Manufacture vs. Construct

By Liv Jenks

Enlarged Manufacture vs. Construct: one-off construction alongside standardized manufactured buildings

Construct and manufacture are synonyms, but they describe two different ways of making things. Construction makes a unique thing, one time, in place: a bridge, a school, a house. Manufacturing makes the same thing many times, under controlled conditions, and can become cheaper and more efficient with scale. Construction is a project, while manufacturing is a product.

Seeing that difference mirrors manufacturing itself: break something into its component pieces, question the assumptions and constraints, and rebuild it anew. That first-principles approach is at the heart of a theme across the USV portfolio: manufacturing instead of construction.

Here's how the two compare:

Enlarged manufacture body 0

But what if we manufacture what we construct? This was the question that industrialization and prefabrication sought to answer. The idea dates back to the mid-19th century, with homes that could be assembled from kits for California’s Gold Rush, and then larger-scale projects like the Eiffel Tower and London’s Crystal Palace. The introduction of iron and glass, which were lighter than masonry and easier to machine and transport, changed the building process. Motivated by rising labor costs, supply chain constraints, and tech innovations like CAD and 3D printing, prefab had a resurgence in the early 2000s, applied mostly to housing.

Enlarged manufacture body 1

Many pivotal project to product transitions happened by changing one of the core inputs in the table above – materials, labor, place, scale. Silicon took computers from room-sized machines to the iPhone in your pocket, as the integrated circuit turned the transistor into a manufactured unit that could be printed by the billion and shrunk with every generation. The camera followed suit: image sensors evolved from chips, so cameras went from bulky standalone devices to a component in every phone. Tesla questioned the assumption that a car needed a custom battery, instead packing thousands of standard, mass-produced laptop cells into a car and riding a cost curve already driven down by consumer electronics.

We’re again in a moment where technology is changing how things get made. Generalized vision models can teach robots without months of custom training, allowing factories to automate. AI-driven design makes it cheap to design, iterate, and realize something. And because standardized units produce comparable data, what one site learns can compound across a fleet rather than staying with one crew. Simply put, AI lowers the upfront cost of manufacturing, so combined with market forces like a shortage of trade labor and the urgency for more power on the grid, it now makes sense for things that never justified the capex of a factory before.

The framework of systemizing both the component parts and the assembly of the end product – which prefab popularized – can extend beyond buildings, and it’s something we've been investing behind at USV for some time.

Our bets

In energy, we've invested in Rune, which mass-produces data centers and places them where there's excess power. As Nick wrote in a recent post, “Rune is ruthlessly minimalist: from the physical footprint (no construction, not even a concrete pad; just a rugged box with server racks and a Starlink) to the choice of compute hardware (initially: a variety of older model ASICs that were available at a discount).” A data center is usually a construction project, but Rune turned it into a modular product.

Radiant does the same for nuclear. Rather than building a massive plant on site over a decade, Radiant makes small nuclear power plants in a factory and ships them. By building a 1-megawatt microreactor, Radiant reimagined what incumbents thought nuclear power required, creating a standard unit that can be stacked for more wattage.

Enlarged manufacture body 2

Isembard applies this to manufacturing itself. It’s a franchise business for critical industries (aerospace, space, and defense), acquiring or partnering with CNC factories that make component parts for those industries, then automating them. Each factory becomes a self-contained unit that feeds into a shared data loop across all of the factories.

A few common themes link these examples:

  1. Question the status quo: Our investments in Rune and Radiant started with most people saying that data centers and nuclear power couldn’t be done differently. Challenge one of the inputs that traditionally constrain it to a project (i.e. labor, materials, or place). Rune rethought place and materials; Radiant rethought size.
  2. Standardize the unit: Both the component parts and the system that makes them, so that each one can become a factory. And AI-driven CAD, robotics, and manufacturing APIs make the unit easier to design, replicate, and improve.
  3. Start small: Small units let you build within constraints that tighten learning loops and iteration. A construction business earns about the same margin on every project, so it grows linearly by adding crews or taking on more projects. A product moves down the cost curve with every unit, as factory costs spread across more output and each round of learning cuts waste. That compounding lends itself to venture-scale growth.

What could be the next project turned product?

A few places we’d love to see this transition next:

  • Bio: Drug manufacturing is one of the most “constructed” parts of pharma: each drug requires a long, slow, and expensive custom process. If bio and drug discovery increasingly look like an engineering process rather than a lab science, then what does it look like to biomanufacture at a small scale for a wide use case?
  • Energy: Transformers and substations are still built like construction projects: custom-engineered for each site, with multiyear lead times. They've become a bottleneck for data centers and electrification. What would it look like to turn grid hardware into a product, with standardized, factory-built substations that ship and drop in?
  • Robotics/physical AI: Deploying a robot often looks like a construction project, with an integrator custom-building each site. As world models generalize to new tasks with less training, the opportunity is standardized kits that drop into many similar environments. We'd start in critical industries where humans shouldn't be in the first place: mines, forest fires, and first response.
  • Private AI: Most personal AI hardware today starts at the high end, with models that only run on expensive computers. Is there a start-small approach here? Maybe it's a focused market where privacy is a requirement rather than a preference, like law firms or clinics, or novel chips designed for cheap, high-volume production for local inference.

If you’re building across any of these – or something else we haven’t considered – please reach out.