Rambus (NASDAQ:RMBS | RMBS Price Prediction) has become one of the more interesting names in the AI infrastructure supply chain, and it remains under the radar for most investors. This is the memory bandwidth IP company sitting underneath the entire AI buildout, yet awareness outside semiconductor circles is limited. That gap between fundamentals and recognition is what makes the research case worth examining.
Rambus designs the memory interface chips and the controller IP that let DDR5, LPDDR5X, and HBM memory actually keep up with AI accelerators. CEO Luc Seraphin put it bluntly on the Q1 call. “Everyone is trying to optimize now the memory subsystems… HBM, DDR, LPDDR… this plays to our strength because this is what we’ve been doing forever at Rambus.” When NVIDIA (NASDAQ:NVDA), AMD (NASDAQ:AMD), or a hyperscaler designs a custom accelerator, the memory side of that chip is Rambus territory.
Reason One: A Hardware Business Running Software-Company Margins
Q1 2026 gross margin came in at 79.73%. Full-year 2025 sat at 75.98%, up from 69.14% in 2023. That margin profile exists because the royalty and IP side of the business is essentially capital-light, and the chip side is a high-value piece of silicon that customers cannot easily second-source.
Operating income went from $91.5 million in 2023 to $260.2 million in 2025. Q1 2026 operating cash flow was $83.2 million, and cash plus marketable securities now sit at $786 million against total liabilities of just $139.9 million. The balance sheet is effectively net cash.
Reason Two: Every AI Chip on Earth Needs This IP
Product revenue grew 15% year over year to $88 million in Q1, with Q2 guided to $95 to $101 million. Seraphin says Rambus exited 2025 with mid-40% share in DDR5 RCDs and is gaining as the market transitions from Gen 2 to Gen 3.
The company also launched the industry’s fastest HBM4E controller, the exact IP block AI accelerators need for next-generation memory throughput. The MRDIMM opportunity alone is a $600 million SAM with the earnest ramp in 2027. Full-year 2025 revenue grew 27.13% to $707.63 million.
Reason Three: Victor Peng on the Board Is the Signal
AMD’s former president joined the board, with the equity grant landing April 1, 2026. Peng ran Xilinx before AMD bought it and led AMD’s adaptive computing push.
He does not take a seat at a sub-$20 billion semiconductor company by accident. Custom silicon for hyperscalers is exactly where Rambus is positioned, and Peng is the operator who has lived inside that customer base.
The Honest Risk
Royalty revenue declined to $69.6 million from $74.0 million a year ago, non-GAAP operating margin compressed from 46% to 42%, and one analyst downgrade flagged tightening DRAM supply.
The CFO also resigned. None of that changes the underlying thesis. Royalty lumpiness comes with the licensing model. Margin compression is funding R&D that grew to $50.23 million, which is the spend that produced the HBM4E controller in the first place. Supply tightness is a demand problem.
Why the Valuation Matters
Shares trade at a forward multiple of roughly 24 for a company growing product revenue double digits, sitting on $786 million in cash, earning 79.73% gross margins, and embedded inside every meaningful AI memory architecture.
Analyst consensus rating sits at seven buys and two holds. The stock is up 170% over the past year. For investors researching under-the-radar AI infrastructure exposure, Rambus warrants a closer look as the next decade of inference workloads gets built.
All that said, I still wouldn’t expect big, explosive gains. This is an IP company and they’re never going to scale into the trillions with IP only. Thankfully, Rambus has moved into memory designing and derives around half of its revenue from semiconductor products. But the non-IP business still needs time to scale significantly and mature.