Nvidia’s CEO says new Vera chip will use SK Hynix’s memory chips

Nvidia said its new Vera CPU will use SK Hynix DRAM, extending an already strong partnership as both companies prepare for a larger second half of the year and 2025. The chip is Nvidia’s first standalone data center microprocessor, positioning it against Intel Xeon, AMD Epyc, and hyperscaler in-house chips. Huang also highlighted ongoing talks with telecom companies, underscoring further AI-related demand.

Analysis

The key signal is not just another supply win for an incumbent vendor; it is NVIDIA broadening its bill-of-materials dependency into a memory partner that already sits at the center of high-bandwidth AI compute. That increases the strategic value of the memory supplier because it raises switching costs across the stack: once the CPU roadmap is coupled to specific DRAM validation, the supplier can gain share not only on next-gen GPUs but also on adjacent data-center platforms. The second-order effect is that memory content per AI server may keep rising even if accelerator unit growth slows, supporting revenue durability for the supply chain.

The competitive implication for CPU vendors is more subtle. Vera entering the data-center CPU market makes the incumbent x86 duo defend share in a segment where platform integration matters more than raw core count; if NVIDIA can use its ecosystem to bundle CPU + GPU + networking references, it can pressure pricing and design wins at the margin. The risk is that this is still a long-cycle qualification story: the market can over-earn the near-term benefit before actual shipment volume ramps, so the strongest P&L impact is likely months rather than days away.

For the memory supplier, the upside is cleaner than for the broader semiconductor group because this kind of design tie-in can improve mix and bargaining power more than headline unit growth. The contrarian angle is that consensus may be underestimating how much of the value accrues to the memory maker versus NVIDIA itself: if AI systems keep becoming more memory-intensive, the bottleneck migrates toward the parts of the stack with the least investor attention. The main reversal risk is any slowdown in hyperscaler capex or a platform shift that reduces per-node memory intensity, which would hit this thesis with a lag of 2-4 quarters.

The broader signal to watch is telco AI infrastructure: if NVIDIA is discussing telecom networks as future AI rails, it suggests an eventual expansion of inference workloads outside hyperscale cloud. That could create a second demand leg for edge and network-optimized silicon, but only after carrier budgets reaccelerate and standards settle. Near term, this is more important as a read-through on ecosystem breadth than as an immediate revenue catalyst.

Key Decisions for Investors

• Long SKM over the next 3-6 months: the design-win optionality plus rising AI memory content supports multiple expansion if hyperscaler capex stays intact; risk is a slower-than-expected qualification cycle, so size modestly and use weakness to add.
• Pair trade: long SKM / short AMD for 1-2 quarters — AMD faces more direct CPU competition from NVIDIA’s platform expansion, while SKM benefits from being embedded as a supplier across multiple compute layers.
• Initiate a tactical long in NVDA on any post-event consolidation, with a 1-3 month horizon: the message is ecosystem deepening, not a one-off partnership, which supports narrative durability; trim if the stock re-rates before order flow evidence appears.
• Use call spreads rather than outright longs in CPU incumbents if you want exposure to competitive response; the better setup is a short-dated hedge against margin pressure, since the market may not fully price in pricing concessions until design-win headlines arrive.