Above the semiconductor industry chain, before digital twins: Analysis of the invisible champions improving yield

If we view semiconductor manufacturing as a system, we will find a long-overlooked position: above the industry chain, before digital twins are truly implemented, there exists a layer of cross-enterprise, full-process “cognitive layer” that has not been fully defined. The value of PDF Solutions comes from here.
It deals not with point data, but with the causal chain spanning design, process, equipment, and testing: a certain design structure, forming specific defects at a particular process step or on a specific piece of equipment, ultimately mapping to electrical failure. A single fab or testing organization may have all raw data for certain segments, but it’s difficult to stably connect these data into reusable causal models. That’s the essence of where PDFS intervenes.
Why haven’t EDA, fabs, and equipment manufacturers done this themselves? It’s not because they can’t, but because they lack the motivation to take that step.
Design companies like Synopsys and Cadence Design Systems can only achieve feedforward optimization, lacking a closed-loop feedback after manufacturing;
Fabs like TSMC and Intel have the most comprehensive data, but systems are fragmented, organizations are dispersed, and cross-process integration costs are extremely high;
Equipment manufacturers like KLA Corporation and Applied Materials control inspection and control, but their perspective is limited to single processes.
Each layer optimizes locally, and cross-boundary issues are left unaddressed. As a result, above the industry chain, a “explanation system” gap naturally appears, and this is exactly where PDFS fits.
The industry chain uses PDFS because of data disconnections—design looks at design, process at process, equipment at defects—but there is no unified mechanism to string these pieces of information into an explainable causal chain. The core role of PDFS is to provide a “unified language” across segments, transforming otherwise unrelated data into structured cognition usable for decision-making.
PDFS reuses an abstract layer: defect classification, feature engineering, analysis pathways, and the “pattern–cause” mapping relationship. This is a form of “cognitive compounding,” not a strong network effect like the internet. The more clients, the better the model.
Why has PDFS reached its current position? Because they started with the hardest tasks.
Initially, they were not platform companies but entered through engineering services, solving the most challenging yield issues. Yield problems naturally span design, manufacturing, and testing, forcing the entire chain to be connected from the start. As projects accumulated, similar problems recurred, analysis methods and data structures were continuously refined, gradually shifting from “people-driven” to “method-driven,” ultimately productized into a platform (Exensio). The so-called “full industry chain coverage” is not a top-down design but a natural expansion driven by problem-solving.
Currently, the company’s moat is still insufficient to automatically evolve into an industry standard. The three conditions that determine its ceiling are currently only one accelerating: AI’s dependence on data structures. As AI enters manufacturing, companies prefer to model on existing data frameworks rather than rebuild systems, reinforcing PDFS’s position. However, standardization of data models remains slow, cross-company collaboration is still in early stages, and the flywheel has yet to close.
Compared to KLA Corporation, this difference is even clearer. KLA controls “what is seen,” with data coming from equipment, tied to the physical world, with a rigid and direct moat; PDFS controls “how to understand,” belonging to the cognitive layer, relying on data structures and accumulated experience. The former is unavoidable, while the latter has alternative paths. Therefore, at this stage, KLA is stronger and more certain; PDFS has a higher ceiling, but its path is not locked in.
From the perspective of Jensen Huang at NVIDIA, the end game for PDFs is digital twin.
Digital twins require a closed loop of real-time data, causal models, and control capabilities. PDFS has already covered the most difficult part—causal modeling and cross-chain data structures—placing it in a very delicate position: above the industry chain but before digital twins. It is responsible for “understanding the world,” so others can “change the world.”
Looking at development paths, PDFS’s next step is likely to converge on three main lines: standardization, AI integration, and embedding. Standardization means turning its data structures into the industry’s default language; AI integration means enabling models to depend on its data system; embedding means shifting from “analytical recommendations” to “production decision-making.” If these three points are achieved, it could cross that boundary from cognition to a true system layer.
Therefore, the most accurate assessment of PDFS is: it occupies a special position—above the semiconductor industry chain, before digital twins.
If this layer is eventually standardized, it could become infrastructure; if not, it remains a high-value tool.
Disclaimer: I hold the assets mentioned in this article. My views are biased and not investment advice. Investment risks are significant; enter with extreme caution.