The AI Chip That Refuses to Go Obsolete

Here is a slightly odd fact about Quadric: it is a semiconductor company that does not sell semiconductors. You cannot buy a Quadric chip, hold a Quadric chip, or find a Quadric chip if you pry open your phone. What Quadric sells is the design - the arrangement of logic that another company licenses, tapes out, and manufactures inside its own system-on-chip. In the trade this is called IP, intellectual property, which is a grand name for "the part of the chip you paid someone else to figure out."

This is a good business when your idea is the hard part, and Quadric's idea is a specifically hard one. Most devices that do AI - detecting a face, transcribing a voice, keeping a car in its lane - contain a neural processing unit, or NPU, a chunk of silicon purpose-built to run neural networks very fast and very efficiently. The trouble with purpose-built things is that the purpose keeps changing. Build a great NPU for the models of 2022 and you may find that the models of 2024 use an operation your hardware simply cannot perform. Now your fast, efficient, expensive accelerator is a fast, efficient paperweight, and your product needs a new chip, which takes years and millions of dollars.

Quadric's answer, sold under the name Chimera, is to make the AI core fully programmable in the first place - to fuse the raw math throughput of a neural accelerator with the general-purpose flexibility of a digital signal processor, the workhorse chip that has handled audio and radio and sensor data for decades. One core, two jobs, and crucially: when a new kind of model appears, you write software for it rather than melting down the hardware. The company likes to point out that when large language models and vision transformers arrived - two model types that stumped a great many production NPUs - it added support with a software port and no hardware changes at all.

Quadric was founded in 2016 by three engineers - Veerbhan Kheterpal, Nigel Drego, and Daniel Firu - who had previously built hardware for 21 Inc, a bitcoin computing company. That lineage matters more than it sounds. Designing custom silicon to mine cryptocurrency is an education in a very specific discipline: squeezing the maximum useful computation out of the minimum power and silicon area, because in that world the margins are the whole game. The trio, with pedigrees from MIT and Carnegie Mellon, took that instinct and pointed it at a different target - the messy, fast-moving problem of running AI on devices at the edge, away from the data center.

The founders still run the company: Kheterpal as CEO, Drego as CTO, Firu as chief product officer. It is a small, deeply technical shop - roughly 66 people - which is roughly what you would expect for a firm whose product is essentially a very good idea rendered in Verilog and C++. You do not need a factory to sell a blueprint. You need engineers, a toolchain, and customers willing to bet their next chip on you.

The name is a small joke that happens to be accurate. In Greek myth the chimera is a single creature stitched from two others; Quadric's Chimera is a single processor stitched from two normally-separate ones. On one side, the wide, parallel matrix math that neural networks devour. On the other, the branchy, control-heavy code that everything around the neural network needs - resizing an image, filtering a signal, deciding what to do with a result. Traditional designs use two blocks and shuttle data between them, which costs power and latency. Chimera runs both on the same core.

The other notable trick is range. A single architecture stretches from 1 TOPS - modest, sip-of-power inference for a sensor or a wearable - all the way to 864 TOPS, the kind of throughput an autonomous vehicle or an edge server needs. That means a chip designer can pick a point on the curve without switching programming models, and Quadric offers automotive safety-enhanced (ASIL-ready) versions for customers whose silicon has to satisfy a car's functional-safety requirements rather than a phone's.

Quadric's customers are chip designers and system builders, so the "you" here is really a semiconductor team deciding what goes into next year's product. If you are building silicon for a car, Chimera gives you one core that handles both the neural perception - reading cameras and radar - and the conventional signal processing around it, in an ASIL-ready package. Tier IV of Japan, a self-driving software company, selected it for exactly this. If you are building an edge server to run language models close to where the data lives, the same architecture scales up to hundreds of TOPS and runs LLMs up to 30 billion parameters without a cloud round trip; an Asia-based edge-server LLM silicon provider is among the design wins Quadric has cited.

And if you are building for the more familiar edge - a camera, an office machine, a consumer gadget - the pitch is durability. The reason a licensable, programmable core is attractive is that your product will outlive the AI model you shipped it with, and Chimera lets you update the intelligence in software. That is the whole argument, really: the model always changes faster than the silicon, so build the silicon to change with it.

The investor list tells its own story. Alongside conventional venture firms - Pear VC, Uncork Capital, Cota Capital - sit DENSO, a giant automotive supplier, and MegaChips and NSI-TEXE, semiconductor companies. When your customers and your investors are the same people, it usually means the product is being validated by the people who would actually put it in their chips. The January 2026 Series C added the ACCELERATE Fund (managed by BEENEXT) as lead, with Volta, Gentree, Wanxiang America, Pivotal and Silicon Catalyst Ventures joining.

Quadric is not alone in licensing AI silicon. The neural-IP business includes the industry's heavyweights - Arm with its Ethos line, Cadence with Tensilica, Synopsys with ARC NPX - plus specialists like Ceva and Expedera, and edge-accelerator startups such as Hailo and SiMa.ai. Most of these draw a line between the NPU and the DSP. Quadric's whole differentiation is erasing that line. Whether "one programmable core" beats "two specialized ones" is the bet the whole company rests on - and tripling product revenue in 2025 suggests at least some chip designers are buying the argument.