Nvidia is reportedly unbundling video RAM (VRAM) from its GPUs for the consumer and gaming market.

According to Weibo posts from leaker Golden Pig Upgrade, the chip giant Nvidia is leaving add-in card (AIC) partners to source their own memory due to supply chain issues.

Nvidia potentially unbundling its services to focus on the GPU core would see large board partners turn to the likes of Samsung, SK Hynix, and Micron with ease, while likely prioritizing memory for the enterprise market.

A possible predicament is what happens to smaller firms unable to absorb increasing costs. One example of a GPU AIC or add-in-board (AIB) partner, EVGA, left the GPU market in 2022 due to a fractious relationship with the firm and increasing prices affecting its profitability. A similar situation saw one-time Nvidia partner XFX switch to AMD GPUs back in 2010.

As supply chain issues heat up, Nvidia CEO Jensen Huang was on the offensive while on a trip to Taiwan, visiting TSMC founder Morris Chang.

“The demand is very strong right now,” Huang told Taiwanese media over Thanksgiving weekend. “We have complex memories … We don't have an abundance of anything, but our supply chain is very strong. We have the largest supply chain in the world, so if customers need supply, we will deliver it.”

Recent weeks have seen VRAM providers such as Samsung reportedly hike the price of its memory wares to meet demand, while HP warned rising memory chip prices would impact its profitability in the second half of fiscal 2026.

Hewlett Packard Enterprise (HPE), meanwhile, had its ratings cut by Morgan Stanley, with the financial giant warning that hardware original equipment manufacturers (OEMs) such as HPE historically face gross-margin compression within a year after memory costs begin rising.

SCADA and AI

While Nvidia may be unbundling memory from its GPUs, it is also offloading AI-related CPU tasks to its chips using Scaled Accelerated Data Access, or SCADA.

Nvidia research published last month presents the GPU-controlled storage I/O framework as able to offload both control and data paths from the CPU to the GPU for faster small-block transfers in AI inferencing, in which a trained machine learning (ML) model generates predictions and outputs via new input data.

As SCADA is distinct and separate from the VRAM function, being built more for enterprise workloads, a core GPU works the same for traditional AI workloads in its absence.

The memory ware SCADA does rely on is NAND, with Nvidia sourcing PCIe Gen 6 SSD from Micron, specifically the 9650.

As reported by Blocks and Files, the flash memory offering is designed for small-block I/O operations, thus accelerating AI pipelines by allowing GPUs to grab data directly from the SSD via the non-volatile memory express (NVMe) protocol.

It is most likely that Nvidia will continue to bundle memory with its enterprise-grade chips as the chipmaker continues to meet ever-increasing AI demand, with Huang confident demand won't outstrip supply.