The Last Mile Isn't Software, It's Silicon: Quadric Demo -- Ride AI 2026

0:04 We're all here because we're betting on autonomy, right? And if you're not, then maybe you're in the wrong place. But it's probably for good reason that you're doing so. The models keep getting better every quarter. And the hardware mostly keeps up if you're willing to pay both in power and in dollars. But most people can't and until that changes where the last mile isn't silk or software, it's silicon. This is what we've been going through for the past decade. Highway assist, urban prediction, end-to-end perception, and now the model is the autonomy stack.

0:42 So the models got bigger, but the pipelines got simpler. And all of these required more from the silicon underneath it. So the industry split into effectively two camps. Fixed function, which is efficient, cheap, and stranded when models change. Or fully programmable GPUs. But they have data center power budgets in an embedded device. AI model generations come every one to two years, but automotive silicon is about five to seven year window. And so the SOC your partners choose today is sold in a car in 2028 that's still receiving model updates in 2038.

1:25 And the models it needs to run then don't exist. The first generation of ADAS silicon was purpose-built for one class of workload. Object detection at highway speeds, automatic braking, etc. So that fixed function hardware was optimized for exactly one era and it was efficient and it was cheap and it worked beautifully for that era. But then the models changed and the silicon couldn't follow. New math, new architectures, new memory access patterns, none of which the hardware was built for.

1:59 Performance doesn't slowly decline, it actually just falls off a cliff. And so, the chip you qualified last year can't run this year's model. And so, what's What do we naturally do? Reach for GPUs because they can run anything, but not at the power and cost envelope that's required for automotive at scale. For L4 robo-taxi fleets, where you own the vehicle, maybe you absorb that. But for 80 million 8 as equipped cars a year, that math doesn't work.

2:30 And so, we're stuck. We need efficient silicon, but it can't run new models, and flexible silicon that blows your power and cost budgets. And so, that top right quadrant of programmable and efficient is empty. And that's the thing we need to solve. And here's how you know it's real. Tesla designed its own FSD chip, Waymo built custom silicon. So, the largest players, the ones with the deepest pockets, looked around and said, "None of this is good enough. We'll build our own."

3:01 And so, that's to tell you something about what's available today. But building custom silicon is a half a billion-dollar multi-year bet. Requires hundreds of engineers, has enormous technical risk, and it's min- minimum of 3 to 5 years. And you're betting typically on a model architecture that might be obsolete by the time you ship. Tesla might be able to absorb that, Waymo can, but most can't. And there are about 30 major auto OEMs.

3:32 Two or three will and can build some custom silicon, but the rest depend on SOC vendors like Qualcomm, Renesas, MediaTek, who build chips from licensed IP blocks. And those are the vendors who need AI compute that actually works. This is how those SOCs get built. CPUs are licensed, GPUs licensed, ISP, safety islands, all licensed. And then for that AI engine, the one that determines what you can run from now until 2038, you're sort of on your own. They're all broken.

4:08 Arm figured this out about 30 years ago. They didn't build phones, but they gave phone makers access to really, really world-class processor cores. And I think that AI compute is at that same inflection point today. We need one core that can replace three separate cores and keeps up as the models evolve. And so, if you remember that empty quadrant, programmable and efficient, we filled that. We built Chimera, a licensable silicon IP core, so that SOC vendors don't have to.

4:42 This is something you drop into an SOC, and your customers can run any model now and when they change. But don't take my word for it. We actually ran Doom on Chimera, the actual game running at 1785 frames per second. So, think about what that means. Most AI accelerators can only run neural networks, but Chimera runs anything. And that means vision transformers, BEV former, robotic action models, lidar voxellation, audio beam forming, path planning, just to name a few.

5:17 These are workloads that typically require three separate processors, and instead one core can handle all of them. And when the model architecture arrives, you just recompile. Kyocera's taping out an SOC with Chimera in it to this year. Uh another unannounced licensee is in tape out right now, and Denso, one of the world's largest tier one automotive suppliers, has licensed the IP. So, consumer electronics, AI compute, automotive, all three with the same core.

5:48 Now, software defines what this hardware can do. So, when CNNs gave way to transformers, our customers just recompiled without any hardware change. And when the next architecture arrives, it'll be the same story. Because the software update is a fraction of the time and cost of a silicon respin. SOC vendors are committing to their IP choices for 2029 and 2030 chips today. The autonomy stack that you'll deploy at scale will run on silicon being designed this year.

6:21 And so, the decisions your SOC partners make in the next 12 months really influence and decide what you can run at launch. The fleets are scaling as we saw. The models keep getting better as we keep and continue to see. And the conference theme this year is it's time to market. But, make sure the silicon your SOC vendors choose can run 2038's models and not just 2026's. The last mile isn't software, it's silicon. Thank you.

6:53 Thank you. Thank you, Matthew.