eMemory Technology Inc. (3529) Earnings Call Transcript & Summary

Good afternoon, and welcome to eMemory's Second Quarter 2026 Webcast Investor Conference. Joining us today is our Chairman, Dr. Charles Hsu; Head of IR, Ms. Li-Jeng Chen; Director of the Finance Department, Mr. Joseph Hsia; and Head of Digital Marketing, Dr. Felix Hsu. The format of today's event will be as follows: First, eMemory's Chairman, Dr. Charles Hsu, will give an opening remark. Afterwards, our Financial Officer, Mr. Joseph Hsia, will present a review of our financial results. Following that, Dr. Charles Hsu will share our business outlook. Next, Dr. Felix Hsu will give a talk titled, securing the next generation of AI infrastructure, the hardware anchor for Caliptra Root-of-Trust to secure Chiplets and Compute Express Link or CXL. Then we will conclude today's conference with the Q&A section where our management team will answer your questions. [Operator Instructions] As a reminder, this conference is being recorded, and a webcast replay will be available after the conference is finished. For more information, please visit the company's website under the Investor Relations section. As usual, before we begin, we would like to remind everyone that today's presentation may contain forward-looking statements subject to risk factors associated with the semiconductor and IP business. Please refer to the cautionary statement on Page 3 of today's presentation. Now I would like to give the floor over to eMemory's Chairman, Dr. Charles Hsu.

Okay. Good afternoon, shareholders and investors. Welcome to our investor conference. First, I would like to report on our latest progress in security IP. Over the past few years, our security IP has evolved from the foundational OTP and PUF into a PUF Root of Trust. Today, as security requirements for AI servers continue to rise, our technology has taken a major step forward, advancing from stand-alone security IP to the design of integration of comprehensive security subsystems. This transition is crucial for us. We are no longer providing clients with just OTP PUF or Root of Trust. We are now delivering fully integrated system-level security IP that combines encryption technology, software, firmware and anti-tampering protections. In other words, our role within our clients' chips has become far more critical, allowing us to deliver significantly higher value. Consequently, the potential license fee and also royalty are expected to increase significantly as well. We have licensed to several major global memory manufacturers to integrate our security IP into SSD solid-state disk systems for AI data centers, providing the security subsystems required to comply with Caliptra standards. Beyond SSDs, we have also expanded into another vital area of AI servers, which is memory expansion, IC. As AI model grows increasingly large, the memory capacity required by AI servers is surging rapidly. As a result, CXL is becoming a dominant interface technology for AI server memory expansion. CXL switch and their related controller chips similarly require a complete hardware security architectures. We are currently working with several CXL chip makers to supply both security IP and the security subsystems. So connecting the dots over the past few years reveal a clear trajectory for our security IP applications. Last year, we entered AI, AGI, CPUs and also BMC. And this year, we expand into AI data center, solid-state disk security subsystems. And now we are penetrating CXL switch for the AI server memory expansion. This means our security IP is expanding from a single chip inside an AI server to multiple critical chip. Furthermore, many of these advanced chips are utilizing cutting-edge process nodes such as 3-nanometer and generating much higher licensing and royalties than corporate average. Another highly favorable trend for us is the Caliptra driven jointly by global tech giants like -- such as Microsoft, Google, Intel, AMD and NVIDIA, we are seeing an increasing number of AI and data center chips and the system adopting the Caliptra security architecture. This indicates that the number of chips requiring Root of Trust, security subsystem and the related security IP will grow significantly moving forward. Therefore, we remain highly confident in the long-term growth of our security IP within the AI server and the data center markets. And next, I would like to also share some updates on another key growth driver we are heavily focused on, which is we call is the Logic Flash technology. Currently, our collaboration with foundries on core Logic Flash technology is accelerating applications extend beyond embedded flash to include stand-alone flash products as well. We believe the significance of Logic Flash goes beyond -- far beyond adding another IP technology to our portfolios. More importantly, it presents an opportunity to leverage a more competitive process and the cost structure to gradually replace a portion of existing legacy flash technologies. Flash memory represents a massive market with exceptionally broad applications. Once Logic Flash enters mass production and begins progressively replacing conventional flash technologies, the market opportunity and the long-term impact on our future revenue and profitability will be far reaching. For these reasons, we have a strong confidence in our multi-year growth outlook ahead. Next, I would like to invite our Financial Officer, Joseph, to present our second quarter performance. Thank you.

Good afternoon, everyone. Now let's begin with our 2026 second quarter financial results. The second quarter revenue was TWD 1,097 million, up 0.2% sequentially and up 17.1% year-over-year. Operating expenses were TWD 431 million, down 0.3% sequentially and up 10.2% year-over-year. In results, our operating income was TWD 666 million with an increase of 0.6% sequentially and an increase of 22% year-over-year. Operating margin also increased by 0.2 percentage points sequentially and increased by 2.4 percentage points year-over-year to 60.7%. Our net income amounting to TWD 580 million, experienced a decrease of 2.8% sequentially, but an increase of 44.9% year-over-year. The EPS for this quarter was TWD 7.77. And next, let's move on to revenue contributions by licensing and royalty. First of all, licensing in the second quarter accounted for 39.2% of the total revenue, increasing 12.8% sequentially and up 35.1% year-over-year. On a U.S. dollar basis, licensing grew by 12.8% quarter-over-quarter and 32.6% year-over-year. Royalty in the second quarter contributed 60.8% of our total revenue, decreasing 6.4% sequentially, but increasing 7.8% year-over-year. On a U.S. dollar basis, there was a 6.7% decrease quarter-over-quarter, but an increase of 8.6% year-over-year. And overall, the revenue increased by 0.2% quarter-over-quarter and up 17.1% year-over-year. On a U.S. dollar basis, the growth was 0.1% quarter-over-quarter and 16.9% year-over-year. And for the first half of 2026, the licensing and royalty revenues are as follows. First of all, licensing in the first half accounted for 37% of our total revenue, increasing 45.2% year-over-year. And on U.S. dollar basis, licensing grew by 45.9% year-over-year. And royalty in the first half contributed 63% of the total revenue, increasing 7% year-over-year. And on the U.S. dollar basis, there was a 9.8% increase year-over-year. And together, the total revenue for the first half increased by 18.5% compared to the previous quarter. And on a U.S. dollar basis, the growth was even stronger at 20.9% year-over-year. And with that, I will comment further on our revenue contribution by specific IPs. First of all, NeoBit accounted for 21.9% of total revenue in the second quarter. The licensing revenue increased 11.2% sequentially but decreasing 22.4% year-over-year, while royalty increased by 9% sequentially and increasing 9.2% year-over-year. And for NeoFuse technology, they accounted for 56.6% of the total revenue in the second quarter. The licensing revenue was up by 24.8% sequentially and up by 55.6% year-over-year. In terms of royalty, NeoFuse royalty decreased by 13.3% sequentially but increased by 4.4% year-over-year. And for PUF-based security IPs, it contributed 10.9% of total revenue. The licensing revenue decreased 13.8% sequentially but increased 133.5% year-over-year. And in terms of royalty, PUF-based royalty increased by 213.9% sequentially and increased by over 1,600% year-over-year. And lastly, for MTP technology accounted for 10.6% of the total revenue in the second quarter. The licensing revenue increased by 40.2% sequentially and increased by 9.3% year-over-year. The royalty from MTP was up 7.2% sequentially and increased by 19.2% year-over-year. And together for the first half of 2026, the revenue by technology are as follows: First of all, NeoBit licensing revenue decreased by 18.5% year-over-year, but the royalty increased by 2.4% and together accounting for 20.9% of the total revenue for the first half of 2026. For NeoFuse, the licensing revenue increased by 37.9% and the royalty also increased by 6% year-over-year, contributing to 58.2% of our total revenue in the first half. And for PUF-based security IPs, the licensing revenue increased by 264.6% year-over-year and the royalty increased by over 1,300% year-over-year and together accounting for 11.6% of our total revenue in the first half. And lastly, for MTP Technology, the licensing revenue increased by 19.6% and the royalty increased by 30.4% year-over-year and together accounting for 9.3% of the total revenue. And now let's take a look at our royalties for 8-inch and 12-inch wafers. First of all, 8-inch wafers accounted for 37.5% of the royalties, up 4.2% sequentially and up 1.2% year-over-year. And on U.S. dollar basis, this represents a sequential increase of 4% and a year-over-year increase of 3.1%. And for 12-inch wafers, it contributed 62.5% of the total royalties, down 11.9% sequentially but increased 12.3% year-over-year. On a U.S. dollar basis, this represents a sequential decrease of 12.1%, but a year-over-year increase of 12.2%. And in total, 156 product tape-outs were completed in the second quarter, and we will provide more information in our management report, which will be released shortly after this earnings call. And next, I would like to invite our Chairman, Charles, to share a little bit more about our future outlook. Thank you.

Okay. In the following section, I will address our future outlook. As far as licensing revenue is concerned, licensing will continue its strong momentum due to robust demand for our technologies from leading edge to the legacy process node security and next-generation Flash technologies. And for the royalty revenue, the royalty revenue growth is expected to accelerate driven by the higher ASP from the new advanced node applications and the new application ramps and also expanding the PUF royalty contribution and a growing mix of higher royalty rate of MTP related applications. And as far as new technologies for the advanced node, OTP and PUF-based hardware security continue to develop and qualify next-generation OTP and PUF-based hardware security solutions for 2-nanometer gate-all-around technology and also sub-2-nanometer nodes and meeting the growing customer demand in the device identity, key protections and secure boot and also hardware Root of Trust. And for the next -- another new technology will be next-generation 1T Flash. 1T NeoFlash technology is advancing across embedded also stand-alone applications and its logic-process-compatible architecture offers greater scalability, lower process complexity and better cost efficiency. And in the future outlook for the business development platforms, which we have 5 items. And the first is Chiplet Security Platform. We continue to work with ecosystem partners on an end-to-end security framework for Chiplet-based systems covering supply chain traceability, identity and authentication and also secure provisioning and also the die-to-die communication and also the key management and hardware Root of Trust, addressing the increasingly complex security challenges of AI and also the advancing packaging and heterogeneous integration among today's geopolitical environment, okay? And the second item is the Data Center Security and Caliptra Platform, targeting data center and AI servers, we continue to upsell expanding PUFrt Root of Trust from hardware Root of Trust solution into a Caliptra-compatible Security Subsystem and integration service that reduce the integration complexity and accelerate customer deployments. And another business platform is on the AI Compute and Root of Trust Platform. We are extending collaboration across CPU, AI accelerator and AI ASIC ecosystem to integrate chip level Root of Trust, Secure Boot, device identity and also secure key protection into the next-generation AI computing platform, strengthening trust and also security from system to system -- from silicon to system. And another business platform we are developing is HSM Edge Server and SECaaS Platform. The PUF-based HSM Edge Server combines device identity, key and certification management and secure OTA updates, signature verification and privacy protection and Post-Quantum Cryptography migration and early opportunities are progressing in the automotive OTA and PKI and also HSM integration with the potential to expand into the industrial control and Edge AI and smart device medical and also the data centers. And the last platform is we call it the Post-Quantum Security Platform. So we continue to strengthen our PUF-PQC portfolio with attack resistant, resistant hardware security, including side-channel protection to support the transition to post-quantum security standards. Okay. So next, I would like to pass to Felix, our Head of Digital Marketing, to share with our feature topic today. Felix? [Presentation]

Hi, everyone. So what you have just viewed in this video is one part of a much bigger movement happening around Caliptra and the hardware rooted security. Before I begin this talk, I'd like to briefly mention some related activities we've just been involved in. Earlier this week in Taipei, we participated in the Caliptra workshop focused on implementation and on bringing the specification closer to production silicon. We also presented at two sessions at OCP APAC, where we discussed Caliptra and Chiplet Security from both the architecture and systems perspective. Those discussions reinforce something we've been seeing more broadly across the industry. These security technologies are becoming increasingly important as AI servers and data center infrastructure become more distributed and more interconnected. So rather than covering everything around Caliptra, I'd like to use the next few minutes to just focus on the underlying reason why this movement is gaining momentum. And instead of starting from security, I would actually like to start from how AI compute itself has been changing. AI systems are increasingly being built across multiple dies, accelerators, memory devices and high-speed fabrics. And as more components are connected together, more of these components and their interfaces also need to be individually trusted and protected. And that's the story I'd like to present today. So AI compute can no longer scale in a single die. When we talk about scaling AI compute, simply making chips larger is no longer enough. In the past, we wanted more performance. We could build a larger die or use more advanced process nodes, but these are practical limits that we will approach, including die size, yield, cost, power, memory bandwidth and I/O all start to become constraints. So increasingly, the system -- the industry is scaling by composition. There are two important architectural trends shown on this slide, and I want to separate them because they're related, but they're not the same thing. The first one is Chiplets, which operate inside the package. Instead of building one enormous monolithic SoC containing every function, we can divide the system into multiple size specialized dies. For example, we have a CPU compute die, a GPU or MPU accelerator dies. We have management dies, a security die or dies with other specialized functions. Importantly, these dies do not necessarily need to use the same process technology. The compute portion may benefit the most from advanced node, but other die such as management security functions, it may be more economical on less advanced nodes. So Chiplets give us modularity, reuse, better economics and another way to continue scaling the processor package. But even if we scale inside the package, we eventually hit another boundary, the processor socket and the server itself. And that's where Compute Express Link CXLs come in. CXL operates at the system level outside the processor package. Here, you can see the servers on the left connected through a CXL fabric switch to a CXL Type 3 memory shelf on the right. Instead of memory belonging permanently to CPU socket, CXL allows memory and other devices to participate in a coherent fabric. That means memory can increasingly be expanded, pool, shared and dynamically assigned across systems. So you could think of it as Chiplet scale the processor by combining multiple specialized dies in one package, whereas CXL scales the system by connecting processors, accelerators and memory across the server. That's a very important shift for the AI infrastructure. We are moving away from scaling only through one large piece of silicon and towards scaling by connecting many specialized resources together. But this composition has security consequences. Every new die link, device controller and firmware layer creates another place where trust has to be established. And that brings us to the next architecture. The same composability that gives us flexibility and scalability also creates a much larger attack surface. Let me separate this again to two architectural domains. On the CXL side, we now have multiple endpoints joining a fabric, high-value memory, moving across links, switches controlling connectivity and a fabric manager potentially deciding how pooled resources are located. That introduces several different classes of risk. For example, if the firmware of a highly privileged infrastructure component, such as the CXL switch is compromised, the impact radius can be very large because now the attacker is not targeting only one endpoint. They may potentially manipulate connectivity, resource allocation or the behavior of multiple devices in the fabric. Another important area is DMA, direct memory access, which is in Panel 5. DMA itself is a legitimate and very powerful capability. The security problem is when a compromised or malicious device receives more access than it should have. If those permissions are not properly controlled, a device can potentially read or modify memory outside its authorized region. And in an environment where memory is pulled or shared, the consequence can become especially serious, data leakage across workloads or even between tenants. Inside the Chiplet package, the security problem looks different, but the principle is the same. Now we're dealing with multiple active dies communicating across die-to-die interfaces. A counterfeit or rogue Chiplet can introduce an unauthorized component into the package or supply chain, severely affecting silicon identity and supply chain trust. A malicious or unauthorized die can introduce hardware Trojans, manipulate transactions, leak data or undermine the assumptions of the entire package. Recent Chiplet Security research specifically highlights malicious Chiplets and hardware Trojans as major system-level threats. Chiplets may also be attacked leading to key extraction and secret leakage as shown in panel 4. If the root key or device secret is compromised, the consequence is much bigger than simply losing one piece of data. An attacker may be able to impersonate the device or undermine the trust relationship built on top of that identity. So I don't expect everyone to remember these attacks, but the key point is the trust boundary has expanded. Security can no longer exist only at the Board level or only around the CPU. What we increasingly need to know is this, is the device or die authentic? Did it boot trusted firmware? Can I trust the keys inside it? Can I trust the communication between components? And can I prove that trust to the rest of the system? This is why the root of trust increasingly has to move closer to individual silicon itself. And this is where architectures such as Caliptra become very important. Caliptra gives us an open source security architecture for establishing a hardware root of trust in modern devices. But there is an important distinction here between a digital root of trust architecture and the physical security foundation underneath it. On the left, we have the Caliptra subsystem. Within the subsystem, the Caliptra core includes things such as the RISC-V processor, cryptographic functions, SRAM and ROM, mailbox and the firmware responsible for implementing the security architecture. But ultimately, all that digital security logic has to anchor itself to something physical in silicon. So as you see on the right, there are three fundamental requirements for this physical hardware anchor, which we also call foundational security primitives. First, a unique device secret, UDS. We need something unique to the individual piece of silicon that can establish device identity and support key derivation. Second is a secure nonvolatile storage. Security critical information such as life cycle state, configuration, seeds or other protective value needs to survive power cycles and remain protected from unauthorized modifications. Third, a trustworthy entropy. All of the cryptography above this layer ultimately depends on good randomness. If the entropy is weak or predictable, the keys derived from it can also become weak or predictable. So Caliptra provides the architecture and framework for the Root of Trust, but the foundation of that Trust still has to terminate in physical properties and protected state inside the silicon. That is the physical security anchor. And those requirements actually match very naturally to technologies of hardware security that eMemory and PUF security have already been developing for many years. This is where we move from the architecture to the actual silicon implementation. For the UDS unique device secret, we use NeoPUF. Rather than simply programming the root of secret into conventional memory, NeoPUF derives device unique information from the intrinsic physical characteristics of each individual piece of silicon. That gives us a silicon-bound foundation for device identity and key derivation. For secure nonvolatile storage, we have NeoFuse, our OTP. NeoFuse provides protected storage for security critical information such as life cycle values, seeds, configuration or other persistent security assets. For entropy, we provide the physical noise source supporting the TRNG true random number generator, giving the cryptographic system, the randomness required for secure key generation and operation. On top of that, we're extending this foundation to the post-quantum era through PUF-PQC. Importantly, we view these IPs as different pieces of the same security foundation, identity, protected storage, trusted entropy and cryptographic agility. Together, these are the types of silicon level primitives that a framework such as Caliptra ultimately needs in order to establish trust. But having the primitive is only part of the problem. But for our customers, the next question is how to integrate all these pieces into a complete validated security system. This is where our strategy goes beyond selling individual IPs. We start with PUFrt as the hardware anchor IP. Around that, we add post-quantum security capabilities. And importantly, we have experience in integrating with Caliptra because from a chip designer's perspective, the difficult question isn't just, can I license a PUF? The real question is, how do I turn the specification into silicon that actually works? How do I connect the hardware primitives to the Caliptra root of trust architecture? How do I validate the interfaces? How do I shorten integration and verification time? That is why we're moving towards a validated subsystem approach. And once that foundation is available, the same security architecture can scale across accelerator boards, AI servers, rack scale systems and eventually data center infrastructure. What starts as a very small security block in silicon can ultimately become the trust foundation for a much larger computing platform. And that brings us back to where we started. This increasingly composed AI system. So the trust must extend across every layer of the AI infrastructure. AI infrastructure is becoming distributed across multiple layers. At the system level, CXL can protect data moving across the link through mechanisms for confidentiality, integrity and replay protection. But protecting the link does not automatically mean that the device on the other end should be trusted. The end still needs an integrity, trusted firmware, secure boot and also needs to be able to prove its state through attestation. The multi-die package follows similar principles. Each active Chiplet may have its own identity keys, firmware, life cycle state and relationships with the other dies in the package. So trust increasingly has to extend from the data center to the rack, to the server, to the device to the package, ultimately down to the individual silicon, and that's why our foundational IP is so important. NeoPUF, NeoFuse and our entropy source and PUF-PQC may be physically small pieces of an overall AI system, but they provide some of its most fundamental security properties, a unique identity, protected state, trustworthy entropy and cryptographic trust. PUFrt bring those capabilities together as a physical hardware anchor, while Caliptra provides the architecture that builds a root of trust on top of them. As AI compute becomes more composable and more distributed, we believe security has to follow exactly the same direction. Trust must also become distributed, but it must remain anchored in silicon. And that is the role we see for eMemory and PUF Security's foundational IP, providing the Silicon Trust Foundation that enables security -- that enables secure Chiplets, secure CXL connected device and ultimately, the next generation of AI infrastructure. Thank you for your time.

This concludes our prepared statement. Next, we will enter the Q&A section. We will now begin the Q&A section. [Operator Instructions] We will now collect the questions and begin our Q&A section.

[Foreign Language]

[Foreign Language] Our first question is, in the previous earnings call, the company mentioned that royalty revenue was expected to grow in the second half of this year. Are you now confident in that outlook? Joseph, please?

Yes, our outlook remains the same. We have already seen several advanced node products, including ADAS, AI accelerators, SSDs, networking and other new products, they begin to enter the mass production phase. And in addition, with foundry wafer prices gradually increasing, we believe the growth momentum will continue to accelerate.

[Foreign Language]

[Foreign Language]

The company has mentioned that the number of customers using 3-nanometer technology have already been licensed or could some of them being contributing to royalty revenue in the second half of this year. Joseph, please?

Yes, we have already seen customers moving into mass production with several hundred wafers. And this indicates that their products have been successfully validated, and the contribution is expected to become much more meaningful next year.

[Foreign Language]

[Foreign Language]

AI data centers are driving upgrades in high-voltage power architectures such as 400 volts and 800 volts HVDC as well as next-generation power ICs. Has the company started to benefit from this trend? Which product lines will be the main benefits [indiscernible]. Joseph, please?

Yes. For eMemory, the upgrade of AI data center power architectures creates opportunities through our IP being integrated into power management, power control and related analog mixed signal chips. And as voltage levels and power density continue to increase. For power ICs, they require much more precise calibration, parameter compensation, device identification and also reliability management and of course, the security functions. And the related IPs we can provide mainly includes OTP, which is mainly being used for analog parameter calibration, trimming, power control setting and, of course, product identification. And for MTP, they can support system parameters and configuration setting that need to be updated multiple times. And our customers have already started to move into mass production, and we expect the contribution to expand further in the coming quarters. Thank you

[Foreign Language]

[Foreign Language]

Could you share how investors should think about the contribution from advanced node product to your royalty revenue today and how you expect that mix to evolve over the next few years? Joseph, please?

In terms of licensing R&D contribution, advanced node projects already accounted for more than 50%. And this is mainly because the licensing fee for a single advanced node project is several times higher than average project licensing price. So looking ahead, as these projects they gradually move into mass production, we expect the royalty contribution to increase meaningfully as well.

[Foreign Language]

[Foreign Language]

Royalty revenue declined in the second quarter compared with the first quarter. Could you please help explain the main reasons behind the decrease? Joseph, please?

So for our Q2 royalty revenue, they correspond to the foundry production in the first quarter. And overall, mature-node foundry capacity utilization remained at a relatively low level in Q1 and some customers are also going through -- we're also going through seasonal inventory adjustment. And starting from the second quarter, overall foundry capacity utilization began to improve, while new applications started to enter mass production and together with subsequent impact from foundry price increases, we believe royalty revenue is expected to accelerate in the second half of 2026. Thank you.

[Foreign Language]

[Foreign Language]

2-nanometer GAA process gradually move toward mass production. Have the competitive advantage of NeoPUF and NeoFuse in advanced nodes become even stronger. Are customers also adopting this technology faster than they did last year? Charles, please.

As process technologies move from 2-nanometer and GAA chip design and manufacturing costs continue to rise and the customers are placing even stricter requirement on area power consumption, reliability and compliance with security standards. In this environment, the competitive advantage of NeoFuse and NeoPUF-based security IP become even more apparent. NeoPUF can create a unique hardware identity at the silicon level for each chip, supporting key generation and hardware Root of the Trust and NeoFuse provides reliable OTP storage and for the chip configuration, core patching and SRAM repair and other functions. This capability can be integrated with advanced logic processes and support the security, reliability and high requirement of advanced SoC. Compared with the past, we are indeed seeing customers start discussion on the security architectures earlier and the number of applications under evaluation is also increasing.

[Foreign Language]

[Foreign Language]