The AI Chip Company That Sells No Chips

There is a peculiar way to build a chip company: do not build chips. Build the part that somebody else will place inside a chip, hand over the blueprints and tools, and allow your name to disappear from the finished product. Expedera has pursued precisely this sort of cultivated invisibility since 2018. Its customers are semiconductor designers and device makers. Its product is neural-processing-unit intellectual property - circuitry and software that can be configured, synthesized and integrated into a customer’s system-on-chip.

This is why you may have used Expedera without knowing it. The company said in 2025 that its technology had shipped in tens of millions of devices. It publicly describes production contracts with multiple smartphone makers and licensees across automotive, augmented and virtual reality, and consumer chipsets. Most names are confidential. In semiconductor IP, anonymity is often the receipt.

The networking people saw packets everywhere

The founding observation was almost playful. Da Chuang, Siyad Ma and Sharad Chole had worked together on networking systems at Cisco. When they looked at deep-learning inference, they saw a familiar scheduling problem. Data had to move through specialized machinery, dependencies had to resolve, and idle hardware was wasted hardware. As an early company post put it, deep learning looked “a lot like networking.”

A conventional accelerator often works through a model layer by layer. The awkward bit is not always the multiplication. It is moving activations and weights between compute and memory, then waiting for the next layer to be ready. Expedera instead divides a neural network into self-contained fragments called packets. Those packets can move through dedicated processing blocks and, where dependencies permit, different layers can execute in parallel. The aim is less memory traffic, higher utilization and latency that a designer can predict before the chip returns from the fab.

The idea produced a striking early number: up to 18 TOPS per watt on a 7-nanometer test chip. TOPS is a blunt measure, and process nodes complicate comparisons, but the supporting details are more revealing. Expedera described roughly 128 bytes of control-sequence memory per layer and a “fire-and-forget” mode that does not require the host processor to supervise every step. Its current materials claim up to 90 percent processor utilization. This is an architecture obsessed with not waiting around.

“Deep learning is a lot like networking and we should be able to do this better.”Expedera’s account of its founding hunch

The first thing the cloud could not fix

The first disclosed production job was beautifully unglamorous: improving dark video. In 2022, a global consumer-device maker shipped Origin IP for real-time low-light denoising of 4K footage. At that resolution, sending the stream to a distant data center and waiting for it to come back was not practical. The workload had to remain on the device, under tight power and thermal limits. Expedera said its implementation beat the alternatives evaluated by the customer.

That deployment explains the actual problem Expedera solves. Edge AI is not cloud AI in a smaller box. A camera, phone or car has a finite battery, a fixed thermal envelope, a bill of materials and no patience for a network outage. A chip team could design its own NPU, but that means hiring rare architects, building compilers, validating silicon and maintaining model support. Or it can license a block and adjust the performance, area and memory profile to the product.

A catalog measured in ambition

The original Origin catalog stretches from E1, aimed at roughly one-TOPS always-on jobs, through E2 and E6 to E8 for demanding automotive and data-center use. The range is not simply small, medium and large. Expedera tunes cores around the customer’s neural networks and power-performance-area target. Unused circuitry - “dark silicon” in the company’s vocabulary - is treated as a design failure rather than insurance.

Origin Evolution, launched in May 2025, adds a broader set of processing blocks and native support for large language models as well as convolutional and recurrent networks. Expedera lists Llama 3, Qwen, DeepSeek, YOLO and MobileNet among supported families. It also supports common model representations and tools including Hugging Face, llama.cpp and TVM. The company says trained models can run without mandatory retraining or accuracy reduction - a practical promise, because a wonderful accelerator that forces the customer to rebuild its model is only half a product.

The car makes the promises expensive

Consumer video proved the engine could ship. Automotive asks whether a company can be trusted for years. In 2024, indie Semiconductor made a strategic investment and agreed to integrate customized Origin processing into future ADAS products. The companies described a path past 100 TOPS for object detection, lane assistance, emergency braking and driver monitoring. Expedera then obtained ISO 26262 ASIL-B readiness certification, followed by ISO 9001 quality certification in 2025.

This progression also answers what changed the company’s mind about scope. It did not abandon edge inference. It expanded the definition. Phones led to cameras, cars and industrial systems; then language models made memory movement even more punishing. Origin Evolution now reaches from edge devices to data-center inference. The common thread is not the market label. It is a workload whose arithmetic is constrained by memory, power or deterministic latency.

The part worth copying

A product builder cannot copy Expedera’s RTL from a magazine profile, happily. But its method travels. Start with the cost everybody accepts as inevitable. In this case it was external-memory movement. Reframe the workload using an adjacent discipline the team understands unusually well. Test real models, not merely peak arithmetic. Preserve the customer’s trained asset. Then sell a right-sized system rather than the largest number that fits on a slide.

There are clear limits. Licensable NPU IP does not help a team that needs a packaged processor immediately or lacks the people to integrate and verify a complex block inside an SoC. Cloud-first workloads with abundant power and memory may value programmability or time-to-market over finely tuned efficiency. Novel operators still need mapping and tool support. And the company’s performance claims must be checked against each buyer’s models, memory system and process node. An accelerator is never efficient in the abstract; it is efficient at a particular job.

That specificity may be Expedera’s soundest idea. The AI industry enjoys universal machines and universal claims. Expedera began with a narrower proposition: the data has somewhere to be, the computation has a deadline, and the battery is not getting any larger. Treat the neural network like traffic. Then clear the road.