3 Under-the-Radar Small-Caps Riding the AI Memory Wave

Everyone is obsessed with the same five AI stocks. Meanwhile, a Bloomberg Podcast segment this week quietly dropped a detail that should have pinged every small-cap hunter: High Bandwidth Flash (HBF) is coming, it targets 3 TB/s bandwidth, and it stacks NAND dies the same way HBM stacks DRAM. The AI memory supply chain is about to get a whole lot more interesting, and the picks-and-shovels plays hiding inside it are basically invisible to the average investor. Here are three worth watching.

$RMBS (Rambus Inc.) is the kind of company that makes engineers excited and retail investors yawn, which is exactly why it belongs on this list. Rambus develops semiconductor interface IP and memory chip architectures, licensing the core technology that makes high-speed memory actually work at scale. Market cap sits comfortably under $10 billion, it trades on NASDAQ, and its patent portfolio covers protocols directly relevant to next-generation memory like HBM and, soon, HBF. The catalyst: as AI chipmakers race to certify new memory standards, Rambus collects royalties on every design that uses its IP. You do not need Rambus to manufacture a single chip to win. The bear case is that licensing businesses are slow-burn plays and royalty disputes can drag on for years. But in a world where AI capex is running at $800 billion annually and heading toward $1.1 trillion, the royalty pool only grows.

$FORM (FormFactor Inc.) makes wafer probe cards, the unglamorous testing equipment that semiconductor fabs cannot skip. Every chip, including every HBM stack and every AI accelerator, has to be tested before it ships. FormFactor holds a leading position in advanced probe cards for memory and logic, and its revenue is directly tied to fab utilization ramp-ups. Under $3 billion market cap, listed on NASDAQ, and almost never mentioned in mainstream financial media. The specific catalyst here is a capacity ramp at major memory manufacturers as they retool for HBM4 and eventual HBF production. More layers stacked per die means more complex testing, which means more FormFactor cards per wafer run.

$CEVA (CEVA Inc.) licenses signal processing IP for wireless, Bluetooth, and, increasingly, on-device AI inference chips. As AI moves from the cloud to the edge, every smart device needs a tiny, power-efficient inference engine. CEVA's DSP cores are embedded in billions of shipped chips annually, and the company earns royalties on each one. Market cap is well under $1 billion, it trades on NASDAQ, and Wall Street mostly ignores it because the revenue line looks lumpy and the business model requires patience. The bull case: edge AI inference is exactly the workload that HBF and next-gen memory are designed to serve, and CEVA sits right at that intersection. The bear case is execution risk on design wins converting to volume royalties, which can take 18 to 24 months.

Three companies, three different layers of the AI stack, zero appearances in your average financial influencer's portfolio tour. The Asianometry breakdown of High Bandwidth Flash this week was a quiet reminder that the next memory supercycle is not going to announce itself with a press release. It rarely does.