In this week’s podcast, I’m thrilled to welcome Veerbhan Kheterpal, CEO of Quadric! Veer and I dive into the fascinating world of edge AI and explore how Quadric’s fully programmable Chimera GPNPU architecture is shaking up traditional AI processing. We chat about the benefits of combining neural network acceleration with conventional computing, the growing demand for on-device AI token generation, and the challenges of running increasingly sophisticated AI models on power-constrained devices. We also investigate why moving AI out of the cloud and directly onto our devices could be a game changer for developers, businesses, and the future of AI innovation.
Links for November 20, 2026
More information about Quadric
More information about Chimera™: The General‑Purpose Neural Processing Unit
Amelia’s Weekly Fish Fry – Episode 708
Speaker: Veerbhan Kheterpal, CEO of Quadric
Air Date: November 13, 2026
Amelia Dalton — Hello there, everyone. Welcome to episode number 708 of this here electronic engineering podcast called Amelia’s Weekly Fish Fry, brought to you by EEJournal.com and written, produced, and hosted by yours truly, Amelia Dalton.
What if the future of AI isn’t in the cloud, but right in the palm of your hand?
As AI models grow more sophisticated, the race is on to bring more powerful, efficient, and accessible AI processing directly to our devices. But how do we make that happen without sacrificing performance, draining batteries, or adding a whole lot more complexity?
In this week’s podcast, my guest is Veer Kheterpal, CEO of Quadric!
Veer and I chat about Quadric’s fully programmable Chimera GPNPU architecture, how it’s shaking up traditional AI processing, and why combining neural network acceleration with conventional computing could be a game changer for edge AI.
We also explore the growing importance of on-device AI token generation, the challenges of running increasingly complex AI models on power-constrained devices, and how bringing AI out of the cloud could open up a whole new world of possibilities for developers and businesses alike.
So, without further ado, please welcome Veer to Fish Fry!
Amelia Dalton — Hi, Veer. Thank you so much for joining me.
Veer Kheterpal — Thanks for having me.
Amelia Dalton — Absolutely. Okay, so let’s start with Chimera. Quadric has developed a fully programmable GPNPU for AI inference. So first up, what is a GPNPU, and how is Chimera different from other AI processors?
Veer Kheterpal — So, you know, without going too much into the jargon, right? Well, GPNPU stands for general-purpose neural processing unit.
But the simplest way to understand it is that AI functionality is starting to make its way into every device that we own and use every day.
So we use chatbots. These are, you know, as everyone knows, cloud subscription services running on these supercomputers in the cloud.
But when it comes to our technology, we work with companies who build chips for various devices—cars, laptops, printers—and we help these companies build AI inference right there on the device.
Which means when you use AI, you’re using the processing capability on the device. And that processing is what’s being carried out by this GPNPU.
That’s what our core technology does. It delivers AI right there where you are, in the real world, on the device. It does not need to be connected to the cloud or use cloud subscriptions for that.
The other question you asked was how are we different?
And the one thing we have focused on as we built this technology is to make sure chips that are built with our neural processing unit core are future-proof.
You know, one of the biggest problems in the AI chip market is that AI chips are very hard to build in a way that makes them immune to ever-changing AI needs.
And that is one problem we solved by taking a software-first approach, where we are giving developers the power to use software to program and create their AI inference applications of many different kinds.
That’s why the word “general-purpose” is how we named it, as in the GPNPU, which is general-purpose neural processing unit.
Amelia Dalton — So, Veer, running AI on-device can get complicated when you need multiple processors to work together. So how does Chimera simplify this process, and why does that matter for performance and latency?
Veer Kheterpal — Great question.
So the typical way this has been done before is to build in specialized processing for special functions on the device, right?
So you take the AI function, break it up into smaller pieces, and then build specialized processing for each of those smaller pieces.
Makes sense, right?
But what happens with that approach is two problems.
One, when AI changes and the breakdown of those pieces changes—these are called operators—it gets very hard because you sort of designed with a certain AI model class in mind. And so it gets hard to run the new type of AI model on there.
But the other problem is, as the industry knows, the energy spend, the power spend in AI model inference, is in the data movement. It’s actually not in the processing itself.
So now when you create many different processing units, you tend to move a lot of data around and burn more power and cause inefficiencies in scheduling the AI function on those units.
What we did was we took a very different approach.
We unified several different units and created a new unified instruction set that is AI-native, which means it’s not using legacy technologies like CPUs, GPUs, DSP cores, and pulling those together.
It’s built as its own ground-up new compiler, new software, new instruction set.
And the core goal there, which is surprising to a lot of people, is not to get the lowest hardware power, but low overall cost of ownership across the software stack, hardware stack, power, performance, and costs.
Amelia Dalton — So, Veer, AI models are getting bigger, but edge devices have limited power and memory. So how does Chimera help run more complex AI models on devices, and what kind of applications would benefit the most?
Veer Kheterpal — I would say here, the trend is our friend.
If you think about what your iPhone did 10 years ago and the iPhone of today, the same power in your pocket does a lot more, right? It’s a much faster, much better camera.
So there are broader technology trends which are enabling more processing, more memory, more capacity to fit in the same form factor over time.
So edge devices actually tend to benefit from that trend. It’s secular and structural, right?
Your laptop today, the model that can run on it, and the laptop from 10 years ago—day and night difference, just by the nature of this technology scaling that’s already occurring and underway, and has been doing that for about 40 years, I think.
And so we’re benefiting from that trend of the better and bigger models coming to the edge just organically.
But specifically, what we do with Chimera, because we are general-purpose and support many different models, we are able to have a multimodal nature to a single processor.
One of our customers is using Chimera to do audio, video, image, and fraud detection all within the same core because it’s programmable.
And that can be on a personal device or an enterprise device in terms of bringing functionality to that device in an efficient manner.
The second part of the question there was, what applications?
The simplest one to understand here is, you know, if you’re not paying attention and driving, and you happen to glance at your phone or away from the road, and someone steps in front of your car, imagine yourself in that situation.
And what we’re doing is helping Tier 1s and OEMs here to build capability in cars that will have the car automatically brake if a pedestrian steps in front of it.
In fact, world governments, as we speak, are mobilizing to make sure that is a mandate in every car in a few years from now. We’ll have to have those features, right?
And that’s right there an AI processor on the car, inference happening there across multiple cameras and detecting that event in real time if you miss it.
Amelia Dalton — That makes sense.
Now, Veer, we’re seeing more about generating AI tokens directly on devices instead of in the cloud. So what makes on-device AI attractive, and what benefits does it offer in terms of speed, privacy, and power consumption?
Veer Kheterpal — I think you’re bringing up this trend of Mac minis and GPUs flying off the shelves.
I’ve seen that. It’s pretty obvious.
I would say in the last six months or so, there’s been this movement of independence or freedom from counting tokens.
It’s very foundational. For a lot of participants, a lot of AI users, tokens have become their next unit of work, literally, right?
To read an email, to write an email, to process a document, you need tokens. And so running out of them is almost like you’re dead in the water, right?
Imagine you have a laptop, you’re working on it, and it’s metered and it shuts off, right?
It’s right there, it exists, but it shuts off until you put more coins in it to continue working.
It’s literally, you know, to a lot of the community, it feels like that.
And we’re seeing this early movement from those communities to start to build and figure out what the local AI stacks look like, right?
What hardware can they purchase today off the shelf? What does it cost? How do you put it together? What models do you run that start to provide equivalent capability as the frontier models, which are incredible, right?
And so I would say the core value, the benefits people are driving, is continuous use.
Users constantly run out of tokens, so they need unmetered, just continuous-running workloads, right? Those types of applications.
Power, I would say, is not that big of a concern there. Whatever fits in your home, in your office, you’ll be able to plug it in.
But cost of tokens is one big driver.
The other one I have seen now in enterprises is another big movement beyond the individuals, the prosumers, and the hackers.
It’s enterprises starting to figure out budgets between a continuous OpEx token cost of agents and token use in their enterprise versus a fixed cost of investment in hardware that they can install and be able to have the right capacity to then use for five years, seven years.
No different than buying computers, laptops, and computing equipment.
So we’re seeing it in enterprise as well.
Several CIOs and CFOs are starting to spec out what that spend looks like to bring this token capability local.
Many of them are driven by privacy. Law firms, hospitals—many of them are driven by their own IP, their own business processes getting privately processed in their own environment.
Amelia Dalton — So, Veer, do you believe that on-device AI could make AI more accessible to everyone? And if so, how could this shift change the way we use AI and open up new opportunities for developers and businesses?
Veer Kheterpal — When you talk about access, right, it’s very close to my heart.
I was personally lucky to get access to a PC in the middle of nowhere when I was 12 years old.
I grew up in India, and whoever thought it was a good idea to get one PC in our school, right?
And so just the matter of that access to one PC opened up huge avenues of learning and curiosity in that region.
And if you fast-forward today, there are many parts of the world, many places where just pure access to frontier-like models, good AI models, good agents, through a local AI deployment, is exactly, to me, what that looks like.
That kid who has never played with an agent before gets access to it in their school or wherever, on-prem, where this gets deployed.
And that just creates that first step that enables all those communities to be able to onboard to this capability, right?
We already have established this incredible capability roadmap that we are on today, right? Frontier models, small models, and everything in between.
So that’s what I feel about it.
We’re working—we’ve announced this with the World Bank—on initiatives to enable this access throughout the planet.
And when that happens, I think we’re going to see a whole new wave, like we’ve seen with mobile devices, right?
In many parts of the world where we never wired internet and fiber to homes, or hardwired connectivity to homes, these places, these developing markets, got onto digitization and e-commerce just with accessible, affordable mobile phones and wireless, right?
These edge-based deployments are not very different from that.
Amelia Dalton — Absolutely.
All right, Veer, let’s throw a little off-the-cuff question at you now.
So, if you could have one meal right now, it doesn’t matter if it’s on the other side of the world, you need a passport to get there, what would you have?
Veer Kheterpal — I would spontaneously fly to Japan and get sushi at my favorite place there.
Amelia Dalton — Very good answer. Absolutely.
Well, Veer, I think that’s all I have time for today. Thank you so much for joining me.
Veer Kheterpal — Yeah, it’s been a pleasure. Thanks for having me.
Amelia Dalton — Well, folks, that’s all I’ve got for this week’s Fish Fry.
If you’d like more information about Quadric, I’ve included a couple links below the player on this week’s Fish Fry page on EEJournal.com and in the description for this week’s YouTube episode as well.
By the way, have you connected with EE Journal on social media yet?
Well, you can find us on LinkedIn, Bluesky Social, and Mastodon.
And don’t forget about our YouTube channel, YouTube.com/eejournal, folks.
It’s packed with all kinds of techie videos, including our very popular Chalk Talk webcast series hosted by me.
And be sure to hit that subscribe button while you’re over on YouTube, too.
Thank you so much for joining me.
If you know of any cool new technology, or heck, you just want to chat, shoot me a line at Amelia—that’s [email protected]—or post a comment on our forums on EE Journal.
For the week of November 13th, 2026, I’m Amelia Dalton, and you’ve been fried!
Source: EE Journal




