Nvidia’s Vera CPU Push Points to Deeper SK hynix Memory Commitments

Nvidia CEO Jensen Huang said the company’s Vera CPU will use SK hynix DRAM, as the two companies prepare for “a very, very large second half of the year and next year,” according to remarks published June 7 by Fortune, citing a Bloomberg report. Huang made the comments in Seoul while meeting with leaders including Chey Tae-won, chairman of SK Group, and Kwak Noh-Jung, CEO of SK hynix. The message ties Nvidia’s newest data-center CPU strategy directly to a key supplier at a time when AI system makers are trying to scale deployments faster than memory ecosystems can expand.

Vera is Nvidia’s first standalone data-center microprocessor aimed at workloads associated with “agentic AI,” where systems not only generate answers but also take actions, run code, and use tools. Nvidia describes Vera as a new class of CPU intended for reinforcement learning, agentic inference, and data processing, and says it is designed to help AI “factories” produce more output tokens per dollar. In a separate Nvidia announcement, the company said Vera is “now in full production” and can deliver 1.8x faster task completion compared with x86 CPUs, targeting the kind of iterative, sandbox-style experimentation that agent systems require.

Nvidia’s product materials also emphasize memory bandwidth as a differentiator for Vera. On Nvidia’s Vera CPU page, the company highlights that Vera uses high-bandwidth LPDDR5X memory and delivers up to 80% faster sandbox environment performance than traditional CPU infrastructure. In a technical blog post about Vera, Nvidia added quantified performance claims including up to 50% faster “agentic sandbox” performance and 1.2 TB/s memory bandwidth, alongside architectural details such as custom “Olympus” CPU cores and Nvidia’s scalable coherency fabric for coordinating movement between CPU and GPUs.

The SK hynix connection is not just about purchasing DRAM for an existing design. On June 7, Nvidia and SK hynix announced a multiyear technology partnership focused on advancing next-generation memory for the global “AI factory” buildout. Nvidia said the agreement is intended to expand supply for advanced memory while addressing extended development cycles, advanced fabrication steps, and capital investments needed to sustain the buildout of AI infrastructure.

In the partnership announcement, Nvidia described it as “codevelopment” work aligned with its AI infrastructure roadmap, from frontier model training to agentic and physical AI. Nvidia and SK hynix said SK hynix will diversify into markets shaped by Nvidia’s computing platforms, including Nvidia’s Vera Rubin AI supercomputers, Vera CPUs, RTX Spark-powered PCs, and Jetson Thor robotics platforms. The companies also said they will use Nvidia software and AI tools to accelerate semiconductor chip design and simulation, including CUDA-X libraries and PhysicsNeMo, and to help SK hynix build “fab digital twins” for more autonomous manufacturing workflows using Omniverse, OpenUSD, and cuOpt.

What the latest comments suggest, taken together, is that Nvidia is using the ramp of Vera and the broader Rubin ecosystem to lock in memory partnership depth as scaling pressures intensify. Huang’s statement that a “very big year” with SK hynix is being followed by a much larger second half and next year implies near-term production and supply scheduling is becoming a core element of Nvidia’s CPU and system rollout planning, not a downstream afterthought.

Even with the added context from Nvidia, key implementation details remain undisclosed. Nvidia’s public materials about Vera emphasize high-bandwidth LPDDR5X and performance targets, but they do not spell out which specific SK hynix DRAM generations, die formats, or exact supply volumes will be used in the initial Vera deployments. Likewise, the multiyear partnership announcement outlines co-development and broader platform coverage, but it does not provide a timeline for when particular next-generation memory revisions will reach mass production. For now, the most concrete takeaway is the direction of travel: memory supply and co-design are being treated as strategic gating factors for the next phase of Nvidia’s AI infrastructure.