What Happens When AI Starts Designing the Chips That Power It | Synopsys at AI Infra Summit

Chip complexity is growing faster than the engineering workforce available to manage it. Multi-die architectures, 3D packaging, and hyperscale system requirements have added thermal, mechanical, and cross-domain challenges to already sprawling design workflows. With teams being asked to tackle larger designs without adding headcount, human-operated toolchains are reaching their limits.

At AI Infra Summit 2026, Brendan Burke (Futurum) and Matt Kimball (Moor Insights & Strategy) sit down with Thomas Andersen, Vice President, AI and Machine Learning at Synopsys, to explore how agentic AI could move chip design from assisted tooling to increasingly autonomous engineering workflows, and what that shift requires from EDA vendors, chipmakers, and foundries.

Andersen separates assistance from autonomy by who takes the action. Copilots suggest, while agents execute. In a more autonomous workflow, engineers would define the specification while AI agents launch tools, resolve timing violations, clear congestion, and run verification, escalating only when design intent is unclear.

Reaching that point will require encoding the specialized knowledge currently distributed across EDA vendors, chipmakers, and foundries. General-purpose models trained on inconsistent public data cannot meet silicon’s accuracy requirements. But once grounded in trusted domain expertise, self-learning systems could dramatically shorten spec-to-tapeout timelines and make custom silicon achievable for much smaller teams.

Key Takeaways:

🔷 The talent gap, not tool speed, is the constraint on chip design. Andersen notes the industry has forecast an engineering shortage for years, and multi-die and 3D-stacked designs now add thermal analysis, stress analysis, and constraint cleanup that would demand exponentially more “humans in the loop” under traditional workflows. Automation is the only path that closes that gap at the pace new silicon is expected to ship.

🔷 Assisted and autonomous AI differ by who takes the action. Copilots and chatbots offer suggestions the engineer still has to analyze and execute. Agentic systems take on individual tasks, multi-agent workflows split a larger problem across agents the way a chief of staff assigns work, and full autonomy starts from a written specification and hands tool orchestration, timing fixes, congestion cleanup, and verification to agents.

🔷 Domain knowledge decides whether agentic EDA works at all. Chip design demands 100% accuracy, so general-purpose models trained on conflicting web data fall short. Andersen points to pre- and post-training baseline LLMs with expertise from both EDA vendors and their customers, then evolving toward self-learning systems that can try new approaches and eventually outperform expert engineers.

🔷 Ecosystem data silos are the blocker to autonomy. Customers hold proprietary design and workflow knowledge that cannot leave their premises, EDA vendors hold tool and debug knowledge, and foundries hold process-node knowledge. Synopsys is building a system that combines all three, and Andersen is direct that the autonomous workflow does not function without that collaboration.

🔷 Agents absorb the tedious work first, then move up the stack. The initial targets are the repetitive tasks expert engineers already hand to junior staff, such as cleaning up data and reconciling timing constraints across partitioned teams. As agents master those, they take on higher-level work, and Andersen expects the separate RTL, verification, implementation, and sign-off roles to collapse into a single chip design engineer who operates engines rather than individual tools.

Engineering organizations that start encoding their design knowledge into agentic workflows now will compress spec-to-tapeout timelines ahead of competitors, and lower the resource barrier to custom silicon in the process.

Learn more about Synopsys at synopsys.com.

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

Matt Kimble:
Hello, and welcome to Six Five On The Road at AI Infra Summit 2026. I'm Matt Kimball, and I'm joined by my friend and colleague, Brendan Burke. Hey, man. For another fun session. Yeah? All right. Hey, today we're going to talk about how agentic AI is starting to impact chip design. Really interesting topic. And we're joined by Thomas Andersen, Vice President of AI and Machine Learning at Synopsys.

Thomas Andersen: 

That's right.

Matt Kimble:

 Excited to have you here. Thank you for joining us in lovely Santa Clara.

Brendan Burke: 

Thank you.

Matt Kimble : 

Thank you. And let's have at it, Brendan.

Brendan Burke: 

Thomas, it's a great place to be meeting back up, because at AI Infra Summit, everyone's setting aggressive design targets in the chip design community. Everyone wants to build not just the chip, but the system that defines the next era of AI. So they're taking on chiplet design, advanced packaging, and aligning with hyperscale requirements, even moving into optics in some cases. But I've been doing some office visits. They're working in small spaces. The teams aren't getting any bigger, even if there's some bigger funding rounds. But with smaller teams and bigger targets, how do you think engineering teams are adapting to these really aggressive goals?

Thomas Andersen: 

Well, they're very lucky because we live in a time of gigantic automation. We've honestly seen this shortage of talent for quite a few years. Right. And over the years you've seen chips getting larger and larger. And the amount of people that it would take if we used traditional design workflows, it would just explode exponentially. So this prediction in terms of lack of workforce has been around for years. And all the examples that you mentioned, where you now have 3D stacked chips, you need to have thermal analysis, stress analysis, all these things complicate things. You would need a lot more workforce. And, you know, lucky for all these startups, we're working on solutions to automate the workflow that essentially elevates the engineer and he has to do much less. And he can essentially delegate the task to these agent engineers who do the work for them. So, that's how we see the future.

Matt Kimble : 

It's interesting, because we've seen for, as you mentioned, some years, these AI-assisted design tools, right? People have been using them, and it's definitely contributed to productivity. We're moving into autonomous engineering workflows, right? How do you actually draw the line between, like, what's the distinction between AI-assisted workflow versus full autonomy?

Thomas Andersen: 

Well, it's pretty easy, I would say. So AI-assisted is essentially you have the same tools as before, but you have some AI helpers, like co-pilots, so to speak, right? Chadbots or assistive tools that give you suggestions. But they don't actually do anything, right? And then the agentic world, I mean, it's been the hot topic for the past, I would say, year. The truth is it's been in the making for several years. Like the idea that essentially you wrap an agentic loop around LLMs and automate workflows has been the dream for many years, and we've been working on that for quite some time. And that's really where the difference is. So in an assisted tool, I simply get suggestions from the tool, but I'm still the person that has to analyze it, take an action, right? It's just the helper function. So the more I move into agentic automation, I can essentially send individual tasks to agents. And then I go to multi-agent workflows, where essentially I just present it with a bigger problem. And then it's sort of like you have a chief of staff. He sends the tasks to different people, in this case agents, and they say, you fix this, you fix this. And then you take that one level higher. You have a fully autonomous workflow, where essentially you just define what the chip should be. You write a specification. That obviously doesn't go away. The creative part still remains. So the human defines what he wants. But then the details behind the scenes, like launching all these tools, setting up the workflows, fixing errors, fixing timing constraints, or congestion issues, or verification flows, I can rely on agents to do that for me. So that is the level of automation that we're looking for with autonomous workflows. Fantastic.

Brendan Burke: 

It seems hard to achieve because you've got all those individual tasks. When we use our coding agents in general application development, sometimes it can be a one-shot process. You have a foundation model, it gets better, it calls its own tools, and it just runs through an entire coding flow and ships a vibe code to production. But it sounds like you're saying that isn't possible with engineering because there's these individual tasks and specific context throughout the life cycle. And when you have a multi-stage platform like Synopsys provides for EDA across the entire chip design life cycle, how much does the engineering specific context from each stage influence how agentic AI is organized, whether it's a single agent or a collection of sub-agents?

Thomas Andersen: 

Yeah, it's certainly compared to say creating a PowerPoint or even generating code. It's a much more complex task. It's much more than coding itself. Coding is also part of the chip design workflow, right? Like Spectre RTL, but there's many much more complex tasks like cognitive abilities and so on. And in all these areas, I think the domain knowledge is extremely important. And if you think about it, like, when you use, say, Gemini or Claude or so, and you just ask a question about some problem that you have, chances are that you get an incorrect answer. I mean, I get that all the time. It's not because the AI doesn't work. It's more that the source data is conflicting or incorrect, right, because it sources all the information from the web. Lots of people say wrong things. And in our world, obviously, we have to have 100% accuracy. So an important part is that the domain-specific knowledge, the experts, gets encoded into the LLM models, right? So there's both pre- and post-training you can do on baseline LLMs. And then you achieve a much higher level of accuracy. And this knowledge really resides both with EDA vendors who build the tools. The knowledge also resides with our customers. So our customers have built workflows. They know exactly the secret sauce about their design, their workflow. So all this needs to be embedded into the model so you can achieve essentially the same level of accuracy or higher that a human has, that an expert engineer has. But it's absolutely doable. It's not easy, but it's absolutely doable. And then if you go a little further, Ultimately, you come to self-learning systems. Now, you don't want to start with a self-learning system that knows nothing because it would just be too expensive to try all these things out and then fail. So, you want to get the certain level of like, hey, what a good human knows. But ultimately, the system actually can become smarter than a human. It can try out new things. It can learn and it can actually get better than a human.

Matt Kimble : 

That's incredible. As we move to kind of autonomy, right, autonomous engineering workflows, this spans infrastructure, silicon, software, it's an ecosystem. Can you talk about, you know, from a synopsis perspective, like how that ecosystem works, how you engage with them, and the importance of it to kind of going from great idea to this is happening in the real world?

Thomas Andersen: 

Yeah, that's exactly sort of what I had mentioned already a little bit about where the domain knowledge sits. I would say in the past, we could live in our silos. I mean, we develop the best EDHOs that we can. We get customer designs, we tune our EDA tools on those few customer designs, and we ship it to them. And then they say, yeah, it improved, but hey, on this new design, maybe it doesn't show what I wanted. That's because there's all these silos in terms of data. And that gets amplified now in the world of AI and agentic AI. So really what needs to happen is that the domain knowledge from all sites is available. Now our customers obviously they have very proprietary technology so they want that their knowledge doesn't leave their premises. What we do is we essentially build a system that has all the knowledge about the EDA tools and how to operate them and how to debug certain things built in. And then we work with customers and also foundries, right? Because technology nodes are also extremely important, all the processes. So we work with them to essentially build a system that has all these pieces of information combined. So that's an absolute key, otherwise it will not work.

Brendan Burke: 

The biggest repository of domain-specific knowledge is, I would say, the engineers themselves. And what's been impressive about the pitches here at the summit is that these engineers are setting their sights really high. They want to focus on high-level design decisions, whether that's dealing with power constraints or adopting open source ISAs and taking on some creative decisions that maybe an agent wouldn't necessarily recommend straight away. But in order to free them up to take on those more design-oriented tasks instead of the individual engineering tasks, how would an engineering leader decide what an agent should be used for? where it can take end-to-end an entire either process or even a whole RTL to tape out and then are humans still necessary at any one of those stages?

Thomas Andersen: 

Yeah, that's a good question. I mean, honestly, I think this will come in stages. First, I think agents will essentially replace like tedious repetitive tasks that humans really don't want to do. Like, when we talk about chip design, it's not as glorious as somebody just writes the RTL, pushes a button, and out it comes, right? They have to spend a huge amount of time cleaning up data. There's all these different teams that have to work together, like they partition up the chip, they come up with different timing constraints, and then they have to work together, because they don't work, and they need to clean up all these constraints, as an example. So there's a lot of tasks that are not very glorious, that actually humans really wouldn't want to do, so they will be very glad to have a system that automates it. Like the way I'm very glad that I have a co-pilot that makes flights for me, because I know the ideas, but I don't want to draw and PowerPoint the boxes, right? So similarly, you can think about in chip design, all these repetitive tasks, they would be happy to have them automated. So you can think of it, if you were an expert engineer, you would today maybe delegate those tasks to a more junior level person. And that's how I see it. Initially, the agents will be used for those junior level, simpler tasks. But ultimately, when they master them, then you level it up higher and higher. And now to your question, will the human always be needed? It's almost a philosophical question. I mean, I can totally see that you can get to a point where you can come from spec to final product, mostly automated. But I mean, the human touch, I guess, in terms of creativity, like I said, it's a philosophical question. Will, at some point, machines be better than a human? That's very, very hard to answer.

Brendan Burke: 

Well, getting from spec to tape out with 100% accuracy sounds like AGI to me, so it seems like we're on the cusp of it.

Thomas Andersen: 

I think it's doable, but it will take a while. It won't be in the next six to twelve months.

Matt Kimble : 

And this is all still amazing, right? But as you start to see kind of a gigantic AI truly being utilized to its fullest extent in engineering, right? Take us four years out from now, five years out from now, right? What does chip design look like? What does the market look like because of this huge compression in time from spec to tape? Give us a sense of what the world will be.

Thomas Andersen: 

I mean, for example, when you look at traditional chip designers today, they partition up the task into multiple steps. There's a front end RTL designer, there's a verification engineer, there's an implementation, there's a sign of engineer. All these are experts and they can maybe operate one or two tools, right? Which is also why, when you look at the EDA landscape, we sort of partition up this entire process into, there's a chain of tools. You're the expert here and you're the expert here. I think these silos will go away. It won't be there's a verification engine anymore. There will just be a chip design engineer. And the tools that are being operated by humans today, they will just become engines, right? So you operate essentially at a much higher level. And you have like an autonomous workflow where you, when problems come up, hopefully the agents can automatically fix it. I do expect that sometimes they come back and they will say, well, is this what you wanted? Because it's also, what's the intent? So there will be an interaction with the human, but I would expect that you don't have to be an expert anymore in the details of the tools, right?

Matt Kimble : 

So let me ask you this. Obviously, there's a faster cadence, right? Do you see a proliferation of more custom silicon? a really creative engineer who maybe doesn't have all the, going back to the original question from Brendan, that doesn't have all the resources, can now all of a sudden get from good idea and spec to much faster and much easier and much cheaper?

Thomas Andersen: 

Yeah, I mean, that's what I would expect. That's what I would expect. And again, the way I see this is initially we seed the system with human knowledge. But ultimately this is not a sustainable path that you always have experts and they're going to continuously teach that system what to do. That doesn't make any sense. Right. It's sort of like I teach it skills. So it's like an engineer and then it becomes self-learning. It's the same way that you would operate as a human. You join a new company and then you learn from the old guy, but you will also find new ways because the old guy is very set in his ways. He does it a certain way. He doesn't want to try new things. So you're the new guy. You will come up with new things. I see the exact same thing happening here with these self-learning loops.

Brendan Burke: 

something that's going to be exciting to every AI company here, the ability to let their engineers do more and ultimately be the one to define that next-gen system. It's amazing that you're making it happen. Yeah, very exciting times indeed.

Matt Kimble : 

It truly is fascinating. It's almost, you know, for folks that maybe don't sit in at chip companies or work in chip design, it's almost science fiction come to life. It's fascinating.

Thomas Andersen: 

I think chip design itself is science fiction, because it's so much more complicated than software development. But yes, you're absolutely right.

Brendan Burke: 

It's been great talking science fiction into science fact with you, Thomas. Thanks for joining us. Sounds good. It's been a pleasure. Thank you. And thank you for tuning in to Six Five On The Road at AI Infra Summit 2026. Don't forget to hit subscribe, follow us on socials, and check out all our coverage at SixFivemedia.com. We'll see you next time.

Thomas Andersen: 

All right, that's good, I don't know how to sound this with all this noise

‍