AudioCodes Ltd. (AUDC) Earnings Call Transcript & Summary

Well, Marina, it's great to have you join you today for this webinar on 3 Trends Driving the Future of AI. My name is Jim Lundy. I'm the Founder and CEO at Aragon Research. And we're really going to do a deep dive into edge computing and AI and why the 2 are intertwined and why they matter. It's a very hot topic in the industry. It couldn't be a better time for this webinar, Marina. I'm really excited about this. So we have a lot to cover. And let's just talk about the agenda, and then I'll let Marina say a few words. We're going to make this interactive, by the way. So -- and this is being recorded in case you missed it. But we're going to talk about key trends and this whole idea of what we call edge, some people call it on-premise, but we really never left edge, but it's back and it's back in a big way. Why edge infrastructure is -- there's a big rationale, including cost, particularly something called tokens. And also that sometimes when you run AI closer to the information, you're going to do it faster. And there are some things around data privacy we'll talk about. And then some practical advice. And of course, we're going to hear a lot about what AudioCodes is up to in this area. So Marina, why don't I turn it over to you for some intro?

Jim, nice having you. It's a great pleasure having you here. My name is Marina Risher. I'm Senior Product Manager in AudioCodes, and we are welcoming everybody who has joined to the webinar. Please pay attention that you have a Q&A session, and you are welcome to write your questions during the entire webinar. We will review them and near the end we'll address them. And whatever questions we will not address, we will release a summary of the webinar with all the questions answered. So let's start.

Okay. Great. So look, so we've got some trend slides. And the idea is to kind of really get you at the end of this webinar, fully kind of aware of the kind of scenario we're looking at. I'm based in Silicon Valley. I've been analyst for a long time. We do use the technology. So this whole era of agents, and you can call them assistants, but we're kind of calling them coworkers or digital workers. It's been moving very fast, particularly in 2026. The era of cowork arrived in February. And literally, there's a new race on. And it's -- the whole idea is that it's not just about having agents, but the idea of an assistant now, even today, things called a personal assistant that can do work for you on your behalf. And of course, one of the things that comes up when you talk about that is something called agentic identity and security. That's a hot trend. The only problem is a lot of the products for agentic identity and security are not shipping yet. So some people have a kill switch. But understanding who the agent is or who the assistant is that is doing work on behalf of a human is very, very critical as is making sure your infrastructure is set. So we're going to talk about that. But the whole idea in 2027 is this whole idea of this full agentic digital worker that can work on your behalf and with token costs going down, can do it for a long time. So that's kind of a little bit of a teaser, but we're going to dive into all this and more. So let's go ahead and keep going. So the return to edge computing. So there's a lot that's kind of going on behind the scenes. If you're watching the news, you're going to think that all AI runs in a hyperscaler cloud, and that's just not true. The real answer is that a lot of AI is going to run on your phone, on your computer or in your local data center. And part of the reason for that is information security, data sovereignty and cost. So there is going to be a massive demand to upgrade PCs, Windows PCs in particular, because local compute power, i.e., GPUs and inference chips, will make your machine, your local machine, your computer or your phone, an AI machine. So if you haven't heard about that, at your enterprise, you probably need to start asking the question, particularly if you're like a managing general manager and you have employees, what is IT doing to refresh PCs? It is no longer about the cost of the PC. It's actually about the power of compute power of the PC. And there's a big race there. There's lots of different architectures, but the GPU-based approach with cores is the new way forward. Private data centers, sometimes called neoclouds, also very, very big. Massive investment. Most data centers today, in fact, about 90% of data centers worldwide are not AI-enabled. And then the last thing, which some deals were struck this week on some nuclear power with Google is how to find power. Now there's next-generation data centers that don't use heat-seeking GPUs, but power is going to be one of the biggest issues for local data centers as well as hyperscaler data centers. So Marina, we try to define this as a quiet revolution, particularly for many regulated industries. What's your comment on that?

Well, with all this race to the computing to the AI usage in AI, which mostly goes to cloud and private data centers, there are some industries which were not meant to go to cloud from the beginning. So we are seeing in government and defense sector that the race to AI was waiting for the solution to be ready for also on-premises as AI capability that can run now locally on the private data centers with local data centers or sovereign clouds is something that now can be used also by government and defense, which were never planning to go to cloud.

Yes, I totally agree. In fact, I know a lot of work with clients and a lot of industries where they never went to cloud. So it's kind of like SaaS applications will still probably be SaaS, but there's many, many enterprise workloads where you don't want to move the data for a lot of reasons. And part of it is cost even before AI. So it's a trend. And that's why, by the way, getting a hold of a Dell GPU server is almost impossible. So that means that people are buying them, and there's a reason why they're buying them. So the other thing that's happening, and this is really next year is robotic edge. And so it's not just fixed data centers. It's actually devices, including robots and drones and transportation vehicles that basically will need edge compute capabilities, a mini network, many cars have that already. But we think that the demand for more capacity there, more video storage, for example, for cars and trucks. So there's actually going to be an awful lot of demand for microprocessors and for edge AI because you actually have to have the AI running on a device for it to function. In fact, 80% of computer vision applications run on a local device versus in the cloud. So that's another trend. A lot of people kind of like, it's not going to happen. But we actually do see this coming very, very fast. And I mean, for example, deliveries with the drone, mainly rural areas like Australia. But now you have -- I'm actually in Austin, Texas right now and the cybercabs, which are -- there's no driver, wildly successful, wildly successful. So it's kind of crazy, right? And we think a lot of the robots will launch in the $20,000 to $40,000 price range for elderly health care in the next 2 years. So that's another thing that basically implies that, that's not going to be a hyperscaler data center at all. And for the business, you need to think about the ramifications of, for example, what can -- how can we automate things and what's it going to do for my company? All right. So we're going to dive further in here talking about a lot of the stuff going on with AI assistant and agents, and it's really a preparation period. This is the time to get smart on it to understand what agents do to understand what agentic agent systems do. And just one of the things on the slide is if you just look at the weekly active users, it's crazy high, and it's going to keep going up. But we have an example in this slide of a sales example where you have a sales assistant for an account executive and then that assistant can invoke subagents for appointments, for follow-up, for proposals and contracts. And I'm going to be honest with you, there's only a few people that are actually doing this today. Some people are actually saying, hey, we've got one sales assistant and then they're actually going to hire somebody to build these subagents. Guess what? You can actually talk to some of these studio tools now and build them yourself. But you got to make sure they're secure. They got to make sure they don't hallucinate. So testing is -- it's like a system deployment. So this is a great use case. Marina, what do you think about the sales use case? Notes, I think meeting notes is one of the most important things in sales. And I know that's something AudioCodes does. So can you comment?

Definitely. We are in the business of meeting intelligence where what happens in the meeting is very important because it includes not only the transcript, but most important decisions, commitments, objections. And it's very important that all this information will eventually be shared among different organization members. So not only creating the summaries or action items is very important, but also make all this automatically inside your CRM system. So as you said, Jim, the automations and the agents which are not only summarizing, but also integrating into the CRM system, the notes. So the next meeting starts from the point where you already have a summary, you already have all the points mentioned in the previous meeting and just not start with blind zero, you are sharing organizational knowledge. So all the people involved can definitely take and benefit from it. And this shared context, which goes directly from the meeting intelligence into the CRM systems is the strongest advantage of meeting intelligence applications. So we do see the usage, and we do see the agentic being transformed from action items and summary into integration part of CRM systems.

Yes. And I would just comment the #1 issue for sales teams is notes. It's the #1 issue.

Exactly.

It's getting notes into -- yes, it's very hard to get them into the CRM.

Exactly. And the part they need to go, open the ticket, make it manually or make a summary by themselves is tedious work and they live on meetings. Sales is based on the meetings and the automation that, from the meeting, you have automatic summary, action items, decisions and that's automatically entered in the company CRM. It's a huge momentum for all those tools.

Absolutely. So yes, so there's a lot of things on the horizon. The world is changing before our eyes. It is time for you to prepare and look at all the different use cases where you can -- basically, these things are going to help to automate. And really what it's automating is knowledge. So to try to net it out, where we were before and where we're going. In the past, it was basically applications where you had to go and know what buttons to push to get information out of the application. And now what you're seeing is this idea of something called headless. A headless experience where the human instructs the coworker assistant to do the task and the assistant goes and presses the buttons. So it's conversational, and you have to have the right connections into it. But just the example Marina mentioned about getting notes into the CRM automatically is still to this day, one of the biggest tasks we hear about because we cover revenue enablement, and it's still the #1 issue. But there's no other issue that's bigger than where are the meeting notes. So I think you're hitting the home run there. Okay. So look, we're going to kind of transition a little bit and talk about technologies, and we have a little video we're going to try to play here to talk about something called an Aragon Research Tech Arc. So I'm going to go ahead and advance the slide and start the video. [Presentation]

Okay, so that's a quick flyby. It does a much better job videos of explaining things than either Marina or I could do in the same amount of time, which was less than a minute. But we just highlighted some of the things in the video, agentic systems, agentic security and privacy, other things that are starting to be integrated into some of these agentic systems, including deep learning and reinforcement learning. But again, these are -- the idea of this is that people are not waiting for the trough of disillusionment. I was at a company that started with the letter G. What we see is waves and early adopters and even middle adopters are jumping in with some of this stuff right now. They're not waiting until it's mature. So it's kind of a great way to kind of look at where you are and are we missing out? For example, if you don't have, for example, a plan for next year for agentic systems, you probably need to get one. So anyway, so like I said, we -- these are about 45-page reports, but the reason we do videos is to make it easier to digest for -- particularly for business executives. All right. So let's keep going. We're going to talk about this whole idea of not only the edge battle, but the discussion that's going on in the industry right now about open weight versus closed models, basically open source versus closed source. And what people have been led to believe is that you must buy closed source, meaning that you're buying a subscription to maybe one of the frontier providers like OpenAI or Anthropic. But there's this other thing called open source. And so when you start looking at right now, the big enterprises are looking at edge data centers with an open source LLM that's clean, probably made in the Western Hemisphere because, first of all, they can do the same things. It does need maybe a little bit more training. But the cost -- let's just say, if you had -- you're spending $10 to evaluate 100 documents, you might spend $1 with open source or even $0.10. So the hyperscaler factory is generally going to be using their models. They're going to guarantee things. However, when you actually start looking at agents and if you're running their agents in their data center, which is digital labor, then a lot of the profit will actually go to the hyperscaler that's in your country. So there's actually kind of a race going on like do we want the hyperscaler in our country? Or do you want to have our own local data centers where we run the AI. So there's a lot to unpack here, Marina. We know that and you mentioned this earlier, there's a lot of firms that have really started to wake up and realize that was the SaaS era, just an era and can we still do it on our own? And I would argue that they can. But give me your take on what you see here.

We definitely see that it's a possibility. One of the solutions we're providing today is the meeting intelligence solution, which is running fully on-premise. And as you properly said, the models are getting smaller, and you can have a smaller model, it's called SLM, small language model. And those small language models coming pretrained or we are making the training for them. And then when it is installed on the part of the application, you are running fully the entire model and the entire solution components like STT, speech-to-text agents, LLMs, and TTS, text-to-speech components, all those components can be running locally on your premises. And we see huge importance of the ability to provide such solutions in government, in health care, in defense sector. And I will speak later a little bit about asylum interviews. So any place where you cannot -- where data cannot leave the premises of the organization and before the edge computing era cannot -- could not benefit from AI. Today, it's possible. Today, we have solutions with models which are adaptable and can be packed up as a solution with proper scaling and fully running on premises.

Yes. It's a big race. And I think that the light bulb is starting to go on in a lot of large enterprises. But remember that with small language models, as you said, Marina, you may be able to run a lot of stuff even on local PCs, MacBooks, Mac Minis and also HP servers, Dell servers. I mean there's a lot of server providers that basically can run SLMs and actually now large language models. So you have a lot of choice. Do not believe what the hyperscalers tell you because they're now going to say, oh, we can put stuff in your local data center. And they can, and they can do that. But you also have to look at cost. But again, a lot of times, it may be a managed service or you may run it on your own. But remember, with a command-based interface, talking to the application is much, much easier than it ever was. And you might be surprised, we've even proposed that like some field engineering teams or field SEs that they have kind of like a jam session and compete against product management about who can make the best agent. And they're like, oh, let's do that, right? I don't think the product management people like it too much. All right. So look, this whole race, and this is just a little bit of background. It kind of talks about this whole race to build. They're calling them AI super factories. And what this chart shows is last year into this year, about basically 85% to 90% of data centers around the world were legacy, meaning they did not have AI compute capabilities. But we're looking at by 2031, a massive build-out where between 47% to 66% may have AI componentry built in in the supercompute data centers. So there's massive spending. And we think there's also going to be massive spending for local data centers. So basically, these are new knowledge engines for economies, and you don't want to miss out. So there is a race. And the fact that hyperscale is going to give you a data center for free, it may not be the best free, right? It may not be the best. So Yes. I would just say, Marina, this is kind of like the race to who owns AI and is it the hyperscaler of the enterprise. So what's your take on this?

We think that eventually, it will be the mix. We don't think that it's black or white. We don't think -- we believe that part of the data centers will stay local and they adopt local models for AI. And companies who are using SaaS models and running in cloud will make a decision what are the workloads that they prefer to leave in cloud as a SaaS model. And some of the data workloads will be running on-premise because it's a hybrid environment that we expect that will eventually hold for the long term, where you don't have to decide for the entire organization and for all the applications, if it's going to be fully on-premises or fully on cloud, you can make a decision per your application and per your usage. And it's not only the -- what is the LLM used and where the model is running is who is control the operation, who is the owner of the software and who is running and decided when the software is upgraded, what is the life cycle of the software. All those issues are way beyond the LLM and how to run the LLM on-prem or in cloud. So basically, we see this trend. And we think that eventually, the hybrid infrastructure will be used for the next years.

Yes, we agree. It's a great race. We also agree it will be hybrid. There will be a mix of both because there's still some things that you may want to have a partner do versus doing it all yourself. But again, the power of choice. But let's peel the onion a little bit more because one of the things that's coming up is a lot of people are railing against these AI data centers, super factories. The big shift is going to be towards custom chips or ASICs or inference computing because it takes lower power, lower water, sometimes smaller form factors. So NVIDIA is the leader in the GPU market, but they also bought Groq, which actually makes ASICs. So they're hedging their design bet. NVIDIA is also really a big supporter of open source. So they're providing models and supporting and they just bought Hugging Face. But this is just so you know, if you're going to say, hey, I want to do a data center, you don't always have to do it with one design. The forward-looking designs are basically these inference processors. Again, these are all basically processors. But again, there's just different design points. And you can look at the graphic on the left, that's a lot of hardware and it gets hot. It's a graphics processor. So that's one of the things you say is like how much heat does it put off. And this is the newer design. So you actually see some of the hyperscalers making their own TPUs, tensor processing units, partially due to cost and partially due to compute power. But right now, NVIDIA is the #1 choice, and they're powering everything and the GPU stuff is -- the TPU stuff is coming on gradually. So that's just a little bit of background. So if you're going to do your own data center, you need to start looking at part 1, part 2, maybe start with GPUs, then transition down to ASICs and inferencing as supply becomes available because right now, you can't get a lot of GPUs, right, because all the hyperscalers are making them and keeping them. Okay. So this is a real important topic that we see. And again, it's really just one of the things where you kind of have to have it. Agentic identity and security. It's tied to data privacy. But basically, agents are kind of like little kids, they can find everything, right? Agents can find everything. So if you've got stuff on your servers, on your PCs, it's going to find it. And so things like getting passwords off your systems, having passkeys and multifactor and also having agentic security and agentic identity. That's why like the personal agent thing just came online is a form of that. So it's a big deal. And you have to kind of when you say, hey, we're going to do agents, the security stuff is not trivial. As an observation, what I would say is we work with many CISOs and they mean well. But a lot of times, the -- getting everything deployed so you have like something close to Zero Trust, they don't have that. A lot of enterprises just don't have that. They have a lot of tools. They haven't deployed them. They think one tool will do it all. So there are going to be many providers that will actually have a guarantee with their products that say, hey, we're not going to -- we won't train on your data, we won't take your data, and we have partners that will guarantee that your stuff is protected. So that's just something to keep in mind. Marina, I don't know if you have any comments on that. I know you guys deal in a lot of regulated industries. I know that's a requirement. Do you have any comments that you have here?

Definitely. We see in the regulated industries, a clear requirement is to keep data -- some of the data should be kept on-premises because those are regulation requirements. The regulated industries have regulation requirements where data cannot leave the premises, and it's applicable for health care when privacy -- people's privacy data cannot leave the perimeter of the server of the organization. And it's applicable for government sector when you have citizen data, which also should be kept in local storages. So a lot of IDs, credit card numbers, et cetera, for government, defense and health care sector are definitely required to be kept on premises.