Preparing For AI

Jobs, markets, and economics are undergoing dramatic changes, but it’s still too early to know how this will all shake out. Key Takeaways: Semiconductor Engineering sat down to discuss how AI will alter what engineers do, how to ensure designs are robust and reliable, and how the economics will shift in the future, with Cindy Cui, vice president of global customer success at ChipAgents; Wally Rhines, CEO of Silvaco; Shelly Henry, CEO of Moores Lab AI; Dave Kelf, CEO of Breker Verification Systems; Vince Wong, head of AI development at Verific; and Ann Wu, CEO of Silimate. This panel discussion was held before a live audience at the recent 2026 ESD Alliance Executive Outlook meeting. To view part one of this discussion, click here. Part two is here. L-R: Breker’s Kelf; Silvaco’s Rhines; Verific’s Wong; Moores Lab AI’s Henry; Silimate’s Wu; ChipAgents’ Cui. Photo Credit: Paul Cohen, SEMI ESDA SE: How will AI affect engineering jobs? Cui: With GenAI, you can generate different options — millions of options very quickly. So human judgment of which options really matter becomes very critical. For future engineering, this needs to be part of an architect’s role, part of system-level thinking, or part of a supervisor’s role to enable different agents to work together. If we look at the industry right now, we still have a definition like a design engineering group, a verification engineering group, layout, and package design. Moving forward, there are no boundaries for these functional groups anymore. So every engineer needs to be able to learn new demands very quickly, and maybe leverage AI to understand what the real end-to-end design cycle could be. Then they can direct all the AI agent solutions with trust. We need to start using AI to understand new domains and gain that knowledge, and eventually we will build a very smooth-running system that includes both AI engineers and human engineers working together. SE: Is the learning curve changing with AI? Rhines: It may change in the short term, but it never changes in the long term. The cost per transistor continues to be on a curve as it always has been. When I was in business school, I did a cost per pound for copper over 1,000 years, and it still held. The learning curve is an ultimate truism we have to deal with, but it may show itself in different ways. We heard the announcements from China about, ‘You can’t measure cost per transistor. You’ve got to measure the cost per computation.’ But ultimately, everything has to come down in cost. It comes down for lots of reasons, like depreciation and experience. And anything that is measurable as an equivalent unit of capability will always be on a learning curve. SE: How do we add visibility into AI? These are basically black boxes, and as we get to agents, there is even more uncertainty about what is going on inside a system or a chip. Have we made any progress? Henry: I don’t think we will ever know what is going on inside. I tried to learn this. This was my first curiosity when I started learning about AI. I wanted to get into the nuts and bolts of, ‘How is it doing this? It’s giving me something amazing, but how is it doing it?’ So I went deeper and deeper into the actual workings of this, like how the embeddings and the matrix multiply and the training work. These numbers just do magic. The first thing that blew my mind was a very simple training example. So it was given a training set of lots of pictures of dogs and cats, and it was asked to classify which one is a dog or cat. It did a fantastic job. That was a training set. Now, when you give it a picture that it has never seen, it still comes up with the right answer. That was when I realized, ‘Wow, it’s not memory.’ It’s not just trying to pattern-match what it has already seen. It is able to pattern-match what it hasn’t seen before. I haven’t figured out how to actually explain that in a logical way. The math was too complex. I believe we will never know. Kelf: Experienced engineers need to understand what the heck is going on to trust it. But I wonder if the new grads will have that feeling. And maybe, like with synthesis, the new guys coming in didn’t necessarily have this need to understand it. It worked. They were trained on how to use it, and off they went and they used it. So maybe there is hope for the new grads, and maybe the older guys will get kicked out. SE: That leads to another question. How do you know the agents haven’t been corrupted or what else they’re tapping into? Wong: You don’t know. That’s why, from the beginning, you have to put in these safety measures. That’s probably more of a preventive thing, as well as a recovery phase. The smartest thing would be to sandbox it. Put some kind of container around it, put in some permissions so that it can’t reach outside of its scope. You want to be able to audit it. If something does go wrong, you want to be able to go back and see where that happened. At the same time, you want at that particular moment to be able to do some kind of recovery. So hopefully there was some savings of the state along the way. And at each stage, ‘If this one passed, no problem.’ But you would still be able to recover if there was a problem. It’s just a safety measure. You don’t know ahead of time. SE: The challenge there is that large language models are meant to mine data wherever it happens to be, be able to make sense of it, and to pull it from all different things that typically in the past didn’t go together. Different data types were separate. That’s not true anymore. Cui: The biggest risk is not that AI generates something wrong. It’s that it generates something wrong that looks correct. It’s very hard for a human to identify whether it’s correct because it looks good. A common question from the customer side is, ‘Hey, if you train the model using a group of data and verify your solution using the same group of data, then how can your solution verify the code generated by your model or your AI agents?’ The answer is that we shouldn’t. So back to the importance of benchmark data. Nowadays, all the large language models are trained using a huge amount of data. But in the meantime, we need to create some different benchmark data, represent the real design case, and more and more complex cases, to be able to verify your agentic AI solution. I come from a customer success background, and I really value the feedback loop from our customers. When we build a tool, we don’t know if that’s reliable or robust at first. But if the tool has been used by thousands of users every day on their real design and succesfully taped out, you can keep evolving from there. And eventually, that’s a mature system you can rely on. But we need to verify that very carefully. Wong: Anthropic mentioned something about secret languages agents developed. I don’t think that’s something they’re built to do. What they’re trying to do is to optimize their communication. So basically you’re getting charged tokens. When they talk to agents, these agents are going to say, ‘Hmm, given this leeway, I want to be able to save on all these different tokens. And the human didn’t tell me that I have to use a specific language.’ So they will come up with something that is optimal, but you can stop that. That is not a good thing because you don’t want to have these points of invisibility. You want everything to be visible. LLMs are all based on instructions. So we tell the LLM up front, don’t communicate in anything other than a human-readable language. Use a human-readable schema like JSON, for example. Usually, that’s how the LLMs communicate anyway, tool-to-tool. As the human interface, when we design these systems, we have to be aware of these things in advance. And if we put these guards inside there, a lot of these issues can be avoided. SE: Is the goal here just to deliver chips cheaper and faster? And if so, who’s going to make money in the future? Is it going to be the chipmakers, the system vendors, or the EDA companies? Kelf: The business model doesn’t change that much in terms of the chips that are produced and the folks that are using them. We’ve always had this value chain, and the people at the end of it have made most of the money. If AI makes the design process that much more efficient, then theoretically the EDA companies may lose money eventually. And we don’t want that, so I don’t see the business model changing that much because of this. Rhines: Historically, EDA, for the majority of its market, makes a percent of semiconductor revenue because they’re paid out of the R&D budgets. The R&D for the semiconductor industry is about 14% over the last 30 or 40 years, and it grows along with the revenue of the semiconductor companies. That’s been the growth. Semiconductor companies have the leverage. Nvidia becomes a multi-trillion-dollar company because it sells a lot more chips at a higher price, and therefore it gets rewarded for this productivity of additional designs. EDA traditionally does not get that reward because we’re a fixed percent of budgets. What you’d really like to see is for that to change. In my own case, I want to go after the manufacturing budget. That’s a great one. Other ways are to have reasons for the users to spend more than just a percent of revenue on their EDA software. Henry: I agree with you. But on the other hand, if you make the chip-building so much cheaper, that will create more competition for people like Nvidia, and they won’t be able to get that much profit margin anymore because there’s so much competition that is enabled by the EDA industry, and by the learning curve. So the cost per transistor comes down. Rhines: Yes, and you don’t get bubbles like Nvidia very often. Kelf: Is it cheaper or more innovative? People are trying to compete with Nvidia, and they’ve failed. Maybe they’re starting to get there now. AMD is getting stronger. So there are a lot of other issues. But if they could produce a much more innovative chip, and AI can help with that, maybe that’s the answer. Henry: That would help, as well. Rhines: There’s another aspect. Traditionally, the differentiation in chips was programmability. We have chips that have a lot of flexibility. But right now, we are so desperate to get the highest performance and the lowest power that we’re doing custom chips for specific algorithms. Once again, that’s great. It generates a lot of chip designs. But at some point, you can’t just continue to spend $200 million per chip design and continue to get back the economics. What I believe will happen is morphable chips. You’ll see chips that can reconfigure themselves in connectivity, memory partitioning, instruction sets, and be able to switch during processing to do different algorithms at different times. Cui: Regarding competition and who will eventually make more money, we are pursuing a totally different market right now. It’s a huge and growing potential market. The existing EDA market is stable, and the budget is fixed, and growth will be limited. But a lot of semiconductor companies have an AI budget right now. There are so many trillion-dollar companies right now — Micron, Marvell, Nvidia, and others. So we are pursuing a huge and growing market. Eventually, if we work together to create something really innovative, everyone will win something from this. Wu: It’s going to be really interesting to see the long-term economic impact. One of the reasons for Nvidia’s insane profit margin is the constraint of chips due to fab capacity. It’s well documented that TSMC took their time in ramping up capacity for the AI market because they weren’t sure whether or not it was going to be a real thing. So there was this constraint caused by the foundry side because of capacity limitations. And then, of course, the tranche is going to the high-volume players. So there is a very real and very exciting potential expansion in fab capacity as costs go down and the cadence of putting out more specialized chips by the systems players. That can actually be a much larger TAM for the overall ecosystem, where everybody makes more money on at least the same trajectory that they were experiencing. Kelf: Traditionally, for the last 20 years in EDA, it’s been very hard for a small company to raise funding. VCs will say, ‘Eh, limited exit strategies. It’s a percentage the R&D budgets.’ But you guys have raised a ton of money. How did you do it, and where did your VCs see the return that smaller companies were unable to do before in EDA? Wu: One of the hard things about raising capital for a hardware startup is that there’s so much risk, and it’s such a long time horizon from getting that initial investment to getting that first test chip that is minimally functional, under-performant, to justify the next tranche of yet another $100 million or more. That’s the risk factor in achieving these very difficult and technical milestones spread over many years. But we build software. We got our first annual license contract in less than a year from when we started. So it’s easier to raise capital as a company that is building software where you ship and you can generate value on the interval of days and weeks, compared to many years. And what helps that, of course, is it’s a very exciting time. When else have you seen all the major news publications have semiconductor-related headlines every other day than over these past couple years? And there is this highlight, and a lot of investors follow the trends. They see what’s exciting and they’re like, ‘Oh, can we help with that?’ It’s really the cross-section of it being a very exciting time for our space, and it’s easier to prove value early on as a startup. Rhines: The VCs were spoiled by the software industry, where it was a low investment up front until you see an actual product and revenue. Hardware is exactly the reverse — a big investment up front before you see customers and a working product. I spent five years chasing VCs for Cornami. As soon as you said you’re involved in hardware, they gave a nice smile and said, ‘I think I have another appointment.’ It’s changed. Nvidia and other companies are now changing the paradigm, and incredible valuations are coming through before there are successful products. Cui: This is the right inflection point for the entire industry. A lot of investments are coming from the semiconductor companies, led by TSMC, Micron, and Ericsson. They all see the value that this new generation AI solution can create, and how this is going to reshape the future of chip design. This is a critical inflection point, which allows us to raise those funds very quickly. Leave a Reply