Jeka Viktorova: Six weeks from dying, then the world came knocking

0:00We were not being true to our DNA, which is very much nerdy, solving problems using technology, talking about science, engineering. engineering. I think that's been the biggest learning for me. Whenever you're trying to do something that's inauthentic, that is not your company's or your technology's DNA, you're not going to raise money. You're going to attract the wrong people and sooner or later, they will realize they're not talking to the real you. They're just trying to sell things. What is LAM? This technology can be used for something that is one of the biggest bottlenecks of AI chips today.

0:31We have enough gray matter in our AI chips to do human-level intelligence. So why is it not possible today? And we think it's the wiring connectivity problem. So we haven't seen the full potential of AI yet. And there's more to come in quantum and neuromorphic that will be enabled by this density of wiring that hasn't been kept out yet. We've built the most powerful AI power plants in history, but the grid connecting them can't handle the load. That grid is the wiring between chiplets. It's old, thin, slow, and it's the reason GPUs sit idle for up to 40 % of the time. Not because AI isn't smart enough, but because the electricity can't flow.

1:14Yekka Viktorova is rebuilding the grid. Her invention, LEM, stamps ultra-dense copper wiring directly onto chips. It's like rewiring a city block with high-voltage cables instead of extension cords. Her invention, LEM, stamps ultra-dense copper wiring directly onto chips. It's like rewiring a city block with high-voltage cables instead of extension cords. Bigger bandwidth, bigger packages, faster communication. The promise of Sienta is that AI can use the power it already has. Last year, Yekka had six weeks of runway and no term sheets. A top US semiconductor giant told her, you'll never hit these milestones.

1:53And three months later, they said they'd never seen progress like it. Four term sheets, three governments on board, and a quiet lab in Redfern, Sydney, is now at the center of the AI hardware conversation. If GPUs are the muscle of AI, Sienta is building the nervous system. The company that fixes the wiring becomes the company that everyone relies on. So let's get into it with Yekka, CEO and co-founder of Sienta.

2:22Can you contrast what the last six months has been and perhaps compare it to the last six months of 2024? Oh my God. Well, let's start at the beginning. Last year, we pivoted pretty hard, not in terms of the technology platform, but market wise. We had a bit of a breakthrough. What I was building during my PhD was such a different technology. It was not scalable. It was incredible. We were doing something one of a kind. But my CTO, he was running a tight ship, and he did something that we didn't think was possible, which is prove out highest resolution combined with scalability and materials properties that were just not matched by anything else in the world at the moment.

3:10And we looked at this and we went to Taiwan. And we started talking to everyone in Taiwan about this unique capability that exists in this random lab in Sydney. And we realized we could do something for AI chips. So we got back and we started the project with the top US. When is this, by the way? So this is advanced packaging. No, when? Last year? Last year, beginning of the year. That was March. We've done the Taiwan trip. We did the Taiwan trip late 2023. We were kind of exploring the market. 2024, we fully pivoted. Yep. We realized that this technology can be used for something that is one of the biggest bottlenecks of AI chips today.

3:56And we started work with the biggest semi-cap equipment manufacturer in North America. They gave us this challenge. They were like, oh, they will never make this. They will be never building this resolution combined with this uniformity at speed and scale. And every month we would catch up with them and share the results on the project. And after three months, and there were many challenges on that road, we realized, and they realized that they've never seen the rate of technological progress that we've demonstrated before. They told us that. At the same time, the money in the bank goes down, running out of money, six weeks away from dying at the end of last year.

4:43And no term sheets on the table. I'm still raising. We have this massive win of that project demonstrating something that's so uniquely positioned to solve some of the biggest challenges in AI hardware today. And two weeks later, we had four term sheets on the table. We got lucky. We worked hard. How does one go from we're raising, we're raising, we're raising, now we have four? Well, I think VC landscape is all about FOMO. And you know this very well. Once you have the first offer, the others come. First, it feels like kilometers and there you're inches away. But I think for our space, telling the story that is so complex, nerdy, and really communicating the impact, the global scale of what our technology can do, it's difficult.

5:34But when you combine that global vision of we're going to change the world because we build material in a different way with essentially here's a rate of progress, both commercially and technologically, that is so fast. That rate of learning is unmatched. If you can demonstrate the two together, that's how you raise money in a very tough environment with a very nerdy tech and so you've shared you got four term sheets contrast that now with the last six months what have you been up to so we've closed the round um late last year we went from nobody will lead because this is too hard to understand to we had four new VCs joining and two more governments actually three governments Singaporean government US government invested through their funds in addition to Arizona state government wow giving us more money in incentives than we've ever raised before just because Arizona sees such value and they see that there's a new revolution coming and it's really interconnect technology rather than transistors.

6:48We've kept out transistors. Now we're focusing on this new space, which is exciting. And it's really anyone's game and everyone's game. And you don't have to invest hundreds of billion, even though they are. You can start with millions and you can go gradually and we'll need innovative solutions. We'll need to think outside the box. So that's really the unique perspective that I think Arizona saw. We're not the only ones, They're investing heavily in other technologies as well. But we've never seen this level of support before. What's special about Arizona? Arizona is going to become the new home of TSMC.

7:28TSMC is incredible at what they've been able to build. And keep in mind, they're using the technology that hasn't been built for Interconnect specifically. They're using transistor-like manufacturing technology to build everything. And it's unbelievable what they've been able to achieve in a partnership with NVIDIA, which gave us the baby that is Chad GPT. And at the same time, we're realizing that sovereignty, second sourcing and supply chain is becoming more and more important. And that's why even TSMC is investing outside of Taiwan. So they're looking into Arizona. They're investing$160 billion to build some parts of AI chips.

8:11So when we talk about these chiplets that are connected by the wiring. So maybe I can start at the beginning, actually. AI has a wiring problem and a capacity problem. A wiring problem is really the interconnect density and a package size problem. So current technology, which is based in traditional processes, hasn't been able to solve for communication between chiplets. and that means that currently running inference chips within the data centers are idle up to 40 percent of the time and that's extremely inefficient totally think about the hundreds of billions of dollars being invested in data center infrastructure how much of that energy and money is wasted just because we're not operating with optimal wiring yeah just asleep Exactly.

9:07That's terrible. So why is that? Nobody really thought about manufacturing on micron scale. Everybody thought about nanometers, which was Moore's law and still is, which was transistors. Wiring needs to be micron scale and it needs to be large area. And this is something that is very important for our industry. And the second part I mentioned is capacity. because we're using these nanometer scale technologies to build micron level stuff everything gets slower your field size gets smaller and then getting high yield and high throughput is impossible using conventional technologies so what does that mean you have people building more and more chips or trying to design more chips ordering more data center infrastructure, but supply chain of AI chips can't keep up with the demand.

10:07And the only people who've mastered the building of this packaging are TSMC. And the only place where they're able to do this is Taiwan. So there is a little bit of magic to what TSMC does. And we're not talking about EUV now. We're really talking about co-op's packaging. This thing that connects two chiplets together. They've only been able to do this in Taiwan. And a part of it is supply chain. And the second part we think is the secret sauce process that they have. And currently, they're unable to replicate it anywhere else in the world. So despite Arizona becoming this hub of activity when it comes to Semicon, they're not going to be packaging the chiplets there.

10:56They're going to be building the chiplets, shipping them back to Taiwan, and that's where AI chips are built. Yeah. And I think the granularity of that narrative, almost nobody understands. There may be hundreds of people that really understand that there's no second source for packaging in the world today. But people who do understand, they start thinking about AI sovereignty and the beast of AI demand, it needs to be fed somehow as well. So even purely commercially, you need to start thinking outside the box. Keep going. You're on a roll. I'm hanging off every word. Does this make sense to you?

11:39Yeah, it makes perfect sense. Like in the back of my mind, I'm thinking like, wow, I assumed I would need to come in and then summarize what you've just shared. Oh, please do as well. If I need to, I will, but I don't need to at the moment. That's wonderful. because it's super clear. Should I tell you how we solve both of those problems? Please. In detail? Yeah. Okay. That's why I was like, keep going, because I was like, you just teed yourself up. You don't need me. I need you. No, no, please. I need you to make me laugh. So the wiring problem. Yeah. There is one thing that the Hogwarts, that is TSMC is unable to do, is to go to high resolution density of interconnects with the existing process.

12:24There's some physical limitations. And this means one of two things. Either you have to build entirely new fab infrastructure, invest even more. But you're still slow. So that's not necessarily solving the capacity issue. So what we do is we allow to break those physical limitations of high resolution interconnects. interconnects in addition to not being reticle limited which is this field size limitation of current lithography we can build packages that are bigger we can build on the panel we are allowing for that maximum reticle utilization that allows you to get your economics and throughput up so that's how we solve the wiring problem denser wiring on larger packages and then you can stack more chiplets on top of it.

13:17So your high bandwidth memory and your, you know, compute goes on there. And then the capacity, the cool thing that came out of my PhD was the scalability and the speed of deposition of material. So that's something that we've maintained and really parallelized since then. And that allows us to triple the productivity, triple the throughput, and reduce the costs at the same time because we gain efficiency by eliminating some of the steps in the current manufacturing process. And walk us through where you sit in the supply chain. Because of the disruptive nature of what we're building, it's a chicken and egg problem.

14:00You kind of have to demonstrate the full chip level capability of this. We're relying on so many other parts and people and pieces. Exactly. so what we've decided to do is to bring the industry together yeah um and i mentioned this big project we did last year and since then we've really rallied the foundries the osats the chip designers the oems the edas all of these beautiful acronyms which means basically people who build the tools build the chips design the chips design the software to build the chips all of them together and what we want to own is the process. This is the secret sauce that is LEM, that is Sienta.

14:43We've developed a unique way of building high-resolution wiring and we want to integrate it into the software. We want to enable the chip designers to use it to break design rules. If you think outside the box of the current manufacturing technologies, you can build things you've never built before. So we are essentially the recipe builders and we give that recipe to the foundry that builds the chips. Yeah. And why, how have you been able to show to everyone that you should be the center of gravity, the magic, the recipe? What was that journey like? In essence, power. TRL progression, really de-risking gradually.

15:28It's very easy to show something works once really translating it into numbers and data to prove that you know you can actually get to high yield high reliability the aspect ratio the material quality that matches what is currently possible so we focus a lot on how we build things but the thing that matters to you know nvidia tsmc is that what comes out at the end is exactly the same as they're used to seeing in terms of material quality and material property um and then showing the scalability and that's something we've been obsessing over since the beginning of sienta in fact my cto ben when he first met me he was a customer of my technology we had a very nice conversation about what i was in the lab at that time he saw a demo and then he asked me, Yeka, this is fantastic but slow, how do we scale this?

16:33And then gradually over the next year we came up with an idea of how to scale it from the get-go. Before there was Sienta, before we incorporated, we had the two product lines in mind. One of them was solving for scale and then over time Ben realized that if he helps me solve my problem, he will solve many others, including his. What is LAM? Localized electrochemical manufacturing. To put it simply, imagine you have a high resolution stamp and then you have a substrate, which is your silicon wafer or glass or any flat substrate you can imagine. And then what you do is you apply electrochemistry and you use the stamp and the substrate as two electrodes.

17:22And what you can do is you combine the deposition of metal and patterning at the same time. And that's the unique property. That's the one of a kind, you know, fast amplifying advantage that LEM has. It's almost like building copper with a stencil, but we use electrochemistry as a trigger. And that really localizes the way we deposit our metal. typically it would take many many steps one of them would be masking step where you put polymer down and then the unexposed areas would be filled with copper so we avoid that first polymer step and then the removal of that layer and that's one of the biggest differentiators of our technology and was this part of your phd yes lem was invented as part of my phd the reason I asked that question is I it's clearly been the thing that hasn't changed out of the product and remains the magic or no yeah yeah yeah and when did you first realize the power of your insight and maybe you can teach us a bit more about um how it works so the journey started um back in Germany um I was actually born in the Soviet Union I was born in Latvia luckily enough when I was living there we joined the European Union so I realized that I wanted to get a great education so I moved to Germany to do a chemistry degree and as part of my master's thesis I was looking for a cool project and I discovered this unique thing called printed electronics I had no idea what it was but it sounded sci-fi so I thought why not use my chemistry background to do something cool.

19:15I applied for this master's project and pretty much three months later I invented my first printed electronic sensor and that became a product that my supervisor decided to commercialize. So we started a startup. I was the first engineer and within the next few months what we really realized is that there's a strong need to build more of these sensors, we started working with Samsung and LG. And my job became not to just build six of those sensors and show how they work, but start thinking about mass manufacturing. So think about building a sensor and your goal is to make them cost less than one euro cent.

20:01There weren't many technologies that would enable you to do that. But I started looking and we realized there's one way is to use roll-to-roll manufacturing, which is a high speed, high scalability process that was very emerging back in the days. It was 2017. The technologies were so immature that I had to redevelop full material stack, rebuild the entire sensor from the ground up. And despite us growing very quickly, we failed. We could never scale the sensors. I built 30. I never built hundreds. We never got to scale. We tried using conventional technologies too. We went into the clean rooms.

20:47We looked at hybrid, basically a part using conventional technology, a part using roll-to-roll. Never worked. So this is when I fell in love with the problem of scalable manufacturing of electronics. and I was especially passionate about the value proposition of printed electronics because it was a lot more sustainable that whole part of just using the material you need rather than building many layers and then removing them only to have a couple percent of that material be functional that was something I was very passionate about and then I thought the next logical step to solve the big problem would be to go and deepen my research roots, carry a project, find something relevant still in material space, but using the new field that I discovered for myself, printed electronics.

21:42I went online and I looked for different PhD projects and I found this one. It was really appealing because it was both polymer chemistry and additive manufacturing and functional materials, which I saw slotting right into that. And yeah, six months later, I moved to Australia and we invented LEM pretty much months after that. And when we invented it, we realized it's going to solve the problem that I was facing in Germany. It was instant. It was so clear to me. Even back when it was a prototype that we built for, you know,$200 and there was nothing but a little dot of copper on aluminium foil, we realized that this will have a life-changing impact.

22:36To the extent that you can share, where is this scaling magic? What is the breakthrough exactly? The breakthrough is just in the way we build material. That magic of using electrochemistry as a trigger and now introducing scalability through the stamping process. That is the innovation and that is the biggest differentiation. And what is electrochemistry? Electrochemistry is moving electrons around to, in our case, go from copper 2 plus to copper zero. So redox chemistry, things that your cell does. and you know electroplating is a very good industrial example of that chroming on your wheels that that is the electrochemistry and you mentioned there was a scaling challenge that you faced I think last year or probably earlier yeah like what was that about that was really about the lack of parallelization in the system so if you have your electrochemistry and it's operating at the resolution that you wanted to operate at what you really need to solve for next is exactly what i described being the challenge in germany really having large area paralyzing that process that took a long time and this is what ben worked on even back when we only were focusing on the non-scalable version for productization.

24:14We've been able to demonstrate using stamps. So really increasing the area of deposition, that was the innovation. And I think when we realized that we could reach the resolution that's never been demonstrated before with that parallelization and match the material quality that is needed for industry, that was the big aha moment where LEM even though it was the same technology it transformed into something that will solve so many industrial challenges in our mind and what does the rollout look like from here and maybe and maybe like give it a bit more context where are we in this AI moment and what does the rollout of TSMC's project look like and your place in it yeah perhaps start there okay well for us the next step is to transition from single stamp all the way to large area demonstration so we're talking wafers and panels that's going to happen next year in arizona and in sydney this is really about engineering integration and demonstrating that the rest of the process works not just the lem part which is the most important part because it allows you to build copper, you still need to ensure that it's compatible with the existing infrastructure.

25:36Because one of the other advantages of having this process that just slots in into the existing infrastructure, you don't have to build new fabs. You can just adopt LAM and extend your roadmap and solve your capacity issues without switching to a new annex to your clean room. So proving out that integration is one. And then demonstrating the efficiency gains. So how do you solve the memory AI wall? You improve the density of wiring and then you improve the chip bandwidth. So actual performance of your chips improves. If you increase the bandwidth 20x you've solved the memory AI wall. Now you have to optimize memory And then you have to focus on transistors again.

26:23But that's a challenge for another decade.

26:30So, 2028, we're going to go into high-volume manufacturing. And that's the biggest impact on the capacity part of the problem that we're solving. This is where most likely high-volume panel manufacturing will occur as well. And then that gap in the current AI chip capacity can finally be filled. through innovative technology. What is the AI memory wall? It is this communication limitation between the two chiplets. This is the limitation of the wiring that leads to inefficiency in the data center infrastructure. And that's what essentially inhibits the 40 % off time. Exactly. Okay. And you've already shown that LAM works, And now you're, just clarify that first stage for me, just so it's clear in my head.

27:22Sounds good. Yeah. So we've demonstrated that LEM works on a stamp level size. Got you. So now what you have to do is place multiple stamps onto a wafer with certain alignment precision. And that's, you know, using existing alignment infrastructure. We're not inventing EUV here. We're operating on micron scale. And then once you've placed your stamps with this alignment precision that is required, you do the same deposition using electrochemistry. So that's really, again, the final step of that parallelization milestone. And then you will be able to go full throttle. Yes. Okay. What do you need to believe over the next year to, like, what needs to go right?

28:05And what's inside, like, what's in your control and not in your control? Sometimes you can accelerate things in different ways. In the beginning of the LEM journey, it was really the time. We needed to try out different things and find the optimal pathway for the technology to take and the TRL maturity to kick in. Right now, the biggest limitation for our scaling is money. And we've been incredibly lucky to find enough capital to get to this stage. And we see that with this final hardware integration and process integration milestone, If we're able to demonstrate the chip level performance improvement, that is the final milestone that we need to demonstrate.

29:02There we are relying on several partners. So there's a little bit of a logistical thing that needs to go right in terms of bringing people together. But I have full confidence because I've seen this industry do something that nobody has done before over and over again. and this is the magic that is asml and this is the magic that is tsmc so there's nothing really holding us back but that accelerating capital which will get us all the way to profitability the journey of setting up your own fabrication with it requiring so many different pieces like my understanding is that the timeline traditionally because there is so few people who can do it all in Taiwan is just like a total dream like good luck the word never comes into play then TSMC decides that the moment is now to take what we know and build that in the US specifically Arizona if that's if that's true it seems as if like this really is a rare moment in time where the stars have aligned that everyone's incentive is also to see you win Is that unfair or fair?

30:17The way that I see it, it's the pull from the top. It's not the push from the bottom. That's how you crack the knot of speed and scale. If you look at the hyperscalers, your anthropics and open AI, the way that they've been able to demonstrate this incredible power of AI without even getting to AGI has been through NVIDIA, AMD, Google delivering the right hardware. And then Google and AMD and NVIDIA are pushing on TSMC to give them the chips. TSMC is pushing on semi-cap equipment people to bring them the right tools. And that's how you leverage momentum. you really go from the top. Who are the owners of the data center inefficiency problem?

31:14It's the hyperscalers. How much are they willing to pay to solve that problem? A lot. Yes. So I think if you use the right stakeholders and you have the right incentives of efficiency, capacity, you can solve the challenges of tomorrow. And we can go one step further. If we really are talking about AGI, if you're talking about quantum and neuromorphic, all of these technologies will have a wiring problem. Currently, AI chips have, in terms of transistors, the same capacity as our human brain. And my CTO loves to talk about gray matter versus white matter. We have enough gray matter in our AI chips to do human level intelligence.

32:03So why is it not possible today? And we think it's the wiring connectivity problem. This is why we need more white matter. And if you communicate that to the hyperscalers, to the chip designers, they really get this problem because that's been the newest thing that changed the entire Semicon world. Everybody's thinking about the wiring now. So to your point, I think we will see enormous acceleration. Yes, we are in a bubble, but also we haven't seen the full potential of AI yet. And there's more to come in quantum and neuromorphic that will be enabled by this density of wiring that hasn't been capped out yet.

32:47How do you avoid getting acquired?

32:52or do you want to be acquired at some point? Like if I'm one of these hyperscalers, and you don't have to answer this question, but if I'm one of the hyperscalers and I'm seeing progress and like either there is a race for specific specialized chips or I can like the missing piece is actually what you've invented, that's a heck of a position to be in. With the current trend toward vertical integration, AI sovereignty, second sourcing, supply chain resilience, it's inevitable that there will be players that will look at this technology and see it as their backup option. Not to rely on TSMC, not to rely on Samsung, not to rely on Intel.