CEVA Sells the Part of the Chip You Never See - and Gets Paid Again Every Time It Ships

The most revealing fact about CEVA is what it does not own. No fabrication plant. No warehouse of chips. No glossy gadget waiting for a consumer launch. The company designs reusable pieces of silicon and software - digital signal processors, neural processing units, wireless connectivity blocks and sensor-fusion code - then licenses those blueprints to the companies that make the actual chips. If those chips find their way into earbuds, cars, routers or industrial machines, CEVA can collect a royalty on every unit.

That makes CEVA an unusually pure bet on the invisible layer of electronics. Its customers do not buy a finished processor off a shelf. They license an architecture, software stack and toolchain, integrate the package into a system-on-chip, send that design to a foundry and hope the resulting product wins a market. CEVA gets paid at the front of that journey through licensing and support fees. The more interesting money arrives later, after the customer ships.

Three verbs, one complicated machine

CEVA explains its portfolio with three verbs: connect, sense and infer. The simplicity is useful because the technology underneath is anything but. “Connect” covers Bluetooth, Wi-Fi, ultra-wideband, cellular IoT and 5G. “Sense” includes DSPs and software that turn camera, radar, microphone and motion data into something usable. “Infer” is the AI layer, where NeuPro NPUs and SensPro DSPs run machine-learning models close to the source of the data.

The customer problem is time. A chipmaker building a connected sensor cannot casually spend years inventing a standards-compliant Bluetooth radio, a vision processor and an AI accelerator before starting on the feature that makes its product different. Licensing tested IP shortens that path and reduces integration risk. CEVA's pitch becomes stronger when a buyer takes multiple pieces together: a Wi-Fi and Bluetooth combination, or an NPU beside a sensing DSP. One vendor, fewer seams, more content per device.

That customer list runs from semiconductor suppliers to the brands specifying the finished experience. Public relationships include Microchip, Renesas, NXP, Actions Technology, Lenovo, boAt and Nothing. Seven of the world's ten largest microcontroller vendors have licensed CEVA technology, according to the company's 2025 filing. The applications are deliberately untidy: a digital car key, a noise-cancelling headset, a robot vacuum, a cellular base station and an industrial camera may share little on a store shelf, yet all need some combination of radio, signal processing and local inference. CEVA sells beneath the category label.

The $18 million hinge

CEVA was created in 2002 by combining DSP Group's licensing division with Parthus Technologies, an Irish developer of platform-level intellectual property. The original strength was digital signal processing, especially technology used in communications. That was valuable, but concentration in handset baseband exposed the company to a market where customers could consolidate, switch architectures or build more internally.

The pivotal diversification move came in 2014. CEVA acquired RivieraWaves, a French supplier of Wi-Fi and Bluetooth IP, for aggregate consideration of roughly $18.4 million. It also committed about $3.4 million to retain the acquired team. The logic was plain: add non-cellular wireless technologies that could ride through a much wider universe of connected products. This was not a glamorous purchase. It was plumbing, which is exactly why it mattered.

A good IP acquisition does not merely add a product. It adds a royalty stream that can survive several generations of somebody else's product.The CEVA lesson

In 2025, Bluetooth accounted for 1.1 billion Ceva-powered shipments. Wi-Fi set a company record at 266 million units, while cellular IoT reached 241 million. RivieraWaves did not eliminate cyclicality, but it changed the shape of the business. Connectivity became a base layer that CEVA could pair with newer AI and sensing products.

What failed first: the clean story

CEVA briefly complicated its own model. In 2021 it acquired Intrinsix, a custom chip-design business. Design services could help customers move from licensed IP to finished silicon, and the deal promised a broader relationship. But services scale differently from reusable IP: more revenue can require more people and more project work. After a strategic review under CEO Amir Panush, CEVA sold Intrinsix's aerospace and defense business to Cadence for $35 million in 2023.

The divestment was the change of mind worth watching. CEVA chose focus over a wider menu, returning to silicon and software IP that could be sold repeatedly. It refreshed its brand around “connect, sense and infer,” bought spatial-audio assets, and concentrated research on wireless platforms, edge AI and sensing. The result is not a miraculous turnaround story - CEVA still reported a GAAP loss in 2025 - but the operating thesis became legible again.

Where the AI bet becomes real

CEVA's AI portfolio stretches from NeuPro-Nano, a self-contained NPU for tiny, battery-constrained devices, to NeuPro-M, which can scale toward more demanding classic and generative-AI workloads. NeuPro Studio supplies the less photogenic but essential part: tools for importing, optimizing, compiling and debugging models. A hardware block without a usable software path is a very expensive rectangle.

In 2025 the company signed ten NPU agreements, and AI supplied more than 20 percent of licensing revenue. Microchip licensed the broad NeuPro family for future products. Other disclosed adopters included ALi and Nextchip. In 2026, CEVA announced a NeuPro-M agreement with a large U.S. software and AI platform company working on custom silicon. Meanwhile, its AI DSP and accelerator technology entered production in Renesas's R-Car V4H platform used in the 2026 Toyota RAV4. These are different ends of the same wager: more intelligence will run where the data is created.

The market position sits between broad processor-IP companies and specialists. Arm, Cadence, Synopsys, VeriSilicon and Imagination offer overlapping alternatives. In AI, a crowd of NPU designers joins them. Yet the toughest competitor is often the customer's own engineering group. A large semiconductor company may license CEVA in one product and use an internal core in another. CEVA wins when proven standards support, low power, a small silicon footprint and faster time to market outweigh the control of building in-house.

The conditions under the flywheel

The licensing model looks effortless only from a distance. After a contract is signed, a customer must finish integration, tape out the chip, clear manufacturing, win a product design and reach volume. Any broken link can delay or erase royalties. Semiconductor cycles are long, standards change, geopolitical export controls can restrict markets, and a successful customer may eventually decide to internalize the technology.

CEVA's 2025 numbers show both sides. Revenue reached $109.6 million, up 2 percent. Licensing and related revenue grew 6 percent to $63.6 million, while royalty revenue totaled $46 million. A record 2.1 billion powered devices shipped, yet the company posted a GAAP operating loss of $11.3 million. Volume alone does not guarantee profit; mix, royalty rates and continuing research spending matter.

The strategy works best where standards are hard, power budgets are tight, integration risk is costly and many customers need the same foundational capability. It works less well when a single buyer has the scale to build everything internally, when an open architecture becomes good enough, or when end products fail to ship. CEVA is not collecting a toll on all edge AI. It is building tollbooths on selected roads and betting that enough traffic will arrive.

By the second quarter of 2026, the bet had fresh evidence: revenue rose 13 percent year over year to $29 million, licensing revenue climbed 21 percent, and 567 million powered devices shipped in three months. The company had also moved beyond individual blocks toward fuller systems, including a Bluetooth High Data Throughput design combining baseband, software and CEVA-developed RF. Bigger packages can raise licensing value and customer stickiness, but they also demand broader execution.

CEVA is selling time: the years a customer does not have to spend rebuilding Bluetooth, a sensor pipeline or an NPU from first principles.

That is the cleanest way to understand the company. CEVA does not need its name on the box. It needs its design inside the chip, its software inside the product and the customer's product in enough hands to make the royalty line move. Invisible businesses can be excellent businesses. They simply require patience - first from the engineers, then from the shareholders.