How Semiconductor IP Licensing Works
When a company designs a system-on-chip, they don't build every block from scratch. The CPU core, the GPU, the USB controller, the PCIe interface, the memory PHY — many of these come as pre-designed, pre-verified intellectual property blocks licensed from third-party vendors. This IP-based design methodology is what makes it possible for a relatively small team to design a chip containing billions of transistors.
The licensing model typically involves two payments: an upfront license fee for the right to use the IP in a specific design, and per-unit royalties on each chip shipped. The upfront fee can range from $100,000 for a simple interface block to $10-50 million for a high-end CPU core. Royalties are typically 1-3% of the chip's selling price, though the actual percentage varies widely based on negotiation, volume, and the strategic importance of the customer.
ARM: The Architecture That Powers Everything
ARM Holdings is the most important semiconductor IP company in the world. Their processor architecture is in over 99% of smartphones, most tablets, a growing share of laptops (Apple Silicon, Qualcomm Snapdragon X), and an increasing number of data center servers (AWS Graviton, Ampere Altra). ARM doesn't manufacture a single chip — they license the instruction set architecture and processor core designs to other companies.
ARM's licensing operates on two main tiers:
- Architecture license — the customer gets the ARM instruction set specification and can design their own custom cores that are compatible with it. Apple, Qualcomm, Samsung, and NVIDIA hold architecture licenses. This is the most expensive tier but offers the most design freedom. Apple's M-series and A-series chips use custom-designed cores that implement the ARM instruction set but share no microarchitecture with ARM's own core designs.
- Core license — the customer gets a specific ARM-designed core (like the Cortex-A720 or Cortex-X4) as RTL or a hard macro. They can integrate it into their SoC but can't modify the core design itself. Most ARM licensees use this model. MediaTek's Dimensity chips, for example, use ARM's off-the-shelf Cortex cores.
ARM's revenue in fiscal year 2024 was about $3.2 billion, which sounds modest until you realize that the chips containing ARM IP generate over $200 billion in annual revenue for their licensees. ARM captures a thin slice of an enormous pie.
ARM's Pricing Pressure
ARM has been steadily increasing royalty rates, which has created friction with some customers. Qualcomm and ARM spent years in legal disputes over licensing terms for the Nuvia acquisition (Qualcomm bought CPU design house Nuvia in 2021, and ARM argued the acquisition required renegotiating the license). The case was settled in late 2024, but it highlighted the tension between ARM's desire for higher revenue per chip and customers' desire to keep costs down.
This connects to the ideas in 3D NAND Flash: Layer Stacking, String Architecture, and the .
ARM's September 2023 IPO valued the company at about $55 billion, with investors expecting significant revenue growth. That growth has to come from somewhere — higher royalty rates, expansion into new markets (automotive, IoT), or both. Customers are watching nervously.
Synopsys and Cadence: The Other IP Giants
While ARM dominates processor IP, Synopsys and Cadence are the leading suppliers of interface and foundation IP. If your chip needs a USB 4 controller, a PCIe Gen 6 PHY, a DDR5 memory interface, or an HDMI 2.1 transmitter, you're most likely licensing that block from Synopsys or Cadence.
Synopsys' DesignWare IP portfolio includes over 2,000 IP titles covering processor cores (ARC), interface protocols, security, analog, and foundation library cells. Their interface IP is particularly critical — designing a high-speed SerDes PHY that meets PCIe Gen 6 specifications at 64 GT/s is a multi-year, multi-million-dollar project. Most chip companies would rather pay Synopsys a license fee than attempt it in-house.
Cadence's Tensilica processor IP and their interface PHY portfolio serve a similar market. In some categories — like MIPI camera interfaces and certain memory PHY configurations — Cadence has stronger market share than Synopsys.
Together, Synopsys and Cadence generated over $4 billion in IP licensing revenue in 2024. That's separate from their much larger EDA tool businesses.
We covered a related topic in ARM Architecture Evolution: From ARMv8 to ARMv9 and Custom C.
The IP Block Economy
Beyond the top three, there's a diverse ecosystem of semiconductor IP vendors:
- Imagination Technologies — GPU IP (PowerVR). Lost Apple as a customer when Apple designed their own GPU, which was devastating. Now focused on automotive and IoT GPU markets.
- CEVA — DSP and AI processor IP for communications and edge AI applications
- Alphawave Semi — high-speed connectivity IP (SerDes, die-to-die links). Growing fast in the chiplet interconnect space.
- Rambus — memory interface and security IP. Their DDR5 and HBM PHY IP is in many data center chips.
- VeriSilicon — Chinese IP company with a broad portfolio, increasingly important for Chinese chip designers who face restrictions on some Western IP sources.
There's also a significant market for process-specific foundation IP — the standard cell libraries, memory compilers, and I/O cells that every chip design needs. TSMC provides some of this themselves, but companies like ARM (through their Artisan physical IP) and Synopsys supply most of it for advanced nodes.
Build vs. Buy Decisions
The fundamental question every chip design team faces: do we license this IP block or build it ourselves? The tradeoff is straightforward in theory but messy in practice.
Licensing is faster and lower risk. A verified USB 4 IP block from Synopsys has been taped out by dozens of customers on the target process node. The bugs have been found and fixed. The compliance testing is done. You might pay $500K-2M for the license, but you save your engineers 12-18 months of design and verification work.
Building in-house gives you full control and eliminates per-unit royalties. If you're shipping 100 million chips per year and paying a $0.15 per-unit royalty on a licensed IP block, that's $15 million annually in perpetuity. At some point, the math favors designing it yourself — assuming you have the talent to do it well.
For a related perspective, see Memory Testing and Reliability: DRAM Retention, RowHammer, a.
Apple is the poster child for the build-it-yourself approach. They've designed their own CPU cores, GPU, neural engine, image signal processor, and most interface controllers. But Apple ships over a billion ARM-based chips per year and has a hardware engineering budget measured in billions. For smaller companies, licensing is almost always the rational choice.
Chiplets and the IP Licensing Shift
The move toward chiplet architectures is changing how IP is licensed. In a traditional monolithic SoC, you license IP as design blocks that get integrated into a single die. In a chiplet architecture, an IP block might be a separate die — a "chiplet" — connected to other chiplets through a die-to-die interface.
This opens the door to a new model: IP-as-chiplet, where the vendor delivers a physical die rather than a design file. UCIe (Universal Chiplet Interconnect Express) is the standardized die-to-die interface that's supposed to make this practical. In theory, you could buy a pre-packaged PCIe controller chiplet from one vendor, a memory controller chiplet from another, and your own custom compute chiplet, then assemble them into a system-in-package.
We're still early in this transition, but it could fundamentally reshape the IP licensing business. Instead of selling design files and collecting royalties, IP companies might sell physical dies at a fixed price per unit — a model closer to component supply than intellectual property licensing.