transferable knowledge versus tacit, undocumented process knowledge held by a small number of specific individuals. A company whose defensibility depends heavily on two or three key engineers who understand a specific manufacturing process carries meaningful key person risk that a diligence team has to price into any acquisition structure, often through extended employment agreements or earn out provisions tied to those specific individuals remaining with the business through a transition period.
What remains an open matter for the industry is whether artificial intelligence tools will eventually close even this gap, by allowing a well resourced competitor to reconstruct tacit manufacturing knowledge computationally from sensor data and production logs rather than through years of direct human experience, a possibility that would erode the very defensibility argument this analysis has laid out. The next piece in this series examines a related force reshaping deep technology investing from a different angle: the migration of engineering talent away from pure software roles and toward the physical categories this analysis has described, and what that migration means for where the coming wave of defensible companies gets built.
Interested in analysis about the intersection of venture capital, tech, artificial intelligence, public policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/