GSI Technology at Sidoti Conference: AI Innovations and Financial Strategies

On Wednesday, 21 May 2025, GSI Technology (NASDAQ:GSIT) presented at the Sidoti May Micro-Cap Virtual Conference, providing a strategic overview of its financial health and technological advancements. The company showcased its innovative AI chip development while addressing funding needs and market challenges.

Key Takeaways

• GSI Technology has invested over $150 million in developing its APU (Associative Processing Unit) technology.
• The company reported $20.5 million in revenue for fiscal year 2025, a significant increase from $4.5 million the previous year.
• Operating costs have been reduced to $5.6 million, with a cash burn of just over $1.5 million last quarter.
• GSI Technology is pursuing government funding and strategic partnerships to support its APU development, particularly for the PLATO initiative.
• The company believes its unique compute-in-memory architecture is underappreciated by the market.

Financial Results

• Revenues: Fiscal year 2025 revenue reached $20.5 million, driven by AI development and GPU support, marking a significant rise from $4.5 million the previous year.
• Operating Costs: Reduced to $5.6 million last quarter, with a cash reserve of $13.4 million and a burn rate of just over $1.5 million.
• Funding: GSI Technology is actively pursuing additional government funding, with $6 million in SBIR grants in the pipeline. The company has a market cap just under $100 million and insider ownership of 27%.

Operational Updates

• SRAM Products: The legacy SRAM product line remains profitable, with high ASPs from the Sigma Quad family. Expansion into aerospace with radiation-hardened SRAM is underway, with ASPs reaching $30,000 per device.
• APU Technology: The APU family, featuring a compute-in-memory architecture, includes Gemini one and two, with PLATO targeting edge AI applications. The company is focused on finalizing the Gemini two silicon and software frameworks in the coming months.

Future Outlook

• Revenue Growth: Anticipated from APU developments and radiation-hardened SRAM products, with production orders expected by the end of the year.
• Funding Needs: Essential for PLATO and Gemini two, with a timeline for securing funds by the end of the year for PLATO and within three quarters for Gemini two.
• Strategic Measures: GSI Technology is exploring equity funding, spin-offs, and partnerships, focusing on short-term government and military sales.

Q&A Highlights

• Market Focus: While hyperscalers represent a long-term opportunity, the immediate focus is on government and military sectors for revenue generation.
• Technology Feedback: Positive responses from edge AI partners, particularly for applications in drones and no-GPS environments.
• Stock Valuation: The company believes its stock is undervalued, with unique IP and technology not fully recognized by the market.

For a detailed understanding, readers are encouraged to refer to the full transcript below.

Full transcript - Sidoti May Micro-Cap Virtual Conference:

Operator: Lillian Chu. I also have the Didier Lazaire. He’s the VP of sales and investor relations. This will be conducted as a presentation followed by q and a. And if you would like to submit a question, you can do so in the q and a function at the bottom of your screen.

And with that, I’ll hand it over to you guys. Welcome.

Didier Lazaire, VP of sales and investor relations: Thank you, Anya. Good afternoon, and thank you for joining us. I’m gonna be quickly going through a little bit of where the company is today, but I’ll be focusing on the future. So, certainly, you know, we’ll we’ll have to put the safe harbor statement in there. The company was started thirty years ago by Leland Shu, our president and CEO.

We went public in 02/2007. Since day one, we’ve been partnered with TSMC. We get all of our wafers from TSMC. We started the company and and continue to offer the highest density, highest performance memories in the market. This product line is what’s funding our AI, which, you know, is the APU family.

We like to refer to ourselves as a self funding AI company. We’ve invested over a hundred and $50,000,000, into this APU development, on our own, and then, I’ll talk about the the road map as well. Well, we finished the fiscal twenty twenty five, year at 20,500,000.0. We have a 25 employees worldwide. We try to outsource all the labor intensive positions like the wafer fab, assembly, sales, and functions like that.

So the majority of our employees are engineers. We have spent a tremendous amount of effort to to get patents. This is important because as you’ll you’ll see in the in the presentation, we have a very unique architecture for our AI chip, and we wanna make sure it’s protected. We have $13,400,000 in cash and cash equivalents. We’ve never carried debt in the company.

Market cap is just under a hundred million, and we have a fairly large insider ownership of right now 27%. Quickly looking at our legacy product line, the SRAM area. So this has certainly been a you know, the stand alone division here is very profitable. We’ve had the majority of our growth coming from the Sigma Quad family. What’s nice about the Sigma Quad family is between the density and the performance, we’re sole sourced.

And so we were able to get a lot of design wins without any competition, which allows us to keep ASPs and margins high. Also talk about how we’ve taken this family and migrated it to, radiation hardened radiation tolerant for the space industry. So the two areas that we’re expanding, into are, aerospace and, the the AI, specifically the edge and inference with our AI chip. But what’s important to understand is that, you know, both of these are highly leveraged from our expertise in SRAM. As I mentioned, the the RET tolerant devices are a derivative of our current product line, while the AI chip or APU is done in an SRAM cell.

So, certainly, we have a lot of synergy between what we’ve done and where we’re going. Both these markets are growing markets. If you look at the space TAM, it’s growing at just under 10% CAGR. And, you know, as we all know, the the AI industry is just going crazy. You know, it’s certainly looking for tremendous growth at over 20% CAGR.

So this will be the last slide we’ll talk about, SRAMs. This is our radiation hardened intolerant family I mentioned. What’s nice about this market is that it’s really difficult to make a a robust part for space. So, therefore, you get rewarded with high ASPs and high gross margins. To give you a feel for what the the the the ASPs look like, if you take our our highest density, commercial part today, it sells roughly between 200 and 50 to $300.

You take that same part and you you robust it for space in in a red hard environment, you get upwards of $30,000 of ASP per device. For the red tolerant, which is a little less robust, you still get 3 to $4,000 per device. So, you know, obviously, you know, every time we we ship it, a dollar revenue in this market space, a a high percentage of that goes to the bottom line. The the market opportunity we’re looking at is about a hundred million dollars, and it’s our intent to get a minimum of 10 to 20% of this market over time. The rest of the talk will be about our APU, our AI chip.

So the APU stands for associative processing unit. What’s, you know, what’s really unique about this technology is that it’s really a true compute in memory. A lot of folks talk about CIM, and they they they toss the the the the acronym around very loosely. What most folks are talking about really are, at best case, near memory processing, and I’ll talk about what the differences are in a bit. We also have a a, you know, a tremendous amount of memory bits, which allows us to have a a hugely parallel processing system.

We have two families available now, Gemini one and Gemini two, and our road map is PLATO, which I’ll get into details on all of those. So if you look at how we compare to a a GPU and a CPU, there are some real fundamental differences. You know, first of all, I mentioned the number of bit processors. So if you look at a a GPU, they have dozens of ALUs, which are algorithmic logic units. You look at GPUs, they have thousands.

And so, you know, the GPUs were really used for AI because they are quote, unquote massively parallel. So you go to the our APU architecture. We have on the on our first generation device, we have 2,000,000 bit processors. So we’re extreme parallelism compared to GPU or CPU. So that’s one of the the differences.

The the but the major difference is really in the way that that we’re architected. If you look at a GPU and a CPU, their title was called the Von Neumann model. So they if they have a task they need to do, a computation or something, they have to go off chip to fetch the data from memory, bring it back, do the processing, and when they’re through, they need to write the data back to the memory. So there’s a constant back and forth flow of data being transferred. With our architecture, the the the the memory and the processing bits are in the same place.

And so we’re not going to go fetch data and bring it back. The data is there. And then so then we use the data for whatever function, whether it’s a search or a computation. And when we’re through using it, it remains there. And so that’s, you know, gonna save you all of the the power of having to go back and forth, you know, transferring data.

The other interesting part of our solution is that we’re a bit processing part, which means, as I mentioned, we have 2,000,000 bits. You can organize those in any method you want. You can have a one bit machine, a 10 bit machine, make up a number between one and two million. And so that’s important because if you look at a GPU, they’re hard coded. You know, you’re buying a 16 bit GPU or you’re buying an eight bit or make up something.

And so so when you got when you get one of those, that’s you’re limited to that. With ours, the customer determines what the resolution or the bit width is. And on top of that, they can change it from cycle to cycle. So in other words, if they say I need four bits now, but the next cycle they need six, they can go ahead and change the six on the next cycle. Not a problem.

And so that really also allows us to really future proof our solution. And what I mean by that is researchers are taking a tremendous amount of effort to try and optimize models and use cases, and they’re finding that different resolution or bit widths are, you know, the most efficient mechanism. And so for us, we don’t really care what that resolution is. It can be anything, five bit, 12 bit, make up a number, and we’re able to address that with with our device immediately. What’s also interesting about our solution is the way we’re architected, it almost resembles a memory.

And and as you know with memories, you can add memories together. You can kinda cascade them together. And so we can scale our solution. We can put multiple boards, APU boards in a server and have them act as one. So the scalability is certainly important when you see some of the sizes of some of these models and databases.

So a little diving down into some of the families. So Gemini one was our first family. It was really our way to illustrate our unique technology. And so, you know, we do wanna monetize it. So we we’re looking at a couple of niche markets, specifically fast vector search on on Earth SAAR image creation and database index builds.

Now in case some of you don’t know what SAAR is, it stands for synthetic aperture radar. It’s, you know, similar to, let’s say, LIDAR, which you might be familiar with, but instead of using beams, it uses microwaves. So it works well at night. It works well through weather. In fact, if you turn on CNN right now and you happen to see some images of something, I don’t know, war in Ukraine or something, the images are most likely, SAR generated images.

Gemini two, which we actually saw first silicon last year, and first silicon looked fantastic. It had, you know, a few bugs, but we were able to do software our software workarounds on all those. We will see the you know, our finalized second silicon chip, which we think will be production worthy within the next couple of weeks. With this device, it’s gonna have similar ish markets in the fact that we’ll be able to do SAR as well with Gemini two. The difference, though, is we’re looking for onboard SAR.

So as I mentioned, Gemini one would be for processing done in a data center or someplace on Earth. While with Gemini two, we’re able to bring that technology and put it on the satellite or the drone or what have you itself. And so that’s important because instead of currently, you’re taking images and data from the satellite and you’re projecting it to Earth to create the images, and and at that point, you’re making a decision. Now you can do the the image creation on the satellite itself and be able to take some kind of action immediately without having to come down to Earth. And then lastly is PLATO.

That’s our our next generation chip. So PLATO is actually going to address a different market. We’re looking at multimodal gen AI along with large language models, but we’re really looking at it on the edge. And what I mean what I mean by that is, you know, certainly, when most people think of LLMs, they think of ChatGPT, and they think of, you know, these these huge models that are sitting in data centers. And if you look at the GPUs today that are being used to address those, they’re hundreds of watts, and then some of the newest ones are kilowatts of power usage.

With PLATO, we’re looking to do these LLMs at the edge, and so a high power budget is not in the works. If you look at PLATO, we’re targeting it to be sub 10 watts. So this is something that could actually be powered by a battery if necessary. If you look at some of the early interesting discussions we’ve had with customers, it’s revolving around areas like drones and and automotive and surveillance and places like that. You know, drones is actually extra interesting to us because there are use cases where, you know, their their GPS navigation, and it’s being jammed.

And so now you have an issue. And so with us, with the onboarding of the PLATO, you can do some SAR imaging. And with the SAR imaging, you can see where you are on Earth by recognizing certain landmarks, and then you can navigate that way. And that’s called a, you know, a no GPS environment, and that’s, you know, very important for, you know, for our customers. If you look at the, the the software frameworks, for Gemini two and PLATO, they’re both gonna be, Python and TensorFlow.

We should have our first release of our the beginnings of our compiler stack in the next month or so. I’ll quickly go through this because I kinda kinda went through it already. Again, SAR on on Earth for Gemini one along with fast vector search. You know, just to kinda give you a feel for the technology and how it works in real life. This is something that was done with our first generation Gemini one part.

Some engineers at AWS a couple years ago looked back and said, hey. What what would it take hardware wise to do a 1,000,000,000 dataset search? And they said it would take 12 nodes of an Intel Xeon platinum. So that would be, you know, a 200 watts a node. It would be 2,400 watts per hour.

So we can do that same search with one Gemini chip that runs at 40 watts. Candidly, we need a host, so you can throw in, you know, a, one, Intel at 200 watts. And so we’re at 240 watts, per hour. So we’re, you know, 10% of their power. So, obviously, you know, that would tremendously lower the operating cost of the system.

That kinda gives you a feel for the for the technology power. Gemini two, as we discussed, will be more on the edge along with onboard satellite and drone applications. We are in the process of going after some some government funding to do the the radiation testing for, Gemini two that that will allow us to have, all the the test requirements done to be able to put it in space. So turning to PLATO here. As I mentioned, it’s, you know, it’s got the LLM.

But what’s interesting about our offering is that, you know, as I mentioned, we have this bit processor. So we can get down to one bit or two bit or whatever, you know, your fancy is. And so there’s a real push to do quantization in the LLMs. You know, these models are so large. It’s it’s just becoming too cumbersome, you know, on performance wise.

So they’re they’re taking some new approaches, and they’re quantizing these LLMs, and they’re putting them in low precision formats. And then they’re able to do that to to speed up, you know, the the the processing without compromising accuracy. And, you know, as we talked about, it’s low precision. We’re a bid engine, and so we we fit in absolutely perfect with that. And as I mentioned with our 10 watt power, we’re gonna be really enabling the the LLMs at the edge.

So the you know, how how you know, what are the what are the challenges and where do we fit in with our APU? So we talked about the high power consumption, and we talked about the CIM, you know, without the transferring of data so we can reduce power. You You know, as these models and these databases get bigger, they’re throwing more and more hardware at it. And it’s and the scalability is, as I mentioned, is a challenge with the GPU. With our s with our sRAM memory like architecture, we’re able to scale very easily.

With Gemini two and with PLATO, we’re gonna be bringing essentially data center performance to the edge. And then lastly, you know, with and I just discussed this with the, you know, the density growth in the LLMs. You know, they they’re going to this quantized model. And with our single bit structure, we’re able to support that, right off the bat. So where, you know, where does a p the APU fit in in our growth?

And and so we’re we’re looking to, you know, like I said, showcase and monetize SAR with Gemini one. It’s also enabled us to be able to to to have Gemini two be successful on the actual onboarding. We’ve also been very successful with SBIRs, which stands for small business innovation research, and these are grants from the government, and I’ll get into more details and some following slides on this. And then, the Gemini two, as I mentioned, is we have a a low power version of that, which will be for, you know, extreme mobile edge applications and satellite. Quick snapshot of our financials.

You know, the the revenues have been growing the last few quarters. A year ago, I wanna say we’re at about 4,500,000.0. So, you know, we’ve had some nice growth in the last year. The majority of this is attributed to the the build out of AI. And what I mean by that is these these s rams don’t go into data centers, but they go into the support of the ramp of GPUs for, you know, the AI build out.

Our largest customer now is KYEC, which we’ve just announced. They do the burn in, which is part of the manufacturing for GPUs, and some of these new GPUs are starting to ramp. And so they’re building out these burn in systems to support that GPU launch. These systems require, you know, some some high end memories from us. Another area that we’re seeing significant growth in is with the the emulation and simulation of some of these GPU hardware designs and any hardware designs for that matter.

So these designs are very complex. And so companies, to save money, you know, they don’t wanna waste money on mass sets and and so on. They’re doing a lot of emulation in software to emulate the hardware. And the and one of these customers who I’m sure will be a 10% customer for us that we’ll announce shortly use uses our highest density and our highest performing part for that, for that emulation. We’ve taken some strategic, measures to lower our costs.

So our operating costs, have come down to 5,600,000.0 in this past quarter. I mentioned earlier that we have, 13,400,000.0 in cash. We burned, just over a million and a half last quarter. And, you know, certainly, if you look if nothing improves at this point, we have about two years worth of cash. But, certainly, we we’re anticipating some revenues growing, so we’re looking to to slow that burn and and get rid of it in the future.

So I mentioned the SBIRs. So in the last year plus, we’ve actually won three of them. One was with the space development agency, one with the US Air Force labs, and the last one recently was the US Army. So the the first two were direct to phase twos, and so one of them is 1,250,000.00, and the other one is 1,100,000.0. So we don’t treat this as revenue.

We treat this as an offset to r and d cost. And so think of it like essentially grants from the from the government to to help pay for our our design efforts. But besides bringing in some grant dollars or r and d dollars into the company, what this has also given us is exposure to these DOD units that will allow us in the future to to get some revenue out of these areas. Now we’re continuously submitting new SBIRs. Right now, we have enough in the pipeline for about $6,000,000.

Lastly, besides these SBIRs, we are also pursuing other funding sources within the government to help pay for some of the deployment and and development of some of these future products. So just quick near term milestones. Gemini two, I I mentioned that we’ll have the the second silicon within the next couple weeks. We’ll also have the respun leaderboard that these Gemini two parts go on. We’ll have that in June.

At that point, we have what we feel will be the production worthy hardware, and it will also allow us to ship some of these boards to a few customers, including the Air Force, which is one of our SBIR milestones. One of the SBIRs was an algorithm development, and it was specifically for YOLO, which stands for you only look once. And it’s a way for folks to identify certain objects or what have you very quickly, real time object detection. And so we’ll be delivering those algorithms this quarter. Now, lastly, if you look at, you know, where we are, you know, certainly, 13,400,000.0 is is not what we are looking for.

So there are certain areas that we’re looking to raise money. So PLATO is one of them. So we’re looking for the, you know, the, funding for PLATO. PLATO, however, we’re looking to fund via customer partners. But for Gemini two, we are looking to obviously launch and promote Gemini two, and we need some other efforts and extra resources for areas like software and and, compiler stacks and so on.

So we are, right now, working with Needham and Company to raise, some equity funds. But with that said, we’re also a publicly traded company, and so we have a a duty to our shareholders. And so we are looking at, you know, other avenues as well as possible spin off of assets or IP licensing. And, obviously, we’re you know, at this point, we would we would be open to mergers and acquisitions. You know, our our first choice, is to raise some funds and to be successful on our own, but we certainly are open.

At this point, I’d like to open it up to questions and answers.

Operator: Okay. Thank you so much. That was a good overview. We have a couple of questions here from the audience. So first, can you talk about the deal pipeline?

How are you developing sales into hyperscalers and defense integrators?

Didier Lazaire, VP of sales and investor relations: Yeah. So we’re so, candidly, we’ve we’ve had the discussions. We’ve talked about this in the past with the hyperscalers. They’re a longer term play, for us, and what we found is, you know, we we wanna generate the revenues as quick as possible, and we’ve seen a lot of interest from the more of the edge folks and, like, especially with the government and the military. So while we’re still talking with the hyperscalers, the majority of our efforts right now, candidly, are around, getting the successes, for the short term sales with the with the government and the military.

Operator: Okay. And I just want to remind the audience, if you want to participate, you can submit your question in the q and a function at the bottom of your screen. We have another question here. How do you see revenues growing in the future? Do you see this, like, flipping a switch, and then we will see revenue grow rapidly to 10 to 20,000,000 a quarter?

We kind of been running in place or going backwards the past two years.

Didier Lazaire, VP of sales and investor relations: Right. So, you know, with the SRAM, as we see that slowly coming back, we also when I when we say the SRAM is specifically the commercial SRAM. Now we also have, the radiation hard intolerant. We did mention on our last earnings call that, we shipped a few parts this past quarter into a customer that’s looking to, launch their system sometime. It looks like it’ll be the end of neck or I’m sorry, the beginning of next year, which means they will probably take product from us at the end of this year.

It’s, you know, it’s a the visibility isn’t perfect as far as the timing, but that would be our first, radiation hardened production order. And so, certainly, that would that would add. As far as the, you know, this exponential growth that that your that your, person asked, that would have to come from APU. And, and right now, you know, the hardware in the next month or so, as I mentioned, we’ll have the production worthy. We still need to get all the software tools out there, and that’s gonna take us, you know, a quarter or two to do that.

So we’re looking to do the some of the seeding in the second half of this year to get some of the pipeline going and looking into next year for any kind of the revenues.

Operator: Okay. And what’s been the feedback from early edge AI partners around latency and deter determinist?

Didier Lazaire, VP of sales and investor relations: I’m sorry. The the the feedback from the AI, the edge AI folks?

Operator: Yeah.

Didier Lazaire, VP of sales and investor relations: Yeah. So, yeah, so the the feedback has been has been really good. As I mentioned, you know, one of the use cases I mentioned was the drone in the in a no GPS environment. You know, that’s just one example. I mean, there’s a lot of use cases that are coming out where they need to be able to do not l l just LLMs, but computer vision as well at the edge.

And and most of the focus by the larger folks in the industry are really focused at the data center LLMs. And so we we feel we feel like we have a nice niche, on the on the edge with this 10 watt part, that we think is gonna be very successful. And, certainly, the the early indication from the discussions we’re having is are very positive.

Operator: Okay. And another question here. What is the timeline to raise money for the initiatives that you had mentioned?

Didier Lazaire, VP of sales and investor relations: For okay. So as I mentioned, you know, there’s two areas we’re looking to raise funds. One is for the development costs of, PLATO, and those those are have been active. And, and, certainly, there’s a a few benchmarkings we need to do on Gemini two for at least, two of the folks, to demonstrate the, you know, the reality of it, and then and then that’ll start that development. So we’re looking for that, funding within, the next two quarters, so by the end of the year.

For the other half, which is the equity raise that I mentioned for really the the launching and the success of Gemini two, Really, we need that within the next three quarters or so is what we’re is what we’re focusing.

Operator: Okay. And I have one last question here. Given trends in the stock price over the last couple of years, what do you think the street may be underappreciating?

Didier Lazaire, VP of sales and investor relations: Think they’re underappreciating the uniqueness of our of our IP and our technology. You know, certainly, I understand that, you know, the the stock doesn’t reflect what we feel the technology is worth. And I and it it’s clearly we need to get a couple wins to show that it’s real. But I I think that the underappreciation is the the value of the technology, the IP, you know, all the patents that protect all that. So, certainly, that’s worth a lot more than than our market cap is today.

Operator: Okay. Thank you. And we time is up, so I just wanna thank everyone who participated. Actually, we have one more questionnaire. Do you believe GSIT is the only true CIM company, and can you discuss what hurdles are, to product recognition in the marketplace?

Didier Lazaire, VP of sales and investor relations: Thing else that is that is doing true compute you know, in memory. You know, we’ve seen a few papers out there. In fact, there was one paper by a very large company. It was sometime last year where the the the headline was CIM. And then as you start reading the paper, it’s clear it’s not CIM, but it’s near memory.

And their their last paragraph, which was a, you know, a recap review of of the paper says, and in our near memory solution so I and like I said, a lot of people talk about it. We have not seen anybody who really is doing what we’re doing. And, certainly, we feel that we are well protected with our patents on on that technology.

Operator: Okay. And now, actually, time is up, and there are no more questions in the queue. So I just wanna thank everyone who participated and to Didier and Lillian who, joined us today. And I’m gonna hand it over to you, Didier, for some closing remarks before we close it off.

Didier Lazaire, VP of sales and investor relations: Again, I think we’re well positioned with our technology. You know, we just have to get over this last little hurdle. And then we do think that some of the milestones will be kicking in in the second half of this year. Thank you for joining us.

Operator: Thank you. Thank you, everyone.

This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.