Transcript: Sviat Dulianinov, CEO of Bright Machines | How AI & Robotics Are Reshoring Data Center Hardware | Startup Project #127

Nataraj: Hello everyone, welcome to Startup Project. Today on the show we have Sviat, the CEO of Bright Machines. Bright Machines is reimagining how we build physical infrastructure using artificial intelligence. It was founded in 2018, and they combine AI-enabled robotics and manufacturing intelligence to build data center hardware. Their investors include BlackRock, NVIDIA, Microsoft, and Jabil. They are, to put it fairly, bringing manufacturing back to the US using AI and robotics. We'll learn more about how they're doing that and their story. Sviat, welcome to the show.

Sviat: Thank you, Nataraj. It's great to be here and thanks for having me.

Nataraj: So for those of us who haven't heard about Bright Machines, what does Bright Machines do exactly?

Sviat: As you mentioned, we're eight years old, and we are rethinking how you design and manufacture complex electronics. We create factories with the approach that you use software first and robotics first before you engage people, so that you can enable production and assembly of electronics anywhere you want to do it. We also say we bring manufacturing to the edge, which means we can build electronics close to the place where they're going to be deployed. So in other words, we are building manufacturing capacity in the United States today, and we are 100% focused on AI infrastructure — all the electronics that goes inside modern data centers.

Nataraj: So you're enabling building the equipment that is put in the data centers. Is that right?

Sviat: Correct. You could think about it as three big components at a high level. There is compute — AI compute, GPUs and CPUs, the previous generation of compute — and then there is networking, and there is storage. All those building blocks you need to build, and then you put those racks inside data centers to create larger systems. So basically we build those compute nodes, and then racks, at the end of the day.

Nataraj: Was that the original idea — to build and power data centers — when Bright Machines first started?

Sviat: The original thesis didn't actually change much, but we narrowed it down. The original thesis was: how can we use software and robotics to enable manufacturing of electronics anywhere you want? Historically, manufacturing has been centered around China and other Asian regions, including Taiwan. During COVID, it became clear that the supply chain is not as robust as everybody thought, and you don't want as many dependencies — you want to move some critical assembly to local markets.

So the thesis was: you need to solve it in two ways. If you think about China or other Asian markets, they have two big things. One is skill, because they've been building electronics for many years. The second is the amount of labor — the scale of that labor. Our answer was: we can solve the skill with smart software — back then machine learning, today AI — to teach robots how to do things, how to navigate, how to use force control, and so on. And on the scale side, you can replicate robots easier than people. So if you want to build in the US and not depend on Malaysia or China, you just use as many robots as you can, with smart software to enable that.

Historically, we started with different types of electronics — anything from coffee machines to appliances to drills. Over time we learned a lot and deployed in different countries for different products and customers. But as a startup, you always need to narrow down; you cannot do everything. So maybe four and a half years ago we made some bets, and one of them was data center infrastructure. That was a pre-ChatGPT moment. When the ChatGPT moment happened at the end of 2022, I think we were in a good spot, and we decided to go all in and continue focusing on that.

Nataraj: You mentioned China. Today, if you think about server racks — a traditional server rack, or a GPU-based or CPU-based one — where are they manufactured?

Sviat: Today, I would say Taiwan is maybe number one, because a number of companies focus on that there. It made sense because you had chip production there by TSMC to start with. In the last several years, some companies started trying to bring production closer to the big markets like the United States. You can find a few production sites around Mexico because it's close to the border, and companies are starting to build out sites in the United States — which we're part of — because you also want to do it as close as you can to the data centers where you deploy it. But historically, Taiwan and a few other nations around Asia, and then a move toward Mexico for the US market. Today we're part of bringing it to the United States and building it right here.

Nataraj: If you think about who is actually procuring the data center racks — is it a company like Dell? Is it Dell with EMC for storage racks? Then you have HP, IBM, and hyperscalers who have their own data centers. Talk to me a little bit about the stack, from the original equipment manufacturer to getting it into the data center.

Sviat: It's a really good question and a good way to think about the market. You need to start with chip designers — think AMD or NVIDIA. They design the chips, and TSMC builds the chips for them. What happens afterward is that you can have multiple players who design the system itself — the compute node and then the rack. On one hand, you can have companies called ODMs — original design manufacturers — who design and very often also build the product. Those companies are very often based in Taiwan, as we discussed.

Then you have OEMs — companies like Dell or HPE — that can take a chip and design a server and a rack. They have two options: they can go find a partner to manufacture it — a contract manufacturer — or they can have some capacity in-house and do it themselves. It depends on their choices. And then the next part of the market is what we call hyperscalers, who already have the critical mass and capacity to build their own designs, and very often their own silicon. They have a similar choice — very often they go find a partner to build it, whether in Taiwan, Mexico, or the US, and rarely they decide to control it themselves. Rarely today, but it's going there — the larger you become, the more vertically integrated you want to become.

So that's how I'd look at the market: you have OEMs, chip designers, hyperscalers that could design a chip or a server, and then partners on the market that help them design (like an ODM) or build (like contract manufacturers, or what we do — manufacturing as a service).

Nataraj: When a company like Dell — which now owns EMC — or HP comes up with a rack design, who are they generally partnering with? I'm talking about the pre-GPU era, before NVIDIA started selling racks. Who are they partnering with to manufacture those?

Sviat: It's a standard list. You have a number of contract manufacturers around the world, both in the US and other markets. If you look at our cap table — you mentioned Jabil as an investor in our Series C — that's one example of a contract manufacturer that is actually an American company. So they have the option to go to companies like that, or to companies like Foxconn, the largest electronics manufacturer in the world — the one that also builds iPhones. You'll remember "designed in California, assembled in China" — that was with Foxconn. So historically, for large-volume production, they could go to China or Taiwan, to companies like Foxconn. Now you also have a few American options like Jabil and Flex and others. Or you can come to us today and build it with robots instead of people.

The big differentiation is that historically all those names use a large factory. If you go to a factory in Shenzhen, one factory can employ 60,000, 70,000, 80,000 people — large facilities with tons of manual labor. The idea was that humans are flexible; you can build different things with them, and it's easy to change from one product to the next. We want to replace that with robots. And I don't think you have many factories in the US where you employ a hundred thousand people on a given day, to be fair.

Nataraj: I think that's the reason the whole thesis of offshoring worked — human labor in the US is costly. If you wanted to produce an iPhone in the US, there were estimates of it costing five thousand dollars or so, and I think it would actually be much higher. Originally the whole arbitrage was the cost of labor. That's why offshoring really worked, and it shows up in every part of the US consumption economy. When I first moved to the US, I lived in Wisconsin, then moved to the West Coast, and I realized the lifestyle wasn't much different, because all the big-box stores were selling the same things at the same price across the country. That's one of the benefits — you could argue inflation stayed around 2% for two decades partly because of that. It had negative consequences too, with certain critical sectors and manufacturing moving out, but the positive side is you can consume a lot more with a constant inflation rate. I attribute part of that to offshoring. But moving on — you mentioned betting early on AI data centers. Talk to me about what the bet was and the thought process back then, especially before ChatGPT.

Sviat: Not only AI data centers, but data centers generally. You brought up an interesting point on inflation and electronics historically not built in the US. When we looked at the market and worked with different devices and customers around the world, we saw that if you build a hundred-dollar piece of equipment — a drill, a coffee maker — there's a different attitude toward how you build it. You look at different KPIs, because a coffee maker or a drill is not that strategic, even at the state level.

A data center unit is different. If you build a server — even a CPU server — one unit could cost around eight to ten thousand dollars. And today's GPU compute nodes could be anywhere from fifty thousand dollars to two hundred plus. That's a really expensive unit. Imagine you build a box, but the box costs as much as a Ferrari, on the line. When you have such expensive components, the way you look at innovation becomes different, because you're driven by different metrics. It's not only about labor. If your labor is 20 bucks per hour to build a $250,000 piece, do you really care if it's 20 bucks or 25 bucks? You really care about quality. You care that you don't lose or damage materials during assembly. You care about throughput, because if you can do it fast and at high quality, you can deploy it fast in the data center and start generating revenue.

So the metrics and drivers of decision-making become very different. That was one of the reasons even back then — if you build a unit at ten thousand dollars a piece, we'd care less about how cheap the labor is in China. Also, if you look at strategic verticals — aerospace and defense, AI infrastructure, pharmaceuticals — those matter a lot to each state, which makes it important not only to produce locally at high quality, but also to have secure operations. If you build critical electronics in a country you don't have access to, with a partner you don't have access to, and it's built manually, it's really hard to guarantee the security level you're looking for. You'll remember incidents with Huawei and the sanctions afterward, and other instances too.

So you look at the key drivers for such a device, and that's why our value proposition was much more aligned with data centers than with building a drill. I'm not saying you can't build a drill with this technology — it's just, is the drill strategic enough for the country to move it to the US? Maybe for some people, but to be fair, aerospace, defense, and infrastructure are. That's how we made the decision. And what happened in the market confirms it 100%. For the last year and a half, since the new administration came into office, they've become much more vocal about building here, leveraging robotics, having secure operations — for the same reasons we were thinking about four and a half years ago.

Nataraj: Once you decide to bet on AI infrastructure, what is the first product you started building, and who were you building it for?

Sviat: Compute servers — not GPU but CPU compute servers, compute nodes — for one of the large hyperscalers.

Nataraj: Got it. And now, what are the different types of products you're building overall, and roughly what does each represent as a percentage?

Sviat: Since we started on this, we built older-generation compute, then GPU compute, and then different types of accelerator compute, because you don't have only GPUs but other types as well. We also built different types of data storage servers that go to the same data centers, and historically we also built some networking boxes. So basically all the flavors of equipment you put in the data center.

AI compute and AI accelerators are the most needed today, and they're the most expensive. There's a lot of attention to that from the companies we work with — to accelerate it and build it at higher quality and higher speed. Everybody wants to deploy as much compute as possible, because today demand is much higher than supply.

Nataraj: Talk to me about how different your method of building a server box is compared to a traditional approach.

Sviat: The traditional approach is not new. When Ford designed a conveyor, you had several steps and you put things together. When you think about electronics assembly, think about assembling Lego. You have maybe 30 different pieces to assemble, at higher precision and quality, but it's like assembling a Lego box. Then fifteen of those go to a rack, and then you have a bigger Lego box that you deploy in the data center. So think about ten to twenty stations, maybe up to thirty depending on complexity.

Historically you have some stations on the line and a few offline, all dependent on manual operations. People build in sequence when they have materials. At the end, you test. Very often the test doesn't go well, so your first-pass yield is a really important KPI. The higher your first-pass yield, the more efficient it is, because you don't need to rerun the assembly of the units that didn't pass.

In our approach, there are two big things. First, in the ideal world for the next generation of a server, we want to start with the design first. When we did our Series C, we started working with NVIDIA and Microsoft, and the idea was to build this collaborating with their tech — we called it Bright Designer. We take the design from standard CAD, move it into our app, and run different tests to give feedback to the design team of that OEM or hyperscaler on how the design could be improved. In the long run, over a few generations, you can move from, say, 60% automation to 80% to 95%.

Why does it matter? Historically the design and manufacturing teams are not well connected. The design team doesn't take into consideration how it's going to be built, and that's why you run into bottlenecks and need more manual labor, because robots can't do everything. So you start with the design, iterate on it, and improve the design, the end-of-arm tooling for your robots, and the flow of assembly. Then you meet in the middle and get from 60% to 70% automation. After that, you run a simulation of the line based on the finalized design, then build and deploy. You can do it in waves — start with 50% automation and add more robots.

So versus 100% manual operations historically, our deployment has many more robots on the line, many more sensors and cameras for inspection and traceability at each station — even manual ones — and still a few people, because we have humans in the loop. It's a very different picture from a conveyor without many sensors and tons of people assembling stuff.

Nataraj: Do you have an estimate of the apples-to-apples advantage — traditional approach versus Bright Machines? Beyond producing in the US, in terms of numbers, what advantage is the customer getting?

Sviat: The number-one thing for many customers today is time to market. Imagine you have a whole stock of inventory — TSMC shipped you the chips, you have memory and motherboards — and you need to move as fast as possible, because the CapEx numbers today are insanely high. A large part of that CapEx goes into those materials. If you have materials in your warehouse but they're not deployed in the data center, that working capital is not actually working — it's frozen. So the faster you can deploy, the better.

You drive that with two big metrics. First, the quality we discussed — if you're really close to 100% initial yield, you take units off the line and start deploying. When you do new product introductions, initial yields can be extremely low, sometimes 30%. We know we can ramp faster, and our yields at the line level are around 98%, which is very high. Second, throughput. Robotics is more consistent and more precise — we can operate at 45 to 55 micron level precision, which is about half the width of a human hair — and it can be faster than human operations, especially if the design is right. In some cases our throughput could be up to two times better than manual operations. Combine those two metrics and that gives you how fast you can deploy your product to the data center.

Nataraj: Talk to me about how a typical assembly line looks, because we can all imagine a Ford or Tesla line. Is there an optimum mix of robots versus humans in the loop? At some point Tesla over-engineered with robots and had to remove some to get the optimum. And does the line change when you switch from, say, CPU server nodes for one customer to a GPU node for another — how much do you have to change?

Sviat: It is true — going back to how design matters for how you build it, sometimes too many robots is the wrong choice, because some things aren't built for robots and it creates a bottleneck. That's the key reason you need people on the line for some parts.

How the line looks: we design it to have flexibility on both hardware and software and to be modular. We had different generations of lines. If you search Bright Machines, you'll likely see the previous generation of large black closed-loop cells — we call them BRC, Bright Robotic Cell. It's about the size of a large fridge. Each cell is one of the processes, and you put several cells together to get a line. Today we've moved to a more flexible, hybrid generation — we can open the doors and put a person in if needed. The cell became a little wider. You still put ten to twenty stations together in whatever form fits your floor, and it runs — process after process. You have people for the parts that aren't good for automation, and robots for the parts that are automatable and critical to quality, because robots are much more consistent.

To give you an idea of how much is automated: you might start with 50%-plus automation and gradually move to 80% or so, but today you still have around 20% done manually, depending on the design.

Now, how reusable is the line for new products? On average, the software and hardware are designed to be as flexible and reusable as possible. But it depends on the sequence of assembly and the physical size. For example, if you build a standard CPU server that is one U — not super thick — and then move to an H100-generation GPU server that's much deeper, the question is whether the robot can do it in that configuration. Maybe the answer is yes, and you don't need to change much — maybe only the end-of-arm tool, because our robotic hands use a specific end-of-arm tool for each process. Your memory, CPU, or GPU from AMD looks a little different from NVIDIA, and from Google TPUs, and so on.

So you look at the number of processes, the sequence, and the end-of-arm tools. If it's the same product family — same brand of GPU or CPU server with just a design change — you usually don't need to change any tooling. Because we use AI to navigate and make decisions live, you can introduce those new products within just a day. But if it's a different platform — AMD versus NVIDIA — it might take a bit more time if the physical design differs and we need to change tooling. We've also built several brands of GPU servers on the same line with just a tool change — when the robot sees this generation is Intel, it changes to the Intel end-of-arm and does the placement; if it's NVIDIA, it changes to the NVIDIA tool. That's my point: it can be hardware- and software-flexible to enable that. Does that answer your question?

Nataraj: Yeah. From a business perspective — you get an order of, let's say, ten thousand racks for a couple of data centers. Once you manufacture and hand them off, are you done? Or do you also have maintenance — if a node isn't working, you go fix or replace it? How does the business guarantee to the customer work?

Sviat: We don't have services today; it could be part of our roadmap. We're focused on manufacturing services. We work with large companies that very often prefer to service themselves for IP and security reasons on the data center sites — that's driven by their choices. Today we provide manufacturing as a service. Each unit we build — whether you define it as a rack or a compute node — is how we provide the service and how we charge customers, for the units we build for them.

Nataraj: What is the scale of operation right now? How many physical locations, how many units do you produce, and how fast is it growing?

Sviat: We're growing this year pretty substantially — more than 3x. In terms of volume, to give you an idea in gigawatts, it's going to be more than half a gigawatt of capacity that we build. I can't give all the details on where we're deployed for customer reasons, but it's substantial. Because we drive quality and throughput, even with one or two sites you can generate a lot, because the lines are very efficient versus manual operations. We can build from forty to fifty units per hour at really high yields — if you have materials and it runs non-stop, you can produce a lot of compute nodes. Forty units meaning a compute node with a few GPUs, not the rack — a rack takes much more time. But you can produce a substantial amount of compute nodes from even one site, just because of the efficiency of robotics.

Nataraj: How do you see the business evolving? Do you see CapEx spending continuing to grow? And how do you approach a year where CapEx spending drops from the hyperscalers — does it affect Bright Machines?

Sviat: It's a question we're very often asked, because everybody's worried about what's happening with CapEx. In the short term — the next several years — we don't see the risk, because most CapEx is already committed and pre-committed. You need to plan materials in advance and build them. There could be some delays, as you see in data center build-outs with energy levels or permits, but where we see a commitment to components and assembly of those server generations, it's not going away in the short to mid run.

In the long run, the beauty of our technology is that today it's data centers and we focus on AI compute, but we can also do networking and storage. We're flexible because this tech is applicable to similar complex electronics — even telecommunication boxes, like 6G boxes, are built with a similar approach. So we can easily go there. But right now, to the mid run, we are 100% focused on data centers, and we don't see any drop in demand. Within data centers, the flavors could be different — today GPUs, tomorrow CPUs, TPUs, data storage — enough flavors to build with the same technology stack.

Nataraj: When we hear "Meta committed X billions of dollars," how long does it take from that commitment to actually being deployed?

Sviat: It depends on a number of bottlenecks. From the very start you need to get chip allocation from TSMC — whether in Taiwan or now in Arizona, which is closer but a newer site. Then you need allocations for all the other components. If you look at NVIDIA and all their partners — from liquid cooling to memory — it's a lot of vendors and partners. That's to start with.

Then you go to the manufacturing part that we solve. When you have all materials and want to move fast, you want to do it with software and robots close to where you'll deploy it. If you depend on Malaysia, Taiwan, or China, it takes time and creates risk. Then you go to the data center, where you need energy, site approval, and liquid cooling before you deploy the racks. So it depends on how you unlock each bottleneck, and it's not one answer fits all. It also depends on the region — some regions are more challenging because there's not enough energy, so you move to nearby regions. It's an ongoing process.

Nataraj: Talk to me about the core IP Bright Machines is building. I can think of the assembly and manufacturing, but robotics should be a huge part, and how you operate that robotics. Is there software orchestrating the whole thing, plus software taking the design and improving it? Are these different things together from Bright Machines?

Sviat: They go together. It's a platform we build. The key component and critical differentiator is the software platform that orchestrates the whole line — the automation, what happens on the line. On top of it we have an AI stack we call "smart skills" for navigation, inspection, and force control — how the robot decides to do things. After this, we collect the data, and you can work with it for analysis and continuous improvement, and you close the loop by taking that data into the design application, Bright Designer.

So three big pillars: the platform powered by smart skills, the data layer, and the design layer, Bright Designer. You can do virtual NPI — new product introduction — and improve your designs. All of it is a pure software stack. On the line, we have our own design of the cells, but we use off-the-shelf hardware. We can use a robotic arm from different partners — FANUC, UR — or build a custom robot; we can use different cameras and conveyors. That's driven by the flexibility of the software stack.

So if you ask, is your secret sauce the robotic arm? The answer is no. My secret sauce is the platform that runs all of it, collects the data, and then improves the design. We use different robotic arms from different partners on the market.

Nataraj: Physical AI and robotics have seen an upsurge — a lot of startups creating humanoids, partly driven by large language models changing how robotics learns. Previously robotic programming was deterministic; now it's shifting to more advanced machine learning and LLMs. What's your perspective on what's happening in robotics generally?

Sviat: It's a great, exciting time. There are big camps. One camp is general-purpose humanoids, which we're definitely not part of. The other is purpose-built robotics for solving real problems. The humanoid space gets a lot of the news and excitement for obvious reasons, but it's nascent — you can't use it to solve real problems today. When will you? We don't know; it'll take time, as it took time with other types of AI. Ten years ago, AI could only maybe say "this is a cat, not a chair," and not in all cases. Today we can write essays and PhD-level content with ChatGPT and Claude.

But for real industrial applications today, our approach is much more mature. You can operate at the precision and quality you need — but you're not going to use general-purpose models or humanoids for that. General-purpose models are good for unstructured tasks like folding a t-shirt. What we need to do is take a memory module and place it at 70-micron precision, at the right force, onto the motherboard, without cracking it, consistently. That's driven by a different approach, different models, and different equipment. We use a cobot or robotic arm, not a humanoid.

Is there an application for a humanoid on the factory floor in the future? Maybe. But even to move materials, you can use AMRs instead of a humanoid. Today we see a lot of return on investment on the technology we use for particular use cases that don't need humanoids. But if you show me humanoids that can help solve tasks on the manufacturing line, I'd use them. They just don't exist today.

Nataraj: Do you see improvements in LLMs or machine learning impacting what you can do with general-purpose hardware? Previous-generation robotics and AI companies were specialized — we had a company called Grey Parrot scanning waste-management lines, developing their own ML models. Now vision has largely moved to a regular API call on OpenAI's API. Are there improvements impacting what you can do with a regular robotic arm that weren't available five years ago, and where is that research coming from?

Sviat: We use different types of AI for different applications — to make our engineers' lives better, to work better with data, to use teleoperation for a particular task or a less-structured pick-and-place with a cheaper component. There are examples where we use a beneficial stack, definitely. But it's not a wide range where we decided to repurpose our stack for 80% of the processes — that's not the case.

For particular examples — how we can create training for a robot faster, learn using LLMs digitally, work better with the data we collect, create hypotheses for engineers, or use teleoperation for a particular test to train robots better — we do that. But it's applicable to a narrow set of use cases versus all the use cases we solve for. For those, we still use the tech stack we built over the last seven-plus years.

Nataraj: One final question as we get close to time. We always talk about the advantages of offshoring. What are the advantages of onshoring — bringing manufacturing back to the US — that people don't talk about?

Sviat: First, I think we shouldn't be talking about bringing it back. It should just be about building the capacity here. Often, if the industry is growing and you're building more racks, you're not bringing them back — they'll still build them in China; you're creating your own capacity to build more because there's growth.

Second, we covered security — many people forget about this and just think about the cost of labor. Security matters. We covered time to market and quality, which matter a lot. It's not only cost — China versus the US. And we shouldn't forget that if you want to build capacity in the US today, by different assessments, around eight million people are missing to do that. So you have to use robots — but it doesn't mean you're not going to create new jobs. You'll still have humans in the loop; you start with robots because there isn't a lot of labor, but you create new jobs where you build those sites. The fact is you don't have enough people, so that ratio works out in a good way.

The last thing: it's not going to be reasonable for every device. Some devices are cheap and not strategic — do you need to build kettles, drills, or TVs here when it's cheaper elsewhere? But for AI infrastructure, aerospace, defense, and a few others, it makes a lot of sense.

Nataraj: Five years down the line, what other products do you think Bright Machines would be manufacturing?

Sviat: Any complex electronics that is complex enough for humans to make mistakes, expensive enough to make sense to build here, and critical enough for strategic reasons. It could be anything within the data center space, telecommunication equipment, or future AI devices. We all know the iPhone will evolve, so we'll see what's next — but it's going to take time.

Nataraj: I think that's a good note to end the conversation. Super excited about what you're doing, and I'm looking forward to what else you'll be manufacturing. Thanks for coming on the show.

Sviat: Of course, thanks for having me.