Providing high-performance computing solutions like AI data centers and CXL… Expanding customer-specific services
AD Technology has started its second leap.
The company was once a leading domestic semiconductor design house, generating stable revenue of around 300 billion KRW as an official partner of Taiwan’s TSMC. However, after transitioning to Samsung Foundry DSP (Design Solution Partner) in 2020, its revenue dropped by around 100 billion KRW.
CEO Park Jun-Gyu emphasized, “In recent years, production projects have decreased, and we had to make significant investments to transition to Samsung Foundry. We also focused on various technical preparations, which led to a decline in performance. However, we expect revenue to rise starting in the fourth quarter of this year.”
In a recent meeting with Thelect, CEO Park shared, “After experiencing the ‘Golden Flow’ while working with Samsung Foundry on 4nm and 5nm processes, we clearly identified the strengths and weaknesses of Samsung Foundry compared to the experience we had built with TSMC.” He further added, “We began actively promoting the transition at the end of last year, and since then, the number of orders has steadily increased.”
AD Technology is focusing on expanding its customer-specific services by providing high-performance computing solutions, such as AI data centers and CXL. The company is also accelerating the development of next-generation AI and HPC chiplet platforms. Recently, it has expressed its goal to enhance its competitiveness in AI semiconductors through collaborations with Samsung, Arm, and Rebellion.
Currently, AD Technology is investing more than 20% of its engineering resources into building platform infrastructure in partnership with Samsung Foundry. As a result, solutions targeting the AI and HPC markets are beginning to take shape. CEO Park noted, “Currently, development revenue accounts for more than 60%, but we anticipate that the proportion of mass production revenue will increase starting next year, creating a virtuous cycle between development and mass production.”
Park added, “Based on our technical advancements, we are aiming to break our record for the best performance in 2026 and further strengthen our global market competitiveness through customization and differentiated technological capabilities.”
### Interview with CEO Park Jun-Gyu
* AD Technology has been in the design house business for a long time. It’s well-known in the semiconductor industry, primarily as an official partner of TSMC?
“Yes, we worked with TSMC for about 15 years, starting in 2005.”
*Now, you’ve transitioned to a new partnership with Samsung, meaning you’re no longer working with TSMC. When did this change occur?
“We officially became a design solution partner with Samsung Foundry around the end of 2019 and into 2020. The market refers to us as a DSP.”
:DSP (Design Solution Partner): A partner that provides various services related to semiconductor design to foundries (semiconductor contract manufacturers).
* In terms of annual sales, there were years when sales were in the high 200 billion won range, and even over 300 billion won. Was that the year when you recorded the highest sales when you were working with TSMC?
“Technically, we started development and mass production using TSMC processes, but the peak revenue came after we began working with Samsung as a DSP.”
* That was likely in 2021 when you recorded the highest revenue.
“Yes, in 2021, we worked together with Hynix on NAND and NAND controller solutions, and as these solutions entered mass production, our output surged. If you look at the history of our company, we faced some challenges in 2016 and 2017, but we started growing rapidly in 2019, 2020, and 2021.”
* Some might be curious about why you decided to shift from TSMC to the domestic foundry ecosystem. What led to that decision?
“The main reasons were twofold. First, our largest client, Hynix, started building its own development and production teams, taking on the tasks we used to handle, particularly for mass production.
Second, the domestic non-memory market lacked many active fabless companies, making it difficult to form a sizable market. After going public in 2015, our shareholders have demanded continuous growth, which added some pressure. Especially when combined with TSMC’s business policies, expanding our business abroad became extremely difficult.”
* Isn’t there a company like a subsidiary of TSMC?
“There is a company that is a partner of TSMC and has shares in TSMC. In the case of this company, we worked hard at the time because there were board members and a lot of direct communication, but I remember withdrawing. So I thought that some kind of decision was needed for long-term growth. If we had continued to cooperate with TSMC at that time, we might have been able to achieve decent results for 2-3 years, but considering the future, we would have been worried. We had to make a big decision for the company’s continued growth, and in a situation where the current leaders in the global foundry market are limited to TSMC, Samsung, and some Intel, we also decided that change was inevitable to increase our competitiveness.
In particular, it was difficult to respond to the increasingly large and complex chip design demands with simple implementation. As customer demands became more diverse, collaborative designs increased, and while we used to focus only on implementation, we now had to change to combine board solutions with various functions. Of course, TSMC has already built a solid ecosystem, but we had limited internal resources to prepare for this. So we had to make tough decisions along the way because we needed an environment where competition was smooth and we could access the global market.”
* I’m not sure if it was a company invested in by TSMC or a subsidiary, but did that make it difficult to conduct sales activities targeting customers in Taiwan or other strategic regions?
” That had a big impact on the overall situation. For example, when we were trying to expand by opening overseas offices, there was a time when we could originally move freely regionally. However, in the years after 2010, as TSMC actively cooperated with China, we worked a lot with design service companies in China, and the situation became complicated as they entered various markets such as the US and Korea with favorable mask and wafer prices. In response, TSMC also clarified its sales policy to prevent market confusion, and overseas promotions were restricted. From then on, each company was required to operate its business centered on the location of its headquarters.
* That would mean there would inevitably be some limitations to growth, right?”
“Yes, often when people look from the outside, they ask why someone would choose to join Samsung when TSMC seems to have a good position. But we had our own challenges, you see.”
* To grow the company, you need to secure customers from overseas markets like Taiwan, the U.S., Japan, etc., right?”
“Companies like TSMC, which are leaders in the industry, can control the market through how they work with their partners. So, I still think we’re not entirely free to operate as we want.”
* Changing partners has resulted in a significant drop in sales. You had around 160 billion KRW in 2022, and last year it dropped to about 100 billion KRW. You also experienced some losses last year. How does this year look, and what about next year?”
“Looking at the essential characteristics of our industry, a lot of development work is required. Development is a key leading indicator for future growth because it leads to mass production. When we first started working with Samsung Foundry, the biggest thing we focused on was understanding the differences in reference flows between TSMC and Samsung. So rather than focusing on quickly acquiring orders by contacting customers, we concentrated on training engineering teams aligned with Samsung Foundry’s processes. We spent about a year and a half to two years working on engineering tasks for products developed internally at Samsung, within Samsung’s RBS environment. Without this preparation, we believed we wouldn’t be able to play our role properly as a DSP at Samsung Foundry.
At that time, significant investments were required, but the SSD controller sales from Hynix started to peak and then decline. And as we worked with Samsung Foundry’s 4nm and 5nm processes and experienced the ‘golden flow,’ we could clearly identify the strengths and weaknesses of Samsung Foundry compared to our experiences with TSMC. Since late 2022, we’ve been actively promoting, and we’ve seen an increase in orders this year and last year.
The interesting thing is that in the last five years, the volume of orders has significantly increased. In the past, we could manage with around 10 billion KRW for a 14nm process, but now just buying IP can cost over 100 billion KRW. As the size of individual transactions has grown, customers are also facing deeper concerns. In the past, discussions would take two or three months, and we would start projects with a small engineering team. Now, because we’re investing more engineering resources, the review time has lengthened.
For example, a startup needs to secure investment for operations, and a portion of that funding must go into projects. However, very few companies can secure 30-40 billion KRW in one go. Even companies valued in the billions of KRW, and even AI-related foreign startups, find this difficult. As a result, the coordination process takes time, and in some cases, it can be delayed by three to six months. Although there have been delays, overall, the process of securing new projects and developing them is progressing smoothly. Internally, we are establishing our business plan for next year, and we believe that last year marked the bottom for us.”
* Last year was the bottom?
“Yes, I believe there is definitely room for recovery this year and next. It won’t be a dramatic rise like in the past, but I’m confident for next year.”
* How does this year look?
“I think it will improve significantly compared to last year.”
* What about profitability?
“The profitability still needs improvement. Since we had a loss last year, it’s difficult to expect a strong performance this year, but we plan to focus on improving profitability in Q4. The projects we’re currently preparing require substantial upfront investment. Traditional design services were about taking on customer tasks and proceeding with design or joint development, but the collaborative business model we’re aiming for now requires proactive investment.
Like many companies are preparing platforms, we’ve been steadily developing our basic platform, and we’re actively preparing additional elements tailored to market demands. One example is the chiplet solution we recently presented with ARM, Samsung, and Rebellion at OCP. To enhance our competitiveness, improving power efficiency and performance is key. While there are many ways to achieve this, technologies like specialized libraries and memory components are crucial for foundry or design services.
Currently, more than 20% of our engineering resources are focused on libraries, memory, and platform infrastructure. As a result of continuous investment in this area as a DSP for Samsung Foundry, we began to see tangible results starting late last year. This has allowed us to build a foundation for collaborating with high-level clients, and about 150 to 200 R&D personnel are continuously working on this project. This is why we’ve faced some difficulties in our management indicators, but the situation is gradually improving.”
* How many employees do you currently have?
Around 800 employees.
* How many are at the headquarters in Korea?
About 450 to 480.
* Are the rest of your team based in Vietnam?
The team in Vietnam is primarily set up with implementation personnel.
* That’s quite a large team..
“Once we start receiving projects in the form we discussed, we’ll be collaborating with those clients. Currently, around 20-30 people are involved in a single project, but in some cases, there may be situations where up to 140 people are mobilized at once. Considering that we have a workforce of around 800 people, we estimate that we can handle up to four projects per year at most.”
* Four projects. What kind of revenue does that translate to?
“The estimated revenue would be around 150 to 200 billion KRW. When the processes are more advanced, development costs can be significant, and when the system complexity is high, IP costs can also increase. So, right now, instead of sticking strictly to turnkey design services, we are being flexible and taking on projects that align with our strategic direction and the needs of our customers.”
* You mentioned that this year’s revenue will improve significantly compared to last year. Do you think you’ll surpass the 160 billion KRW you recorded in 2022?
“We’re doing our best to reach that number. There may be some differences depending on the timing of the orders and when the revenue is recognized, but compared to last year, we’re definitely seeing improvement.”
* How is the revenue composition looking?
“The development fees have improved significantly compared to last year. This is because we secured many new orders. As for mass production, there isn’t much at the moment. I’d say the proportion of development fees is over 60%.”
* Are development fees considered one-time revenue?
“No, they’re not. If you look at the revenue structure of design service companies, it often takes more than a year from contract signing to mass production, after the final validation from the customer. Revenue is recognized as the project progresses, based on each stage’s progress, so it’s not just a one-time event. For example, if a customer is developing products with a long-term roadmap, it’s not a one-off. On the other hand, a project like IP validation, which is completed in one go, would be considered one-time revenue.
In most cases, we use MPWs (Multi-Project Wafers) to develop products, or we conduct single runs based on our platform for the project. These approaches provide us with continuous opportunities for orders. Instead of being one-time projects, they are part of a continuous cycle where new products keep coming in, and we’re seeing this as a very positive sign.”
* You mentioned you’ve started planning for next year’s business. What kind of revenue do you expect?
“It’s not finalized yet, but I expect it to be significantly higher than the 160 billion KRW we saw in 2022.”
* Do you expect next year’s revenue structure to be similar to this year?
“I think it will be similar until next year. When we had the highest revenue, mass production revenue accounted for almost 90%—that was when controller sales were high. But, in reality, while the numbers may look good, from the perspective of profitability and sustainability, that’s not always ideal. A cycle where the ratio of development fees to mass production revenue is about 50-50 or 40-60 is much healthier. That’s because it leads to development turning into mass production, and creates a positive cycle where new development projects keep coming in. Right now, since there’s very little mass production, the development fee ratio is high, but some of the products we’re working on are expected to go into mass production at leading-edge nodes next year, which will convert into revenue. So, considering that, I’m very positive about our outlook.”
* So, you expect the development revenue to make up about 60% of your total revenue next year?
“Yes, that seems likely. However, if mass production revenue comes in faster next year, the ratio of development revenue could decrease.”
* Do development contracts usually result in mass production as well, or are they completed once the development phase ends?
“When working with Samsung Foundry, especially to gain the trust of overseas customers, we don’t insist on only taking on turnkey projects. If our technology is what the customer needs, gaining their trust is the most important thing. Therefore, we’ve been actively providing engineering support. Additionally, if the customer is targeting a large market or a relevant application, we also actively engage with them. So, instead of taking a passive approach, we’ve been actively preparing for the market, and we expect this preparation to be nearly completed by the end of this year.”
* How many customers do you currently have?
“In terms of customers we’ve done design work for, there are about 10 or so. And we’re currently in discussions with a little over 20 customers.”
* How is the split between domestic and international customers?
“The majority is overseas. That’s just the nature of the business.”
* Is it because there are very few domestic customers?
“It’s not exactly that. In the last three years, there has been a lot of semiconductor development in Korea centered around AI. Many good ideas have come out, but companies with market dominance, like NVIDIA or Tesla, have quickly incorporated these into their products, which means they often get released faster. Because of this reality, domestic startups and SMEs are finding it harder to keep up with rapid development, and the number of development cases has gradually decreased each year. Outside of AI, there aren’t many companies in Korea with significant market dominance, and the number of products they develop is quite limited.
Also, working with large corporations like Samsung or LG, which already have fully integrated systems, doesn’t align with our business model. As a result, we naturally have to focus more on overseas markets and create new opportunities there. Samsung Foundry is actively supporting us in this regard.”
* Isn’t it good for Samsung Foundry that companies like AD Technology do well in sales, develop products, and handle mass production?
“An interesting aspect is how foundries strategize their operations. The reason they prefer large customers is to increase production efficiency, as running small-scale production with diverse products is not easy. While many companies need to be appropriately distributed, if you look at TSMC, the revenue share from DSPs or small- and medium-sized customers like us is actually in the single digits. However, because the overall market is large, our share becomes noticeable.
One of the important roles we play is to create new momentum in specific application markets and collaborate with promising companies that can lead the way with technology. Helping them succeed at Samsung Foundry, like what we’ve seen with companies like Marvell, is one of our key roles.”
* Which process nodes do you primarily focus on for development?
“With 5nm and 4nm processes, especially 4nm has seen an increase. As reported by the media, there are cases where we originally secured orders for 3nm products but later switched to 2nm. In the case of the Neoverse V3, which we’re collaborating on with Arm, we’re starting with a 2nm process.”
* Is that a development project?
“Currently, it has an R&D nature, but the key will be how the market reacts to this technology. There has been significant market response to NVIDIA’s Grace Hopper performance, leading to private interfaces like NVLink, but we’re trying to develop better products using open architecture interfaces like UCIe and universal technology, and the expectations are high.”
* In the past, while working with TSMC, 90% of the revenue came from mass production and 10% from development, with one large customer contributing significantly to revenue. Although you now collaborate with multiple customers, investors or the market may expect a few major clients. Do you have a large customer of that scale?
“It depends on perspective. It’s a matter of whether they are tier 1 in terms of technology, or if they can control wafers at the level of a tier 1 customer. If we consider major customers who collaborate directly with the foundry as tier 1, TSMC also has a small portion of revenue from small and medium fabless companies or VCA businesses. For them, tier 1 may differ.
The important thing is that if we aim for revenue targets of 300 billion, 500 billion, or even 1 trillion won, we need to secure the possibility of tier 1 customers. This means using advanced processes and entering areas with sufficient market scale.
One area we’ve been focusing on is developing new libraries and design kits. This market requires low power, high performance, and a compact profile, which includes applications with periodic demand, like Bitcoin. We expect a market to emerge in the next year or two that meets these demands, so we’re actively preparing to develop library-related technology, as we believe it has the potential to become a cash cow.
For us, tier 1 revenue potential is in the fields mentioned above. And from a technological tier 1 standpoint, or in terms of company value and future growth potential, we are looking at the possibility in chiplet solutions we are currently working on. Data centers are facing challenges as they shift from traditional enterprise forms to supporting AI applications. For instance, it’s reported that services like ChatGPT incur significant operational costs as usage increases, to the point of becoming unprofitable.”
* The infrastructure investment and electricity consumption are enormous.
It’s gotten to the point where unused nuclear power plants are being restarted to meet power supply agreements. This shows how rapidly demand is surging, and NVIDIA’s solution has been the one to address it quickly. Even if you order NVIDIA products now, you can’t get them right away; it could take up to three years, even for large orders.
So, it’s important to examine whether this approach is just conceptual or something the market genuinely needs. Given the market demand, this seems like the right direction, even if the methodology may not be perfect.
For example, with our 2nm process, we could arrange Arm’s general-purpose CPUs in a tiled format. The number of tiles could be 32, 64, or, if companies like Microsoft or Google desire, it could go beyond 128 cores through hardening. We’re preparing this technology in a 2nm process and plan to invite accelerator developers to collaborate. We intend to work with various accelerator companies, leveraging existing chiplets.
Accelerator companies feel compelled to take this approach because it simplifies complex validation processes and facilitates chip provision. It’s an essential solution in this scenario. Of course, there are technical hurdles, but this approach could lead to various positive effects. For instance, if data centers are customized for AI through chiplet structures, there could be significant opportunities.
Through this hardening process, we’re developing chiplets that incorporate Arm-defined infrastructure with core-level performance. Previously, products like DP or smart NICs handled specific functions, but now servers are transitioning to a structure where multiple chipsets distribute the tasks previously done by a CPU. Reliable-performance CPUs are needed in this structure, not necessarily for edge computing but within the main structure.
Moreover, there’s an approach to configure the composition in response to market demand, using dedicated CPUs or inference-specific structures, rather than chiplets. This approach was also demonstrated by major companies up until last year. We’re sharpening our technological prowess through these developments, combining it with market experience to build a competitive edge.”
: ADTechnology recently entered into a partnership with Samsung Foundry to combine its 2nm process technology, Arm’s IP platform, and Rebell AI accelerator REBEL to develop an AI CPU chiplet platform for cloud, high-performance computing (HPC), and AI/ML training and inference.
* It sounds like you expect the current chiplet development project to bring results in the future. What is ADTechnology’s role in chiplet development? For example, do you design pathways to facilitate data flow between Rebellion or Arm architectures so that NPUs developed by other companies can be integrated into a platform?
“That’s the highest-level concept. To achieve this, those developing chips that influence lists and selections need to consider a die-to-die (D2D) connection. Unlike going through a PCB, this approach involves directly attaching the silicon dies, so we need to consider how signals connect physically when additional chips are added. While this might look simple logically, it’s more complex physically. Currently, some of our influencer partners are using appropriately sliced cores, attaching two for some applications and four for others. For our CPU die to support these requirements, the chip must be designed for flexible application. Solving this is a major technical challenge for us, but it also opens up opportunities to expand the chiplet business.
Also, if we consider the value of the Neoverse CPU, it differs from Arm’s previous approach of providing CPU solution IP. Arm is now in the Neoverse CSS (Compute Subsystem) business, where they’re preparing designs and samples in advance more than before. Instead of simply releasing RTL and providing guidelines, they are preparing for easier customer implementation by accounting for the complexity of hardening and deployment. This process is not as simple as connecting multiple CPUs together. There are many factors to consider, from architecture design to hardening, with significant performance variance based on setup and implementation.
There’s a program called Arm’s Total Design (ATD) Partnership. We were already preparing under this program license, and we were recently automatically registered in line with market changes. Currently, two companies offering design services in Korea have also joined this program, and they are working on similar challenges related to CPUs.
It’s very important to ensure that customers can reuse IP. When this is in place, it enables access to specialized functions like DPUs or smart NICs and provides a foundation to overcome technical barriers in chiplet configurations. If you push the limits with CPUs alone, it could be challenging to achieve configurations with 128 or 200 cores. Preparing for this is very meaningful for us.”
* If you succeed in technology development and acquire clients, how will ADTechnology earn revenue?
“That’s a very important question, as our business model is a bit different. If a customer wants a monolithic chip design with our CPU integrated, we would need to develop differently on a case-by-case basis. Since this is custom development, we would charge separate development fees. In contrast, the chiplet solution allows for reusability, so we would prepare and optimize everything except for the customer’s specific silicon in advance.
Customization is possible even with chiplets, but in a chiplet-only model, we aim to create a de facto standard. In this model, the customer’s silicon combines with our developed chiplets to create a complete solution, which is the general business model. In this case, we would negotiate with the client to share the development costs based on the engineering work we’ve done.
If a customer wants to develop Neoverse CPUs and customize them as chiplets or monolithic designs, the development costs alone could exceed hundreds of millions of dollars. But with our universal solution, they can significantly reduce these costs since we’ve already built it. This positions us beyond just implementation and gives us a competitive edge in the market, thanks to our prepared technology and accumulated expertise.”
* I think you provided guidance on this year’s and next year’s performance; when do you expect to surpass past record performance?
“I expect it could be as soon as 2026. Next year, we’ll have development and preparations for mass production, and there are markets where we can quickly recover revenue. Taking all these factors into account, we’re targeting 2026 as our goal.”
* In the past, ADTechnology achieved substantial development and mass production sales in the SSD controller field. Is there potential for new development sales in this area? I noticed a recent disclosure as well.
“The flash market related to controllers is changing rapidly. For example, companies like Kioxia, Samsung, Hynix, and Intel have achieved significant success in the memory business by providing storage in chip form. However, with flash memory, adding a controller optimized for specific applications and packaging it into a thumb drive or SSD form factor can create significant added value. This shift has led Samsung and Hynix to focus on the solutions business.
As solution businesses began in earnest, companies that previously led the flash drive market faced a shrinking market position. Though Hynix followed a slightly different approach, it previously had a business model where foundries and design service companies co-developed and shared mass production. Currently, however, both Samsung and Hynix independently develop and sell flash solutions, taking a large share of the global market.
Initially, we could share sales through collaboration by complementing each other’s needs, but now we primarily focus on an engineering-only model, collaborating primarily on development projects. We continue to discuss opportunities for turnkey projects as they arise from these projects.”
* You announced a 12.1 billion KRW) supply contract on October 10, but the contract date was early this year. Why was the disclosure made so late?
“There were various factors at play between the companies, but the main reason is that the contract was for ongoing development work we’ve been carrying out.”
* Should we consider this contract a one-time revenue event?
“There are cases where we receive development fees by providing ongoing support over a certain period, and cases where we charge for individual projects. This contract was term-based in relation to memory controllers, and we expect term-based and individual contracts to be more flexible from next year.”
* What does “term-based” mean?
A term-based contract is where you agree to develop a certain number of units over a year.
* So will this continue next year as well?
“The customer’s division is continuing to expand its controller solutions, and there are themes such as CXL (Compute Express Link). Memory is evolving from a simple unit product to a solution form, so I think controller development will continue. This trend will create more opportunities for us.”
* You expect to break past record performance in 2026, and with growing development revenue, as well as increased mass production sales, can we assume that the company’s scale will continue to grow?
“As I mentioned earlier, with our current workforce, we can handle about four large projects annually. Of course, this may vary based on process sophistication and product type, but increasing the number of projects indiscriminately could pose risks. So, even if the volume decreases, we believe it is necessary to maintain a certain level of project quality and revenue scale. We’re carefully managing our engineering resources to make prudent decisions.”
* Samsung Foundry has several design service partners, such as CoAsia, Gaonchips, and SemiFive. What is ADTechnology’s competitive edge compared to these companies?
“Our competitive edge lies in our ability to provide extensive customization. As I mentioned earlier, while many companies have developed specific platforms, we have been working in this area for a long time and have further advanced these efforts. Currently, we are building platforms that meet the demands of AI data centers, power performance, and chiplet solutions, offering solutions with diverse development elements beyond a simple platform. Rather than waiting for customers to come to us, we proactively reach out to explain our offerings.
This approach will be especially beneficial to customers targeting AI applications or data centers. We offer customization that can reduce silicon size or power consumption and achieve relatively higher operational performance on the same process. While major performance improvements are challenging, a 3-5% improvement can still secure product competitiveness and allow for better pricing through mass production. This strength is highly appealing to international customers, and we have branded our customized reference floor as ‘Capella’ to actively promote it.”
– Thank you for your insights today.