NVIDIA's Rubin CPX Stalls; Supply Chain Cites No Orders as Groq LPU Takes Over Inference Role

NVIDIA's planned Rubin CPX GPU, which relies on GDDR7 memory, appears to be shelved, as suppliers report a complete lack of orders.

This shift follows NVIDIA's integration of Groq's LPU technology, an alternative AI chip that now fills the CPX's intended role for fast AI inference.

As a result, the industry is pivoting even more heavily toward HBM memory and advanced packaging, diminishing the near-term server market for GDDR7.

NVIDIA's anticipated Rubin CPX GPU, once slated for a 2026 release, now appears to be quietly shelved.

The most direct signal comes from the supply chain, where key Korean memory and substrate suppliers have reported receiving no orders for the GDDR7 memory or related components tied to the CPX project. In the fast-moving semiconductor world, a lack of orders this close to a planned launch is often interpreted as a cancellation or a major redesign. This development has effectively shrunk the immediate market opportunity for GDDR7 memory in AI servers.

So, what caused this sudden change? The story traces back to a strategic pivot by NVIDIA. First, NVIDIA struck a major technology licensing deal with a startup named Groq. This gave NVIDIA access to Groq's innovative Language Processing Units (LPUs), chips specifically designed for extremely fast AI inference—the very task CPX was meant to handle. With Groq's technology in hand, the strategic need for a separate, GDDR7-based product like CPX diminished significantly.

Second, this shift was made public, albeit subtly, at NVIDIA's GTC 2026 conference. The Rubin CPX was conspicuously absent from the company's updated roadmap slides. In its place, Groq's LPUs were elevated as a key part of the 'Vera Rubin' platform. This sent a clear message to the industry: NVIDIA's inference strategy had evolved.

Finally, broader supply chain dynamics have reinforced this decision. The entire AI hardware industry is currently bottlenecked by the production of HBM (High Bandwidth Memory) and advanced packaging technologies like CoWoS. Memory manufacturers like SK hynix and Samsung are investing billions to ramp up HBM production, as it's the critical component for high-end training GPUs. Prioritizing a new GDDR7-based server product, which doesn't solve the core HBM bottleneck, became a less attractive path for both NVIDIA and its suppliers. In essence, the market has voted with its capital, betting on HBM as the future of AI memory.

HBM (High Bandwidth Memory): A type of high-performance memory where DRAM chips are vertically stacked. It offers much wider data paths and lower power consumption compared to traditional memory, making it ideal for high-end AI accelerators.
GDDR7 (Graphics Double Data Rate 7): The next generation of memory primarily used for graphics cards. It offers high speeds on a simpler, board-level integration but lacks the bandwidth of HBM.
LPU (Language Processing Unit): A specialized processor architecture, developed by Groq, designed to excel at the sequential, memory-intensive operations required for running large language models (inference) with very low latency.

NVIDIA's Rubin CPX Stalls; Supply Chain Cites No Orders as Groq LPU Takes Over Inference Role

NVIDIA's planned Rubin CPX GPU, which relies on GDDR7 memory, appears to be shelved, as suppliers report a complete lack of orders.

This shift follows NVIDIA's integration of Groq's LPU technology, an alternative AI chip that now fills the CPX's intended role for fast AI inference.

As a result, the industry is pivoting even more heavily toward HBM memory and advanced packaging, diminishing the near-term server market for GDDR7.

NVIDIA's anticipated Rubin CPX GPU, once slated for a 2026 release, now appears to be quietly shelved.

HBM (High Bandwidth Memory): A type of high-performance memory where DRAM chips are vertically stacked. It offers much wider data paths and lower power consumption compared to traditional memory, making it ideal for high-end AI accelerators.
GDDR7 (Graphics Double Data Rate 7): The next generation of memory primarily used for graphics cards. It offers high speeds on a simpler, board-level integration but lacks the bandwidth of HBM.
LPU (Language Processing Unit): A specialized processor architecture, developed by Groq, designed to excel at the sequential, memory-intensive operations required for running large language models (inference) with very low latency.