← 返回

2026.09.30 | TB | Memory | SSD Controller Industry: Fadu & Sandisk

原标题：SSD Controller Industry – Fadu & Sandisk – Senior-level Role at ScaleFlux Inc

#memory #semis

Third Bridge · 2026-09-30

Client-led • 22 Sep 2026

Companies: A005930 KR, MU US, 285A JP, A000660 KR, SNDK US, MRVL US, A440110 KR, ALAB US, MSFT US, INTC US, META US, NVDA US, AMD US, 8299 TW

Expert Senior-level Role at ScaleFlux Inc

Agenda

Questions

00:10:45] Who are the most relevant players in the SSD [solid-state drive] controller market for enterprise SSD data centre applications?

Expert [00:11:50]: The controller market can be easily divided into two big segments first. One is the internally developed controllers, let's say, Samsung, Micron, SK Hynix, Kioxia and Sandisk. They have their internally developed controllers. Then merchant silicon controllers, which can be divided into traditionally speaking, Marvell, Phison, SMI, Fadu, us on the western-facing world and then Innogrit and a few others focused only on the China market. Within western-facing merchant silicon controllers, you can divide into two sections. One which is entirely focused on the clients side and other which is focused on the data centre. In the data centre, the biggest player used to be was Marvell. Unfortunately, or fortunately, due to this AI boom, Marvell scaled down their merchant silicon business for the NAND controller significantly.

The second company was Fadu or is Fadu. We are the third one. Now the SMI, Phison, they are specifically focused on the client side. They have been trying to break into the data centre, but unfortunately, despite their last three years of commitment, they have not been able to win any customer in that space. Now there is an outlier in this one, and that is Astera Labs. Astera Labs has a controller only and only for the CXL. Unfortunately, that's not their primary business. For all intents and purposes, we could say we have Fadu and us, which is us in the enterprise or hyperscale data centre. SMI and Phison on the client side, and these are the western-facing, Marvell is pretty much out, and I'll explain to you why I'm saying that. Two in the client side and two in the hyperscale side.

That's the market right now on the merchant silicon. On the IDMs or NAND focused, you have Micron, which is shipping PCIe gen 6 controller as of this year. I think they started shipping in very small quantities earlier this year for the qualification and then Samsung. Everyone else in the IDM side is currently shipping gen 5, and they will be looking at purchasing a merchant silicon controller from one of the vendors such as Fadu, us down the road. Internally developed controllers outside of Samsung and Micron have not been shipped or have not been even taped out yet in some cases. That's far behind. That's how the controller for NAND industry looks like. I'm going to stop it here. Please feel free to interject if I'm going to the direction which is not relevant to you.

Client [00:16:32]: Are there two main camps, consisting of the memory makers who design their own controllers for the data center world, which are largely Samsung and Micron for Gen 6, and the merchant silicon side, which for data center applications is largely represented by Fadu and you?

Expert [00:17:03]: That's correct, yes.

00:17:05] What percentage of the enterprise SSD market is using merchant silicon vs the memory players' in-house controllers? How do you think that will change going forward?

Expert [00:17:33]: Very good question. Let's make it very tangible. The reason Fadu got their controller accepted [is] because they worked with Meta for a long time and they built their firmware as per Meta's requirements. Once they completed that, they went to Sandisk and said, "Hey, we have a customer who wants to buy our controller and we have tested your NAND on it and let's partner." At that time, Sandisk controllers for gen 5 were far behind and they did not have the right specifications. Now since then, Fadu has been trying to break into a couple of other accounts. By all indications, they have not won other major accounts yet. The emphasis is on the yet, but I know that they have been testing this at Google and Microsoft. My prediction says that for the PCIe gen 6, in addition to Meta, they are likely to win Google. Microsoft, TBD, I think Microsoft has a lot more stringent requirements.

The Samsung and Micron have been shipping in low volumes, their PCIe gen 6 with TLC and high capacities today. I'll be very careful what I say about us because GLG (ph) prohibits me to say too much about our company. Market is shifting up to be that Fadu with the help of Sandisk will likely have Google in addition to Meta sometime in late 2026 or 2027, and we will have some others in the same space. That leaves us literally in the merchant silicon business, two customers who are going to gain market share where Samsung and Micron cannot service. For the purpose of this discussion, I'm excluding YMTC or CXMT, I'm not bringing that, and we should have a separate commentary on that.

Kioxia is in this. To answer your question, for gen 5, Fadu has whatever the market share you have in Meta, that's it. Everything else was driven by IDMs or Samsung, Micron, Kioxia. For gen 6 Fadu in the western world for the merchant silicon and for Samsung and Micron with fully in-house developed controller and SSD. That's how the market share is going to break down. Now to answer your question specifically, the share, I think the share, if you have access to how many SSDs are being shipped by Samsung and Micron, that will give you a very clear indication of number of units. If you have a TrendForce SSD shipment units data, then you can get market share for Fadu and others.

My view is Samsung is the leader in the SSD and NAND business, followed by SK Hynix, which is SK Hynix and Solidigm. Shipment data, but I think the merchant silicon will have probably 30-40% of the market share with PCIe gen 6. Now I'm not giving you a year right now because Samsung and Micron have started shipping gen 6 products this year. They have gained a lot of market share. Compared to gen 5, the transition has not been completed, but the real deployment for gen 6 products will take place when the Intel DMR and AMD Venice will be deployed at scale. That's when the gen 6 products will be adopted. Market share at that time, I would say, at least 30% of the merchant silicon adoption, at least 30%. Hopefully, this is a long-winding answer, but it provides you the context and a bit more nuance how you can estimate the market share.

00:23:55] You mentioned that Samsung, by SSD shipment volumes, is the leader, followed by Hynix or Solidigm. I guess Solidigm is one of the two shops, similar to Kioxia, that could be a potential customer for you or Fadu. Which vendor is it using for Gen 5? Does it still have an in-house Gen 5 controller?

Expert [00:24:41]: Excellent question, and that goes into the acquisition and now proposed divestment of Solidigm. Solidigm used the gen 4 controller that was internally developed. Gen 5, they asked a couple of vendors for the bidding, and that process is not completed yet, I'm sorry, gen 6. Gen 4 was internally developed. [For] Gen 5, SK Hynix was granted to them.

For hyperscalers, the product definition goes way above what is the OCP requirements. We believe that both from the timing, product feature, availability and pricing, we are very, very competitive at this stage.

00:27:27] When you said Fadu was likely to win the Google account soon and potentially Microsoft in the future, were you referring to Gen 5?

Expert [00:27:52]: No, it is gen 6. Here's the reason why they will likely win. Sandisk has entered into sale of the NAND to Google and Microsoft and others. As part of that, there is a commitment between [the] two companies that Google and Microsoft will evaluate. If [the] controller meets their requirements, we will purchase SSDs. The controller that is used for those evaluation is Fadu controller, so that's how Fadu is going to win. It is not that Fadu actually went to Google or Microsoft and try to win the qualification, no. It is the Sandisk that is shifting a lot of their bits from the client side to the enterprise or hyperscale data centre, and they are planning to use or they are promoting PCIe gen 6 Fadu controller. That's how Fadu is getting into Google or likely into Microsoft. This is a very nuanced answer. It is like, "Hey, is Fadu directly winning the account at Google and Microsoft?" No. "Is Sandisk winning the account and they are going to be using Fadu controller?" Yes, that's the way to look at it.

Client [00:29:40]: That means that Fadu also already has a Gen 6 controller, which is probably undergoing testing or sampling, right?

Expert [00:29:52]: They don't have it yet. They are, I think, starting to give samples sometime early next year, but that is the business model where this is going to go. Please look at this thing. Sandisk is largely shipping gen 5 today. At Microsoft and Google, the gen 5 was not qualified, and any qualification takes at least 12-18 months. Right now, they are getting qualified for gen 5 so that they can start shipping their SSDs now. As part of that, they will move on to the gen 6 when gen 6 becomes available, and that's where Sandisk will use Fadu's controller. Because today, Sandisk doesn't sell any SSDs into Google or Microsoft. Now as part of the NAND deals, they started qualification and qualification takes a long time, anywhere from 12-16 months. For gen 6, I think qualification could take at least 16 months. That is going to come along with BiCS10 when Sandisk starts shipping that.

00:31:27] Are most of Fadu's hyperscaler wins driven by Sandisk's customer wins?

Expert [00:31:45]: That's absolutely correct. I'd like to make one comment. That qualification of an existing controller at a new customer takes at least 9-12 months, of an existing controller. There are inflection points where the qualification and adoption of a new controller or new PCIe gen controller and new media virtually takes long time and it's very expensive, very, very costly. Companies like Fadu, which are not based in the US, most like 98% of their team is based in Korea.

They have some cost advantage to some degree, but not having team here in the US is a major hurdle, and it could take longer time, but the blessing for Fadu is they don't need to do a lot of this heavy lifting because they need to satisfy Sandisk and then Sandisk will interface directly with these big customers and provide the feedback to Fadu to fix all the problems with the controller. That's why the new media, new PCIe gen, new server will require a longer qualification time and it's very, very costly.

Client [00:34:15]: Is it fair to say that Fadu is now working on qualifying the Gen 5 controller through Sandisk with Google and Microsoft? Once that is done, the gen 6 qualification should be quicker compared to doing a new controller for a new customer because it's an existing customer.

Expert [00:34:54]: That is a fair assessment. That is the expectation. That's how it should work. Assuming everything meets the OCP requirements and then on top of that, the Google and the Microsoft requirements because each of these customers will have their own requirements on top of standard OCP requirements.

Client [00:35:15]: Would you not try to acquire Fadu, as it is based in Korea and working so closely with Sandisk on American hyperscalers?

Expert [00:35:48]: I cannot be very specific. Actually, there's one other thing. There's a good and bad thing of working with NAND suppliers because NAND suppliers will need you when they cannot meet time to market, and they will dump you when they have internally developed controller. Sandisk is working on PCIe gen 6 of their own controller. You can Google it and find out. Fadu and we are basically competing with their engineering team for the slot.

That is a problem of working with Micron, Samsung, SK Hynix, Kioxia and Solidigm, but it is not a problem of working with the hyperscalers or neocloud guys because they can source NAND independently and work with their contract manufacturer and we qualify our controller. That's a different way of looking at the same problem statement because in that case, we are responsible for all the upfront qualification evaluation qualification and we charge the value. We get to capture the full value of the controller vs we give to Sandisk and they mark it up for another 70%, and we have to compromise on the cost. We have to compromise on the margins, okay? It's an interesting problem, but the business model works slightly differently.

00:38:39] Sandisk is also developing a controller in-house in parallel. Are you saying that even if Fadu's controller qualifies with Google and Microsoft through Sandisk, the company could still end up using Sandisk's in-house controller?

Expert [00:39:10]: Absolutely, absolutely. If you Google it, you'll find out they are actually implementing some very fancy features in their SSDs, which Fadu's controller doesn't have, okay? Because Sandisk has been working on those features for a long time. Fadu is not [a] very well-capitalised company. Therefore, building these additional, very fancy features takes second or third priority, whereas Samsung, SK Hynix and Micron have large teams and are very well capitalised. They have more appetite to invest in that so they can build those fancy features. Fadu they are working only on one controller. Their fixed cost is high. The variable cost is low, but even the fixed cost is lower because they're based in Korea.

00:41:20] What do you think about Fadu?

Expert [00:41:21]: Fadu are a good player. I like Fadu. They have gone through some very tough times in recent past, and Sandisk helped them tremendously, and I want them to succeed in the marketplace because when we have more players in the marketplace, it's good for the industry, but there are pros and cons of being in Korea vs US. You would imagine that they are based in Korea, they're next to Samsung and SK Hynix, they would win easily their business. Unfortunately, that has not happened, and there is some really bad reasoning for that, but again, I think Fadu is a darn good company. I think [it] will have solid business. It is not going to be in the billions of dollars in revenue because that is not this industry, but I think they will have solid products. Please don't take my criticism of Fadu that I'm against the company. No, that's not my intention.

00:42:30] I guess Samsung largely does it in-house. What about Hynix/Solidigm? What were you referring to when you mentioned there's a reason why Fadu hasn't made any progress with the Korean NAND vendors?

Expert [00:43:01]: This is public knowledge. I don't have anything private to share with you. What happened was that SK Hynix had invested in Fadu. At the time of IPO, and I don't understand the Korean Security Exchange rules and regulations fully. I will apologise on that note, but somehow, Fadu claimed to have a very big potential revenue in their S-1 filing. A lot of these things are in public domain. You can Google it, you'll find it. I know a lot of these things because I go to Korea to meet with Samsung and SK Hynix at least once in a quarter.

Because they are my direct competitors, so I've been asking them all these questions. That's my take on this one, but I think that is behind them. I think Fadu today has recovered, revived. Despite their limitation, I think they are going to win business at least for gen 6, if not more, at least one more hyperscaler, but that is going to be due to the help of Sandisk because Sandisk has committed volume and NAND volume, that is going to be a critical factor in winning the business at Google.

00:45:49] You keep emphasising how Sandisk is helping Fadu. Does that not contradict Sandisk developing its own controller?

Expert [00:46:21]: Very good question. Very good question. Companies are run by humans and humans have psychology. Also, humans, even within teams disagree with each other. You asked me this question previously, how come you didn't win this business? Fadu helped Sandisk to win at Meta. Sandisk needed that revenue really badly at that time. This is before Sandisk was separated and things were not all that great for Sandisk in those days.

00:50:25] It doesn't make much sense for a memory house to design its controllers in-house – a process that needs a lot of resources, large engineering teams and foundry relationships. Why would that be worth pursuing vs buying from companies such as Fadu that work entirely on 1-2 products and have the expertise and customers?

Expert [00:51:27]: We look at the fully burdened cost, okay? That's the actual cost plus depreciation. On fully burdened cost. On a fully burdened cost, it does not make sense for Samsung, SK Hynix, Micron, Kioxia to build internally developed controller. I'm convinced of that. Two, if we look at the controller pricing as a percentage of SSD pricing, even when we normalise the dollar per GB NAND pricing, okay, the controller pricing for 8TB is about 12%. You go to 60TB of SSD, it is about 1%. You go over 60TB, it's slightly less than 1%. How much CAPEX do you need to put to generate that 1% or less than 1% of ASP as a percentage of the SSD. That's a mind-boggling number. You answered the question why companies are doing it. It's a fear that if I build a controller with third party, they will not give me help on time, and I will lose the market share to my competitor.

Let's go back to Marvell. Marvell used to be very big in NAND controllers. Going forward, I don't believe it makes a lot of financial sense for the company to build internally developed controllers because they don't add much value to it, and they're always behind. Big companies have a capital allocation process, which is very cumbersome. It moves very slowly and the projects get perpetually delayed vs Fadu and us where we make decision based on half an hour notice and we move fast, really fast. We also develop multiple controllers, so IP costs can be amortised. I think as we go forward, we will see more and more merchant silicon coming out, but you put yourself in the shoes of a hyperscaler who says, hey, you have a NAND, I want SSD from you, you give me the controller with it. Companies like Fadu, we have opportunity to work directly with the hyperscalers as well as the memory partners.

Any customer, Microsoft buys NAND from SK and Samsung. Google buys Sandisk and Micron. Our controller will work with them, so will the Fadu. Fadu will basically be more limited because they're just focused on Sandisk. These are the different ways to look at it, but purely from the cost point of view, it is highly, highly inefficient allocation of resources for the memory suppliers to build internal controllers. The only reason they do is fear of missing out to their competitors. That's it. There's no logical financial reason for that. A case in point, if you look at Samsung, Micron, SK Hynix, they don't build clients' controller anymore. They always purchase it from third party. I think the time will come sometime in the near future when they will look at the economics of building the controller and say, you know what, I'm going to buy from third party.

Client [00:56:34]: I guess in the mid to long term, this should significantly grow the addressable market for you and Fadu, right?

Expert [00:56:51]: That's right, yes.

00:56:55] What is the eSSD [enterprise solid state drive] controller market's share and growth rate?

Expert [00:57:15]: Again, growth is tied with the platforms, whether it's Nvidia's new platform or AMD, Intel new platform is tied with that. I wish I had known that you were looking for the forecast of SSDs. I could have looked at a couple of my data points. We do some forecast ourselves for the unit forecast. We believe that the controller market will grow differently for different capacities. Higher capacity growth will be slower, but 2TB and 8TB drives number of unit growth will be much higher. That's for NAND. For CXL, I tell you one thing about 50% of the fleet will have a CXL device, at least one device attached to it. That is a pretty solid number.

Now NAND, each server will have multiple devices. If you look at Nvidia's viral platform and see how many SSDs are attached to that and what type of SSDs are those, similarly with TPU platform and how many SSDs or boot drive plus other SSDs are attached to it. That will give you a very, very good benchmark to see where the SSDs are going. Then you can say at least 40% of that will be merchant silicon. That's one way to do it. I don't have actual numbers on front of me, and I don't want to quote you any wrong data here.

00:59:35] What is CXL [compute xpress link]? What does it do and why is it needed? Who are the main players?

Expert [01:00:01]: Let's start with the server. The whole problem is the CPU, whether it's AMD or Intel. As the number of cores have grown over the last 10 years, the bandwidth per core has declined. If we look at CPU from hyperscale usage point of view, the usage has gone down, and there is an unused compute power that they cannot monetise, okay? That's the problem statement. In order to solve that, previously, we had three or four different interfaces where we could add additional memory to the CPU. If we look at CPU, Intel and AMD, they have fixed number of slots, DDR slots where you can put the number of RDIMMs on it. The objective is we need to have a cache coherent memory attached to the server could be slower, and we are okay with that. CXL solves that problem. It is a cache coherent memory on PCIe bus, you can add more DRAM as an E3 or add-in card. Now you can utilise the unused compute power and sell that to your customers. Hyperscalers can do that. That's the first use case. The second use case is now KV cache. It has become a very big and important use case.

CXL solves the problems of underutilisation of the CPUs by adding more memory on the PCI bus. The fancy thing is it's a cache coherent, data is not lost. It is slower. If we look at how much of the memory is hot memory, warm memory and cold memory, industry has been saying this for a long time, about 35-40% of the memory on server is hot. 30-40% is probably warm and 10-15% is cold. Cold and warm tiers are very good candidates to be placed on CXL. That's where the second use is. The third use is still there is some memory underutilised. Let's pool it together. We pool it at the rack level. Now you can utilise most of your memory and sell that as a service to your customer. The reason for doing that is if you look at how much DRAM is purchased by hyperscalers, how many CPUs are purchased by hyperscalers and the AI infra or infra, IT infra is underutilised by 40%. 40% of USD 1tn, that's a lot of money and dedicated to CPU and memory. Yhat is a lot of money that is sitting idle in the hyperscale data centre. CXL helps to solve that.

01:04:15] Who else are you competing with in the CXL space?

Expert [01:04:24]: In CXL space, we have only one player, and that is Astera Labs. They are not very strong because their primary business is retimers. Marvell has good products in CXL. One very, very important distinction here. Anybody who builds SoC. Building SoC is about 40% of the actual work. 60% of the work is building the firmware and getting it qualified. Any company whose primary business is not building firmware and getting it qualified is going to take 2-3 years to get that deployed. Difference between Fadu, us vs Marvell or Astera is, Fadu has a good number of firmware team. We believe we have probably the largest team in the world. That gives us a competitive advantage over others or internally developed controllers because this is our core business, and we have invested heavily in that, whereas others have taken their firmware's team away when the market was not all that great. That differentiates Fadu and us in NAND side and then us in the CXL side vs Marvell or Astera Labs.