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

Afternoon, and welcome to eMemory's First Quarter 2023 Webcast Investor Conference. Today, we have our Chairman, Dr. Charles Hsu; President, Mr. Michael Ho; Head of IR, Ms. Li-Jeng Chen; and the Director of the Finance Department, Ms. Teresa Kuo. The format of today's event will be as follows: First, eMemory's Chairman, Dr. Charles Hsu will give an opening remark. Afterwards, President, Mr. Michael Ho, will summarize our operations in the first quarter of 2023, followed by our business outlook. Next, Dr. Charles Hsu will talk about how PUF-based solutions secure ChatGPT and AI. Then we will go into Q&A session, where our management team will be ready to answer your questions. [Operator Instructions] As a reminder, this conference is being recorded, and a webcast replay will be available within 3 hours after the conference is finished. Please visit the company's website, www.ememory.com.tw 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 the 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. I was told that some of the voice did not go through. Let me start it again. Okay. Good afternoon, everyone, and thank you for attending our conference call today. Although the first half of this year was affected by foundries overall low utilization rate, we are still very confident about our long-term growth. With the rapid increase in security demand, our license will be driven by PUF-related solutions. Since we have more than 1,000 new product tape-outs over the past 2 years, royalty will regain its growth momentum. As these 2 new existing holdups enter mass production last we are developing 3-nanometer with key customers and 5-nanometer customers will tap out soon and 6-, 7-nanometer customers adoption is accelerating. So all of this will speed our future growth, okay? So next, I would like to invite our President, Michael to share our first quarter performance and the future outlook.

Thank you. Good afternoon, everyone. Now let's begin with our 2023 first quarter financial results. The first quarter revenue was TWD 668 million, down 26% sequentially and down 8.2% year-over-year. Operating expenses were TWD 301 million, down 18.3% sequentially and down 4.2% year-over-year, mainly attributable to the decrease in bonuses and reverse. Operating income was TWD 367 million with a decrease of 31.3% sequentially and 11.2% year-over-year. The operating margin decreased by 4.2 percentage points sequentially and decreased by 1.8 percentage points year-over-year to 55%. EPS for the quarter was TWD 4.2 and ROE was 38.9%. Next, let's move on to revenue contribution by licensing and royalty. Licensing in the first quarter accounted for 21.4% of the total revenue, down 3.3% sequentially and 25.3% year-over-year. Royalties in the first quarter contributed 78.6% of the total revenue, decreasing 24.8% sequentially and 2% year-over-year. Total revenue for the first quarter decreased 26% compared to previous quarter and decreased 8.2% compared to the previous year. With that, I will comment on our revenue contribution by specific IPs. NeoBit accounted for 19.7% of the total licensing revenue in the first quarter, decreasing 22.6% sequentially and down 19.8% year-over-year. Its royalties accounted for 30.5% of the total royalty, down 31.1% sequentially and down 23.2% year-over-year. NeoFuse accounted for the 49.6% of the total licensing revenue in the first quarter, down 28.5% sequentially and down 42.9% year-over-year. In terms of the total royalty revenue, NeoFuse royalties decreased 21% sequentially but increased 16.3% year-over-year, accounting for 67.8% of total royalties. PUF-based security IPs contributed to 11.5% of licensing revenue, decreasing 60.7% sequentially, but increasing 138.2% year-over-year. Its royalties accounted for less than 1% of total royalties, which is down 93.8% compared to the previous quarter and down 46.3% compared to the previous year. MTP technology accounted for 19.2% of the total licensing revenue remained the same sequentially but increased 8.4% year-over-year. Royalty from MTP decreased 19.8% sequentially and 58.2% year-over-year, accounting for 1.7% of total royalties. Now let's look at the royalties for 8-inch and 12-inch wafers. For 8-inch wafers accounted for 47% of royalties, decreasing 25.5% sequentially and 9.2% year-over-year. 12-inch wafers contributed to 53% of royalties, decreasing 24.1% sequentially but increasing 5.3% year-over-year. In total, 137 product tape-outs were completed in the first quarter. We will provide more information in the management report later. In the next section, I will address our future outlook. We expect the second quarter to be a top of the year, mainly due to the continued decline in the foundry utilization rate with new customers tape out moving into production, we expect revenue to pick up momentum in the year's second half. For licensing revenues, licensing will significantly grow during the rest of the year, driven by PUF-based security solutions for the royalty revenues with more than 1,000 new product tape-out in the pipeline, royalty will regain growth momentum as new tape-outs move into production. Moving on new IP technology and business development. One, this year, the PUF-based solutions will be adopted in 5, 6 and 7-nanometer for the CPU, DPU, AI and automotive-related applications. 2, we continue working with foundries to develop the NeoFresh to increase the penetration rate of the mature process 3, lastly, we continue developing PUF-based security solutions in the most advanced process with our CPU partners. This concludes my comments. Next, I will pass the time to Charles. Thank you.

Okay. So next, I would like to talk about how our PUF-based solutions secure ChatGPT and AI. Recently, AI's application in ChatGPT has been a hot topic just last month, Samsung employees extensively like proprietary information, make companies worldwide aware of some of the risk associated with ChatGPT. Today, I would talk about how ChatGPT or even AI operations can be hacked attacked, then we will discuss how our PUF-based technology can help secure AI operations. So turning to the Page 12. Yes. ChatGPT is a great tool that can answer even the most context questions. That's why we decided to challenge it and see if ChatGPT is aware that it needs security and how this program goes about it. I will use this as an example to talk about how PUF-based solution can protect AI securities. So as you can see from the screen shot on the right, when asked, how do you secure ChatGPT, the answer is the list in the slide. Although it may not be completed, we just used it as an example. And from these answers, we have highlighted some of the key words we thought were important and the list of how our company can help solve the problem according to the bullet point --according to the point - each bullet point numbers, okay? The first, the answer is to protect the data encryption between the user and the ChatGPT. So to encrypt the data between the exchange between the user and the ChatGPT, our PUF can be used to generate keys for the encryption. And the second point is the mentioned authentication and the access control need to be managed in order to protect the ChatGPT. So we can use a PUF as a unique ID and they can generate secret keys for authentication. And the third point is how ChatGPT identifies security bridges. So by doing the auditing and also the lock. And this will also require the system to be able to identify user. So PUF can be utilized to generate such ID for the -- each users. So then in the future, we can test users, ID and also their log data. And for the first part is how ChatGPT regularly updates and fix its system. And this is typically done by over the air to upgrade the OS or the system parameters. So PUF-based IP is capable of signing firmware and the software updates, so through the over the air to update the systems. And the last one is the -- about the security awareness training and by adopting the hardware security, we can limit the chances of human error during the operations. So with these points in mind, it is easy to see how our security solution can be applied to AI applications. And I will further explain how it is done in the flowing few slides. So before understanding how our security solutions work, we must first understand how some of the popular attack that can happen, not only in ChatGPT but also in AI system in general. There are 4 types of major attacks. The first one is called poisoning attack. A poisoning attack happens when the attacked tempers to training temps to the training data, which will then affect the training model, allow the hacker to manipulate the results through the model. The second attack is called backdoor attack, hackers may also replace the model during a backdoor attack by adding additional neurons to the model. And the third attack is called evasion attack. So even after a model has been trained and hacker can still manipulate data resulting in an evasion attack leading incorrect output because of modified input data. And the fourth one is stealing attack, and this attack applied to all stage of the AI model or the training data model parameters, user data, et cetera, or variable know-how and also vulnerable to be stolen from AI. So basically, these 4 are the major attack to the AI systems. And this -- the next, I would like to talk about the -- each attack, how do you mitigate those attacks because each attack will require different protection methods and the solutions. So for the poisoning attack, the problem with poisoning attack is the trending data may be corrupted. And so we need to sign the training data to ensure the integrity of the data. So by signing the trending data to make sure the integrity of the data we will use the PUF-based solution for the signing. I think I will talk about that in the next section by how to use our PUF-based solution to do the signing of signature. And about the backdoor attack. The backdoor attack, the biggest problem is modifying replacing or even stealing the model. Not only does this require signing the model to protect its integrity, but it also needs additional encryption and the key management to prevent hackers from gaining access to the model. And therefore, the evasion attack, how to protect the evasion attack, unlike the previous method that target trending and the model evasion attacks cause a problem with input data. Attackers can tender with input data to influence the output of results. That's why we must authenticate and provision users use the unique ID and equip the asset. And about the stealing attack, the last attack is the theft of all the AI assets to avoid the attacks stealing valuable data and the know-how from the AI hardware. We need a unique ID, signing, encryption and anti-tampering or well designed to dissuade the attackers. So of all the problems we have mentioned, our PUF can play a very important role in solving these issues. We have developed 2 type of security IP. One is hardware rule of trust and the other is crypto co-processor. The rule of trust IP is also called PUF IP combined PUF which is so-called the physical uncomparable functions and secure IP and a true random number generator besides we have implemented multiple anti-tampering techniques to defend against attacks. On top of that, our crypto co-processor, which is called PUF CC comes with all of the above and have cryptographic hardware accelerate building on it. With all the industry standard crypto algorithm supported, we also have a firmware and the software such that our crypto co-processor can enable advanced security protocol and applications for the users. So after understanding the risk of all -- each stage of AI used to operate ChatGPT, it is easy to see that the PUF-based solution are efficient for all security requirements throughout the AI product life cycle from training data to influence results, the AI developers know-how must be protected from copying and stealing. PUF-based solution are secure and cost-effective and the IP blocks make them best suited for securing AI applications. And this concludes my remarks. Next, I will -- we will enter the Q&A section. Thank you.

Thank you, Charles. This concludes our prepared statement. We will now begin the Q&A session. [Operator Instructions] [Interpreted] So the question asked was compared to quarter 1, there was still no improvement in PUF in quarter 2, and the inventory level of design houses remained high. Since the company's royalty revenue is a quarter behind that of the PUFs, what is the outlook on royalty in the second half of the year, Michael?

We have seen that the earliest customers who adjust their inventories started wafer production in February. In addition, the new tape-outs over the last 2 years were for new product applications and will begin mass production in the latter part of the year and will drive the royalty growth. We expect the royalty to grow sequentially in the second half of the year. Thank you.

[Interpreted] So the question asked was eMemory has been very successful in the development of DDI, PMIC and analog applications. May I ask which new applications will drive the company's royalty momentum in the future? When do you expect to start seeing noticeable contributions? Michael?

Besides the DDI and the PMIC, we also have applied it to other applications such as various sensors because the die size of sensors is related to small, the amount of wafers used is much less than that of DDI and PMIC. So it is not as obvious. As we develop into more new advanced process nodes, for sure, we will be adopted by even more applications on new process nodes once qualified. For example, when 28-nanometer was ready, customers like ISP, Wi-Fi, TWS, SSD controller, KCOM and other high-speed IO start to adopt our solutions from one customer to many all allow the pattern of DDI and PMIC and will eventually become industry standards. Regarding which application will be the next driving force for royalty revenue. We believe the ISP-related will be the next, followed by the Wi-Fi TS and the networking related. As we accumulate production records for 7-nanometer and the progress to final meter and the 3-nanometer, we are confident that other processor-related applications will be the next following areas. Thank you.

[Interpreted] Okay. So the question asked was since the largest portion of the company's royalties come from DDI, have we seen a recovery in DDI customers. Michael?

Customers started wafer production at the end of Q1. Besides from the recent report, we see the 2 big foundries in Taiwan with the higher DDI ratio, the April revenue is pretty good. And another big DDI company in Taiwan, they also have a quite good Q2 forecast. So I think that could be the good index about the DDI market.

[Interpreted] The question asked was are there any royalty contributions that reflect the company's progress in automotive and AI customers? Michael?

For automotive customers from like a 7-nanometer ADAS and the 28-nanometer networking related, they are ready for to the mature process such as DDI sensor and the PMIC-related our mass production. For most of the AI customers, they are still in the tape-out stage only minimum volume production.

[Interpreted] So the question that was asked was, you mentioned before that most companies are your customers. Why is the market share for 8-inch capped at 20% if that is the case, Michael?

Besides our large MVM technology, there is another 8-inch embedded memory non-multi-memory technology, coated embedded fresh. It was developed by like SST or ISSI in the 1990s and many used in the MCU applications. Its market share is also around 20%, with the remaining hours used the foundry e-fuel solution. We believe that the penetration rate of mature process have room to grow because we still see customers switching from the EVs to our OTP due to the density and the performance requirement. In addition, our MTP related technologies provide a better cost performance than the embedded fresh. For example, compared to the embedded fresh, our new fresh doesn't need additional investment in the equipment and use fewer masking layer with much lower wafer cost. It is also suitable for specialty process. For that part, the embedded fresh is not available and is currently being developed with many foundries' customers. For our new E square, it has been adapted for PMIC in the DDR5 modules and will start big volume production at the end of this year. The above reasons we'll continue to increase our penetration rate in mature process. Thank you.

[Interpreted] So the question that was asked was are Amazon and Google pushing CPU and GPU partners to adopt PUF-based confidential computing. Michael?

Customers are currently correcting with us for security-related applications.

[Interpreted] So the next question is security related. So the question is modern processors are often attacked, can PUF Protect them, Charles?

Okay. The process of protecting information usually is divided into 3 categories. One is to protect data in storage and the next is to protect data in transit and the third one is to protect data in use. So to attack the process to attack the processor is to attack the data news belong to that category. And in order to accelerate the speed of information processing, current high-performance computing uses multicore architecture and shared virtual memory to facilitate the application of multi-tasking by multi users. Because of the shared virtual memory, the attacker will have the opportunity to manipulate the other programs through the virtual memory. And the solution is to ensure that each program can only use -- that's netted virtual memory during operation and not the assets other virtual memories. So PUF can protect the data in use in CPU by providing a key tech to computing programs and their corresponding virtual memory. So each computing program can only use its own key tech to assess their own virtual memory, and this will prevent the attackers from gaining access to other programs operation through the virtual memory.

[Interpreted] Is this year, a licensing fee is the focus of eMemory's growth. What are the main process and applications that are adopting your technology? Charles?

It will be mainly driven by a PUF-based security solution and 5, 6 and 7-nanometer application in CPU, DPU, AI and automotive-related applications, which are the major -- our major growth on the license fee.

[Interpreted] So the question that was asked was, what is used to protect the hardware security of AP processors and mobile phone factories? And follow-up, what is the benefit of switching to eMemory solutions? Charles?

Okay. Currently, the hardware security mobile phone manufacturer use is the e-fuse to store the public and the private kids. It does not use a chip fingerprint to protect the entire hardware rule of trust. So our solution can obtain the best hardware rule of trust, while the top chip fingerprint protect the whole security system. At the same time, the public and the private keys from the chip can significantly reduce the security concern and the cost for an injecting key.

[Interpreted] So the question that was asked was a little complicated. Let's try to repeat it in English. So in the past, the encryption protocol RSA played an important role. How will NeoPUF replace it? In addition, the bios can also be put into the security authentication secure boot. Is this mutually exclusive from NeoPUF? Charles?

Okay. I say is the algorithm which can generate -- which can generate public key from private key, okay? And the key include PUF key, private key here are mainly used for authentication and the signing of the security functions, okay? And the NeoPUF is the chief fingerprint, which can provide the unique private key. And this private key is input to the assay and the output is a corresponding public key, okay? So once we have the key payer include a public and the private key, so we can use this key payer by using the private key to protect the files. The system will use the private key to protect the BIOS and send this to the user. And this -- the user will use the public key provided by system to decrypt the BIOS protected by the private key and then we will then start the [indiscernible] product securely put the process. So this is 3 different 3 items: BIOS, NeoPUF and then BIOS, there are 3 different things, okay? So their relationship is as what I just explained.