← 返回

2026.10.05 | TB | SemiCap | Aehr Test: AI Burn-in Demand

原标题：Aehr Test Systems – AI-driven Burn-in Demand & Wafer-level Adoption

#semicap

Third Bridge · 2026-10-05

Analyst-led • 28 Sep 2026

Companies: AEHR US

Warren Wartell (WW) Co-founder, SVP & Chief Intelligence Officer at SemiSphere LLC

Garrett Waters (GW) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

GW [00:00:02]: Welcome to Third Bridge's Interview entitled Aehr Test Systems – AI-driven Burn-in Demand & Wafer-level Adoption. I am Garrett Waters, and I will be facilitating today's Interview with Mr Warren Wartell, former Senior Director, Global Test Services at Amkor Technology.

Warren, before we start today's Interview, please state I agree or I disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

WW [00:00:42]: I agree.

GW [00:00:44]: If you would, could you please begin with a brief introduction to your background?

WW [00:00:50]: Warren Wartell, and I've been in the semiconductor industry over 30 years, a lot of that time in product development. I've worked in the fab, a lot of new product development with a few start-ups we successfully acquired. Over the last four-plus years, was with the major OSAT, Amkor, focused on growing their test services business worldwide. We had a lot to do with burn-in for some of our big automotive customers and several other AI customers and quite familiar with that market space.

Questions

00:01:29] How would you describe Aehr Test Systems today, and how do you think its business has evolved from SiC [silicon carbide] for EVs towards AI processors and silicon photonics?

WW [00:01:52]: Aehr is a specialist in the product reliability, production reliability screening, basically burn-in equipment. They're a mature company. I think they're going on 40 years in the industry, a long industry reputation as a trusted, certainly package-level burn-in systems provider, maybe not top tier recently just based on some of the OSAT selections for the more high-power burn-in system. It seems that might be changing with some of their newer capability and this wafer-level burn-in plus testing is a novel concept that's gaining some traction. I would suggest they've had a strong reputation for doing package burn-in plus testing, not just a static burn-in, and strong reputation, a trusted supplier. I think right now, as they're evolving from silicon carbide for EVs, that was their foray on this wafer-level space, where I think they've gotten some traction on some early AI processor, at least one hyperscaler is committed and another was evaluated, and they're getting in on some silicon photonics space. Their Fox-XP system is this wafer-level burn-in plus test novel concept, and exciting to see what's going on there and how that might drive the revenue. It certainly has spiked their stock attention on the market.

00:03:33] How would you say burn-in requirements are changing with each new generation of AI processors? What is becoming more complex and demanding, specifically in the test process?

WW [00:03:55]: Burn-in has obviously been going on for a very long time. I think it was pretty much a foundation on the wafer level for memories, but it's always been or had always been a more static burn-in where you would programme the wafer up, these memories up in a specific memory state and look for leakage. Bake for anywhere between one and three days at a specific temperature, and you're looking for any substrate defect or oxide leakage path that might corrupt that memory during the bake, one way to find out early fail studies. With these AI chips, obviously, the process nodes have gotten so aggressive. The packaging interconnect as well has got very elaborate and higher power. Certainly, a lot of thermal management demands with a highly integrated CoWoS-type or 2.5D packaging. More compute tiles, these HBM stacks are growing, dense interconnects, finer pitch, just less room for failure when you get one trillion transistors in a die.

I think that's obviously driving more opportunity for error in the same given space and maybe more protective nature. Certainly, there's things happening with the level of integration that your value count in the package has gone up enormously, very expensive substrates, a lot of die being packaged together, whether it's the HBM stacks or multiple CPUs and interface die and the other processing that might go on package level to make all that possible. It's a very expensive product, and a built-in desire, if you can weed out any defects early, man, that's a good economic advantage. Whether that's really the case or not remains to be seen. We can probably dig into that a little bit more. Certainly, these fine-pitch, high-count contacts, the thermal management is definitely growing with complexity in the AI space and ever-growing integration, more die in the HBM stacks, more memory, more of everything in a given form factor.

00:06:26] Could you quantify how much more burn-in capacity a current generation AI processor might require compared to previous generations?

WW [00:06:40]: That's a good question. I think there's a level of coverage that you want to try to do, but it's not typical that you're going to do super aggressive testing in a burn-in environment. You want to run the product in a way that mimics what might be going on to get the processing power and heat up to a certain level, but you're not going to be testing all the functionality during this burn-in stage. You're trying to get it to more of a performative state and then stress it in a way that hopefully accelerates these potential defects to weed out earlier under the high-temperature conditions. On a time-per-part basis, I'm not sure that it's really growing with what's going on, but certainly, it scales linearly, especially the products and companies that choose to do 100% burn-in or even a sample burn-in is going to grow with volume. On a wafer-level situation, one of the unique things with the Fox-XP system from Aehr is it's touching every die at once in these cartridges, these cassettes.

That doesn't necessarily mean it's testing every die at once. My understanding, it goes through cycles. I think it can test one-quarter of the wafer, and that's going to be depending on the power in one go, and it needs to cycle through all of them to get through a run. The more power a die takes, it might have some impact on throughput, but I think that might be a lower order. I think the time per product probably doesn't scale like it would in a normal die-level test or package-level test. This is more of a speciality reliability test that I don't know that the test time per se is growing. Bottom line, scaling more with volume than with complexity driving a lot more test time.

00:09:03] You mentioned that the value of the finished package has gone up enormously. Could you quantify that? What is the rough dollar value of a current AI package vs prior generations? How could that change the economics of wafer-level burn-in?

WW [00:09:23]: We run this gamut of what is an AI chip. You have AI ASICs that might be a little bit less complicated. It might use similar type of packaging, but it's not going to be the workhorse, these big GPUs or heavy compute accelerator-type products. Those might be in, probably on the low end, thousands of dollars, where the high-end heavy GPU, high-end accelerator-type compute chips can run up into the USD 10,000-20,000, I've seen estimates of USD 40,000 of stuff that's coming down in package value once you get through all these expensive components, the HBM stacks, the die themselves and expensive processing and a very expensive substrate to stitch everything together. I thought top line right now is in the USD 20,000-25,000. I'm sure there's price points that are different depending on volume, but that might be the top end of the chain right now. You can see if you've got a potentially USD 20,000 product, you don't want maybe USD 800 or USD 1,000 die taking out that whole package if it could have been weeded out and not coupled with the rest of that content. You're getting it early in line, you're screening it out. If it really is a die-level defect, boy, you really would love to be able to not take down the whole ship. I look at it as a flat tyre taking out the whole car.

00:11:19] As we see the integration of more compute dies, HBM [high-bandwidth memory] and advanced packaging, how should we think about the cost of discovering a single defective die once the package has already been assembled?

WW [00:11:39]: To me, this is the big question. If you do this wafer-level burn-in, they're not doing it on… Part of the processing they do is they're attaching die on substrate or they're attaching one die on top of another die. The type of wafer-level burn-in they're doing is only on the processor portion, the big GPU, the TPU, CPU accelerator processor chip. The big question is, are your defects, are the reliability concerns contained in that die and not in the rest of the package. If it's true, and you can catch those types of defects early with the wafer-level processing and you're screening them out from ever getting populated with the rest of the package, that could be a very big savings. Just the example, if it's a USD 15,000-20,000 package, there's opportunity as well. They might be selling it for USD 50,000-75,000 yielded out to the customer. That's a substantial savings if you could have saved that because there's not really… As far as I know, there is no recovery. You can't find the die and replace it once it's already in the package. The amount of reworkability is very limited on these very advanced packages. Screening it out early, that could easily be a 10x, even up to 100x savings in potential recovery of the very, very expensive package that got yielded out.

00:13:35] Could you help us better understand the TAM for wafer-level burn-in within AI processors? How large could this become over the next 1-2 years? You mentioned how this opportunity is growing.

WW [00:13:49]: It's a great question. The big if, and right now, it's focused seemingly on a single hyperscaler customer that's evaluated it, running it in some level of production and clearly have ordered follow-on systems in the AI space. That's not the only space that Aehr is participating, but that's certainly the eye-raising opportunity because of the explosive growth in AI and the revenue that it drives. The economics are very different to anything else right now. Photonics has a chance to be a similar situation, but right now, AI is the big headline on potential growth and the potential economics, if the failures that you're catching on wafer level are really eliminating the need for this package-level burn-in. If that's the case, you may even justify a two-pass flow. You still keep package burn-in. Maybe it's not 100%, but you leave it in as a sample rate. Certainly, if that's adopted and it's demonstrated and convincing to other hyperscaler customers, particularly players in the very high-end compute space, I think they would run at the chance. I think it's been a pretty slow spread rate. Right now, there's only one other customer as far as I know, that's adopted from an evaluation standpoint, but hasn't yet placed any production orders.

I still think it's very much two big unknowns in my eyes. One, the existing customer, is it really 100% production-type situation, where they're really planning to run it on 100% every wafer that goes through their line, and it really is replacing production package burn-in or if they're still keeping package burn-in at some level, a monitor and still evaluating the concept. That's one big question, or they're doing it in a sample rate, and every nth wafer, maybe it's a few per lot, who knows. It's not going to scale anywhere near the same rate, and there may be a strong potential for them to further tweak the sampling rate of not having to run it as frequently as they do now. The other big question is, will it spread to other AI, big chip compute GPU-type players? I think it's too early to tell. If it does, boy, I think the volume could go through the roof. Certainly, that would be hugely beneficial to Aehr. I think they're trying to prepare and have some production capability to generate systems to make this a step-function change in a new need and an economically viable need for this type of wafer-level burn-in system.

00:17:21] Regarding wafer-level burn-in, if the market for AI processors grows at 20% annually, how much faster do you think Aehr Test Systems' revenue could grow if it gains market share and still expands within its existing customers? If the market grows at 20%, by how much could the company grow?

WW [00:17:43]: I think for the existing customer, it's going to probably grow pretty linearly with volume. I think the big traction is if it spreads to several other hyperscalers, especially some of the big volume guys, and they're convinced that it's saving them a lot of cost by screening out these loss of expensive packages later in life earlier, and then I think they'll be very aggressive to adopt it. I could easily see it growing. It takes a certain amount of time to get these things folded in and teach the OSATs how to use it properly. There's new steps involved. It's maintenance, it's installation. Mainly, it's just qualifying the system and the flow for these new players. I could see it easily a 5x expansion just for a few other AI guys coming in, potentially 10x of the systems that they've either shipped and have bookings to show for. It's been slow. This isn't a new story. The wafer-level burn-in systems have been around with silicon carbide for several years. Even in the AI space, I think it's going on two years. I'd say it's still been pretty slow to get any rapid adoption. The bookings have come in, but system deliveries lagged last year from what I can tell. They're still trying to catch up with orders, or maybe the customers slowed down delivery need date could be a few things going on there.

00:19:48] From an OSAT [outsourced semiconductor assembly and test] perspective, what would you say are the biggest risks or barriers to adopting wafer-level burn-in?

WW [00:19:58]: OSAT perspective is very much revolving around something called fungibility, which is how flexible is your equipment to run multiple products, multiple customers, ideally multiple applications. Nobody likes in the OSAT world to be stuck with a custom, very expensive piece of equipment. That's one problem. There's not enough players that use this for it to be an attractive desire from an OSAT perspective. Second, it's a pretty expensive and timely. It's an extra process step in the line to run this wafer-level burn-in. You can make the argument, well, it's getting rid of package-level burn-in potentially, but still it's a new process. The alignment is a slowish process. You don't know really what your consumable bleed is going to be on these WaferPaks as they are consumable items that only have so many touchdowns. I believe there's still probably a learning curve going on, on what it's touted to be and what it really is and when it starts to impact yield. I think there's awful lot of concern about potentially overrejecting because maybe the technology is not so mature yet. There's still learning going on. Those two big items, probably three big items, the newness of it, the cost is a barrier. The fungibility is probably the bigger barrier and just getting your process flow aligned and being convinced that it's really helping, and not just adding an extra step and potentially an overrejection step in your flow.

00:22:00] What determines whether production burn-in is performed at the wafer, package or system level? At what point does wafer-level burn-in become economically necessary rather than optional, if at all?

WW [00:22:15]: Another difficult question. I think nobody wants to do burn-in. You do burn-in because you have to. You've got defects that are escaping. You've seen early fail indicators, and you need this level of stress to try to catch these types of failures before a product ships, or sometimes it's causing some shifting in some parts that you want to have well-understood and tested after the fact, so it doesn't fail any other technical specs for the device. It's a lot of data collection. It's understanding your failure criteria or failure root causes and then being convinced that those root causes are die-related and not package-related, and you can focus on doing this type of a wafer-level burn-in to basically simulate those conditions in the package to get those failures to be detected. There hasn't been a lot published on the types of failures that are being detected. That's a concern to me because sometimes a customer will adopt this protective type of test when it's a new process.

The lack of disclosure on who the hyperscaler customer is, the type of process that they're using and the types of defects that they're catching, hard to be super convinced that it's a broad need yet. It could be, but it doesn't seem that they've done a great job convincing the huge players like an Nvidia that they need it because I'm fairly convinced that if they were economically convinced it was saving them material, especially very, very, very expensive packages, they would jump at the chance and they would be aggressively qualifying and implementing it in production. I think there's some special wrinkles with this lead customer that may make their situation somewhat unique. It might be a different process, maybe a different supplier, but that would be my engineering intuition. I don't know that there's anything publicly released. They certainly haven't disclosed who the customer is. To me, that would make more sense.

00:25:01] Aehr Test Systems said that its lead AI processor customer moved all production burn-in screening from the system level to the wafer level. How much downstream screening cost and complexity do you think that shift realistically eliminates?

WW [00:25:22]: I saw that announcement, and I have to assume that system level really means package level, that they're no longer doing 100% production package burn-in and they're focusing their production burn-in on the wafer level with this Aehr system. I believe they still have some level of package-level sample burn-in to monitor for a period of time, probably for a long period of time, at least a certain threshold of volume before they're convinced. I don't know that it's any proof. It's a specific customer claim that it's going to be a widespread benefit for the other players in the space. Certainly, the potential savings is you've got expensive boards, expensive burn-in ovens on the package level and that can be phased out, no longer needed. I don't know that it's on a hardware-for-hardware basis, it's not clear to me that the wafer-level system is a more efficient cost way of testing.

The economics pay off when you catch the problem earlier and not having a very, very, very expensive package thrown in the trash can. I think if you look at a like-for-like basis, it may even be a little bit more expensive to implement the wafer-level testing, expensive systems. These WaferPaks are expensive. They're consumable just like the burn-in boards and probably more costly than the burn-in boards. I don't know that they wear as long. Certainly, there's some learning, as I mentioned. You don't know how much use you're going to get out of them before it might start to shift your yield at wafer-level burn-in plus test and potentially, we talked about the risk of overrejection. That's something I think they're probably monitoring very closely.

00:27:36] You mentioned that adding a few more AI customers could potentially drive a 5x or 10x expansion. If that scenario plays out, how could that translate to annual revenue for Aehr Test Systems' AI business? How many customers would need to be added?

WW [00:28:01]: I think the story seems to be that they are going to get several hyperscaler customers to come in and adopt it on a production level. It's going to drive that 5x-plus equipment revenue. I think they need to do that to get their margins up to a level that I think can support the valuations that the stock price certainly is trading at now. Usually, the burn-in equipment vendors isn't like cream of the crop margins compared to other equipment key to semiconductors. My understanding is they had a pretty major factory expansion capability to be able to supply a lot more. Certainly, they want to run their factory at high utilisation to lower their cost on a per system basis, and that can drive a lot more revenue and higher margin on a per system basis. I think the story requires them to probably win another two AI customers that are probably as big, if not bigger, than what they've already got commitment for already, and that the story also suggests that there's going to be follow-on orders at the existing company to keep up with volume. I think those are big ifs. It's very much a monitoring perspective from what I can tell and tracking not just bookings but system shipments really closely and paying very close attention to what they disclose on customer names and specifically how they're using it in production.

00:30:08] You said that margins are worse off for Aehr Test Systems' equipment. How could the company improve its margins? How is Aehr Test Systems positioned in relation to competitors?

WW [00:30:25]: I thought the suggestion was… I read on some reports that their costs have gone up in securing some of the materials to build their system and certainly the WaferPaks. I think that may strain… Also, I think they've had a lot of cash outlay to build up capability of the new factory. A fuller factory is going to be a much more efficient situation to lower cost per unit. I got to believe that's going to be… The numbers I had seen was somewhere in the high-teens to very low-20s on margins, which would be not that out of whack for burn-in equipment. Might be some of the higher end tend to be a little bit higher margin. I thought that there was some room for improvement once they build out their factory.

00:31:26] What are some of the biggest levers Aehr Test Systems has to lean on to improve margins? Is it just the expansion of the factory? Do you have any other thoughts on that?

WW [00:31:37]: It's a number of things. I think once you get more system commitments, you can try to negotiate for better volume on your own supply chain. You get volume discounts there. Certainly, being able to do just output more for the same factory, you have a lot of fixed costs in a factory, especially in a recently expanded factory, and just trying to recoup that by pushing more systems out. I think the more systems they get and the more in production, they start growing what's probably higher-margin WaferPak cassettes because there's not really a third-party player that can replace those WaferPaks. You're pretty much tied in, I think, with Aehr to provision that. As they start going through production use, those are wearable consumable items and those WaferPaks, I think, go from USD 60,000 up to a few hundred thousand and probably is a good revenue, and not so much revenue because the systems are going to generate a lot more revenue, but good margin drivers for Aehr on the replacement piece of the system.

00:33:07] In August 2026, Aehr Test Systems announced a USD 22m follow-on order from its lead AI processor customer. From your perspective, what's a realistic ceiling on Fox-XP and WaferPak spending for a single AI processor programme once it reaches full production?

WW [00:33:26]: That USD 22m order followed the, I think, what was announced in February, the USD 14m order. I don't think that's the programme's lifetime pack, more than likely, it grows with volume. Best of my understanding is the management suggested that these systems that are generally sold with an alignment piece, an aligner and some initial order of WaferPak. You're probably talking about a USD 5m per system type of… USD 5m-6m per unit. It all depends if the lead customer has what they need on a sample basis or if it really is 100%, and it's going to grow with volume as they grow volume and need more systems. I think that should be pretty clear, but there really hasn't been much disclosed on volume that the lead customer is running. One would think there would be follow-on orders within 3-4 months, even beyond what they've already committed to keep up with growth rate because I got to believe lead times are not instant.

It's probably somewhere in the order of 4-6-month lead time for these systems. A lot of times, they'll get in orders before the end of the year, planning ahead, and that would be one of the things I'd be watching for. These things tend to come in multiples as the volume expands. Usually, you order a couple of systems, you never want to be leaning on just one system for evaluation. Usually, you'll order two or three, and then the next step as things go into volume might be another 2-4. This seems to have followed that flow, and about six months apart from the orders. You expect another order to happen right at the end of the year or maybe early next year to be more convinced that it's a production volume growth and not just a sample-based wafer-level burn-in.

00:35:57] In July 2026, Aehr reported approximately USD 100m of effective backlog. How should we assess the quality and visibility of that backlog? What are the biggest risks to converting those orders into revenue on schedule?

WW [00:36:19]: I have no reason to doubt that they were real orders. They just came in after the close of the previous financial quarter. I don't think they made up, "Oh, this is verbal." I think they really did get an order. It just came in late. The big question and it really doesn't disclose with too much granularity is when the customer put in the delivery date and how much shuffling there's been in the delivery date. I think there was already signs of it pushing out, and it was probably a lot more customer-led than Aehr ability to ship the system. They're still building these things. It's not like they've been building hundreds and hundreds of them. There could be some delays on the build time.

My guess is the customer is a lot more empowered to push out or pull in when they need it. I don't know that Fox-XP has slipped in their ability to build and ship the system. I think it's much more likely customer demand than it is Aehr ability to build out. What could cause the customer demand to slip, and boy, it could be a whole host of things, but you think they'd be super aggressive if their volume had grown and this was really part of the production volume, it would be an absolute priority to get it in ahead of need. My guess is maybe the ramp slipped a little bit, and it just hasn't grown quite the level that they'd hoped for other downstream reasons, either the die needed another round or their end customer demand wasn't quite what it was when they thought.

00:38:27] Let's discuss the competitive landscape. Who are Aehr Test Systems' most credible competitors in wafer-level and package-level burn-in today? Where would you say that the company is differentiated?

WW [00:38:42]: I would divide it into three categories. One is we talked about the static wafer-level burn-in. Lower tech. There are companies like Pentamaster, and I think there's a company called Semi that participates in that more traditional static burn-in. I don't really think Aehr competes there. Then you have the traditional package burn-in. I would point to two. There's a third that's not so much of a third party, but it's built in-house. At KYEC, they have a lot of their proprietary burn-in system that I think they've used for a long time, they have the most capacity in the industry. On the package level, I think it really boils down to MCC, the Micro Control Company. AEM has a high-end AI processor burn-in system and their AMPS system. Those would probably be the two biggest guys competing against their Sonoma on the package level. On wafer level, what they're doing is pretty unique. There's only two guys that I know of that have announced a wafer-level burn-in system. They both seem to be focused on silicon carbide, which has a little bit different needs in the AI space. I don't think it's quite the power level that you need in AI.

Semi, they've announced a system, and there's another company called Nexus, I believe. Is that right, Nexus? I believe so, that recently announced a system. There's supposedly some IP concerns and maybe Aehr is stepping up and challenging. I don't know all the details on that. There's really not another wafer-level burn-in plus test player out there. Nobody is in production in the AI space. Could there be? Yes. I think there are companies capable of pairing if it becomes a legitimate need. I could see a Teradyne or Advantest entering the space with a partner. I don't see that. I don't believe MCC has announced or AEM has announced a competing wafer-level solution. Either they're not convinced of the technology, I think they would probably be capable of developing it. There'd be a lead time. I don't see any of the technology with them finding a partner to be insurmountable within maybe 1.5 years if it really became an ensconced up. We absolutely need this for these different applications. No two ways about it. It's hugely economically impactful.

00:42:27] Where does Aehr Test Systems' Sonoma system specifically win or lose vs MCC [Micro Control Co] or AEM?

WW [00:42:37]: I think it's probably with the same… It could be with the same class of customers. They're both very targeted high-end systems. I think MCC has been more successful in selling their high-end systems to a lot of the automotive processors and other players in the AI space. That may be changing based on some of the new orders coming in for Aehr. I think Aehr's really tried to focus on the very high end, very high power. I don't know that they have a sustained advantage over MCC. I thought at least for now, they have the highest-power system available in the space. That was my understanding.

00:43:29] In July 2016, Aehr Test Systems said that high-volume Fox-XP production test cells typically sold for about USD 4m-7m, so a pretty broad range. However, management also discussed wafer-level systems at roughly USD 5m, including WaferPaks, with some configurations above USD 6m. How comparable are those systems? How much do you think throughput or economic value to the customer has improved over that period?

WW [00:44:05]: The latter is difficult to know because it just hasn't been much publicly announced. You got to assume that the company involved and in bed with them sees the advantage or else they wouldn't have committed the dollars that they had to order follow-on systems and make the announcements they have about either turning off production package burn-in or maybe leaving it as a monitor, but they seem to be full-bore committed and catching whatever defects they believe they're catching. If those defects are shared with other products, not clear. It's also not clear if they are able to eventually address root cause so they can mitigate the need for burn-in or maybe go to a backed-off sample level. Let's take it for granted that it was convincing they're running with it. They've got follow-on orders. The variance in pricing, I think, has a lot to do with whether that alignment system is included and what type and how big your initial load of WaferPaks is on the initial deliverable, which is going to drive something on the low end that could have been… I've seen as low as USD 3m-3.5m all the way up to USD 6m might be a heavy investment in WaferPaks and certainly the alignment system.

00:45:38] Aehr Test Systems describes individual die temperature control through the WaferPak contactor as a key competitive advantage. Why is that capability so important for the wafer-level burn-in? Do you think that truly gives the company much of an advantage over competitors?

WW [00:45:59]: We've seen in the test space even on just traditional probing of these wafers that that individual die control and having a very complex, well thermally managed wafer chuck is crucial for these high-end AI processor GPU-type die. You've got to avoid these hotspots and risk the overstressing of the devices. The die doesn't heat equally. You can get these really big spikes in certain parts of the device. You've got to have a thermal mass and a very active chilling system to manage that as the part goes into various forms of operation. It's the same thing what they're doing in burn-in. They're still operating the die. It's a dynamic test. It's going to get locally very hot in different spots. You've got to make sure that your technology, this wafer cassette is dissipating that heat and spreading it. One die, maybe on one edge of the pack or in this case, one edge of the wafer, isn't getting significantly hotter than another. Then you have an unbalanced condition, and you're not stressing the parts equally. They have to have a multi-sensored approach. They have to be monitoring it, and they have to have a very strong, actively controlled cassette for thermal management.

00:47:48] Aehr Test Systems' management refers to WaferPak as a consumable input that drives recurring revenue from the installed base. What does consumable really mean in terms of usage, replacement frequency and useful life relative to the underlying Fox system?

WW [00:48:10]: I mentioned before, I think that's one of the learning items. There's just not enough proof of concept to know what it's going to be. I've seen numbers as low as 5,000 touchdowns all the way up to 20,000 touchdowns. There's different interpretations of lifetime. A guy who's focused on the testing, you want to make sure that's nothing shifted outside of your testing window. Even though it may not be destroyed, it was enough to cause a shift and that could push you over the edge on being usable. I think that is something that's being analysed very closely, and you wonder if Aehr has committed some level of warranty for early adopters to make sure that they're getting the lifetime they've guided.

Practically, it's just difficult to know. Who knows that they've even gone through several sets of replacement cycles in the AI space. I did see a big order for silicon carbide on a bunch of WaferPaks that probably got to end of life on some more mature products. Thousands of touchdowns on a wafer, that's a lot of wafers going through. It's hard to know what your real wear is, and maybe they had to realign it, what constitutes a touchdown. They may have to do several per wafer while they're going through some characterisation. These are difficult questions to answer now. There hasn't been a lot of public data released. I would suggest that 5,000-10,000 quantity touchdown is probably a realistic baseline, which is a pretty decent starting point.

00:50:15] What do you think is driving silicon photonics customers to move from engineering qualification into high-volume production burn-in? How should we translate that into demand for Aehr Test Systems' Fox systems and WaferPaks as optical capacity ramps?

WW [00:50:37]: I think there's a breakout coming, but it's probably another 1.5 years down the line in these co-packaged optics. Even these pluggables, there's still, I think, potential need for these reliability screens, if they can do it at wafer level and catch the issue there, it's going to be maybe not quite the same economics in the AI chip space, but still strong and a potential catalyst for adoption there. I think they just got commitments on their initial production system, they may have shipped a few into the space already on the Fox systems into photonics. It's really the same story. If you can convincingly demonstrate that it's catching those defects at wafer level and you're not catching other package-level defects, and it's saving you money by not having a bad laser take out the whole integrated device, economics are going to look pretty similar. You could see it being proven and following the same type of order pattern where one guy proves it, then other guys get interested, study it more. It might still be 1-1.5 years in design cycle, but it could be a wonderful add-on to their total market.

GW [00:52:15]: Could you quantify that add-on?

WW [00:52:18]: I think it's going to follow the same pattern. It's going to be low-single digits at first. If it is demonstrated to be useful and they want to do it on 100% level, then it's going to grow with volume. You could see that quickly rolling out into dozens of Fox-XP systems shipping into the photonics space within 1-1.5 years.

00:52:49] In our pre-Interview conversation, you mentioned that you do not view SiC as a major driver for Aehr Test Systems over the next 12 months. However, management recently pointed to signs of a recovery in SiC demand and highlighted roughly USD 8m of new SiC WaferPak orders. What makes you more cautious than management for the SiC outlook?

WW [00:53:14]: The WaferPaks is the consumable. They probably got to end of life and there was a need for replacement. It's a good test point that they're still using it. They're still running it and maybe volumes are picking up. Until new systems are ordered and it grows out, that would be the bigger uptick, where they've actually got system orders for the Fox units into silicon carbide opportunities. Supposedly, there was a Taiwan customer that was announced as, I think, a potential evaluator, but it hasn't driven system orders yet. That to me is the more litmus test on silicon carbide recovering. The auto market has been down, EVs have been… I think that was a big player in these potential need for the systems. I don't know that there's been strong signs of the EV market recovering recently. That could be a big tie to additional systems into this space, but haven't seen it yet.

00:54:40] How do you expect Aehr Test Systems' revenue and profitability to trend into FY27? Which customer programmes or end markets are likely to be the biggest drivers of upside or downside?

WW [00:55:00]: I'd say in the next 1-1.5 years, it's still all about AI. That's where I'd be focused on both the Sonoma and the Fox systems. I think it's already generating a significant portion of the revenue, and trying to get proof points of that second customer adopting and getting data disclosed on how the lead customer is using it. Is it really 100% production? There's follow-on orders even beyond what was placed in August. Those would be the things that I'd be focused on, and that this USD 100m in effective backlog is really deliverable in the coming year, and it doesn't get pushed out beyond some of these quarterly boundaries, which seem to be what happened in the last fiscal year. That's what I would be focused on. I think the exuberance in the stock is very much based on pretty rapid expansion into a lot more systems and several more customers. I'm just naturally a little bit more sceptical in seeing that proof point that it really becomes a fixture among several AI players. I just haven't seen the convincing evidence yet.

00:56:28] Aehr Test Systems' management is guiding USD 130m-150m for the company's FY27 revenue. How does that compare with your own expectations?

WW [00:56:39]: I think that's based on basically what they've got in backlog and then it follows through. I think that seems to be pretty good guidance. I find it challenging. If management is guiding that, and the look ahead is suggesting, I think, their valuation based on a lot more adoption, something is off. They've got to get a lot more bookings in the next 2-3 quarters that demonstrate a broader adoption. If it doesn't happen, to me, that's a big negative. I'd be very closely monitoring the next few quarters on not just follow-on orders, but brand-new systems into new customers.

00:57:37] Aehr Test Systems' five largest customers accounted for roughly 70% of its FY26 revenue, including three customers at approximately 26%, 14% and 11%. How significant is the company's customer concentration risk today? How do you see these large customer relationships affecting Aehr Test Systems' pricing power?

WW [00:58:05]: Yes, that really scares me. Very much scares me just because one or two of them change gears for whatever reason, it's a significant chunk of your revenue. AI is concentrated. I don't know that it's too surprising. I'd still be concerned about it. You want to see that the story bleeds over to other players in the space. As importantly, that there's no evidence that this production wafer burn-in is really just a sample monitor and not 100% of the wafers going through and that it's persistent. Concerning, again, because the AI space is concentrated and AI really is exploding, and it certainly seems all the major players are pretty much sold out one year ahead. As long as it is, the hyperscaler company is a major player. I'd say that's a good still strong, solid sign. It bothers me that it's been hard to get any public confirmation on who it is and how they're using it.

GW [00:59:33]: We will now end the Interview. Let me close by saying thank you and goodbye.

Transcription ends at 00:59:43 of the recorded material