Needham's 27th Annual Growth Conference. My name is Shawnee Mbwali, and I'm on the semiconductor team here at Needham. It's my pleasure to introduce GSI Technology. GSI provides high-performance memory solutions for AI, HPC, and aerospace and defense end markets. Joining me from the company are CEO Lee-Lean Shu and VP of Sales and Investor Relations, Didier Lasserre. Lee-Lean and Didier are going to take us through a 30-minute presentation followed by a short Q&A session. Thanks for joining us, and take it away.
Thanks, Shawnee. Again, my name is Didier Lasserre, Vice President of Worldwide Sales and Investor Relations. I'm going to take you through the company. I'm going to spend a little time upfront talking about our legacy product line, but the focus of the presentation will be on our new APU AI chip. Obviously, we'll be making some forward-looking statements. So, obviously, we have the Safe Harbor statement in. If you look at the company, we were founded, in fact, by Lee-Lean Shu in 1995 in California, Silicon Valley. We went public in 2007. We outsource the majority of our manufacturing. Our wafers are done by TSMC. We've been using them since day one. We have a very close both business and engineering relationship with TSMC. In fact, we were the technology partner with TSMC on one of their process nodes years back. And so that's really developed a nice relationship.
So we've had wonderful support over the years. We started the company as an SRAM company, and we'll talk about that a little bit more, but we certainly have the highest density, highest performance products in the industry. This legacy product line has been funding our new APU AI initiative. And so we kind of view ourselves as essentially a self-funding AI company. We've been building our patent portfolio. We're up to 136 patents that are specifically for the SRAMs and our APU architecture. That APU architecture, we're fully committed to it. We've invested over $150 million over the last nine years into this product line. If you look at our revenues, which are essentially 100% SRAMs, our last fiscal year, we were just under $22 million. Our employee headcount is actually, I'd say, down to 122.
We had a reduction in force in April of this past year just to help cut back some of the cost during our launch of our APU, and so we now have 122 employees. We have just over $18 million in cash and cash equivalents with no debt. We've been hovering just under $100 million in market cap. Obviously, we've had a run today on our stock, and so we're over $100 million market cap. We have a high insider ownership on the shares of the company. We're hovering around 27% today. I'll spend a little time going over our legacy SRAM product. As I mentioned, we're a leader in this market. We have probably the largest portfolio in the market space, but we certainly have the highest performing and the highest density in this market.
In some cases, depending on the competitor, we have at least a one- to two-generational lead against them. And so when we're talking about density, that means we have either twice to four times the density size of our nearest competitor. The majority of our revenues come from the Sigma Quad and Sigma DDR families. And these families are mostly unique in nature. And what I mean is that we're sole-sourced. And so all the new designs we're getting, we don't actually share with a competitor. The older generations, they might have second-sourced us, but on the new generations, we're sole-sourced. So we get 100% of the business. If you look at the two markets that we're going to be entering, they both actually leverage our expertise in the high-performance SRAM market. One of them is closely related. We're doing radiation-hardened and radiation-tolerant SRAMs, specifically for the space industry.
And then we're also, as I mentioned, entering the AI market. And if you look at our APU chip, it really uses an SRAM bit, an SRAM technology. And so it's really highly leveraged with what we're already experts at. If you take a dive into the actual markets, the space semiconductor market is a growing market with just under 9% CAGR. And I'm sure you're all familiar with the tremendous growth happening in the AI semiconductor market. I mean, this has a greater than 20%, excuse me, greater than 20% CAGR. And so this is, we're at the right place with the right product at the right time. This will be the last kind of SRAM-ish slide we'll be talking about. And this is specifically for the radiation-hardened and radiation-tolerant chips. We're targeting satellites and space applications.
These are difficult products to make in order to be able to withstand the harsh conditions of space. It requires a lot of robustness in the devices. And for that reason, the gross margins and the ASPs are extreme. We have ASPs up to $30,000 a part with gross margins north of 90%. The market we're addressing on an annual basis is just over $100 million. And it's our intent that we want to capture at least 15% of that market going forward. We have seeded the market over the last couple of years. We've sent in many prototype devices to several prime contractors for several different programs. And we're just waiting for some of these to go into production. Now I'm going to spend the time talking about the APU. And so we truly have something unique here.
This is something that competes in the market with GPUs and CPUs. But we have a true compute-in-memory architecture. This term is used loosely by a lot of folks in the market space. They say they have this CIM architecture, but they truly don't. What they have is they have near-memory architecture, which basically they're bringing their processing elements closer to the memory is all they're doing. While our architecture, the processing elements and the memory are one. They're shared. And so we truly have the compute-in-memory. We have a very high-density solution. If you look at the memory elements, we have two million memory bit processors on our first-generation Gemini-I product. As I mentioned, our first-generation device, Gemini-I, we've had for a couple of years. It really was here to illustrate the uniqueness and the advantages of our architecture. With that said, we want to monetize it.
We're focusing on two market segments, and I'll talk about later, SAR and fast vector search. Gemini-II, which we got first silicon this past calendar year, looks really good. In fact, first silicon, except for a few bugs, we're able to have software workarounds to start going through some of the benchmarking and software development. That device, which I'll talk about more extensively, is really geared towards data center performance at the edge. And so if you look at a lot of the markets that we're working with, it's AI models with computer vision applications. We do have a roadmap for this APU family. We call it Plato. And Plato will actually go after a different market. It'll go after the large language model market. So how are we different, the APU versus a GPU or CPU?
As I mentioned, our architecture has the processing elements and the memory shared, unlike a GPU and a CPU where they're separated, and so with those architectures, they need to go off-chip to memory to fetch data, to bring it back, to do whatever process or search they need to do, and then when they're through with that piece of data, they need to write it back to memory, so there's a constant back-and-forth transfer of data. That takes a lot of time, but it takes a tremendous amount of power. With ours, we have an in-place processing unit, which means the data is in the same place as the search and the processing happens, so we're not fetching data, and then we're not writing it when we're through with it. It remains in place.
We've eliminated all this back-and-forth data transfer, which, again, increases the performance and tremendously lowers the power. Where we're also very unique, and this is key in the fact that it also kind of future-proofs our solution, is that we're a bit engine processor. What I mean is if you look at a lot of the GPUs, they're hard-coded or hard, I guess, coded into being, let's say, a 16-bit processor or a 32-bit processor. With ours, as I mentioned, we have two million bit processors that can be done any way you want. It could be set up as a 16-bit, an 8-bit, a 12-bit, make up a number between one and two million. What's happening is a lot of these models, they're finding out, researchers are finding out, are more efficient at certain resolutions or at certain bits.
No matter what they are, if they zero in on five bits or 12 bits or make up a number, our part is there. The customer can just program it to be, let's say, 12-bit. Guess what? In the next cycle, they can change it to five-bit. That's how flexible we are. Also, within our solution, we're very scalable. The GPU is limited in scalability. With the APU, our solution, we can put multiple boards into a server, and those boards act as one. You can also hook up several servers together. As these databases are growing, we're able to react to that with this scalability. The AI market has a lot of needs and a lot of challenges. One of them is, obviously, these data centers are consuming a tremendous amount of power.
I'm sure you've read where some of these data centers are using as much power as some small countries. And in fact, recently, they've talked about trying to reactivate some of these nuclear power plants in order to be able to drive some of these larger data centers. Obviously, that's an issue. So our CIM, our compute-in-memory architecture, as I mentioned, without all the back-and-forth transfer of data, we're able to tremendously reduce power consumption. Also, because these models and these databases are growing, there's a demand for more and more hardware. As I mentioned, we have a scalable architecture to be able to address that growth. What we're seeing, too, is everything used to be at the data center.
Now, a lot of these models are being transferred more towards the edge, which is a problem because a lot of these GPUs are extreme power consumers, as I mentioned. So with our newest generation, the Gemini-II, we're able to bring that data center performance to the edge. And then lastly, there's the LLMs. If you're familiar with all the ChatGPT and everything, they're growing tremendously. And what they're finding out is if you get into a single-bit structure, it actually better supports a lot of these LLMs. And as I mentioned, being a bit engine, we are already there ready to address that. So I'm going to jump into the actual families individually now. As I mentioned, Gemini-I was our first entry into this market, and it was really to showcase the technology. And we're really looking only at a couple of markets for this.
SAR is one of them. SAR is synthetic aperture radar. Think of it like LiDAR, except instead of using lasers, it uses microwaves, which is actually a superior technology. It works at night. It works through weather and clouds. So it's really the go-to technology. So we've written a very nice algorithm for this market. We are actively in negotiations with a couple of large customers to be able to adopt this technology. Nothing's complete yet, but hopefully, we can have something more that we can discuss a quarter from now. Another area that this technology has done very well, our first-generation part, is to build indexes for e-commerce. I mean, if you look at these databases, they're always changing. And so you need to update the databases and also the indexes. A lot of times, indexes aren't updated as often as they should be because it's time-consuming.
There may be some obsolete products or products that are new that aren't in the database. With our technology, we're able to tremendously increase the time, or I should say we should increase. We can make it so that they can update their databases and the indexes more often because we can do them much faster. I mentioned the power consumption. Here's actually an example that some AWS engineers put together a couple of years ago. But it's a great example of our technology. This is basically what it takes hardware-wise to do a billion-item search? If you were to use the Intel Xeon Platinum CPUs, in order to be able to do that search on that size of a database, it would take 12 nodes. Each node takes about 200 watts.
And so you're talking about an operational cost of about $54,000 a month. Now, we're able to do that same search with one node, which takes about 40 watts. I mean, granted, we need a host. So the host is about 200 watts. So our total solution is 240 watts per hour. So you're now dropping your operating costs by 80%. And so, I mean, this is a tremendous power savings and tremendous cost savings for the data center. I'm now going to jump into Gemini-II. As I mentioned, Gemini-II, we've had since last year, first silicon. First silicon has looked fantastic. There are a few bugs in it, but so far, all the bugs we've been able to work around with software patches. And so we're able to continue the software development, and we're going to start doing the benchmarking this quarter.
We're really focusing on the edge. Some of the interests and some of the SBIR wins, and I'll talk about what SBIRs are in a bit. We won for these edge applications. But some of these will be satellite applications or drones. Some will be under ADAS. It's unlimited the use cases that are available for the Gemini-II. We're actively putting in the software framework for them as well. We're focusing on the major platforms out there, the PyTorch and the TensorFlow. We have written and are planning to write some algorithms for specific applications, but we just don't have the manpower or the resources to cover everything. So it's critical that we get a compiler out. And so we're actively working on getting a compiler so that folks can write their own algorithms for our technology. And lastly, I want to quickly just introduce Plato.
So Plato is a future roadmap device that we have. As I mentioned, it'll be catered specifically towards LLMs. And a lot of these LLMs are going more towards the edge. This architecture we're looking at, we're looking at doing models like Llama 3.2 up to 90B. And we're looking at our technology to work at somewhere around 10 watts. So, I mean, it's going to be unparalleled power budgets for some of these large models we're looking. We're actively right now securing and getting funding for this. We want help in funding this roadmap. And so there are a few customers that we're actively talking to on being initial users and funders for this technology. Just a high level, kind of the applications and the software development. The Plato is a bit of a derivative off of Gemini-II, even though it's a completely different application.
And so it should be a fairly quick turnaround for that design. And we're using a 12-nanometer process for that. So the roles for the Gemini in our growth and the company's growth, as I mentioned, we plan to monetize Gemini-I specifically in the SAR. As I mentioned, we have a couple of customers that we're talking to right now, large ones for the SAR specifically. We've also started the discussions with the folks that are doing the index builds. And the feedback we've gotten on our benchmarking has been very positive. So we're looking to take the next step with them. Also, Gemini-II , we've had three wins now, SBIR wins. And I'll talk about them more specifically on the next slide, which is helping, A, offset some costs and, B, enable some future business on that.
Also, off the Gemini-II, we have what's called the Gemini-II L, which is a low-power version of the Gemini-II. So the Gemini-II runs around 60 watts, which is still very good for that kind of technology. But there are some applications that are really extreme edge or satellite that need something even less. We have, off of the Gemini-II, an offering that will be running between 5 and 10 watts. It can be essentially battery-backed if needed. As I mentioned, we're also looking to get some strategic partners, both for the funding of Plato, but also for any kind of licensing with our current technology. I mentioned the SBIRs. If you're not familiar with SBIR, it stands for Small Business Innovation Research. It's essentially a way for the U.S. government to help fund technology with smaller companies.
So there's been two SBIRs that we've won that we've already announced and discussed in the past. The first one was a direct-to-phase two with the Space Development Agency. And it was valued at $1.25 million. We've closed out about half of that right now. And essentially, that was a mechanism for the SDA to understand how to use our hardware in their environment. We then won, again, a direct-to-phase two SBIR worth $1.1 million with the Air Force Research Labs. And that one is just about 25% complete and paid. We announced this morning that we won our first SBIR with the U.S. Army. And this is a phase one for, again, Gemini-II to understand. It's really the phase one, which is valued at $250,000, is to really understand where Gemini-II can be used in U.S.
Army applications and then identify what algorithms would make most sense for the U.S. Army. We're anticipating or hoping that this would turn into a phase two. The phase two with the U.S. Army has net worth up to $2 million, and so besides the fact that it helps bring in funds into the company, really, it's for us to be able to enable a market for both the U.S. Army and for entities outside of the DOD to be able to use our architecture. We've recently submitted another direct-to-phase two for a space-level board, and again, with a worth up to $1.25 million. In general, if you look at the SBIRs that we have either recently submitted or plan to submit, they have a total value right now of about $6 million. Besides the SBIRs, there are other mechanisms and avenues to raise money with the U.S.
Government in actually larger amounts, and we're actively pursuing those programs as whether as partnerships or as just pure investments into the technology by these other elements, and so it's critical that we bring in more funds into the company, so if you look at the financial overview, if you look at the quarterly sales, which I mentioned have all been SRAMs, we sort of plateaued in the mid-fours. We have seen some upticks in that business. I'm not going to update our December quarter. We will be announcing in the next couple of weeks, but just to give you a reminder, we guided a midpoint of $5.1 million for that market. We're seeing the increase twofold. Number one is some of our existing customers have kind of depleted their inventories and are coming back and buying some more.
Then secondly, we have seen one of our long-term customers bring on a new program. We talked about this in our last earnings call. It's really to build out some manufacturing capabilities for a very high-volume, large chip company out in the market space. And so on that manufacturing system, they require some of our high-end SRAMs. And so we're seeing some very large orders and forecasts from this company through or at least into the first few quarters of calendar 2025. As I mentioned, we had a reduction in force in August. And so we were able to generate a savings on an annual basis of $3.5 million. And then lastly, we talked about the cash sitting at just over $18 million. So we've discussed this over the last couple of quarters. The board recognizes that GSI is not seeing the investment that we've put into this technology.
And candidly, we've been burning through some cash. And so we are exploring alternatives to bring in funding into the company. It could be in the form of, and by the way, we have Needham working with us in this endeavor. And it could be in the form of a sale or spinoff of one of our product lines or assets. We've been actively looking at licensing some of this technology for folks that want to use some of it for some of their ASICs. We've also opened up possible equity financing. We do have an active ATM, though we haven't used it. It's available. And we're also looking at other areas for funding. And then lastly, this technology is just too important and it's too unique that it has to come to market. So whether we merge with another company or we get acquired, that's certainly on the table.
We have to make sure that we put the shareholders upfront. And so just high-level summary takeaways. As I mentioned, the SRAM business over the next 12 months looks like it's going up. Certainly, the orders are in place for the next two quarters. So that looks promising. And certainly, the rest of the year is rounding out as well. The Gemini-II, the hardware, as I mentioned, we still have to fix a couple of bugs in hardware, even though we have the software workarounds. But really, it's the software that's going to be critical for enabling the revenues in the future. And so we'll be coming out with a couple of the algorithms by the end of this quarter and the beginnings of the compiler in the first half of this year. And so that allows us to start doing some industry benchmarking as well.
and as I mentioned, obviously, exploring strategic options.
At this point, we're open for Q&A. Great. Thanks for the presentation, Didier. I see we have a couple of questions in the chat. So yeah, let's get started on the Q&A. The first question is, what is the timeline for the compiler? And are there any different phases for different functionalities?
Yeah, so that's a great question. Yes. So the first couple of algorithms we'll get out will be in conjunction with one of the SBIR wins. So we're doing what's called a YOLOv3 and a YOLOv5, which is real-time object detection. And so those will be the first two that come out. And as I mentioned, one of the benefits of the SBIRs is not only do we get some funding into the company, and it also gives us a mechanism to enable orders within the DOD, but it also allows us to use what we develop outside as well for the commercial level. So the first ones will be YOLOv3 and YOLOv5. The compiler is an ongoing program. We'll have some of the beginnings of it in the next few months, but it's an ongoing program. It takes a little bit of time.
Unfortunately, we don't have enough resources in the company in that area to really get it out sooner. We have this balancing act of controlling costs and trying to get the resources in. So I'm looking at the first half of calendar 2025 to get some of the compiler out.
Great. Thanks for all the color on that. And then the next question is, can you guys talk about the award received today from the U.S. Army? Can you talk about the application specifically for this award and what milestones the company needs to achieve to transition into phase two?
So we're still defining that. So right now, think of it almost like a marketing effort with the U.S. government or, I'm sorry, the U.S. Army. So what we're doing is we're identifying where this G2 hardware can fit in for some of their applications. And then once we've defined that, is what algorithms are required for those applications. And so it's almost like a fact-finding mission, this phase one. And as I mentioned, we're hoping it turns into a phase two. And phase two is when you get into the exact application and then develop either the physical hardware board and/or the algorithms required for that specific application. So that's where really the rubber hits the road is in phase two. The phase one right now is really identifying where the best fits are.
Got it. Thanks for that, and I'm seeing no further questions in the chat. Didier and Lee-Lean, unless you guys have anything else you want to touch on, thanks for joining us, and for everyone in the audience, thanks for joining Needham's 27th Annual Growth Conference.
Thank you.