Good morning, everyone. I'd like to introduce our next presenter of the day, Didier Lasserre, VP of Sales and Investor Relations with GSI Technology.
Thank you. Good morning. Can you hear me okay? Great. You know, as I was introduced, Vice President of Sales and Investor Relations at GSI. Been at the company coming up on 28 years, been there from the very beginning. See a nice transition over the years. I'll be talking about the company overall, and then really focusing on our AI product line, which we're starting up now. Since it is a new product line, I will be making forward-looking statements, I do have to include the safe harbor. Just quickly, looking at the company, we started the company over 30 years ago as an SRAM company. SRAM is a memory, a little different than DRAMs in that it's a really low latency made for super high performance applications.
Since day one, we were partnered with TSMC as our fab partner. We have had actually a very close technology relationship with them. We were not the 0.15 technology partner, but we had the first working silicon on 0.15. For that reason, they did choose us as their 0.13 technology partner. That close engineering relationship has also turned into a nice business relationship. As I mentioned, we've been doing business with them for 30 years. During that time, you know, there's been periods of allocation and tightness. We've always gotten the wafers we've forecasted over that period of time because of the tight engineering relationship. What I'll be focusing the majority of this presentation on is the computer memory, which is our AI.
You know, some of you folks that were in here before, so you know, we saw a lot on the data center, we're focused on the edge, we'll talk about why and the applications. To date, we have self-funded $175 million in R&D for this APU development. Our SRAM division, you know, is very profitable on its own, it's helped offset these R&D costs. With that said, it is an expensive endeavor, we did raise net $47 million October of last year. If you look at the revenues, the trailing 12 months, about $25 million, the majority of that is SRAM, that represents about a 20%-22% increase year-over-year.
We outsource a lot of the labor-intensive functions in the company, like fab, like assembly, like sales, and so we're able to keep our headcount to mostly design engineers and software engineers. We have 126 employees worldwide. I'll be talking about the APU. We have a very unique technology, and we need to protect that IP, and we have really been aggressive on filing for patents. We have 147 in the company, of which 88 are for the APU, and we're continuing to file for more, and we'll get more granted. We have just over $67 million in the bank with no debt. I put this presentation together last week. Last week, we were, you know, about a $450 million market cap.
You know, there's been some softness the last few days, so we're south of that. We do have a strong insider ownership. Management owns 20%, and we also have a little bit more insider. Some of our original angel funders still hold shares, so we're certainly north of 20% controlled. As I mentioned, we started the company as an SRAM company. You know, this memory, as I mentioned, is really unique in that it's very low latency. It's really used for the highest performing infrastructure for the networking and military infrastructure. We were introduced to this AI technology about 10 years ago.
When we looked at it really made sense to acquire it because to implement it would be best done using a SRAM-based cell, which is what we're an expert in. So we then, you know, introduced this APU, Associative Processing Unit, and we do something different. I'll get into more detail, but we're a true what's called CIM, Compute-In-Memory. You know, that's something where you hear a lot lately, and people are starting to say they're doing it when they actually aren't. They're doing what's called near memory compute, which is they're actually bringing the processing elements and the memory elements closer together, and I'll talk about that. If you look at the markets we'll be going after, as I mentioned, we are not going after the data center.
You know, NVIDIA has been doing a very good job there. We're really focused on the edge, and we're focused there for a few reasons, which I'll talk about in detail. Specifically, the markets we're addressing will be right now, it's close to $10 billion markets growing very quickly. You know, looking at about a 20% CAGR in the markets we're addressing. If you look at some of the applications that we're actively working on now, autonomous drones, satellites, especially with SAR imagery, also smart city. In fact, if you get a chance, we put out a press release yesterday about a POC that we've just won on a smart city, and I'll talk about that a bit more.
Let's just kind of look at the challenges right now in this market. You know, the problems are right now that data is constantly being moved. When you're moving data that, it's delay, it's latency, that's an issue. If you look at GPUs, you know, NVIDIA GPUs, they're constantly transferring data, and I'll talk about that on the next slide. When you're moving data all over the place, it takes a lot of power. I'm sure you've heard that these data centers use tremendous amount of power. As I mentioned, we're focused on the edge. On the edge, you have a very limited power budget. You know, the data centers, you know, they're complaining that they're using too much power, but you have power plants that are running those things. You don't have that on the edge.
You don't have that luxury. If you have high performance or, I'm sorry, a high power part, you have to really scale back the performance, which is a problem. The solution is our APU. Because we do the compute and memory, we are not moving data. In fact, the next slide will give you a better feel for that. On the left-hand side is really a GPU and a CPU. It follows what's called a non-von Neumann model. What happens is, you can see that, you know, when the compute elements need data to make whatever calculation, it has to go fetch the data from memory in DRAM. The data has to transfer into the L2 cache through the L1 registers before it gets to the compute.
Once it's used here, it needs to be written back through the same process. This is a tremendous amount of power to do this. If you look on the right-hand side, this is our solution, much simpler. What we do, which is very, very unique, which is what all the patents are written around, is that we actually have the data reside where the processing elements are. We actually do the processing in the memory array itself, and that's very unusual. What happens is the data is where you need it to be processed, so we don't need to fetch it. When we're through with it remains there. We don't need to write it back. There was a study done by Cornell.
They actually had one of our original Gemini-I boards, and they compared it to an NVIDIA GPU, and they were using it for a RAG application. If you're not familiar with RAG, it's essentially used for LLMs to bring in relevant data. LLMs, if you're, if you're familiar with them, very expensive, very difficult to train. Once you have the model trained, you don't wanna continue to retrain it, instead you use RAG to bring in pertinent information to keep the, you know, the AI from hallucinating and giving you erroneous data. When they did this experiment on the same performance, the Gemini-I versus the NVIDIA GPU, we were 98% less power. It really is a game changer.
Another area, here's another application where this is a POC that we won for a drone surveillance application. This drone manufacturer originally looked at an NVIDIA Jetson family, and because there were two critical criteria. The system had to be less than 50 W, and it had to have a time to first token of three seconds. Time to first token is basically the second that something happens, how long does it take before you get an answer? That's your time to first token. The Jetson was able to do the three seconds, but their power was actually 160 watts to do that, so it blew away the power budget. Why is that important? Well, this is going on a drone.
If your power is high, the drone goes up, it goes out for a very short period of time. It has to come back to get recharged. That's a problem. They looked at the Snapdragon from Qualcomm. Qualcomm gave them the power, but their time to first token was 12 seconds. 4x slower than they could take. That's when they looked at GSI. At GSI, we gave them the power and the performance. We gave them originally the three seconds. It's since been optimized to 2.7, and we feel that we can get it actually under 2.5. Three second is really the threshold because that's the time it takes a human to react, and so you have to at least be at that level.
We won this bake-off, and we are now the partner for this drone activity, and I'll talk about it a little bit more in detail. This gives you an idea that, you know, we have 300% better performance per watt measurement than both folks, you know, listed. If you look at our roadmap and what each family is going to go after, we have the Gemini-II. The Gemini-II right now is what we've won this. It's called the Sentinel. It's a POC by the Department of Defense or Department of War, depending on what you call them, along with another unnamed government agency that's overseas.
This is a drone surveillance where it takes fixed cameras that are on a perimeter and brings in drones that work together to take out all the blind spots. It's autonomous in the fact that it will determine what the emergency or alert is and then give you a recommendation. It can fly around and see somebody behind a truck with a gun, recommendation, shoot. It is a recommendation because it's still a human that has to make that decision at the very end, but it gives them all the data up front. It's not somebody sitting in front of a monitor trying to determine what's happening. It's being told for them. The smart city, this POC that we won that we just announced, it's phase I.
Phase I will incorporate 20 cameras. These are cameras that are already installed at this particular county in Taiwan. It will take the video from those cameras, and it will identify what the problem is. Is it a fire? Is it a riot? Is it a car accident? Then recommend an action. Right now, the cameras just record, and later on, people can go back and say, "Okay, what happened on that day?" That doesn't do you any good. You know, it needs to be proactive. The p hase I will take a couple months, and then it goes to p hase II . Phase II ups it to 80 cameras and then brings in audio. They're looking to deploy some of these in the city, but also in areas like the schools.
In schools, they've been having some abuse problems, and it's not just physical abuse, it's also verbal abuse. It needs to pull in the audio as well. In p hase III , which will be into next year, that we're obviously hoping to win, will now be a full-scale 6,000 camera system. Just to give you an idea of what that means in hardware sales, one Gemini-II chip controls four cameras. You can do the easy math. Other deployments with Gemini-II have been, we've had some successes in what's called SBIRs. These are grants from the government.
The latest one that we announced last quarter was a phase II, a $2 million phase II from the US Army, and is to create a ruggedized edge node that can do object detection can do SAR, you know, multiple applications. We're particularly excited about this SBIR because this is the first SBIR that really can turn into a product that we can sell for revenue. The other SBIRs have been, you know, nice R&D offsets and give us a recognition within the DoD elements, but this is something that can turn into a true product. Our next generation part, Plato, which we started the design last year, we're due to have it completed about March or April of next year, is really now focused even further at the edge and truly for LLM.
You know, Gemini-II we're using for some LLM applications, but it wasn't designed for that. Plato is. What we've done with Plato is we've really beefed up the bandwidth to go to DRAM. Now, I just got through telling you we don't do that, right? Well, that's the Gemini-I. For LLMs, for this particular application, there's no getting around it because the models are enormous. I mean, LLM stands for large language models, they cannot fit on a chip. We beefed up the bandwidth so we can get the data in and out. What we've done differently is we've brought down the power significantly.
The maximum power on the chip will be 10 W, and then the size is gonna be a quarter of the size of the Gemini-II. Now we can get further out into the edge where it needs even lower power and even smaller form factor. Lastly, we've actually started conversations already with funding partners to define what comes after Plato. As these become more public, we'll let you know, obviously. To give you an idea of some of the awards that we've had for the SBIRs, and again, these are grants that aren't recognized as revenue. They're recognized as R&D offsets. We have three active ones right now. I mentioned the $2 million phase II.
Again, this is something that we're gonna be finishing by early 2027. It's a just over a year POC. At that point, we'll be looking to make revenue from this platform. I mentioned the POC with our drone partner, G2 Tech. This summer we'll have a kind of a mini demo that shows the technology. By the end of the year, we'll have the full-scale demo. GSI, we have actually finished our deliverables. We had to do the software for the time to first token and a few others. We're done. It's just now G2 Tech, you know, debugging their actual drone. Lastly, another active phase II was an extension that we got from the Space Development Agency.
They actually want us to put our part under radiation, under heavy ions, to see how robust it can be for space applications. This is important because GPUs do not do well in space. They're looking for solutions for SAR applications, object detections, what have you, to put on satellites. This is the first step in to show that area or that segment what our part can do. We're also going after larger funding opportunities. Some of these are STRATFI and also BAA. These are opportunities that can bring anywhere between $5 million and $20 million of non-dilutive funding into the company, and we're actively pursuing those.
So, I just want to take a quick step back. You know, to let you know what's funding the company right now is our SRAMs. So what's interesting about our SRAMs is that we have the highest performance and highest density in the market space. [Currently] we have frozen our roadmap because all of the R&D effort is on the APU family. The good news is, all of our competitors have also frozen their roadmaps. So, we have a one to two generational lead on any competitor. That's important because all the new designs we've been getting have been on these sole source generations nobody else makes. So our ASPs and our margins have gone up overtime with our SRAM family.
With that SRAM family, we've taken a little extra step to make them more robust, to put them in space. You know, I've been talking about this for probably coming up on two years now. So we've done what's called radiation-tolerant and radiation-hardened. Radiation-tolerant is good for any LEO orbit satellite or any satellite or any probe or rover that goes away from the sun. Radiation-hardened is more for GEO. Anything that gets closer to the sun has become an issue, so we have radiation-hardened. To give you an idea of the ASPs, you know, our highest-end SRAM sells for a few hundred dollars. The radiation-tolerant, the same density, sells for a few thousand dollars. The same density in radiation-hardened sells for tens of thousands of dollars, up to $30,000.
Obviously, very high gross margins, and it is a growing market. You know, certainly it takes a while to get into this market. As you know, as any of us or any of you that know the story, it's been taking us a while. We've had a lot of protests go out, and we should start getting our first production orders, well, certainly this calendar year. Once that happens, it's an ongoing, you know, revenue. It's certainly something that we wanna go after. High level overview, you can see our revenues have been increasing. As I mentioned, we're year-over-year, we're over 20%. You can see, in the middle there, the operating, you know, number has gone up.
As I mentioned, we started the design on Plato last quarter, or I should say in the December quarter. We had to bring in some IP that we had to purchase for that. We've also kicked in a design team from Synopsys to help with some of the design. All the core design we're doing ourselves, some of this, as I mentioned, we're gonna be doing some DRAM interfacing. The DRAM interface and the PCIe interface, this is Synopsys IP. They know their IP better than we do. We've actually contracted them to do that work.
There's, you know, there is the contract work and the IP, and you can see we'll have a bit of an elevated R&D or expense with that contract work for the next 4 quarters. The cash, as I mentioned, $67 million. It went down a bit quarter-over-quarter because of the IP that we had to purchase. In summary, we kind of view ourselves as an AI startup, because we're just getting into AI, but we're not a startup. What I mean by that is, some of the other, you know, AI guys are coming out. They're coming out with a great idea. They've never manufactured a part before. Remember, we've been doing SRAMs for 30 years.
We've shipped over 100 million SRAMs over that period of time. The manufacturing model for SRAMs is identical to the APU. Same fab, TSMC, same assembly house, ASE, all the same contractors. The team that does the SRAMs will be doing the APU. We can ramp very quickly. As I mentioned, we've been really doing a lot of non-dilutive R&D between our revenue from our SRAM and also the SBIRs. We're able to really kick in this technology. Again, going back, this computer memory really is a game changer. You know, I showed you some of the performance versus power. This is something that is not lost on our customers. This is something that's gaining a lot of traction. We have proven victories because of that.
I mean, between the SBIR wins, the Cornell paper, and the bake-offs, it shows, you know, where we are on that. At this point, open it up for a couple minutes of Q&A. Yes.
Why did you guys decide to go standalone?
Okay. The question is, why do you standalone versus?
Well, there was an announcement about four months ago that you were going standalone versus a partnership, I think with Taiwan. Are you still partnering with Taiwan?
I, no, we're still with Taiwan. I think what you're referring to was we had a strategic initiative in place. That was at the time, you know, we needed additional funding, so we were looking for the, you know, for partners. As I mentioned, we raised $47 million in October, so there was no need for that. With that money, we can be able to achieve all of the, you know, the certainly the short-term goals without any additional funding.
Okay. One more.
Yes.
Cerebras is not your competitor?
The question is, Cerebras is not our competitor. That's correct. Cerebras is data center. Cerebras, you need a lot of power, you need it water-cooled. That is not our game. We're at the edge. Cerebras and so we do compete with NVIDIA, but we compete with NVIDIA on the Jetson family, not on the Blackwell or Hopper, 'cause that's all data center. We're, again, focused on the edge, not in the data center. Yes.
Your new initiative is gonna try to get you in the data centers. Is that right?
Excuse me?
The new initiative, the new chip you're working on here.
No. The new chip is Plato. That will actually take us further away to the edge. As I mentioned, you know, that's gonna have a 10 W power. I mean, you can look at these. The single light bulb above here, it's gonna use less power than that. It's gonna be the quarter size, so it's gonna be very small. It's really to try and get into true edge, into robotics, into, you know, any of these smaller drones. You know, the drones we can do now, but, you know, our power budget, as I mentioned, is 30 W. Now you can get into ones that are less than 10 W, and so you can have really long sortie times. No, it will not be in the data center. It will definitely stay on the edge.
Certainly data centers have a power problem, and they would love to be able to get their power consumption down.
They do. As I mentioned, when I showed you the roadmap, I said, after Plato, we've already started talking to some funding partners, and one of the areas they are looking at is the data center that they would like us to make a part for. As of now, everything's on the edge.
We have enough time for one more question.
When do you expect to get revenues?
So.
How quickly can you ramp up?
Yeah.
You said this year.
Yeah, yeah. Right now, we'll get some minor revenues from the POC on the drones and also on the smart city. The drone, we can get possibly some additional revenues on the smart city if we win phase II by the end of the year. Really, it's 2027. If you look at the, as I mentioned, the 6,000 camera deployment, that's in 2027. The demo for the DoD will be at the end of this year, the full demo, field demo. It'll be a 2027 contract. It's really some prototyping this year with revenue volume in 2027. Yes, Brian?
What about for the rad-tolerant application?
The question was the Radiation-tolerant revenue. We've been sending, as I mentioned, prototypes for the last few years. At least two of them should go in production this calendar year. Hopefully in 2027, we'll see more of them being deployed. You know, satellite deployment takes forever. I mean, there's a couple of satellites that we were already supposed to get orders for a year and a half ago, and they still haven't launched the satellite. It's really a slower market as far as that goes. I'm out of time.