RAISE Summit '26 | Gilles Backhus, Tensordyne

Clips
More from RAISE Summit '26

Gilles Backhus

Co-Founder

Tensordyne

play_circle_outline Cross-continental co-design: Silicon Valley hardware, Germany software teams.

play_circle_outline Tensordyne exits stealth; chip in production, systems expected before Christmas

play_circle_outline Tensordyne exits stealth; first production chip and systems announced.

play_circle_outline Compact 13U 72‑Chip Compute System: Four Units per Rack, 30kW vs 150kW, Cable‑Free Vertical Networking, Juniper Partnership

play_circle_outline Compact, Energy-Efficient 72‑Chip Scale-Up System: 13U Any-to-Any Copper Networking with Juniper Integration

play_circle_outline Competitive positioning: faster, smaller, cheaper alternative to NVIDIA, Groq, Cerebras.

play_circle_outline NVIDIA-like Migration and Multi-Vendor Integration: A Smaller, Faster, More Energy-Efficient Alternative to Groq and Cerebras

play_circle_outline Target customers: NeoClouds, hyperscalers, large data center operators, edge and on-premises

Info

Gilles Backhus, Tensordyne

Gilles Backhus

Co-Founder Tensordyne

search

John Furrier

>> theCUBE, connecting, Silicon Valley and Wall Street. I'm John Furrier, co-host of theCUBE, here with Dave Vellante, my co-host. Welcome to the RAISE Summit coverage brought to you by Solidigm with support from Tensordyne, d-Matrix, and Argentum AI. Welcome back to theCUBE live coverage here in Paris, France. I'm John Furrier, host of theCUBE. The RAISE Summit, day one. We had our physical AI event yesterday, MACHINA Summit. Day 1 is this sidecar event there, Day 0, great event here. We got the innovators, the breakout stories, the leaders are here, global group gathering. We got the co-founder from Tensordyne here, he's co-founder, Gilles Backhus, great to have you. We had RK on in the studio recently coming out of stealth. Congratulations. Okay, you're out of stealth, you're the co-founder. Now, RK, your co-founder's in Silicon Valley, you're in Germany. You got the AI team, they got the silicon team. Together, powerhouse.

Gilles Backhus

>> Yep.

John Furrier

>> Is Tensordyne.

Gilles Backhus

>> That's right. Co-design, cross-continental co-design.

John Furrier

>> So talk about coming out of stealth. You must have been itching to get out. Explain what you guys do and how the teams are organized. We'll set the table.

Gilles Backhus

>> YeahSo we finally put our chip in production and we said, okay, this is the time for us to announce. And so we got the chip and the hardware development in Silicon Valley. And then we got the software development in Germany. And the software development consists of many different pillars. Like we had at the lowest level, we get our compiler development and our LogMath innovation. We've done a lot of innovation on math. And then on the math that the chip basically uses to do all the multiplications and additions. And then we got data center software like your inference server, like how to send tasks to the pod and then retrieve it back. And then the application layer. Yeah. So we've got all of that in Germany.

John Furrier

>> And the status of the chip is what?

Gilles Backhus

>> Status of the chip is it's now in production for the first batch. And later this year we'll get chips back from TSMC in Taiwan. And then, yeah, first systems will be up and running before Christmas.

John Furrier

>> one of the things I was impressed by was the size and form factor.

Gilles Backhus

>> Yeah.

John Furrier

>> Of the actual solution. Energy and power, people talk about this all the time. You got a high-performance inference machine. It's a box.

Gilles Backhus

>> Exactly.

John Furrier

>> It's like going back to the God box for AI. It does everything. Explain why, what was some of the first principle design criteria? What was the focus? What did you guys optimize for?

Gilles Backhus

>> Yeah, so we went straight to the box. Right? We didn't do PCI Express cards or OCP OAM modules. We went straight to compute trays with 9 chips on these trays and then 8 of those trays in a system. And then in our case, the, like NVIDIA systems, they have these networking trays. Our networking trays are actually vertical at the back. And so there's no cables. It's just PCBs plugging into other PCBs through connectors. Very reliable. All that came together because we realized, okay, we need to solve the networking side of things. Because we already had a chip before the ChatGPT moment. And then we realized, okay, you can't serve models on a single chip. It's going to be too slow. So you have to serve it on ideally as many chips as possible. And my co-founder, RK. He was one of the first at Juniper Networks. And so we said, okay, why don't we try to collaborate with them? And they agreed. They invested and agreed to develop that product together, especially collaborating on the chassis component side. So cooling and power, those things we also get from them. We don't have to develop those ourselves. And then the chassis, and then we basically take their networking cards out, we put our cards in. That's the concept. And that allowed us to immediately put 72 of our chips in one box, exactly the same amount of chips as an NVIDIA system, just a much smaller box, only 30 kilowatts instead of 150 kilowatts. And it's actually still

John Furrier

>> outperforming—Kilowatts.

Gilles Backhus

>> Kilowatts, yeah, 30 kilowatts.

John Furrier

>> That's nothing.

Gilles Backhus

>> Yeah, exactly. It's tiny.

John Furrier

>> And the size of the rack is what, a third?

Gilles Backhus

>> It's yay high, I think, 13 rack units. Yeah, that should be a correct measurement. And you can put 4 of those in a single rack. So you can have 288 chips from us in a single rack if you want to, and then you would be looking at 120 kilowatts.

John Furrier

>> Take me through that scale-up. If I wanted to scale up to a full rack, what does that entail? Copper, optical?

Gilles Backhus

>> Yeah, so in our case, thank God it's purely copper. So our scale-up has a degree of 72. Any-to-any connectivity, no torus or anything like that. And the way it works is, our compute trays, they have 6 connectors at the back and they slide into 6 vertical cards. And then every card has a bunch of switches. Right, and so in scale-up fabric networks, you basically need to make sure that every AI chip has at least one link to every switch chip. And if you can ensure that, then everybody can talk to everybody.

John Furrier

>> Gilles, I want to get your thoughts on this, 'cause RK and I were riffing on theCUBE in Palo Alto about this, and I asked him about the Juniper experience. And if you remember at that time, oh, I do, the internet was here, so Cisco connected Companies, one data center to the other, campus basically networking. Packets moved from point A to point B. There really wasn't a lot of performance overhead, it was just get it there. And then the internet hit with ISPs. Juniper purpose-built the box that was built for the big data centers like Exodus. These were companies that were serving the internet at large scale at that time. They ran the table. So, okay, Juniper, large-scale players like the Exoduses of the world, must be considered like a gigawatt factory today. Then you got basic people using Cisco. Cisco was the standard, but Juniper came out of nowhere as number 2 and came to number 2 instantly because they focused on performance. So, okay, let's translate that metaphor again today. There's an inference demand.

Gilles Backhus

>> Yeah.

John Furrier

>> And it's not saying give me the best cost and slower speeds.

Gilles Backhus

>> Yeah.

John Furrier

>> The Exoduses, the Digexes, the MCIs that were the big providers of the internet days for companies are now the AI factories.

Gilles Backhus

>> Yeah.

John Furrier

>> So the question is, is it the same parallel performance need? Is that the key? Because no one wants to buy slower inference. That's like dial-up.

John Furrier

>> Yeah. It's even now you can see even an OpenAI versus Claude, some are slow and waiting, thinking. Even a second, I'm pissed.

Gilles Backhus

>> Yeah.

John Furrier

>> So, okay, I'm it takes minutes. I'm pissed.

Gilles Backhus

>> Absolutely. You're not the only one.

John Furrier

>> So I think there's a real parallel to an opportunity is speed will win.

Gilles Backhus

>> Yeah, absolutely. You can see it on those platforms like artificialanalysis.ai that are measuring cost versus intelligence or cost versus speed. And you can see there's a clear linear correlation. The faster you can serve a token, the faster the token is, the more people are willing to pay for that one token. And then a huge thing that I think Juniper also did right back then. I was alive back then, but I was very young. But I think what I understood though is They didn't try to change the rules of the game. They innovated within the rules of the game, but they delivered exactly what the customers wanted. Which was faster, cheaper, and whatnot. And we're trying to do the same.

John Furrier

>> And that was market forces, by the way. It wasn't that Cisco didn't have a good product. It's just they were selling like hotcakes to companies, but not the specific use case of high speed.

Gilles Backhus

>> Yeah, yeah, so where we're quite different than, for example, Cerebras and Groq, who also serve tokens very fast, is we still try to deliver a system where it's very easy to migrate to. Because otherwise, if you have to rethink your entire inference deployment, people are going to have that kind of hurdle to get over in terms of migration and integrating your solution as part of a multi-vendor data center. And so if you look at our spec sheet, you'll see we have 72 chips in a scale-up. NVIDIA has 72 chips in a scale-up. We have HBM, they have HBM. So it's very similar from that first glance. It's just that the box is smaller, consumes less power, and is faster.

John Furrier

>> basically the NVL72, that's monster rack.

Gilles Backhus

>> Yeah, yeah.

John Furrier

>> It's got multiple NICs in there to manage the copper. You guys have a third of the size.

Gilles Backhus

>> Yeah, or a quarter of the size actually. Yeah, exactly. And then otherwise within that, we try to innovate as much as possible, but always making sure that the user still stays happy. One perfect example for that is also our innovation on the math side. Our logarithmic math. It's completely under the hood. From a user point of view, from an SDK point of view, you don't even notice it. It just looks like normal floating point math. It's just that the engine under the hood is more efficient.

John Furrier

>> What's the LogMath do?

Gilles Backhus

>> So LogMath is essentially a concept to eliminate multiplier circuits, which are really expensive, much more expensive than adder circuits in silicon. So our silicon, if you zoomed in, you would see almost no multiplier circuits, but we're still doing multiplications. And basically, in LogMath, instead of representing numbers as an exponent and a mantissa like in floating point, you now only have an exponent, but it's a fractional exponent. So 2 to the power of, I don't know, 13.58, for example. And then if you want to multiply two numbers, suddenly what you do is you only add the two exponents of the two numbers. You don't add the two numbers. That's the basic idea. Now you could ask, why isn't everybody doing this? Because the next part is really hard to solve, which is now you have to go back to a linear representation to now do the addition, which also needs to happen. And that transformation from a logarithmic representation to a normal one, that is normally so expensive, but we found a way to do this extremely accurately and still very cheap.

John Furrier

>> So from an operational standpoint, you mentioned Groq, what are some of the benefits? Just lay them out.

Gilles Backhus

>> Yeah. So Groq is pretty memory limited. I think in their newest generation they have 128 gigabytes per rack. If you assume 4-bit weights, let's say that's the state of the art, you're going to have 250 billion weights per rack. We can have 20 trillion, actually we can have 80 trillion parameters per rack. 80 trillion versus a quarter trillion. So, there's a huge difference. So, they need 10 racks. To run a single state-of-the-art model, we can do it in a quarter rack.

John Furrier

>> Okay. You're the co-founder. I have to ask the question. I'm an investor or I'm a pundit and I say, well, guys, the chip won't be out until the end of the year. Well, NVIDIA might come out with something better. They got this. I'm not too sure. What's your answer to that?

Gilles Backhus

>> look at the news. Look at what SemiAnalysis and others are writing about NVIDIA. They're having huge problems right now scaling their systems.

John Furrier

>> Right.

Gilles Backhus

>> Rubin Ultra will be massively delayed, even Rubin delayed. So I would say, take a bet on us. Just reserve an early beta system, evaluate the heck out of it. If you're happy, place an order. You will just win. If the system works, you will for sure get more revenue out of every rack in your data center, and you'll save tons of

John Furrier

>> power.How do they reserve that beta?

Gilles Backhus

>> So they get in contact with me or our VP of Sales, and, it's an easy process from there on out.

John Furrier

>> Tell them theCUBE sent you, get a 10% discount on the beta. It's

Gilles Backhus

>> free. ExactlyThe code is CUBE2026.

John Furrier

>> Alright, so this has been a big thing. Also scale across on data centers has been, scale out has been one. Scale across, these are concepts that are kicked around. Scale up I think is a huge opportunity on scale up. Even NVIDIA, they got a lot of NICs in there. They could probably squeeze more GPUs. People want more systems, so that's going to be on the large side. The second thing I think that's important, I want to get your thoughts on is, a lot of action at the edge, where you can have 5, megawatts and you've got to put some racks in there. So you fit the threshold on pretty much all edge configurations.

Gilles Backhus

>> Yeah. Yeah. So the great thing about us is in a single rack, and I can explain later why technologically that is. But in a single rack you can have basically a state-of-the-art data center. Obviously the total throughput is not going to be super crazy. But at least we can serve the largest models at the highest speeds in a single rack. You don't need multiple racks. So your threshold, your entry threshold now to be part of the most premium tier of serving tokens is now a single rack. Right. And obviously the CapEx for that is also not going to be crazy. So yeah, if you have a single rack and let's say 200 kilowatts, you're in. With Tensordyne, you can now be someone who serves tokens with the largest models at the highest speeds, like 1,000 tokens per second per user. Yeah. And then if you have more racks, you can just scale it up. But yeah, we're a big fan of the whole idea of modular data centers at the edge. Like smaller scale data centers, very integratable. We're air-cooled also. So even if you have an old small data center in the middle of a city and it doesn't have liquid cooling infrastructure, we can still be integrated in that.

John Furrier

>> And also the on-prem is a home run too.

Gilles Backhus

>> Yeah, absolutely right. Yeah, we're open to all of them.

John Furrier

>> All right, so since you came out of stealth, not a lot of people know what's going on. Put a plug in, why Tensordyne? Who should be buying it? Explain the target audience and why they should buy it.

Gilles Backhus

>> Yeah, so currently what we see practically is that most of our customers are NeoClouds. But we have a couple of things with hyperscalers. But anybody who buys hardware for a data center can become a customer. You will just— we really try to be super religious about being as similar to an NVIDIA experience as possible. I think that really differentiates us from others so that migration is super easy from day one. We don't want people to first relearn how to deploy on our system for months and then only they can do things. It's going to work from day one. That's the goal.

John Furrier

>> Yeah.

Gilles Backhus

>> So anybody—

John Furrier

>> if I'm an IT estate, I'm running a big estate. I got a lot of x86, I got some GPUs, I got AMD, Intel, NVIDIA. What's in it for me? I just drop it in?

Gilles Backhus

>> Yeah, you drop it in. Yeah, you set it up, the setup time will not be crazy. You'll enjoy a very reliable uptime because we're using components that are used for the backbone of the internet. Yeah, and that's it.

John Furrier

>> All right, what are your goals? Obviously get the chip out, get customers, people looking for the beta, the beta program?

Gilles Backhus

>> Yeah, beta program, yeah.

John Furrier

>> How big is that going to be? Give some stats on that early engagement and how do people engage? Obviously they can DM you.

Gilles Backhus

>> Yeah, so the beta program is gonna be around the first few thousands of chips. So if you divide that by 72, you get the number of systems, let's say 20 systems or so. So it's early batch. still a couple thousands of chips.

John Furrier

>> Who's the ideal person to get in that beta?

Gilles Backhus

>> I would say NeoClouds or hyperscalers, just large data center operators who are, looking for an alternative to NVIDIA and AMD. Yeah, so if they have a little bit of budget for that, a couple engineers who have some free time for that. it's a short process. Yeah.

John Furrier

>> I want to get your thoughts on something we've been covering a lot on SiliconANGLE and theCUBE. Disaggregated.

Gilles Backhus

>> Yeah.

John Furrier

>> Access.

Gilles Backhus

>> Yeah.

John Furrier

>> You've seen that a lot on distributed computing environment. You fit there nicely.

Gilles Backhus

>> So that's the thing. Right now you see NVIDIA Rubin collaborating with Groq. Or you see AWS Trainium3 with Cerebras. With our system, you don't have to go— we can do disaggregation, but the way it works is that one pod from us will be for pre-fill for input processing, and then another pod again from us will be for decode processing. That's our approach to disaggregation. Half a rack, you can run the largest models at the highest speeds in there. So you don't need these behemoths anymore, which consists of I don't know, 14 racks or so. You just have one rack and then you can just scale it up.

John Furrier

>> Upside is better footprint management.

Gilles Backhus

>> Absolutely.

John Furrier

>> You get all the performance.

Gilles Backhus

>> Yeah, absolutely. And you only manage a single vendor, not two.

John Furrier

>> That's a big claim. Yeah. I love it. they did it in networking. And again, the market's very similar. The large scale, at that time internet wasn't that large scale, but at that time that's where the centralized service was.

Gilles Backhus

>> And it really almost breaks down to one key innovation that we have, which is, if I can just explain it for a second. So the reason why NVIDIA Rubin, which if you look at their data sheet, it's actually a pretty amazing rack. But what they're lacking is the low latency in chip-to-chip communication. And this is their bottleneck. They can't go to 72 chips and run a giant model over 72 chips because you're going to have lots of tiny messages that the chips exchange all the time and you're going to be bottlenecked by the latency, that it always takes, like the 10 microseconds that it always takes. And you're going to have hundreds of these communications for every little token that you serve. And in our case, our latency is 10 times lower. It's around 1 microsecond. And so we're not bottlenecked. And this is why we don't need a Groq to basically extend the per-user speed more to the higher end. Yeah.

John Furrier

>> And also they got a lot of hops on their NICs too, and they lose energy.

Gilles Backhus

>> Yeah, we have a single hop for everything. For up to 72 chips, single hop.

John Furrier

>> Yeah. All right. Tensordyne again, innovating on the hardware and the software, smaller form factor, smaller, faster, cheaper. That is the way the world worked in the Moore's Law generation. We're starting to see signs of that happening in the AI infrastructure. Again, it's such a hot area, seeing innovation and optimization for footprint, energy, capital, all that's going on here at the RAISE Summit. Of course, we're doing our part on theCUBE. I'm John Furrier, thanks for watching.

Gilles Backhus

>> Thanks, John.

Gilles Backhus, Tensordyne

search

John Furrier

>> theCUBE, connecting, Silicon Valley and Wall Street. I'm John Furrier, co-host of theCUBE, here with Dave Vellante, my co-host. Welcome to the RAISE Summit coverage brought to you by Solidigm with support from Tensordyne, d-Matrix, and Argentum AI. Welcome back to theCUBE live coverage here in Paris, France. I'm John Furrier, host of theCUBE. The RAISE Summit, day one. We had our physical AI event yesterday, MACHINA Summit. Day 1 is this sidecar event there, Day 0, great event here. We got the innovators, the breakout stories, the leaders are here, global group gathering. We got the co-founder from Tensordyne here, he's co-founder, Gilles Backhus, great to have you. We had RK on in the studio recently coming out of stealth. Congratulations. Okay, you're out of stealth, you're the co-founder. Now, RK, your co-founder's in Silicon Valley, you're in Germany. You got the AI team, they got the silicon team. Together, powerhouse.

Gilles Backhus

>> Yep.

John Furrier

>> Is Tensordyne.

Gilles Backhus

>> That's right. Co-design, cross-continental co-design.

John Furrier

>> So talk about coming out of stealth. You must have been itching to get out. Explain what you guys do and how the teams are organized. We'll set the table.

Gilles Backhus

>> YeahSo we finally put our chip in production and we said, okay, this is the time for us to announce. And so we got the chip and the hardware development in Silicon Valley. And then we got the software development in Germany. And the software development consists of many different pillars. Like we had at the lowest level, we get our compiler development and our LogMath innovation. We've done a lot of innovation on math. And then on the math that the chip basically uses to do all the multiplications and additions. And then we got data center software like your inference server, like how to send tasks to the pod and then retrieve it back. And then the application layer. Yeah. So we've got all of that in Germany.

John Furrier

>> And the status of the chip is what?

Gilles Backhus

>> Status of the chip is it's now in production for the first batch. And later this year we'll get chips back from TSMC in Taiwan. And then, yeah, first systems will be up and running before Christmas.

John Furrier

>> one of the things I was impressed by was the size and form factor.

Gilles Backhus

>> Yeah.

John Furrier

>> Of the actual solution. Energy and power, people talk about this all the time. You got a high-performance inference machine. It's a box.

Gilles Backhus

>> Exactly.

John Furrier

>> It's like going back to the God box for AI. It does everything. Explain why, what was some of the first principle design criteria? What was the focus? What did you guys optimize for?

Gilles Backhus

>> Yeah, so we went straight to the box. Right? We didn't do PCI Express cards or OCP OAM modules. We went straight to compute trays with 9 chips on these trays and then 8 of those trays in a system. And then in our case, the, like NVIDIA systems, they have these networking trays. Our networking trays are actually vertical at the back. And so there's no cables. It's just PCBs plugging into other PCBs through connectors. Very reliable. All that came together because we realized, okay, we need to solve the networking side of things. Because we already had a chip before the ChatGPT moment. And then we realized, okay, you can't serve models on a single chip. It's going to be too slow. So you have to serve it on ideally as many chips as possible. And my co-founder, RK. He was one of the first at Juniper Networks. And so we said, okay, why don't we try to collaborate with them? And they agreed. They invested and agreed to develop that product together, especially collaborating on the chassis component side. So cooling and power, those things we also get from them. We don't have to develop those ourselves. And then the chassis, and then we basically take their networking cards out, we put our cards in. That's the concept. And that allowed us to immediately put 72 of our chips in one box, exactly the same amount of chips as an NVIDIA system, just a much smaller box, only 30 kilowatts instead of 150 kilowatts. And it's actually still

John Furrier

>> outperforming—Kilowatts.

Gilles Backhus

>> Kilowatts, yeah, 30 kilowatts.

John Furrier

>> That's nothing.

Gilles Backhus

>> Yeah, exactly. It's tiny.

John Furrier

>> And the size of the rack is what, a third?

Gilles Backhus

>> It's yay high, I think, 13 rack units. Yeah, that should be a correct measurement. And you can put 4 of those in a single rack. So you can have 288 chips from us in a single rack if you want to, and then you would be looking at 120 kilowatts.

John Furrier

>> Take me through that scale-up. If I wanted to scale up to a full rack, what does that entail? Copper, optical?

Gilles Backhus

>> Yeah, so in our case, thank God it's purely copper. So our scale-up has a degree of 72. Any-to-any connectivity, no torus or anything like that. And the way it works is, our compute trays, they have 6 connectors at the back and they slide into 6 vertical cards. And then every card has a bunch of switches. Right, and so in scale-up fabric networks, you basically need to make sure that every AI chip has at least one link to every switch chip. And if you can ensure that, then everybody can talk to everybody.

John Furrier

>> Gilles, I want to get your thoughts on this, 'cause RK and I were riffing on theCUBE in Palo Alto about this, and I asked him about the Juniper experience. And if you remember at that time, oh, I do, the internet was here, so Cisco connected Companies, one data center to the other, campus basically networking. Packets moved from point A to point B. There really wasn't a lot of performance overhead, it was just get it there. And then the internet hit with ISPs. Juniper purpose-built the box that was built for the big data centers like Exodus. These were companies that were serving the internet at large scale at that time. They ran the table. So, okay, Juniper, large-scale players like the Exoduses of the world, must be considered like a gigawatt factory today. Then you got basic people using Cisco. Cisco was the standard, but Juniper came out of nowhere as number 2 and came to number 2 instantly because they focused on performance. So, okay, let's translate that metaphor again today. There's an inference demand.

Gilles Backhus

>> Yeah.

John Furrier

>> And it's not saying give me the best cost and slower speeds.

Gilles Backhus

>> Yeah.

John Furrier

>> The Exoduses, the Digexes, the MCIs that were the big providers of the internet days for companies are now the AI factories.

Gilles Backhus

>> Yeah.

John Furrier

>> So the question is, is it the same parallel performance need? Is that the key? Because no one wants to buy slower inference. That's like dial-up.

John Furrier

>> Yeah. It's even now you can see even an OpenAI versus Claude, some are slow and waiting, thinking. Even a second, I'm pissed.

Gilles Backhus

>> Yeah.

John Furrier

>> So, okay, I'm it takes minutes. I'm pissed.

Gilles Backhus

>> Absolutely. You're not the only one.

John Furrier

>> So I think there's a real parallel to an opportunity is speed will win.

Gilles Backhus

>> Yeah, absolutely. You can see it on those platforms like artificialanalysis.ai that are measuring cost versus intelligence or cost versus speed. And you can see there's a clear linear correlation. The faster you can serve a token, the faster the token is, the more people are willing to pay for that one token. And then a huge thing that I think Juniper also did right back then. I was alive back then, but I was very young. But I think what I understood though is They didn't try to change the rules of the game. They innovated within the rules of the game, but they delivered exactly what the customers wanted. Which was faster, cheaper, and whatnot. And we're trying to do the same.

John Furrier

>> And that was market forces, by the way. It wasn't that Cisco didn't have a good product. It's just they were selling like hotcakes to companies, but not the specific use case of high speed.

Gilles Backhus

>> Yeah, yeah, so where we're quite different than, for example, Cerebras and Groq, who also serve tokens very fast, is we still try to deliver a system where it's very easy to migrate to. Because otherwise, if you have to rethink your entire inference deployment, people are going to have that kind of hurdle to get over in terms of migration and integrating your solution as part of a multi-vendor data center. And so if you look at our spec sheet, you'll see we have 72 chips in a scale-up. NVIDIA has 72 chips in a scale-up. We have HBM, they have HBM. So it's very similar from that first glance. It's just that the box is smaller, consumes less power, and is faster.

John Furrier

>> basically the NVL72, that's monster rack.

Gilles Backhus

>> Yeah, yeah.

John Furrier

>> It's got multiple NICs in there to manage the copper. You guys have a third of the size.

Gilles Backhus

>> Yeah, or a quarter of the size actually. Yeah, exactly. And then otherwise within that, we try to innovate as much as possible, but always making sure that the user still stays happy. One perfect example for that is also our innovation on the math side. Our logarithmic math. It's completely under the hood. From a user point of view, from an SDK point of view, you don't even notice it. It just looks like normal floating point math. It's just that the engine under the hood is more efficient.

John Furrier

>> What's the LogMath do?

Gilles Backhus

>> So LogMath is essentially a concept to eliminate multiplier circuits, which are really expensive, much more expensive than adder circuits in silicon. So our silicon, if you zoomed in, you would see almost no multiplier circuits, but we're still doing multiplications. And basically, in LogMath, instead of representing numbers as an exponent and a mantissa like in floating point, you now only have an exponent, but it's a fractional exponent. So 2 to the power of, I don't know, 13.58, for example. And then if you want to multiply two numbers, suddenly what you do is you only add the two exponents of the two numbers. You don't add the two numbers. That's the basic idea. Now you could ask, why isn't everybody doing this? Because the next part is really hard to solve, which is now you have to go back to a linear representation to now do the addition, which also needs to happen. And that transformation from a logarithmic representation to a normal one, that is normally so expensive, but we found a way to do this extremely accurately and still very cheap.

John Furrier

>> So from an operational standpoint, you mentioned Groq, what are some of the benefits? Just lay them out.

Gilles Backhus

>> Yeah. So Groq is pretty memory limited. I think in their newest generation they have 128 gigabytes per rack. If you assume 4-bit weights, let's say that's the state of the art, you're going to have 250 billion weights per rack. We can have 20 trillion, actually we can have 80 trillion parameters per rack. 80 trillion versus a quarter trillion. So, there's a huge difference. So, they need 10 racks. To run a single state-of-the-art model, we can do it in a quarter rack.

John Furrier

>> Okay. You're the co-founder. I have to ask the question. I'm an investor or I'm a pundit and I say, well, guys, the chip won't be out until the end of the year. Well, NVIDIA might come out with something better. They got this. I'm not too sure. What's your answer to that?

Gilles Backhus

>> look at the news. Look at what SemiAnalysis and others are writing about NVIDIA. They're having huge problems right now scaling their systems.

John Furrier

>> Right.

Gilles Backhus

>> Rubin Ultra will be massively delayed, even Rubin delayed. So I would say, take a bet on us. Just reserve an early beta system, evaluate the heck out of it. If you're happy, place an order. You will just win. If the system works, you will for sure get more revenue out of every rack in your data center, and you'll save tons of

John Furrier

>> power.How do they reserve that beta?

Gilles Backhus

>> So they get in contact with me or our VP of Sales, and, it's an easy process from there on out.

John Furrier

>> Tell them theCUBE sent you, get a 10% discount on the beta. It's

Gilles Backhus

>> free. ExactlyThe code is CUBE2026.

John Furrier

>> Alright, so this has been a big thing. Also scale across on data centers has been, scale out has been one. Scale across, these are concepts that are kicked around. Scale up I think is a huge opportunity on scale up. Even NVIDIA, they got a lot of NICs in there. They could probably squeeze more GPUs. People want more systems, so that's going to be on the large side. The second thing I think that's important, I want to get your thoughts on is, a lot of action at the edge, where you can have 5, megawatts and you've got to put some racks in there. So you fit the threshold on pretty much all edge configurations.

Gilles Backhus

>> Yeah. Yeah. So the great thing about us is in a single rack, and I can explain later why technologically that is. But in a single rack you can have basically a state-of-the-art data center. Obviously the total throughput is not going to be super crazy. But at least we can serve the largest models at the highest speeds in a single rack. You don't need multiple racks. So your threshold, your entry threshold now to be part of the most premium tier of serving tokens is now a single rack. Right. And obviously the CapEx for that is also not going to be crazy. So yeah, if you have a single rack and let's say 200 kilowatts, you're in. With Tensordyne, you can now be someone who serves tokens with the largest models at the highest speeds, like 1,000 tokens per second per user. Yeah. And then if you have more racks, you can just scale it up. But yeah, we're a big fan of the whole idea of modular data centers at the edge. Like smaller scale data centers, very integratable. We're air-cooled also. So even if you have an old small data center in the middle of a city and it doesn't have liquid cooling infrastructure, we can still be integrated in that.

John Furrier

>> And also the on-prem is a home run too.

Gilles Backhus

>> Yeah, absolutely right. Yeah, we're open to all of them.

John Furrier

>> All right, so since you came out of stealth, not a lot of people know what's going on. Put a plug in, why Tensordyne? Who should be buying it? Explain the target audience and why they should buy it.

Gilles Backhus

>> Yeah, so currently what we see practically is that most of our customers are NeoClouds. But we have a couple of things with hyperscalers. But anybody who buys hardware for a data center can become a customer. You will just— we really try to be super religious about being as similar to an NVIDIA experience as possible. I think that really differentiates us from others so that migration is super easy from day one. We don't want people to first relearn how to deploy on our system for months and then only they can do things. It's going to work from day one. That's the goal.

John Furrier

>> Yeah.

Gilles Backhus

>> So anybody—

John Furrier

>> if I'm an IT estate, I'm running a big estate. I got a lot of x86, I got some GPUs, I got AMD, Intel, NVIDIA. What's in it for me? I just drop it in?

Gilles Backhus

>> Yeah, you drop it in. Yeah, you set it up, the setup time will not be crazy. You'll enjoy a very reliable uptime because we're using components that are used for the backbone of the internet. Yeah, and that's it.

John Furrier

>> All right, what are your goals? Obviously get the chip out, get customers, people looking for the beta, the beta program?

Gilles Backhus

>> Yeah, beta program, yeah.

John Furrier

>> How big is that going to be? Give some stats on that early engagement and how do people engage? Obviously they can DM you.

Gilles Backhus

>> Yeah, so the beta program is gonna be around the first few thousands of chips. So if you divide that by 72, you get the number of systems, let's say 20 systems or so. So it's early batch. still a couple thousands of chips.

John Furrier

>> Who's the ideal person to get in that beta?

Gilles Backhus

>> I would say NeoClouds or hyperscalers, just large data center operators who are, looking for an alternative to NVIDIA and AMD. Yeah, so if they have a little bit of budget for that, a couple engineers who have some free time for that. it's a short process. Yeah.

John Furrier

>> I want to get your thoughts on something we've been covering a lot on SiliconANGLE and theCUBE. Disaggregated.

Gilles Backhus

>> Yeah.

John Furrier

>> Access.

Gilles Backhus

>> Yeah.

John Furrier

>> You've seen that a lot on distributed computing environment. You fit there nicely.

Gilles Backhus

>> So that's the thing. Right now you see NVIDIA Rubin collaborating with Groq. Or you see AWS Trainium3 with Cerebras. With our system, you don't have to go— we can do disaggregation, but the way it works is that one pod from us will be for pre-fill for input processing, and then another pod again from us will be for decode processing. That's our approach to disaggregation. Half a rack, you can run the largest models at the highest speeds in there. So you don't need these behemoths anymore, which consists of I don't know, 14 racks or so. You just have one rack and then you can just scale it up.

John Furrier

>> Upside is better footprint management.

Gilles Backhus

>> Absolutely.

John Furrier

>> You get all the performance.

Gilles Backhus

>> Yeah, absolutely. And you only manage a single vendor, not two.

John Furrier

>> That's a big claim. Yeah. I love it. they did it in networking. And again, the market's very similar. The large scale, at that time internet wasn't that large scale, but at that time that's where the centralized service was.

Gilles Backhus

>> And it really almost breaks down to one key innovation that we have, which is, if I can just explain it for a second. So the reason why NVIDIA Rubin, which if you look at their data sheet, it's actually a pretty amazing rack. But what they're lacking is the low latency in chip-to-chip communication. And this is their bottleneck. They can't go to 72 chips and run a giant model over 72 chips because you're going to have lots of tiny messages that the chips exchange all the time and you're going to be bottlenecked by the latency, that it always takes, like the 10 microseconds that it always takes. And you're going to have hundreds of these communications for every little token that you serve. And in our case, our latency is 10 times lower. It's around 1 microsecond. And so we're not bottlenecked. And this is why we don't need a Groq to basically extend the per-user speed more to the higher end. Yeah.

John Furrier

>> And also they got a lot of hops on their NICs too, and they lose energy.

Gilles Backhus

>> Yeah, we have a single hop for everything. For up to 72 chips, single hop.

John Furrier

>> Yeah. All right. Tensordyne again, innovating on the hardware and the software, smaller form factor, smaller, faster, cheaper. That is the way the world worked in the Moore's Law generation. We're starting to see signs of that happening in the AI infrastructure. Again, it's such a hot area, seeing innovation and optimization for footprint, energy, capital, all that's going on here at the RAISE Summit. Of course, we're doing our part on theCUBE. I'm John Furrier, thanks for watching.

Gilles Backhus

>> Thanks, John.