The next semiconductor superpower will be the one that learns fastest.

The frontier is moving.

The United States is rebuilding semiconductor manufacturing at a scale not seen in generations. New fabs, advanced packaging facilities, supply chains, equipment, and supporting infrastructure are taking physical form. The Department of Commerce projects that the U.S. could produce at least 20 percent of the world’s leading-edge logic chips by 2030, up from effectively none in 2022.¹

This buildout is happening while semiconductor manufacturing itself is changing. New materials, device structures, packaging architectures, process steps, and equipment are moving toward production. Some will extend existing manufacturing platforms. Others will require new ones. In either case, the processes that make these technologies manufacturable still have to be developed, transferred, qualified, and brought to yield.

A fab can be built from specifications. The capability inside it is accumulated much more gradually. Engineers learn how materials and process steps interact, how equipment behaves in a particular flow, where variation enters, and which conditions remain stable as a process moves toward production.

The same classes of tools are available to multiple manufacturers. What separates the best is what they have learned about using them together: the sensitivities of equipment, the behavior of materials, the interactions between process steps, and the thousands of decisions required to turn a promising process into repeatable production.

That knowledge takes years to build. Once accumulated, it shapes how quickly the next generation can be brought online.

Manufacturing advantage compounds.

TSMC is the clearest expression of this dynamic. Its manufacturing advantage was built over decades, across successive technology generations, as engineers developed processes, solved yield problems, transferred learning between products and nodes, and carried that experience forward. Each generation began with a deeper base of manufacturing knowledge than the one before it.³

Physical assets, scale, suppliers, talent, and customer relationships all matter enormously. Beneath them sits an accumulated understanding of how to make increasingly difficult processes work in production. That is a difficult advantage to reproduce with capital.

A new fab does not inherit thirty years of process history when its tools are installed. Its engineers still have to learn how those tools behave in a particular flow, how materials interact, where the process window closes, which measurements are predictive, and which apparent improvements survive qualification and ramp.

The moving frontier makes this especially consequential. New materials introduce unfamiliar behavior. New equipment creates different operating regimes. Advanced packaging and heterogeneous integration create interactions that did not exist in the same form before. Processes transferred between tools or sites encounter conditions that may differ in subtle but important ways.

As some manufacturing questions become new again, accumulated learning changes the starting point. The organizations that carry the right knowledge into the next transition will move down its learning curve faster.

Every new process begins with an information problem.

Semiconductor manufacturing already generates extraordinary amounts of data. Decades of process runs, metrology, equipment traces, maintenance history, failure analysis, qualification results, and engineering studies sit across fabs, equipment companies, materials suppliers, and R&D organizations. The industry does not suffer from an absence of observations.

Yet a new process can still begin remarkably data-poor. The exact combination of material, equipment, geometry, integration flow, and target may never have been run before. A process space can contain trillions of possible parameter combinations while a development team may be able to test only on the order of a hundred. Each physical experiment consumes wafers, equipment time, metrology, analysis, and days or weeks of calendar time.²

So the central question becomes which experiment to run next.

Experienced process engineers draw on prior work to decide which regions are worth exploring, which mechanisms are plausible, and which results deserve skepticism. They remember how similar materials behaved, where a process became fragile, and what happened when a tool previously entered an unusual regime.

Computation can add leverage to that judgment. In a modeled semiconductor process-development study, experienced engineers navigated the broad early search more effectively, while sequential optimization became more useful after the search entered a constrained region. Combining the two reduced modeled cost-to-target by roughly half relative to the expert benchmark.² The study points toward a practical division of labor: engineers contribute physical context and judgment, while computation helps allocate the next scarce experiment.

The ingredients are already emerging. What remains harder is turning them into a system that learns across campaigns. Most analytical methods begin after someone has assembled the relevant data, chosen the variables, defined the constraints, and decided what prior experience should inform the problem. Those steps contain much of the organization’s manufacturing knowledge.

The larger opportunity sits between abundant historical experience and scarce local evidence. Each new campaign should benefit from what the organization has learned before, without assuming that yesterday’s conditions still apply.

Most process history was not built to be reused this way.

A wafer run leaves behind data. What an engineer learns from the run is richer.

The useful context may include why a condition was chosen, which hypothesis it tested, what changed from the previous run, what was unusual about the equipment, which measurements were trusted, what mechanism seemed to explain the result, and what the team intended to try next.

Parts of that history live in structured systems. Other parts remain distributed across reports, notebooks, scripts, presentations, conversations, and experienced engineers themselves. This is manageable while the same people remain close to the same process. It becomes harder when knowledge crosses a boundary: from one campaign to the next, from development into qualification, from one tool or site to another, or from an R&D group into production.

Each boundary forces a new judgment about what can be carried forward and what must be learned again. Previous results cannot simply be copied into a changed environment, but treating every transition as an entirely new problem discards information the organization has already paid to acquire.

The opportunity is larger than better recordkeeping. A manufacturing organization should become more capable because it has run the previous thousand experiments. Historical campaigns should improve the starting point for new ones. Qualification should deepen the understanding available to future development. A process transfer should leave the organization better prepared for the next transfer.

Over time, the learning curve itself can become an asset.

This is what Serial is building for.

Serial is building process-learning systems for the moments when semiconductor manufacturing is changing: process development, transfer, qualification, and early ramp.

We are starting with experimental decision-making. The aim is to bring prior process history, equipment and material context, physical knowledge, engineering judgment, and evidence from the current campaign together so that process teams can make better decisions about what to run next.

Similarity alone is not enough. A previous result may come from another tool, material lot, process generation, or integration flow. Its value depends on what changed, which underlying relationships are likely to persist, and how much uncertainty those differences introduce.

Process engineers remain central to that loop. They carry physical intuition, knowledge of equipment behavior, skepticism about particular measurements, and an understanding of which results are likely to survive manufacturing. Serial is being built to give that judgment greater leverage and to preserve more of what is learned for the decisions that follow.

The ambition is for each process campaign to improve the starting point of the next one. The next generation of semiconductor manufacturing will require new tools, materials, architectures, and processes. The countries and companies that learn how to make them work fastest will have an advantage that is much harder to buy.