Qualcomm Auto Summit 2026: 10B+ On-Device LLM – Is the Car Truly Becoming an AI Agent?

Colleagues and partners, I'm very honored to attend this annual meeting of Qualcomm. I should say, as an automobile enterprise and a world-leading semiconductor company, it is a trend for us to have many projects. It is a trend towards the future. So what I want to introduce to you today is actually about chips. As I mentioned earlier, there is a Flex architecture 8775 chip. As for BAIC Group, last year we mass-produced the world's first solution based on the Flex architecture, integrated cockpit and ADAS solution. We are on the Jihu Alpha T5. Here, I want to share with you, we tried the world's first solution with this. Our entire development process. Let's turn the clock back to the fourth quarter of 2024. In fact, the problem we faced at that time was that high-level assisted driving had been mass-produced. Most of the models equipped with it were all above 300,000 RMB. So we understand that assisted driving, especially high-level assisted driving, the urban NOA we often talk about, should be standard equipment. Whether it's a 500,000 yuan car, 300,000, 200,000, 100,000, or even cheaper cars, should be standard. In the history of automobile development, we have seen that the ABS and ESP that we thought were optional features are now standard. So based on this idea, we thought, for the price range of 100,000 to 200,000, especially with 120,000 as the entry price, to mid-range family cars of 200,000, and high-level assisted driving, this is a huge challenge for us. We know that good things are expensive. Hardware takes up a lot of cost: the cockpit and autonomous driving, two chips, two pre-controls, and two power supply systems. This is a huge challenge.

In the fourth quarter of 2024, we encountered another phenomenon at that time: from BEV to end-to-end, whether it's two-stage or one-stage, from rule-based and code-based to a model-based pricing scheme, to change again. At that time, we heard many industry experts or colleagues say we were considering whether to delete the code, whether to continue with rule-based or end-to-end. Meanwhile, in the automotive electronics industry, there has always been a theory to guide us forward, which is the evolution of electronic and electrical architecture. It should be based on central computing, and the combination of software and hardware. So in that environment, we held a technical seminar in the city of Wuxi, which strengthened our determination and confidence in continuous progress. So we can see that we should have set our goals and we should find solutions. So we started to look for, from chip pricing, pre-control to project management. We also did some work ourselves. BAIC, in terms of intelligence, we have done a department integration ourselves. We know that many car companies have the cockpit and intelligent driving as two teams, two independent R&D teams. We have always insisted on being one team. We have a director in charge. So on this basis, we believe that with the 8775 chip, to promote the cockpit-integrated solution. So we formed such a four-party development and cooperation team. BAIC is responsible for demand definition, project control, integration, and verification. Qualcomm provides the solution of the underlying chip. Chelian World, as a pre-control explorer, and strongly promote the project's implementation. We also collaborated with Zhuoyue and City NOA to make the model smaller and make the experience better, and jointly plan.

As you know, the NPU dense computing power of the 8775 chip is only 72 Tera. In the current high-level autonomous driving, this solution is required. How can 72 tera operations per second? This is also a huge challenge. So our four-party joint team launched the project on February 19th last year. On October 30th, the first mass-produced delivery of Jihu Alpha T5. It took us eight months. The whole process was very tangled and challenging. Many friends in the industry told us, don't touch this: two underlying architectures, two systems, and a large amount of data communication and handshake. Vehicle high-level assisted driving requires millisecond-level, but the cockpit doesn't need it. This is an impossible path. We thought that since we chose this route, we must make it work. So the four-party team has realized mass production. Now we not only have Arcfox Alpha T5 and Alpha S5, including the recently released Wendo V9, are all equipped with the 8775 solution. So we can see that we doubled our performance in Jihu last year. Our consumers' experience with this solution, especially the experience of urban NOA, is very pleasant. And we have also controlled it well. Especially in the face of chips and memory prices rising, we share a memory system. We also developed standards for enterprises and accumulated experience in standard processes.

So in the future, we still need to focus on chips. We believe that the next generation of VLA-based, trained on world models, high-level assisted driving will further require computing power. So the cockpit fusion route we think is a viable path. At the same time, this platform should be continuously upgraded. The increased capabilities should not only include but also AI capabilities. So let's imagine the new generation of cars. We think it should be at the end of this year, in 2027, we should be the standard of the new generation. In terms of assisted driving, the car understands you better. With you, the city, and the environment are smooth and fluent interaction. This uses language to guide driving. This should be no problem. At the same time, we should put a smarter brain in it. Maybe like Jarvis to Iron Man, and omnipotent to assist. So we also hope that this is an ecosystem, not just starting from today, from solutions to the definition and development of vehicles, including the user's usage environment, to form a definition and development and continuous iterative working environment, to provide consumers with more complete, more satisfactory and safer solutions. Thank you.

I'm very happy to have this opportunity today to share with you what we are doing at Banma. You must be wondering why am I representing Banma to talk about this? Because in about 2014, Alibaba, at that time, we proposed a concept called Internet car. At that time, the Internet had reached its peak state, especially Alibaba's own e-commerce was also developing very well. And at that time, artificial intelligence represented by deep learning, this technology was just emerging. At that time, there was also a background: the OS, especially represented by mobile phones, various operating systems was also on the rise. So at that time, I had an idea: whether we could make cars become more interconnected and connect with a huge life platform like Alibaba. I remember very clearly, at that time, our first idea was we wanted to use maps as the core of an intelligent cockpit, and we spent a lot of effort and then we made our own OS system of Alibaba. Of course, at that time, I thought that various hardware and conditions were still very restricted. My cockpit was also like Qualcomm's, which was originally migrated from the mobile platform, unlike today. Today, I think the most important thing is that Qualcomm has launched a dedicated solution for smart cockpit and also has a dedicated solution. This is a huge progress in the industry. So since we launched this smart cockpit in 2015 to today, it has been about ten years. The industry has also undergone earth-shaking changes, but the technology has also undergone earth-shaking changes. Well, everyone knows that I am also an engineer. I also come from an engineer background. I code. In terms of software and systems, it has changed a lot. Whether it was the Internet or something else, it was actually data-centered and database-centered. We were just reading, processing, and writing back information. Everyone, essentially, we were doing the same thing. Most of the interactions were command-based. It was probably such a system. So in that era, it was very simple. Think about it. If there is such a pattern, who would be the best in this field? Right, like databases. It must be the database, right? Because all systems revolve around it. Today we are all talking about artificial intelligence, right? Artificial intelligence has undergone a great change. I think for the software industry, or future systems, will bring a fundamental change. Because the current entire system is not built around this database or something like that, or some other basic building blocks. It is built around an intelligence, the so-called AI. What is the most important ability of AI? Reasoning. So it is based on this reasoning ability are built. So I think in the future, all systems will be reorganized around reasoning. There is a reasoning here. It's more powerful than previous systems and your human ability. So all systems need to be restructured. I will take the part of my original system parts that require reasoning, I have to hand over to this so-called large language model. Right? So what did I do before? What I used to do, I need to do it at night, right? If I need to, and then complete the specific work. After completing the special work, how to do it better? Specialized work. Now there are many new words invented: skill, choreography, and crawl and so on. Just now, this brain hole and many previous sharing, they all raised a very good question. I have very good brain and I have very good technical facilities. So how can I build such a rich application at the top? This is what I think is the value such service providers should exist.

As we all know, the most important contribution of modern industrial civilization is division of labor. Everyone does what they are good at. But I know that there are many car manufacturers in China. They are vertically integrated and they may be very successful. But I think division of labor is not about sharing the cake, right? Division of labor definitely creates new value. We empower merchants, making the industry more prosperous and make the industry more innovative. This is what I think is our purpose. Although Alibaba also makes chips, and our DAMO Academy is also making chips now, but we don't make car chips. We think it is a very professional industry. So what do we want to do next? What do we think we should do? Let's look at the automotive intelligent cockpit in this field, including brackets, what changes have occurred? Because it is a car after all, right? I am in the loop. We can divide the system into people in the loop, and people are not in the loop. The system with people in the loop, it needs to perceive the world. So our car as a physical carrier, it will definitely have more and more perception abilities, including voice, including air conditioning temperature. Anyway, many of them have this perception ability. After having so many perception abilities, how can I make it more intelligent? I think the first step is to integrate all these perceptions, right? Second, I need to reason, right? And decide what to do. This is what a language model should do, what a language model might do. Third, I need to provide a service. How can I use the perception I have obtained as an intelligent reasoning, what kind of result I can get. All these things, I think we need a new system. The core purpose of this new system is to shield the complexity above. And for the application development above, to provide more valuable, cost-effective, innovative means to allow our customers to develop faster. As Liu from BAIC just mentioned, we might have taken eight months and it will take a long time, but now it may be shortened to three months, right? Because we can do a lot of work. From this perspective, we are thinking that what's the essential difference between cars? What is the essential difference? I think in the car, it is very, very important. It needs a very low latency and a very natural interaction. As I mentioned earlier, I think we haven't fully achieved it yet. We are working hard in all directions. You can't wake it up. You can't just say a word to him and wait for a while for a response. You can't always take the initiative to initiate everything. There are many similar things. So we call it active intelligence. This is DS. And my model is different from the Dayu model. The Dayu model, everyone knows, now the manufacturers that provide the model are all about depth and breadth. Because they need to complete long tasks and do long reasoning. Their current goal is to work continuously for 24 hours on one task and continuously finish it. But obviously, this is not the case. A car may be I provide services related to the purpose of travel, and the travel itself, including emotional value, including some destinations for travel, including entertainment and services. This is different from the optimization goal. So the goal we propose is we need to provide a very sophisticated model on the side. Its core is to serve the travel itself and the car itself, and provide extremely low latency and cost. This is the core. We have also launched our own, called Auto Command, this model.

And then, our next step, the fans I just mentioned, right? The cooperation between Alibaba and Banma, Banma's cooperation with various car manufacturers. Our core foundation, we still provide an AliOS, right? We hope to shield the complexity of the hardware to provide a very good operating system. Today, this operating system, Qualcomm is also talking about we need to integrate, right? We need to integrate computing power, and isolation in security, and intelligence, to support large models. Of course, there is also data privacy protection and so on. This is what we hope to provide: a basic OS and then provide a basic ability on it. On top of that, we hope to provide some of the native applications, including integrating services. There are many Internet services, right? There are also many entertainment services. We don't need every manufacturer to connect with us. We hope to integrate these standard and Internet application services. Thirdly, we hope that during this event, our partners can develop their own personalized and customized applications. This is what Banma hopes to do. We can achieve. Well, today, I think it is also very meaningful to have this. I'm very grateful for this opportunity to meet so many colleagues in the automotive industry. After listening to the previous sharing, I feel more confident, right? And Qualcomm, today, we provide a very good computing platform. Like Banma, provides a very good operating system and intelligent native and future-oriented operating system. Then we provide basic applications, and then together with our partners, to enrich the entire automotive application ecosystem and accelerate innovation. Thank you all.

Thank you, Mr. Nakul, and all the distinguished guests for the wonderful sharing. Next, let's welcome Qualcomm Technologies Vice President and General Manager of ADAS and Robotics, Anshuman Saxena. Let's welcome Anshuman Saxena.

Good morning everyone, everyone who's visiting us here at the Qualcomm Automotive Summit. It's great to be back, fourth year in a row, and looking at the stage, looking at everything that is happening over here, it's just so exciting. I talk about scaling the intelligent driving. Nakul covered a lot of what we are doing in the overall automotive, but I'll focus a little more on the intelligent driving, which is basically ADAS and beyond kind of a thing. But before I start any of that, I just want to thank everybody, all the automakers that have given us an opportunity, embraced our solutions in China, globally picking up Snapdragon Ride solutions. It was in 2023, we came here, I came on the stage. We of course had a cockpit business that was going on, but it was new for Snapdragon Ride, and the China ecosystem was basically good, but what have you got? And since then to now, we are there in majority of the automakers shipping based on Snapdragon Ride. And it's not just because the product is good, but it's a humble feeling that the ecosystem allowed us to give us an opportunity and trusted us, which we are delivering now. So really appreciate that. Thank you for that.

With that, how are we seeing the whole AI evolution inside the company? And this is important because that kind of size drives how we are approaching the problem statement across the board. We in the ADAS world particularly started with the traditional perception solution, classic AI, and trying to put the cars on the road with active safety level two kind of functions. At the same time, there was a lot of announcement that was happening on the interaction with the cockpit with natural language processing, basic AI capabilities. But now, more and more Gen AI, large language models coming in to interact with the vehicle, that kind of applications. But the idea of sensing with reasoning, being context aware, and doing actions on that, that's really what is happening. The technology is changing. We are embracing the innovations from the industry and continuously innovating on our side on different compute blocks, different solutions leading to automated driving, automated vehicles eventually going towards a level three, level four system. And not stopping at that, we are expanding physical AI towards the next era, which will be with the robotic solutions actually. I had an opportunity, a privilege to connect with a lot of the robotic companies as well over here. But that's a logical extension for us to move from sensing to now going towards a full physical AI implementation.

If I step back, zooming in to the ADAS side of things, the automated driving side of the things, we used to do all these individual sensor solutions going towards the multimodality BEV applications, putting it together in the car, and doing some action planning based on the rules-based implementation. That was actually not too far ago, being in 2020, 2024 over here and in the global solutions, that's really what was happening. But with the change that has been happening, multimodal perception language models coming in, the vision language models coming in to start doing much more with the same sensors is making a big difference. That's the next level of intelligence that was embraced in the automated driving system. But actually, what is happening now, and I'll talk more about it during the rest of the presentation, end-to-end deployments with vision language action models coming together, connecting with the agents that are there in the car for other reasons like cockpit, and actually feeding it into the system for automated driving. That's really what the technology solution is looking like. And we are building our parts, our solutions, our semiconductor solutions based on that. Not just that, the whole software ecosystem that we are putting together around the SoCs is designed to bring in these capabilities into the car. And the fastest way to see this is actually here in China, where we do these developments and deploy it on the Snapdragon Ride Elite platform, which we'll discuss in more details.

Now if you look at it, what is happening in the ADAS world, we just talked about it. But actually in the cockpit systems, Nakul covered a lot of it. There is still a scalable range of solutions that is required. We have vehicles which are having limited display capabilities and audio capabilities, all the way to multi displays. Some of the esteemed guests who talked about it, they showed what all they are bringing in the capabilities of the cockpit of the vehicles, all the way to eight displays. But actually, displays are just the surface. What is happening really is what you heard in the Banma Intelligence presentation by Mr. Zhang, that there are these agnostic implementations, agnostic frameworks that are coming together, and they have to be co-hosted into one kind of a system. That's what is driving the next level of compute that is there. Same way, if I look at it from the ADAS perspective, we know that there are reasoning solutions that are there, VLAs that are getting deployed. But there is a difference of having a VLA with a small token weight to a VLA which will be really implementing the action. The compute requirements are very different, sensor requirements are very different. We are going towards a level three, level four system. Sensor requirements are different, the resolutions are different. So overall, there is a scalability that is coming in. One central piece in all of this is software investments are the maximum investments in this development. So our goal is to maintain scalability from the left hand side to the right hand side, where we have got entry tier solutions to the premium tier solutions based on a single chassis kind of a platform. And how do we do that? We came here talking about Snapdragon Ride Flex a couple of years ago. Last year, I was here, we talked about Snapdragon Ride Elite processor primarily targeted for ADAS, and similarly for the cockpit, the Cockpit Elite processor. But as of now, we will see actually a lot more deployments on Snapdragon Ride Flex solution based on these Elite platforms, where we'll have all the way to city navigation based on VLA-like implementations, and agile AI, the multi agents running in the car at the same time on a single platform. How do we do that? I'll talk about it briefly. But this is really the future that we see.

And as we go into that, the way we understand the automakers, the ecosystem is looking at it is they are looking at again the scalability across the board. As I said last year, we were talking about Snapdragon 8775 as a Flex solution. It will be coming to production. You heard Mr. Liu talking about BAIC's Arcfox as the MPVs that are some of the first deployments based on the Flex solution, which was perfect to bring in a cockpit system based on our AD 155 and the other ADAS solutions merged together into one AD 775. That was or that is the baseline today beyond our standard ADAS and cockpit system. Now at the same time, we started working on Snapdragon Ride Cockpit Elite and the Snapdragon Ride Elite. Nakul mentioned that we had this deployment on the Leapmotor D90 dual controller, one simple controller. It is already going to be offering in the order of 2000 effective TOPS, equal to many other platforms that exist. That's the extreme of the platforms that are being deployed in 2026 in China, and a similar thing will be repeated in the global ecosystem too. It can host VLA-based automated driving end-to-end implementations as well as the agentic multi-agent flows that are going to be available on the platform. So it replaces big compute of cockpit and a really big ADAS compute into a common platform. But the beauty is here, we are basically working on bringing it down to a single 8797, which is back to the Flex architecture. The same hardware development cut down by half, same software development that can be merged together without really investing a significant effort can be brought into Flex implementation on the same 8797 device. Why? Because we designed these solutions from get-go to host a cockpit and ADAS at the same time. And to fill up the gap on the lower tiers, we have got the new additions on the Elite family where we will be bringing in all these capabilities into our next 87 87 like process.

If you look at it, it's one thing to merge these things together, but it's not the complete story. The idea of Flex is to bring differentiated experiences, giving you a lot more performance. 8295 has been a great platform for cockpit here, but if you go to 8787 or 8797, we are talking about the two platforms merged to one, giving you anywhere from one and a half to two times performance increase in the mid tier, and three times performance increase in the higher tier, going all the way to a dual controller. So one scalable unified compute platform roadmap, reusing all of the software and hardware infrastructure. By the way, this is the same story that we see that when you go global with the China vehicles, this same solution can work, but you have different set of stacks outside. How is it all possible? Again, big thanks to the very ecosystem of stacks, the automated driving stacks that we have been bringing and enabling on our platforms. As recently as earlier this year, we have one of our flagship OEM vehicles, the GSC and N60, driven by the Ride stack on AD 650, has been on the leaderboard for the best scores for the performance tests. This shows that we can bring in a lot of capabilities on our current Snapdragon Ride platforms, and imagine this all can be enhanced multi-fold when we go to the Ride Elite and the Flex platforms. DeepRoute, that is a platform, a solution that is running on our Elite platforms already on the vehicles, deploying VLA stacks already in the car. We announced with Wayve a partnership which is for the global ecosystem. Momenta has been a great partner. So again, really engaging partnership with all these tech suppliers, and then partners from ZYT, Telecom, et cetera, deployed at scale on the Snapdragon Ride platforms. And this is working on a very varied ecosystem of the hardware, tier one partners bringing in sensors and all the capabilities. The most important piece is all these systems are getting deployed in six to nine months from the time the award is announced to the time the cars are on the road. So that's the ecosystem that we have developed, and again, thanks to all the partners over here.

Now this is not a new picture that you might be seeing, but important piece to understand on this is the more we work closer to the deployments of the vehicles, the more we understand and learn what can be improved. I'm not going to talk about the Orion core CPUs and the great NPUs and the thousand TOPS effective, etc. Think about it from a system perspective. What do we take away from all these engagements? The learnings that we do every month, every second month over here, we have been investing on bringing in capabilities of sensor processing. You need to implement safety implementations independently on these Elite processors. We had put small compute blocks for sensor processing. They are completely independent, not counted in the big AI engine which many GPUs etc might have. This is a dedicated block where you can have a lidar pipeline or a camera pipeline to do a separate implementation and deliver a safety use case. All this goes in the planning as we make these compute units into our Elite platforms. And this leads to this mixed criticality approach actually. A good example of how the camera, the HMI for example, will be driven most of it by the cockpit using the GPU clusters, CPU clusters for having a very engaging HMI. We in the Elite process added extra GPUs to do the safety systems as well alongside the cockpit. The AI-powered cockpit, like for all those examples, they are running today on our Snapdragon Cockpit platforms. They are relying on these NPUs again, high tier like the full NPUs or the small NPUs, audio engines. Everything comes together when we put it into the system. ADAS, separate pipeline for the cameras, lidars, radars, doing the point cloud processing, all that is put together on the dedicated crossing blocks which will be shared between the ADAS and the cockpit system, but maintaining freedom from interference. And eventually, everything comes together because when you talk about agents, it has to be tied together into one common platform using the information from data sensors where they are supposed to do a specific job of driving the car, but get that information. You saw that in the great videos from Thunderstorm that was there in the food presentation, how a camera which is designed for doing an emergency braking is using that information to identify a person who was in that black and gray trousers around the corner. This all comes together not just because of the SoCs. We are enabling the agile AI operating system orchestration, all of the deployment bringing on these platforms. We don't need to identify applications. You all are really super innovative in identifying what is the next big thing that you would want to put in the car. Our goal is to be open and enabling the whole ecosystem to bring in this multi-clause, multi-super agents running together, either on the car most of the time, or even have an orchestration going back to the cloud. The cycle of what an agent is going to do for execution, how do you orchestrate multi agents, how do you plan for what is required from an agent, which could be for an interaction or it could be even an input going back to the driving, and eventually giving the instructions to the car, which might be interactive UI or an interactive control system going back to the ADAS. And think about it, this all is designed in a way that it runs on the Flex systems where you can use all these implementations as an input to the safety critical applications as well. This comes by experience, this comes by what we have been doing and the industry has been telling us based on the multi deployments.

Now there is one important thing that keeps coming up, and we have been embracing this, working again with many stack solution providers today in China globally. The idea of how do we bring more compute, we just discussed that. There is a lot of capability requirement for reasoning to come in. Reasoning could be why did the car hit particular motion during the construction zone, why did it nudge out of the lane. But when you bring VLA into the action loop, you're talking about the token rates to be much higher because you need the action to happen in a specific interval of time, not a long range reasoning, but a reasoning to do the action so that you can predict what is going to happen based on the whole world model. The language model needs a lot of compute. Putting compute into one solution and saying this is like a high end single SoC is one way to solve the problem, but actually still might not solve the complete system problem. So we had been investing significantly to deliver this 2000 TOPS equivalent solution with dual 8797 processors. And it's not just trying to tie it together, it is basically to build a system design where you can have a full blown VLA solution which can aim to do a level three kind of a system independently with the compute split across the two SoCs, and have your end-to-end stack which is already deployed in most of the cars today bring it onto one of the processors so that it generates a secondary trajectory so you can bring a chain of thought reasoning. The VLA models can be multi mixture of experts. We have actually more than 30 billion parameter mixture of experts that we are deploying right now. This whole system can generate now a primary trajectory, a secondary trajectory, and the safety guardrails that either have already been built over here in China or we have our own active safety stack as well. This is a solution which scales from a single 8797 for a level 2 plus plus system along with a cockpit, expanding to dual 8797 where you can deploy a standalone level 3 solution. And this is where we are working with a few lead partners to bring the level 2 plus plus and level three kind of systems on the road very soon.

As I come towards the end of my presentation, and really appreciate all the time you guys spent over here, what are the key takeaways? Basically, we have been in the auto industry for quite some time now, and the learnings from giving us an opportunity in connectivity, giving us an opportunity in cockpit, and more recently in the ADAS, all just know how is coming together into real world production. That's what we do day in, day out, doing it in the automotive space and going towards more physical AI implemented. Qualcomm brings one platform that you can deploy here in China, you can deploy as you go outside China. Global readiness with lot of stacks that are available for specific regulatory requirements outside that is all available, whether it is for the Flex solution or for the ADAS specifically. And as we do all this, we are on the path to higher levels of autonomy. There is a lot more that we are doing which we would be willing to talk about in very near future. We have been discussing with some of our lead customers and partners on how we are expanding significantly on our computing capabilities. And by the way, it's not just computing, the knowhow of system is super important because more than computing it's the DDR bandwidth etc that we need to worry about, and that's what we are working on. I will hold it for a later day. But again, thank you for your time. It's a great show, lot of demonstrations and vehicles outside, and please enjoy. Thank you very much. Thank you.

Thank you, Mr. Anshuman, for the wonderful sharing.