PDF Solutions CEO 解析 AI 对半导体行业的效率挑战

This post is based on John Kibarian’s keynote given at the PDF Solutions CONNECT event on October 15, 2026. The recording and slides of this presentation can be found here: https://www.pdf.com/users-conference/connect-2026/next-generation-pdf-solutions-platform-for-the-global-semiconductor-industry/

AI is the most consequential shift the semiconductor industry has faced, not because it is driving an upcycle, but because it is changing how the industry works.

The industry has seen big transformations before. Mobile and the internet reshaped what semiconductor companies sold, but they did not change how fabs, fabless companies and their supply chains operated. AI does both. It drives unprecedented demand for silicon, and is simultaneously reshaping the tools, data and collaboration models used to design, manufacture and test silicon.

We began reimagining our portfolio around an AI-first premise roughly four years ago, when ChatGPT first appeared. Today, the result is a next-generation platform that spans three long-standing areas of our work. The first two are accelerating yield ramps for new technologies and applying the latest IT technology to manufacturing data. The final area is connecting manufacturing operations across the supply chain, a thread strengthened by the 2025 acquisition of secureWISE.

Efficiency is today’s watchword. Efficiency is what matters. A manufacturing executive told suppliers recently that he could not get more tools than had already been promised. His only path to meeting demand was getting more out of the tools he had, a sentiment now heard across the industry, and shapes three challenges AI poses for semiconductor manufacturing:

1. Energy efficiency: getting more out of the AI data centers being built.
2. Data and compute efficiency: leveraging AI to create insights from manufacturing data at an acceptable cost.
3. Supply chain efficiency: operating a globally distributed and increasingly complex supply chain.

Challenge 1: Energy efficiency

Growth numbers are no longer in question, as the semiconductor industry reached a trillion dollars earlier than projected. Looking forward, McKinsey & Company’s base case has device revenue rising from $975 billion in 2025 to about $2.3 trillion in 2030. That’s an 83.5% increase in volume, with 58% of that growth coming from data centers and servers, the one vertical compounding at 29% a year.

Translated into physical capacity, the figures are striking. Installed AI data center capacity of 45–60 GW in 2026 is projected to reach 100–160 GW by 2030. At 150,000–200,000 wafers per gigawatt, half being HBM, that implies 9.6–17.5 million additional wafers and 16–29 new fabs of 50,000 wafer starts per month. This estimate could be conservative, as no company yet builds 50,000-wafer fabs at the leading edge. A modern leading-edge logic fab runs closer to 20,000 wafer starts, so the real fab count could double.

With AI revenue projected at $3.9 trillion in 2030 and margins at roughly $30 million per megawatt, the temptation is to treat this as a capital problem. Nation states look at the numbers and conclude they can write the check. Anyone who has run a fab knows better. Building capacity is only part of the answer. The harder problem is getting more compute out of every data center.

That pressure is already reshaping the landscape, and efficiency is coming from new designs. For years, most of our new customers came from Asia, where fabs and fabless companies were being started. Today, foundation model companies and hyperscalers are designing their own silicon with energy efficiency as the headline metric. Workloads are also being tuned for efficiency. Training keeps a chip saturated at its design point, while a GPU doing inference can spend over 90% of its power on memory updates and idle time.

Using that measure, the industry still has about a factor of a million of headroom, according to Richard Feynman’s lectures from the 1980s on the fundamental limits of computing.

What that means for manufacturers

Fabless companies previously could not justify manufacturing software because they were not manufacturers. They are now. What happens after a wafer leaves the front-end fab is where much of the industry’s action now sits.

A few examples illustrate the rapid change. Energy-efficient, bespoke designs carry operational consequences that land squarely on manufacturing and test with smaller production volumes across more products. Custom designs are proliferating rather than a few high-volume parts.

The breadth of new product introduction (NPI) testing carries through into production for accelerated yield ramps with wider datasets and test programs running to millions of test items.

And new technologies such as co-packaged optics and similar integrations mean more than electrons are moving through the production line, requiring new process controls.

PDF Solutions’ portfolio answers these pressures. Characterization Vehicle and DirectScan address co-optimization of design and manufacturing at sub-7nm nodes using more design information to know where and how to look, including e-beam technology designed to find buried systematic defects in 3D structures. Exensio Aurora, Exensio Test Operations and Exensio Scalable Analytics handle rapid analysis of wide test data and deploy rules and AI workflows at the edge. A Python-first, natively parallel compute engine moves analysis from the client to a scalable back end, delivering 20–40x the performance of the same hardware spend through more efficient use of cores. Sapience and Exensio Process Control extend advanced process control, application integration and supply chain orchestration.

Challenge 2: Data and compute efficiency

According to an interview in The New York Times, Nvidia’s Jensen Huang noted, “Only in the last six months, after 15 years of trying, has AI become useful.” It’s been useful somewhat longer, but the framing raises the right question for semiconductor manufacturing. What does useful mean?

For the semiconductor manufacturing industry, it means combining common-knowledge industry data with company-specific data; efficiently using all available manufacturing data; efficiently managing the compute cost of AI for ongoing analysis; and deploying agents at the manufacturing edge across applications and the supply chain.

Each is harder than it looks.

Large language models already know a great deal about semiconductors. Ask one about wafer maps, parametric test, bins, lots, SPC charts, etch-rate drift or failure mechanisms, and it reasons competently. It selects methods, generates code, interprets results and summarizes unstructured data. Give a general-purpose LLM a yield loss in lots and ask what analysis to run and it will produce a decent first pass, roughly what a junior engineer would propose.

The problem is what comes next. A less experienced engineer runs the analysis, sees a signal they cannot explain, dismisses it as noise and reports that there is nothing in the data. A senior engineer looks at the same chart and sees what is going on because they carry context the model does not. What LLMs miss is company-specific data and IP, the historical baseline, product and process specifics such as recipes, design rules, test flows and bin definitions, large volumes of statistical computation, and validation.

Using available data is its own challenge. Our tools worked through many yield crashes, and the post-mortem almost always finds the data was already there. An engineer cannot know which are relevant until there’s an event. Getting efficient at working with wide datasets is what enables a reaction before the crash.

The scale of private company data

The most important factor in applying AI to semiconductor manufacturing is accessing and processing complex private company data.

The scale is formidable. Leading fabless and giga-scale foundries hold petabytes of data per site across more than 20 data domains, each with 100 or more file formats. A new 3D-IC can carry 2–10 million test parameters at NPI. Mixed-signal companies that consider their test programs modest, perhaps 10,000 parameters, still accumulate around 200 TB of analytics data. All of them have more products than employees, let alone product and test engineers. At one foundry, the task is analyzing 2–3 trillion sort parameters across roughly 2,000 distinct products every week, with 100,000-plus features per unit.

That’s why PDF Solutions built Exensio Aurora using 30 years of semiconductor manufacturing knowledge for AI context. We have petabytes of manufacturing data managed on isolated customer cloud sites, more than 1,000 test chips taped out for process characterization in the past 15 years, over 660,000 analytics templates and best practices built by customers across those sites. We also have an extensive semiconductor ontology for collecting and aligning data that is never commingled across sites, and neither are the agents that run on it.

The LLM cost problem

Tokens have become the single largest and fastest-growing IT expense. One developer ran out of tokens at 2 p.m. The work left in the day would take five hours by hand. Once the cap was lifted, it took five minutes with tokens, illustrating the value and cost of LLM compute.

Manufacturing has a problem with the cost. Manufacturing analysis is repetitive by design, but running an LLM on every analysis is also expensive in a way that does not improve with scale. Consider the goal of repeatedly analyzing a 100,000-row, 10-column data sample:

The repeated run costs of the LLM-created workflow are roughly a thousandth of the energy of a complete LLM approach. For repeatable analytical tasks, LLMs do not provide economies of scale, and the cost of using them for every analysis can be prohibitive.

LLMs write workflows, workflows do the work

The conclusion we reached in building Exensio Aurora is that LLMs need a language, and that language is Python expressed as workflows. The approach has four elements:

1. Align and enrich data with a semiconductor ontology so the model and engineer share the same definitions of lots, wafers, bins, tests and tools.
2. Run analysis on a highly scalable compute engine designed for semiconductor data rather than pulling data out through APIs and asking the LLM to crunch it.
3. Use LLMs to create workflows and agents. The model is asked not to perform the analysis but to build it. This is basically code generation, which is what LLMs do best.
4. Use workflows to encode context and repeatable analysis. The engineer inspects the result as visual workflow blocks rather than lines of Python, executes it, and if it works, generalizes it and stores it in a library. The next time, the system runs the stored workflow and skips the 20,000 tokens.

This architecture also air-gaps customer data from the LLM. Customers can experiment with different models, including ones that do not run on premise, without their data leaving the premises. Captured knowledge can be revisited, which is how hallucinations are prevented. The net result is predictability, repeatability and cost efficiency. An engineer cannot keep track of 10 million parameters. An algorithm can.

Challenge 3: Supply chain efficiency

Material moves around a great deal. In the world of hybrid packages, all manufacturing is a complex supply chain execution problem. Wafer fabrication, die preparation, wafer bonding, die stacking, TSV and interconnect, advanced packaging, and test and metrology are distinct stages done in a different facility by a different company in a different country.

This is why PDF Solutions acquired secureWISE, the largest secure network connecting fabs and their equipment vendors, deployed in 3nm fabs in and moving exabytes of data between equipment suppliers and factories. Several equipment vendors’ own AI solutions are built on it by collecting data across their installed fleet, drive learning, and push improved algorithms back to the tools.

secureWISE’s differentiation is control. Traffic is double-encrypted and contents of the packets are not shared. The factory team decides what data leaves, when it leaves, what data and algorithms come back in, what virus scanning is applied to anything entering the facility, and keeps a full audit trail of all of it.

From human collaboration to agentic collaboration

The semiconductor industry is an industry of collaboration that’s becoming agentic, raising the bar for how tightly, and how securely the industry is interconnected.

secureWISE was created to let humans collaborate. In practice, most traffic on the network, and most of the revenue associated with it is data moving out and software or algorithms moving back. It is more driven by agents and AI than by people. That balance will keep shifting.

One platform across design, manufacturing and the supply chain

The efficiency pressure will not go away. Data centers need more compute per megawatt. Manufacturers need more insight per token and per core. Supply chains need to move data and algorithms increasingly between agents as securely as they once moved people.

Our architecture ties together energy efficiency, data and compute efficiency, and supply chain efficiency on one platform. In new-product development, manufacturing, and at supply chain partners such as OSATs, the same stack repeats. Exensio Aurora and enterprise applications run on Sapience with apps and agents on top. In the fab and at the OSAT, Cimetrix connects equipment, including eProbe, to manufacturing systems and enterprise data warehouses. secureWISE provides secure edge deployment and control across all of it, enabling data, rules, workflows and agents to move between sites under the control of the people who own the data.