And just quickly going over this slide again for the sake of the people that are online. I was saying, ABC doesn't know how to do math when you have 164 million in the bank and the whole fiscal year '26 we spent less than 20 million. I don't think it's so difficult to see we're not really going bankrupt anytime soon. You can see the receivables that we had this quarter, which, as I always say, they're not the real indication — they depend on milestones. Nevertheless, it's nice. You really should be looking at the revenue, which we didn't report this quarter, but that's recognized over time, and that's the real indication of where we're heading.
So, looking at where we stand, there's really so much activity going on right now in the commercialization space. We're working very hard with DB HiTek now that we're qualified. They are presenting us to their customers, and we're pushing forward to bring more customers on board. As we talked about in the past, for a foundry, unlike an IDM, a foundry normally their product customers are more risk, or cautious, let's say. They want to see that you get to qualification before they really engage. But now we're qualified and we're pushing forward. We have a few customers that already engaged in the past. As you know, one of them even already made it public, Overlord. They made it public. We'll talk about them, I think, in a couple of slides. But they're really moving forward towards mass production now.
On Semi moving forward nicely. You guys know how to do the math. Again, we've been working with them for almost two years. This is the time when they're already working on qualification. On Semi, as an IDM, and TI is similar — the products that they manufacture are their own. You can imagine they won't bring a technology in if their products don't want it. So, in the case of an IDM it actually goes in the other direction. The product teams are the ones who actually decide we want the technology, and then they bring it into the fab. So, with them, when we started working, products, you can imagine, have already started moving forward. Their goal is to qualify in the next several months. I won't go — again, it's their internal milestone — but obviously they too want to reach mass production in 2027. They want to push forward as fast as they can. So, you're seeing now that Overlord and others are pushing forward in DB Hi-Tech. On Semi is pushing forward.
2027 is going to be really a very, very exciting year for us, as the most important milestone of them all, going to mass production. I really, really believe and hope that we'll be making it. TI progressing very nicely. Last quarter we mentioned we can't specify which customers, but customers with an S, and we don't have that many, have been expanding their agreements. I want to emphasize: customers that are already our customers, they are diving very deep into the technology with our R&D team. They're doing technology transfer. They're learning the details. They're working with the R&D team. They know not just the sales guys or management, they know the team. When they decide to expand their agreement and bring in more business units and ramp up on more projects, for me that's a huge vote of confidence. And I think that's really what you can see here: things are progressing. We might have not signed new agreements, but in big companies when you bring in a new business unit, it's like a new customer. So, things are progressing and are expanding and are moving forward. We're pushing as much as we can on all of these fronts, and very excited with each and every one of these fab customers.
Now, I talked about expanding the agreement, and again, from the moment they signed the license agreement, we did the technology transfer, they dove deep into it, built confidence. So, again, the fact that they expanded their agreements, I'm really excited about that. To me, it gives me — I always had confidence in our technology, don't misunderstand it — but now we have the proof.
Talking about the product customers, and that's really another fun stuff that we finally can talk about product customers and moving forward. So, that's moving to the next step. It's not just fab customers anymore. Overlord are really pushing strongly forward. It is a startup, and people can say, "Oh, that's a small company, whatever." But it's a startup like many startups of industry veterans, well-known people, respected. They know what they're doing, and they are just pushing forward really strongly with an intelligent power management solution. Today power management is really a very, very critical element in all the different systems. Everywhere you look, people are talking about what's the power consumption. How do you lower the power consumption? How do you control it? Well, that's the power management that you need. When you have a cell phone: A, you want not to have to charge it more than once a night, right? B, you want the battery to live longer — when these batteries are worn out, you have to throw away the whole phone. You can't replace just the battery, right? So, you want the batteries to last for a longer period. All of these things relate to power management, and that's really important. So, these guys are doing something very smart. They have some AI in it. We can't talk about who their customers are, but they have some company customers that I'm sure you guys would recognize.
And we are doing everything we can to enable them to get to mass production. At the end of the day, it's the customer's decision when to go officially to mass production. But we want to make sure that there's nothing about the ReRAM that causes any delays there. So, they're very bullish, and I think it's going to be really exciting when they get to mass production. Obviously, once they're in mass production — and in the beginning, by the way, they're already giving some initial samples to customers — and as they start ramping up, they'll be giving samples, and then their customers need to go to production in order for things to really ramp up. But we know that they have customers already with sockets waiting for these chips and everything. So, that ramp up will be good, eventually leading to the royalties that everyone's waiting for.
Beyond Overlord, we have other customers or products that are being designed with ReRAM in them. Now the number of those projects is growing, and it's much more than just two customers here. We're seeing more and more products that are having ReRAM designed into them. We have another customer that already has silicon in their hands, that are running software, that are ramping up. So, this push to get to mass production is moving forward, and 2027 is going to be a very exciting year. We're really looking forward to it.
So, I talked about Overlord. I won't spend a lot of time here, but really great to see such a great team of engineers working, and they're working around the clock now to push this to mass production. Really good cooperation here.
Beyond that, I believe some of you guys might have noticed that in the quarterly we also started talking about the AI more. We had a target for calendar year '26 to get a first AI customer. This quarter we had two of them sign up. I want to be clear: we're talking about customers that are paying to be able to evaluate our AI technology. So, the AI technology, and you can see here on the bottom, we have a US-based company that took our technology and did an analysis. And these are the results that they reported: 80x lower power consumption relative to an Nvidia GPU. We're talking about 1.25% of the power consumption. You hear people talking about power consumption in data centers and all these other places. This is dramatic. And at the same time having a 100x lower inference latency, basically meaning they're running much, much faster than a GPU. This is a great potential, and I'm emphasizing it shows you where we can go. We still have some technical challenges in order to make this commercially viable, to really get this to a commercial solution. We're working hard on that. Again, when we did the raise, we said part of the funds are going to be used to making this technology happen, but it's really good to see that this is the potential that we have. We have a very good team. I believe the team will overcome these technical challenges and will be able to eventually get this to a commercial level.
So, moving forward, we talked about it already in August, but again, people are always asking me, "So, what's going on? You have such a great technology. Everyone's admitting it's a great technology. How come you're not signing more deals?" It's very, very important to understand: for a fab to sign with Weebit, it is a huge strategic decision. It is a major decision. They're basically making a commitment, and I said it and I'll repeat it, literally for decades. They are telling us — the On Semi VP was saying, "Of course you're going to be here in 20 years. When we ramp you up, that's it. We're stuck." Kind of. He didn't use the term stuck, but basically that's kind of what's happening.
You go in, and especially our IP goes into their process. Now, Flash, by the way, has been around for 50 years, probably. You can see how long it's taken to the point where people decide to actually look for a different technology and bring up a different technology. So, this is a major decision for them. When they are swamped — the fabs, when we talk to them, they have 50-100% more demand than the capacity they can actually deliver. They are being hammered by their customers who are really mad because they're not getting the chips that they ordered. And you can imagine what's going on there now. It's a hell house. They're going crazy trying to keep their customers satisfied. And when you're in such a crazy crisis mode, it's very hard to focus on bringing a new technology and ramping up. So, I repeat what I said then, but it's very important to understand: there's not a single fab that told us, "ReRAM isn't interesting. We're not interested in having ReRAM in the future." They all understand they need ReRAM. It's a matter of, just let us get over this craziness, let us capitalize over this opportunity. We're raising prices like hell and still people are ordering, so they're just busy, but it's going to happen. These deals will happen. We are, by the way, making progress. It's not that we're not making progress. We are making progress with some of them, and even some good progress.
The deals — yes, I am disappointed myself. I wanted to have a deal signed by now. I believed we would. If you would have asked me in January, I would have told you I believe we'd have a deal signed by September. But they're going to happen. This is life. Good news is we have the cash in the bank. We're not going bankrupt even if these deals happen in a few months, in half a year, we're still going to be in a very good position. So, that's kind of where the industry is and what we're doing.
Yeah, I repeat because not everyone understands the timelines that we go through. In the AGM of '23, we had the timeline shown in a different way, but it's important to understand: the top line here that says foundry/IDM, basically a fab. A fab, when we engage with them, technology transfer is about 12 months. You've already seen it with DB HiTek, you've seen it with On Semi. Then they have roughly another year of test, qualification, all of that. Again, orders of magnitude. By the way, after they're qualified, things continue. A lot of work continues because you constantly want to improve the manufacturing process. You need to understand, in semiconductor manufacturing, you never have a situation where 100% of the chips manufactured are operational. There's always defects. In the beginning, the yield — we call it the yield — can be 80, 85, 90%, depending. You want to get as close as you can to 100%, obviously. And there's constant work, analysis. When you get to mass production and you have a lot of chips manufactured, you can actually do analysis on big data and start understanding where you can tweak here, tweak there, improve the yield. So, the work is going to continue all the time to make improvements there.
But, once you're qualified in a foundry, and that's the second line, when you're qualified in a foundry, then the product customers are finally open to, okay, let's start doing our design, let's engage, let's move forward. Now, Overlord, they were risk-takers. It's a young company. They want to take advantage of being first in market. So, they engaged actually earlier, but most of the product companies, they waited to see the qualification, and now it's starting to move. We're starting to see the funnel, the pipeline getting filled. They need to do — again, it really depends on the complexity of the chip. They need to do the design. Once it's designed, they need to do testing and fixing, and for simple designs it can be six months. For more advanced designs it can be 12 or even 18. So, orders of magnitude. And then, finally, they get to mass production, and then we start seeing royalties, which is the holy grail, right? That's what you really want to get to.
Royalties are practically 100% margin. They go straight to the bottom line, and that's what all of the investors always like to see. We commented — I think we did the analysis of other semiconductor IP companies just to give you guys a feel. The multipliers are normally like 50 times — the market cap is normally an average of 50 times the revenue. So, it's a big multiplier that you have in IP because of this phenomenon. So, you want to have about 50% of your revenues come from license fees, which are feeding the pipeline, because you want to always have more products come in that eventually will give you royalties. And so, you want to have constant flow of new projects coming in, but at least 50% is royalties and moving forward.
IDMs, and I talked about On Semi: in an IDM the situation is different in the sense that they start the chip design almost immediately when you start doing the technology transfer. Again, everything was triggered by the design teams. So with them, and that's why you can see On Semi catching up to DB Hi-Tek so fast, because in DB Hi-Tek things were more serial. Here things are more parallel. So both companies now have products that, again, we really are pushing forward, and we believe are going to be getting to mass production in 2027. In calendar 2027.
So, Weebit overall: we raised the money, we're building the infrastructure to be a strong, solid company. There's a lot of stuff going on in Weebit just making ourselves really ready to support all of these massive customers in terms of the way we do the R&D, using more advanced tools, automated tools, building databases that we can analyze, bringing in a lot of AI that now we use to do things more efficiently. So really a lot of stuff — if it's the tools, if it's advancing the technology.
And one thing which again I mentioned in the past: ReRAM is a very complex technology to develop. I think you all know, as far as we know, TSMC is the only other company that has qualified ReRAM. They are manufacturing; you can find products in the market that have ReRAM from TSMC in them already. Of course for us that was a great thing that happened because the market can see that ReRAM is really working in mass production — not with Weebit yet, very soon also with Weebit — but it is out there. I have really given the team a challenge: we just need to be the best ReRAM in the market. We are totally focused. Unlike — TSMC, great company, huge resources and everything, but that ReRAM is not what they make money off of. This is not their big focus. Their big focus is on manufacturing, making their manufacturing more efficient and everything. For us, I gave the team the challenge: on every technical parameter, I want us to be an order of magnitude better than anyone else.
Obviously that's a general statement, but if you look — and not going into the details — when we talk about performance, read time: the standard in the industry, people that are trying to develop ReRAM, or TSMC or whoever, they're talking about read time which is above 20 nanoseconds. Weebit is already focused on going below 10 nanoseconds. The industry is at 10,000 cycles, moving to 100,000 cycle endurance. We already qualified, as you know, last year, more than a year ago, at 100,000 cycles. We're moving to a million cycles. I won't go into all the details here. Obviously temperatures: we already have done 150, we're demonstrating already 175. Most of the other ReRAMs, either TSMC qualified or others that are wannabes, they're talking about lower temperatures. So, we're very bullish. We're pushing very hard. We're not letting go of the gas pedal, just pushing extremely hard to constantly improve. People ask me, "When are you going to be break even?" I can just lower R&D expenses now, we'll be break even, but we're kind of putting a cap on ourselves. We don't want to be there. We want to — there's this big vacuum now called the ReRAM market. We want to fill that vacuum. We want to take over as much as we can, as fast as we can. Make it so obvious to everyone: you want ReRAM, you don't even think about it, you call Weebit. And that's where we want to be, and that's what we're pushing.
So, you guys already know we address so many markets, and really working hard on many of them. And I'm going to hand it over to you to talk a little bit about all the stuff that we're doing in AI and in-memory compute. Yes.
All right, let me talk about — I'll get to drink some. Some water, only water. All right, yeah. So, when it comes to AI, this is probably — excuse me here — the next frontier for Weebit. Of course, today we're very much focused on embedded ReRAM. As Coby mentioned, this is our focus here. We want to make sure it's commercially available, we have customers, customers are shipping in production, etc. But at the same time, we're seeing this memory-centric computing domain where actually the memory, or non-volatile memory, is not only used to store data but also to run some of the computations. And this is especially interesting in the AI space.
So, specifically when you look at the AI market, in general, you can divide this into two parts. You have the training and you have inference. Training is where you actually train the model on big data. This is what data centers excel in. You run this for days and weeks on large GPUs. It consumes a lot of power, a lot of reading and writing and updating of the neural net weights. Eventually, you get to the point where you have that network available, and the next stage is inference. Now you take this model, you put it in another chip. It could be in data centers. Mostly now it's probably propagating into end devices — smartphones, IoTs, industrial applications. And there you run the inference. You just infer. You use these coefficients, you use these network weights again and again to run this AI inference. Interestingly, this year 2026 is the first year where actually inference is overcoming training in terms of the amount of investment being put there. It's just a few companies, of course, running training with GPUs from Nvidia, of course, but inference runs everywhere, and many, many other companies are in this market. We are focusing more on the inference side rather than training. Inference is where non-volatile memory such as ReRAM really excels, because when you do inference, you mostly read this data, these weights. You don't have to update them as you do in training. NVM such as ReRAM really excels in reading — and Coby mentioned here the fast read that we're now developing. Just another advantage that we have. A lot of accessing the data and not so much updating.
Specifically, and if that's going to be the next slide. Yeah. What we're seeing in this AI inference market is that memory wall. If you look at these AI architectures, you have the CPU or GPU, whatever processing unit, and you have the memory. And you have to read from that memory continuously a lot of information. And that actually consumes a lot of power, a lot of time. We've seen a report showing that Nvidia GPUs sometimes 50% of the time they just sit idle waiting for data to come in. That's horrible. That's really not the best approach. And the reason is these two are separate — sometimes separate chips, separate dies — and you have that bottleneck getting data in, and it consumes a lot of power. One of these new architectures that's now happening in the market is what's called in-memory compute, where you actually bring the computing onto the memory itself. So, the memory not only stores these weights, these neural net weights, but it actually runs the entire computation within the memory itself. So, we have here a video that explains this, and for that I will need to do some kind of this here.
Version of what you saw in August. So, it's a new version.
Oh, okay. All right. So, wait a minute. Excuse me here. I will need just to run this here quickly. First of all, I'll do the share. There. It's working.
Artificial intelligence is rapidly becoming part of everyday life, from cloud services to edge devices, vehicles, factories, and robots. Behind every AI application are enormous numbers of computations involving vast amounts of data. In traditional computing, data must constantly travel between memory, where information is stored, and processors, where computations occur. As AI models become larger and more complex, moving data consumes more power and becomes a growing constraint on AI performance. As AI workloads continue to grow, so do the power, cooling, and infrastructure needed to support them. What if instead of moving the data to the processor, we move the processing to where the data already lives? In-memory computing performs calculations directly within the memory itself, dramatically reducing the need to move data. Emerging memory technologies such as RRAM enable memory arrays that can both store information and perform the computation. AI inference relies heavily on a type of operation called vector-matrix multiplication. At each stage of a neural network, incoming data is multiplied by a matrix of learned weights. The learned weights represent part of the model's knowledge, helping it interpret the input and produce a meaningful result. But here's the catch: conventional memory can store the weight matrix, but it can't perform the calculation itself. To get a result, the weights have to be read from memory and transferred along with the input to a separate processor. Only then can the input be combined with the weights to produce the output. And because this happens over and over throughout inference, huge amounts of weight data move back and forth between memory and the processor.
RERAM enables a different approach: analog in-memory computing. In RERAM, each weight is stored directly in a memory cell as a specific conductance level. The input values are applied as voltages to the rows of the memory array. Within each cell, the input is multiplied by the stored weight. The resulting current is proportional to the input voltage and the cell's conductance. The currents from all the cells in each column automatically combine into a single sum, exactly what is needed for vector-matrix multiplication. In this way, RERAM brings storage and computation together in the same memory array. The weights stay in place, and the calculations happen right where they're stored. This can significantly improve energy efficiency and performance while enabling more capable AI systems. This enables more powerful AI in systems where power, performance, and real-time decision-making are critical — from vehicles and factories to robots, aerospace and defense, to intelligent edge devices. As AI expands from cloud infrastructure to intelligent edge devices, in-memory computing has the potential to become a foundational technology for a new era of efficient AI computing. Weebit Nano, bringing a new memory technology for the AI era.
Mute. All right. I hope it's better now. Okay. Yes. So, two AI customers have taken the license and got this in-memory compute architecture using our resistive RAM specifically for this type of application and running exploration. One of them actually showed the performance gains that are pretty incredible, I think. We're not going to see this again. Let's move on.
Yes. So, I've mentioned this: in-memory compute has been in research phases for quite a long time now. The existing solutions are either based on SRAM, which is a volatile technology, which means that it does not keep the weights once the power is plugged off, or trying to use a flash memory, which is of course non-volatile, but it has this scaling limit — cannot scale to the more advanced nodes where you actually want to run embedded AI. And so we are really positioned to solve all these problems.
Yeah, Coby.
Hey. So, in the last AGM we defined the goals for 2026 calendar year revenue. As you know, we achieved that, and actually achieved much better than we expected. We had the first product tape-out — seems like so long ago already — and now we achieved the target of having first AI customers. So again, we're feeling very good about where we are. Working hard to get additional agreements, which will happen. I can't say when with the current status, but we are progressing. And I think at this point we can probably turn it over to questions and open up the questions. We'll start off with questions from people who came here and are sitting in the room, and then we'll expand to questions that are coming online.
I think you should repeat the questions so that people online can hear as well.
Okay, I'll repeat the questions. Yeah, so any questions in the room? Yeah.
With the in-memory compute, my understanding is that you are predominantly for edge devices, right? So it's not a target of data centers and LLMs at the moment.
It's — I would say first phase is going to be more edge AI, but definitely for inference anywhere this can be very useful also for data centers.
In data centers, this might need to run in a more advanced process node, so that might take a bit longer, but that's also definitely a target for in-memory compute.
Also, there's no, I guess, theoretical or technology limitation that would limit it in data centers. It's just an engineering problem.
Exactly. Exactly. Yeah, by the way, this year the data centers are running more inference than training already. So, I mean, inference is growing. At the edge devices, of course, it's mostly inference already. But even at the data centers, you train once and you infer all the time, right? You keep going. So over time, I believe inference will grow and grow and grow in proportion relative to training. And yeah, we'll start off with the edge devices, but the goal is to also go into the data centers.
Yeah, two AI customers starting with exploration — is there a rough time scale for that?
So, yeah, I'll repeat the question. The question was about the two AI customers and what's the time scale. So right now, what we have is an initial technology that kind of demonstrates the abilities. These guys are paying us to actually be able to evaluate — I want to make it clear — but they are doing analysis. It's not a commercial product that's going on. We still have technical challenges that we need to overcome. The team is now working on it. We have the initial signs that we know how to overcome them, but there still is a lot of work to do. I mean, to be honest, I think it's going to take maybe a couple of years to really get this to the point where it can be commercial. Once we finish all of the analysis, we'll have a much better idea of where we are, what the solution will look like, etc. But it's very exciting to see the potential of the technology and the fact that people are paying now to actually be able to evaluate it.
Yeah, I just want to add also another point. These customers, these two AI customers and others that will come in the near future, we also get their feedback about our technology. And that's really helpful, because at the end of the day we're a limited set of people in the company — very smart, but working with all these other companies and research institutes and getting them a feel of the technology and getting their feedback actually makes it much more efficient and will help us accelerate our own developments towards a commercial product.
Yeah, I know you'll say hundreds, but when it comes to ON Semi and Texas Instruments, it's long-term — what kind of capture, to say a number of chips that they're currently using? And I guess if you could speak to what, say, 50% of chip capture would look like in terms of annual royalties, if we got to that point versus 100%, or if you just give some approximations.
So the question is regarding ON Semi and TI and how many products, and I guess what's the revenue expectations and things like that. So I won't go into specific details, but I'll mention, you know, TI is — I think we already said — they have 80,000 products in their price book. And that's before they acquired, what's their name? Forgot the name of the company. It was another 8,000 products in their prices... SiLabs. So, I mean, they have a lot of products. Now, not all of them need non-volatile memory. Some of them are very simple sensors or stuff like that, but I mean, I even say 10% and I think it's more than that need something like ReRAM. You can understand the potential and the scale here that we're talking about. And they are looking at moving forward to the next step from Flash. So they are definitely focused. I can't speak for them. I can't make commitments and I can't repeat what they told me in private discussions. But I believe, you know, that's going to be a major source of revenue over time with all of the products.
Similarly, ON Semi is a very big company, and I'll repeat because a lot of people don't know — ON Semi used to be called Motorola Communication. It's not a small company. It's a very big, serious company. I don't know how many products they have. Thousands. Yeah, I mean, they have thousands of products obviously. A lot of products. They're number two in power management worldwide. They're number, I think, number five, or in the top five in automotive. So very big players with a lot of products. And beyond them, DB HiTek has roughly 200 customers, more than 200 customers. Some of them small names like Toshiba and Qualcomm and Intel. So, you know, and those customers have products that are manufactured at DB HiTek. So even with the ones that were already signed, you can start sensing how the number of products... again, without going into details, and there are many different royalty models and stuff, but even if you say royalties are just a few percent of the sales price of these products, you can do the math already and look at it and what's happening.
So I again don't follow. So you're saying that maybe full penetration might be 10 to 20%?
No, I'm not saying that. I'm saying order of magnitude. We need to look, and it's going to be — as far as I'm concerned, these are big numbers. I can't commit. Again, these are decisions of the customers. I know what they're planning, what they're thinking, they told me. A lot of these things they tell me confidentially, and I really don't want to mess with PI or, you know, get in trouble with these customers. So it's their plans, it's their decisions. But you can imagine if they're making such a huge investment into bringing up this technology, it's not just for a couple of products.
Okay, other questions from the room, or we can also... so, maybe online.
A bunch online. Okay, so first one, regarding Elon Musk Terra Fab plans to make radiation-exposed spaceships in-house. Does ReRAM radiation tolerance make fabs like that direct licensing targets, what do we think? And how far is ReRAM from advanced nodes like Intel's A14?
Yeah, of course. Yes, absolutely. Radiation hardening or radiation tolerance is one of the key advantages for resistive RAM. For Weebit ReRAM, we did our own analysis and found out it to be, you know, 350 times more radiation tolerant than flash memory. So clearly it is a target for some of our customers in the aerospace and defense market. Possibly also for Elon Musk Terra Fab, for sure. The second question was about the Intel...
In terms of small geometries.
So, we're working on some smaller geometries, and we've done some analysis on FinFET nodes. This is the 10 and 18 nanometer. And we currently don't see any physical limitation to go down and continue scaling towards even more advanced process nodes. We haven't looked at, you know, the angstrom level, but at the moment there's no immediate need over there. We're still working on these 20s and 10s. Yeah.
Mhm. Thanks, Eran. Next one. There's been considerable increase in headcount over recent quarters. Would you characterize the work being done by these employees as R&D, evaluations for potential customers, or NRE activities for contracted customers?
So it's actually pretty difficult to distinguish between R&D and work that we're doing for customers. It kind of goes together. I mentioned, for example, lowering the read speed. You know, the trigger was a customer who came and said, 'You know, we just cannot work with 20 nanosecond. You guys have to have less than 10 nanosecond, otherwise there's no deal.' And you know, initially the engineers came over and said, 'These guys — I don't know what they're smoking, but you know, they're just totally out of their mind.' And you know, Israelis, they don't know the word impossible. I told them, 'You're not going to give up on this customer, are you?' You know, and so they sat and — and here again is that big advantage of having all of these disciplines sitting in one room and working together, because each one of them found, 'Oh, I can achieve, you know, another one or two percent advantage here.' And then this other guy said, 'Oh, I can give you another millisecond if I do this, and if you do that, I can do this.' And they ended up finding a way to make it happen. So now is all of this work R&D work or is it for customers or whatever? It's like today it's all in one. We work with ON Semi. It's the first time we're working with a copper fab. So is all of the development on making the technology run on copper R&D or is it customer? It's really very much together.
And I think it's going to be that way for a long time because almost everything we do is a first. It's the first time we do a 300 mm wafer. It's the first time we do copper. It's the first time we have a customer who's telling us leakage is insanely critical and we have to have such low leakage. So the customer demand is feeding R&D, and it's kind of working together.
Thanks, Gabby. What is the cost of not going ahead with ReRAM? I assume it's a big commitment in cost for them. And it's not something they do without confidence.
Yeah, so obviously, I mean, we talked about all the advantages of ReRAM. You guys know that slide I think by heart by now, and it's faster, it's lower power consumption, it's cheaper. And I think the cheaper part is of course the most important one for so many people. If people try to continue to go with flash, they're going to have a more expensive solution, much slower, much higher power consumption. They're not going to be competitive. That's the simple answer.
What is the heat impact of ReRAM?
The heat impact. So, ReRAM is lower power consumption.
Yeah, it's lower power consumption. It's using lower voltages compared to flash, and that actually means lower heat dissipation compared to flash memory or other technologies. So actually it's on the contrary.
Yeah. So, we're jumping about here, but do you need to reduce costs until revenue flows?
Well, we're just about bankrupt, right? No, I mean, to be serious. I've seen so many people who won the lottery and then after a couple of years found themselves on a bench in the public park, you know. So we need to be very careful. We did a raise, we have money in the bank, but you know, I'm very cautious. It's my shareholders' money, right? And when we don't close deals, the good news is when you have such an experienced team and your VPs are so experienced, they don't even come to me to hire new employees when they know deals are not happening — they automatically control themselves and hold back. So we kind of intuitively lower our spending and adapt, and we're very, very cautious about spending. The team knows I'm extremely strict on any expenses. By the way, I look at every PO that goes out, even if it's just a few hundred dollars. I literally sign every PO, and I check and I verify that that's according to plan. So we are very conscious about budget. You can see the results. I mean, at the end of the day, you can see how much cash burn we have despite all of the activity, despite all of the work that we're doing. I think the numbers speak for themselves.
Thanks, Gabby. Difficult to give a specific answer to this question, but how is engagement going with automotive companies? So, how many automotive companies have we been engaged with?
Well, you're talking to a lot of them, so...
Yeah. Well, first of all, I want to say, yeah, ON Semi and TI, as we mentioned previously...
Top five, and they are engaged with us with regards to automotive. So that's a very good proof that we're capable of that. And also, we have qualified the technology for automotive, for AEC-Q100. In terms of the pipeline, I would say we see a lot of interest there. I think Coby mentioned some of the existing resistive RAM technologies, mostly by TSMC, are not yet at the most advanced automotive level, what's called grade one and grade zero. We have proven that we can do that, and that actually brings a lot of interest, a lot of opportunities specifically for automotive.
Automotive, one other thing I want to say: automotive is usually a bit more advanced in terms of process node. If you look at other IoT consumer devices, etc., automotive requires much more processing power, and that means that these vendors usually go down the process nodes towards 28, 22, and the 10 nanometer faster than, let's say, smartphones or IoT. And over there they don't have flash memory; it's not an option. So that's maybe another reason why we're seeing more and more opportunities in the automotive field.
Yes, sir. But I do want to add one comment. You need to realize automotive — a lot of the automotive devices are life sensitive. I mean, you know, they impact human lives. And therefore the regulation and the qualification and all of that, there's much stricter regulation in the different countries. So when you engage with an automotive company, getting to mass production does take longer, because they do need to go through much more rigorous qualification or certifications. And by the way, every industry has its own level of certification. When we say qualification at the fab, that's one level. The products have to go through their levels of qualification/certification, depending on the industry. Some, you know, it's nothing. Some of them, like automotive, it takes a few years to actually get the product to the point where people say, 'Okay, we can actually have human lives depend on this.'
Okay, next one. Having talked a lot about memory and compute, in-memory compute, both today and over recent times, while progress is evident, has the company defined the strategy to mass production commercial offer? Is there an expected timeline for rolling it out?
I think, as I was saying, we're still analyzing and understanding the technical limitations and what we need to work with.
