On the desk today · Rambus
Two engineers invented the way memory talks to the processor. Every chipmaker on earth pays for the conversation.
| NASDAQ · RMBS |
The Architecture Inside
Every Memory Chip
In March 1990, a Stanford electrical engineering professor named Mark Horowitz took a leave of absence. He and a computer architect named Mike Farmwald had identified a problem that nobody else seemed to care about: processors were getting faster every year, but the memory that fed them data couldn’t keep up. The gap — the “memory bottleneck” — meant that even the fastest processor spent most of its time waiting. Horowitz and Farmwald believed the fix wasn’t a faster chip. It was a faster interface — a new way for the memory to talk to the processor.
They raised $2 million from venture capitalists, gave up 50% of the company for it, and started building in a lab in Mountain View, California. They called the company Rambus.
I was reading a technical brief last month about DDR5 memory — the standard that runs inside every new server, every data center, and every AI training cluster. The document described how data moves between the memory module and the processor at billions of transfers per second. The interface architecture — the signaling, the timing, the protocol — traces directly to inventions patented by Rambus in the 1990s. I checked the licensing page. Intel, NVIDIA, Samsung, Qualcomm, Micron, SK hynix, Broadcom, Western Digital, Toshiba — all listed as licensees. You could build a server with any combination of chips from any of those companies. The royalty to Rambus would still apply.
Most people think of Rambus — if they think of it at all — as the company from the “memory wars” of the early 2000s, when it battled nearly every memory manufacturer on earth over patent rights. That reputation is not wrong. But it is incomplete. Rambus today is not just a licensing company. It is a semiconductor company that designs and sells its own memory interface chips — the physical silicon that sits on DDR5 memory modules and controls the flow of data. It also licenses silicon intellectual property — the design blueprints for HBM4, GDDR7, and PCIe 7.0 interfaces that other chipmakers embed in their own products.
In 1996, Intel licensed Rambus’s RDRAM technology as the next-generation PC memory standard. The industry resisted — RDRAM was expensive and proprietary. Competitors developed DDR-SDRAM as an open alternative. A decade of patent litigation followed. Rambus won most of the major cases and settled with nearly every memory chipmaker. The royalty stream that emerged from those settlements has been flowing ever since.
Rambus went public in 1997 on the Nasdaq, raising approximately $54 million. Today, it employs roughly 791 people from its headquarters in San Jose, California. CEO Luc Seraphin has led the company since 2020.
In fiscal 2025, Rambus reported record revenue of approximately $707 million. Product revenue — the DDR5 memory interface chips — accounted for 49% of the total. Royalties from patent licensing accounted for 40%. Silicon IP licensing and engineering services made up the remaining 11%. Operating margins ran between 35% and 38% across the year. Cash from operations hit a record $360 million. And diluted earnings per share reached $2.11.
|
Your Temporary Download Link Is Expiring
If you still haven't downloaded my "Safe Trade Options Formula" book...
...please take a few seconds and download it right now before your new temporary download link expires.
I eventually plan to charge for this book, so do yourself a favor and download it now...
That way, no matter how much it is in the future, you'll have a copy on your computer already.
Make sense?
P.S. Go here to save a copy of my "Safe Trade Options Formula" book to your computer before I start charging for it.
That 40% royalty share tells you something important about the durability of the business. Rambus invented foundational memory interface technology. It patented it. And every generation of memory that followed — DDR, DDR2, DDR3, DDR4, DDR5 — incorporated elements of that architecture. The patents don’t cover one product. They cover the way an entire category of products works. When a new memory standard is adopted, the royalty continues — because the interface principles Rambus invented travel forward with each generation.
The stickiness is architectural. A memory chipmaker that licenses Rambus’s patents has embedded the licensed technology into its manufacturing process, its chip designs, and its product roadmap. The license typically runs for multiple years and renews because the alternative — redesigning the interface from scratch to avoid the patent — is impractical. The technology is too fundamental. The cost of working around it exceeds the cost of licensing it. In fiscal 2025, the 10-K confirmed that Rambus “successfully secured and extended key patent licensing agreements, providing a stable foundation of royalty revenue.”
The product side is where the growth accelerates. Rambus’s DDR5 memory interface chips — the physical silicon that controls data flow on the memory module — generated 49% of revenue in fiscal 2025 and grew 37% year over year in the fourth quarter. Every AI server, every cloud data center, and every high-performance computing system running DDR5 memory needs these chips. As the world builds more data centers to train AI models, demand for DDR5 modules rises — and every module needs a Rambus interface chip.
The flywheel compounds through AI. Rambus is developing silicon IP for HBM4 — the next generation of high-bandwidth memory used in AI accelerators like NVIDIA’s GPUs. It is also developing GDDR7 interfaces for graphics and PCIe 7.0 for the connections between processors and storage. Each new standard requires new interface technology. Each new interface becomes a product Rambus sells and a patent Rambus licenses. The company that started by solving the memory bottleneck in 1990 is now at the center of the next one — the massive demand for data movement that every AI workload generates.
Forget SpaceX. Elon's next BIG bet is a radical "light-speed" device that's turning AI into "Accelerated AI" – and making it 100 times faster and 100 times more energy efficient.
The mainstream hasn't caught on yet. But "Accelerated AI" stocks are breaking out as we speak: 133%, 217%, or even 320% or more… and it's just getting started.
WHY THIS WORKS
The architecture travels with each generation. Rambus invented foundational memory interface technology. DDR, DDR2, DDR3, DDR4, DDR5 — each generation carried the architecture forward. The royalty renews with every new standard.
40% of revenue is patent royalties from the world’s largest chipmakers. Intel, NVIDIA, Samsung, Qualcomm, Micron, SK hynix, Broadcom — all licensees. The royalty base is diversified across the entire semiconductor industry.
49% of revenue is chips — growing 37% in Q4. DDR5 memory interface chips ride the AI and data center build-out. Every server needs them. Every memory module depends on them. The product side compounds alongside the licensing side.
791 employees generate $707 million in revenue. Nearly $900,000 per employee. The company is lean, IP-dense, and capital-light — the economics of a licensing firm with the growth profile of a semiconductor company.
In 1990, Mark Horowitz and Mike Farmwald left Stanford with $2 million and a thesis: the bottleneck wasn’t the processor. It was the memory. They invented a new interface — a way for data to move faster between the chip and the module — and patented it. The industry adopted the architecture. Then it tried to use it without paying. A decade of litigation followed. Rambus won. Thirty-six years later, every major chipmaker on earth is a licensee. The servers that train AI models, the data centers that run the cloud, and the smartphones in your pocket all move data through interfaces that trace back to a Stanford lab and two engineers who saw a gap nobody else was looking at. The bottleneck was the opportunity. The patent was the toll.
