Speed over scale: Samsung pulls ahead with Nvidia’s HBM4

Nvidia is considering a dual-track adoption strategy for high bandwidth memory in its upcoming Vera Rubin AI accelerators, a move that could further cement Samsung Electronics and SK hynix at the top of the global artificial intelligence memory hierarchy, industry sources said Thursday.

Under the proposed “dual-bin” structure, HBM4 chips would be divided into two performance tiers. Premium Rubin systems would use memory operating at up to 11.7 gigabits per second, while lower-tier versions would deploy chips running at around 10 Gbps.

The strategy reflects both explosive AI demand and widening technical gaps among suppliers — and it places South Korea’s two memory giants squarely in the highest-value segments.

Samsung is widely expected to secure a role in Nvidia’s top performance tier. Its HBM4, which recently entered mass production, reaches speeds of up to 11.7 Gbps — 46 percent above the 8 Gbps baseline set by the Joint Electron Device Engineering Council. The company is also the first to apply sixth-generation 10-nanometer-class (1c) DRAM to HBM, claiming stable yields and signaling potential speed expansion to 13 Gbps.

The move marks a notable rebound for Samsung after lagging in the previous HBM cycle. By positioning its 1c-based HBM4 in Nvidia’s premium bin, the company is regaining technological leadership in the AI memory race.

SK hynix, the current HBM market leader and Nvidia’s primary supplier in earlier generations, is expected to account for roughly two-thirds of total HBM4 supply for Rubin. Its HBM4, built on fifth-generation 10-nm-class (1b) DRAM, exceeds Jedec standards and is likely to anchor Nvidia’s mainstream tier alongside Micron.

Even within a dual-tier structure, SK hynix remains central to Nvidia’s volume strategy, underscoring its entrenched dominance in the HBM ecosystem.

Pricing is expected to reinforce the hierarchy. Samsung’s top-bin HBM4 is projected to command around $700 per unit — roughly 20 to 30 percent higher than HBM3E — while SK hynix’s HBM4 had been quoted in the mid-$500 range last year. Nvidia is expected to adjust accelerator pricing accordingly to manage rising memory costs.

Market tracker TrendForce said on Feb. 13 that Nvidia is likely to source HBM4 from Samsung, SK hynix and Micron, noting that no single supplier can meet total demand for the Rubin platform. It projects Samsung’s HBM share to rise to 28 percent this year from 20 percent last year, narrowing the gap with SK hynix.

Industry observers say Nvidia’s approach underscores a broader shift in the AI chip market, where performance capabilities are increasingly outweighing pure mass production competition.

“With competition in AI accelerators intensifying, performance requirements for memory are becoming more critical than production yield,” an industry official said.

Another industry source noted that the market’s recent focus on large-scale production races may ease, allowing technological competitiveness to play a greater role in determining market influence.

Earnings momentum for both Samsung and SK hynix is expected to continue throughout the year, amid rising memory prices. Global investment bank Morgan Stanley forecast Samsung to post 245.7 trillion won ($189 billion) in operating profit this year, and SK hynix to reach 179.4 trillion won, representing on-year increases of 464 percent and 280 percent, respectively.