ASML's AI Diagnostics Raise New Data Control Questions

How Do You Share the Judgment of Someone Who Knows the Equipment?

ASML makes the lithography systems that etch semiconductor circuits onto wafers. A machine’s performance depends not just on precision components and design, but also on the accumulated experience of tuning the equipment and troubleshooting problems on-site.

Even with a manual in hand, when an unfamiliar problem comes up, you end up calling in a veteran engineer. The knowledge embedded in that kind of experience—hard to fully capture in words or documents—is what we call tacit knowledge. Collecting a large volume of field records doesn’t mean you’ve captured all the judgment behind them.

Watching how ASML is applying AI, this is exactly what caught my attention. If records can help solve problems faster, then who handles those records, and where, becomes just as important.

ASML’s Disclosed AI Use Cases

In a statement released in June 2026, ASML explained that it’s integrating AI not just into individual features but across multiple workflows and systems. Its partnership with France’s Mistral AI is part of that process.

The first application area is mask pattern correction—adjusting the design to reduce the distortion that occurs when a circuit is transferred using light. ASML says it has used AI for these calculations for more than 10 years, and that in January 2026 it completed early validation of a generative AI feature with a customer.

Tuning EUV light source settings and electron-beam inspection are also included. The goal is to cut down trial and error in the light source and improve inspection images and inspection positioning to boost overall equipment performance.

What caught my attention in particular is fault diagnosis. In early testing that analyzed error logs and service records, ASML reported identifying errors in some subsystems with the same accuracy as engineers, but more than 70% faster. It also says it explored over 12,000 design alternatives.

We need to be careful about the scope of these numbers. A result about identifying a specific error can’t be read as a reduction in downtime across an entire factory. And having reviewed many design candidates doesn’t by itself tell us anything about the quality of the design that was ultimately chosen.

The follow-up metric I want to see is verification on actual equipment. Even if a diagnosis was correct, we still need to know how long it took to repair and restart the machine, and what performance gains the chosen design actually delivered.

ASML itself emphasizes the role of engineers in defining problems and validating results in real-world settings. It’s hard to read this as an announcement that AI has wholesale replaced field experience.

What the €1.3 Billion Investment Tells Us

ASML had earlier announced a long-term partnership with Mistral on September 9, 2025. Alongside an agreement to apply AI to products, R&D, and operations, ASML invested €1.3 billion and said it would secure roughly 11% ownership on a fully diluted basis.

CFO Roger Dassen agreed to take on an advisory role on Mistral’s strategic committee, weighing in on strategic and technical decisions. This is a far deeper form of collaboration than simply buying API access.

When I look at this investment, I think about data control alongside technology development. That said, ASML’s stated purpose for the investment is joint innovation and improved product performance. We can’t conclude this was a deal made purely to secure data jurisdiction.

Mistral offers deployment options that let companies run its models on their own servers or in a private cloud. For companies handling sensitive field data, that option can matter a great deal. But that doesn’t mean data automatically stays within Europe across all of Mistral’s services. You need to check the specific product, deployment method, and the terms for any connections to outside services.

Holding equity can help in discussing long-term plans with a partner. But an 11% stake and an advisory seat don’t equate to the right to unilaterally decide where data is stored or which laws apply. That has to be worked out in the contract and in actual operating practice.

It’s worth getting the timeline straight when comparing this to ASML’s investment in Zeiss. The acquisition of a 24.9% stake in Carl Zeiss SMT was announced in 2016 and closed in 2017. A long working relationship and the length of time holding equity are two different things. Both deals strengthened ties with an important partner, but that doesn’t mean the rights secured were the same.

What to Check When Feeding Field Records to AI

Using records in an AI system and training a model on all of that content are two different things. A model can retrieve relevant documents to reference in an answer, or the material can be used for training. ASML’s announcement alone doesn’t tell us that a customer’s entire body of process knowledge has been folded into a single model.

Once data is organized so AI can use it, it may become easier to share within an organization. Employees won’t need to ask the same veteran the same question over and over, and it could help in finding similar problems from the past.

That makes access control just as important. You need to check who can see which records, whether one customer’s data gets mixed in with another’s, and whether responses might expose sensitive content. What exactly needs to be protected—the original documents, the retrieval index, the model itself, or usage logs—depends on how the system is set up.

None of this means precision components or manufacturing experience lose their value. It’s more accurate to say that AI operations and data management are now added tasks on top of the technical assets companies already had to protect.

Oswarld’s Lens

In my work building GTM strategy, I’ve seen the same pattern come up again and again. When companies review adopting AI, they often start by building a performance comparison table—comparing benchmark scores, response speed, and price.

But in the contract negotiations I’ve been through, the terms that ended up causing problems were never on that table—where the data is stored, whether it’s used for training, which country’s courts have jurisdiction if a dispute arises. I’ve watched deals fall apart in legal review more than once, well after the performance evaluation was already settled.

So when I look at ASML’s partnership, I’m just as curious about the terms under which data will be handled as I am about which model was chosen. That’s not to say the investment solves all of that. My point is that long-term technology partnerships and data-use terms need to be designed together.

I think Korean companies need to ask the same question. If the debate over “sovereign AI” is reduced to which country’s model you use, you can miss the actual contract terms. Whether the model is domestic or foreign, you need to specifically check where data is stored, who operates the system, who has access, and whether the data is used for training.

Field records in particular are assets built up over a long time. Before adoption, you should define the scope of data you’re willing to allow into the system and check what you can get returned or deleted once the contract ends—and also think through how you’ll respond if the company changes hands or its terms of service change.

I’d love to see these items built into the performance comparison table from the start. Some of the failed rollouts I’ve witnessed firsthand happened simply because these terms were discussed far too late.

---

References & Further Reading

• ASML, The machines behind the machines — June 12, 2026. Explains AI use cases and early test results.
• ASML–Mistral strategic partnership announcement — September 9, 2025. The source for the investment amount, equity stake, and strategic committee role.
• Mistral, introducing Le Chat Enterprise — Explains the deployment options available to enterprises.
• Mistral, data storage location guide — Lets you check the default storage region and the potential for cross-border transfer by feature.
• ASML, regulatory approval for the Zeiss stake acquisition — Helps distinguish between the 2016 announcement and the 2017 closing.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.