Key Takeaways
Nvidia's next-generation Vera Rubin platform depends on a single memory supplier for the majority of its high-bandwidth memory.
SK Hynix has secured between 50% and 70% of Nvidia's anticipated HBM4 orders, making the South Korean memory maker the dominant supplier for the Vera Rubin AI platform that entered production this quarter.
"Nvidia will be the only customer of HBM4 for quite some time," Jensen Huang, chief executive officer of Nvidia, said at CES in January. The company's Vera Rubin architecture relies on fourth-generation high-bandwidth memory (which stacks DRAM dies vertically to deliver data at speeds conventional memory cannot match) to power its next-generation AI supercomputers.
The multiyear partnership, announced in early June, deepens a relationship that began with Huang asking SK Hynix to accelerate its HBM4 timeline by six months as early as 2024. SK Hynix plans to double its wafer capacity by 2030, though Huang has warned that even that expansion may not keep pace with AI demand. "AI factories are the engines of the next industrial revolution, and advanced memory is essential to their performance," Huang said at the time.
The concentration of HBM4 supply with a single vendor creates both opportunity and risk. SK Hynix stands to capture a disproportionate share of the memory revenue tied to Nvidia's next-generation platform, while Nvidia faces single-source exposure at a time when KeyBanc analyst John Vinh identified an unconfirmed HBM4 qualification risk that could slightly affect the Vera Rubin rollout. Vinh maintained his annual shipment forecast of 1.7 million to 1.8 million Rubin chips for 2026, alongside 5.5 million to 6.0 million Blackwell units, and raised his Nvidia price target to $330 from $310.
How HBM4 powers the Vera Rubin leap
Vera Rubin pairs 72 Rubin graphics processors with 36 Vera central processors in the NVL72 rack-scale configuration, connected through NVLink 6 at 3.6 terabytes per second of GPU-to-GPU bandwidth. Each Rubin GPU uses 288 gigabytes of HBM4 memory, a fourfold increase in capacity per chip compared with the HBM3e used in the current Blackwell generation. Nvidia projects up to a tenfold reduction in inference cost per token versus Blackwell, though that estimate has not been independently verified.
The memory bottleneck has become one of the most critical constraints in AI infrastructure. Training large language models requires transferring enormous volumes of data between hundreds of thousands of GPUs at extremely low latency. Conventional DRAM cannot keep pace, leaving compute resources underutilized. HBM4 solves this by stacking DRAM dies and connecting them through vertical interconnects, delivering bandwidth far above traditional memory while consuming less power.
Competitive stakes and supply chain risks
The HBM4 supply chain involves three manufacturers capable of producing the memory at scale: SK Hynix, Samsung, and Micron Technology. SK Hynix's early lead in qualification and its exclusive partnership with Nvidia give it a structural advantage, but the concentration also creates risk. Any delay in HBM4 qualification could ripple through Nvidia's production timeline, even as the thermal lid manufacturing issue that briefly slowed Rubin GPU production has already been resolved, according to Vinh.
Nvidia's competitors are not standing still. Amazon Web Services' Trainium3 accelerator is already shipping for AI training and inference, with UltraServers scaling to 144 chips. Google Cloud's Ironwood tensor processing unit is generally available for large-scale workloads. Both platforms offer cloud customers immediate capacity alternatives to waiting for Rubin's projected efficiency gains.
Nvidia reported Q1 FY2027 revenue of $81.6 billion, up 85% year over year, with data center revenue reaching $75.2 billion, up 92%. The company guided for approximately $91 billion in Q2 revenue, excluding China. Nvidia shares trade at roughly 35 times forward earnings, and the company returned $20 billion to shareholders through buybacks and dividends in the first quarter alone. The August earnings report will provide the first concrete data on the Blackwell-to-Rubin transition and any revenue impact from the company's new Asian customer white list, which removed more than half of its approved buyers in Southeast Asia to prevent chip diversion to China.
For investors, the SK Hynix-Nvidia partnership represents one of the clearest revenue tailwinds in the AI infrastructure theme. The memory maker's outsize allocation of HBM4 orders could drive durable earnings growth as Nvidia ramps Vera Rubin through the second half of 2026. But the single-source dependency also means any disruption in SK Hynix's qualification or production could delay the platform that Nvidia is betting on to maintain its dominance in AI compute.
This article is for informational purposes only and does not constitute investment advice.
