Second-order map
Exploratory — reasoned, hypothetical relationships for research, not investment advice.
- NVDACatalyst
NVIDIA Corporation
Dominant GPU/AI accelerator supplier whose cost and power overhead drives customers toward custom edge-inference silicon, opening royalty runway for IP licensors
- CEVASpotlight subject
CEVA, Inc.
Pure-IP licensor of NeuPro-M NPU IP; as custom-silicon designers seek to avoid GPU cost/power overhead, CEVA royalties into each device shipped rather than competing with customers
- CDNS
Cadence Design Systems
EDA tools are required to design the custom AI silicon that integrates licensed NPU IP; wider custom-silicon movement could increase design-tool usage
- SNPS
Synopsys
Provides EDA and complementary IP blocks; if software platform companies design more custom chips, demand for design/verification flows could broaden
- TSM
Taiwan Semiconductor Manufacturing
Custom silicon built on licensed IP must be fabricated; a proliferation of platform-company chips could feed foundry tape-out volume
- ARM
Arm Holdings
Sibling pure-IP licensor whose CPU cores often sit alongside NPU IP in the same SoC; benefits from the same custom-silicon licensing trend
Custom AI silicon design-service / ASIC vendor (e.g. Broadcom, Marvell)
Software platform companies designing custom AI silicon frequently partner with ASIC design houses to implement chips, making them direct operational beneficiaries of the trend
- AVGO
Broadcom Inc.
Leading custom-ASIC partner for hyperscalers; as platform companies build bespoke AI chips, custom-silicon engagements could expand
- MRVL
Marvell Technology
Custom-silicon and connectivity provider; broader move to bespoke inference chips could increase custom-compute design wins
- ALGM
Allegro MicroSystems
More lateral: edge-inference chips proliferating into devices/automotive could raise demand for surrounding sensor/power components (conditional on end-market overlap)
Edge-inference memory / interface supplier
Custom edge NPUs still need memory and interface IP around them; wider custom-silicon adoption could pull through complementary components
- MU
Micron Technology
Edge inference chips require low-power memory; if custom silicon spreads across devices, on-device memory content could grow
- LSCC
Lattice Semiconductor
Low-power programmable logic used in edge systems; a broader edge-inference build-out could expand attach opportunities for glue logic
Low-power IP / interconnect licensor
Hypothesis: as custom SoCs multiply, licensors of interconnect and interface IP could collect incremental royalties per design, mirroring CEVA's model
Take it further
Copy the analysis below into your own AI tool to pressure-test the reasoning and push it further.
Every earnings cycle, the crowd floods into the dominant GPU supplier — and for good reason. NVIDIA commands approximately 80% of the AI accelerator market , and Wall Street is watching closely as it prepares to report tomorrow. But the enduring structural question is not whether the incumbent captures this cycle's spending; it is which non-obvious node collects royalties on every chip that gets built in response to the GPU's cost and power overhead. That node is a pure-IP licensor sitting at the foundation of the edge inference stack, and it is not a name the GPU hype machine mentions.
The quarter included one of the most strategically significant AI licensing agreements in CEVA's history, with a leading global AI and computing platform company selecting its NeuPro-M NPU IP for next-generation custom AI silicon.
The agreement extends CEVA's customer base beyond traditional semiconductor companies and device OEMs to include software platform companies that are increasingly designing custom silicon to optimize performance, power and area and the overall user experience. As a pure IP licensor, (CEVA) does not compete with its customers — it royalties its way into every device they ship, which means the broader the custom-silicon movement spreads, the wider CEVA's royalty runway grows.