Nvidia Shakes Up HBM-Dominated Market, Adopts GDDR7 in Next-Gen AI Chips to Signal Diversification

Strength in Processing Speed and Cost Efficiency
Surging Demand Set to Boost Graphics DRAM Market
Ripple Effects Loom for Memory Suppliers

Nvidia is preparing to launch its next-generation artificial intelligence (AI) chip with a surprising twist: the integration of GDDR7, often dubbed “half-price memory,” alongside high-bandwidth memory (HBM). The move has jolted the industry, with analysts predicting it could erode HBM’s near-monopoly as the default choice for high-performance chips. Samsung Electronics is expected to emerge as a key supplier, potentially reshaping the global memory market’s competitive landscape.

Pre-Fill with GDDR7, Decode with HBM4

According to industry reports on the 15th, Nvidia’s Rubin CPX, scheduled for release next year, will incorporate GDDR7 optimized for speed. With this design, Nvidia cut Rubin CPX’s total memory cost to one-fifth of its predecessor’s while preserving 60 percent of the R200 chip’s performance. Far from a simple memory swap, the architecture reflects an attempt to rebalance cost and performance in AI inference workloads.

It is rare for GDDR7—typically used in consumer graphics cards—to appear in high-performance AI chips. But Nvidia is breaking that boundary. In Rubin CPX, GDDR7 will handle compute-intensive pre-fill tasks, while HBM4 will be deployed for data-heavy decode stages. This division of labor expands GDDR7’s role into high-end AI applications, positioning it as a strategic experiment to raise inference efficiency.

Behind the shift lies HBM’s high cost and supply constraints. While HBM offers unmatched bandwidth and efficiency, its complex manufacturing makes it expensive and yields less predictable. Market researcher SemiAnalysis estimates GDDR7 is about 50 percent cheaper per gigabyte than HBM. For Nvidia, this cost advantage lowers data center operating expenses and provides a platform for scaling concurrent workloads.

Samsung is expected to supply a significant portion of GDDR7 for Rubin CPX. Nvidia has already placed quarterly orders of 30 to 35 million 16Gb GDDR7 chips from Samsung for its Blackwell GPUs, and similar procurement patterns are anticipated for Rubin CPX. With SK Hynix and Micron focused on HBM3E (12-layer) production, Samsung is seen as having a clear advantage in GDDR7 supply capacity.

Industry Embraces Hybrid Strategies

Rubin CPX is not Nvidia’s first foray into GDDR7. The company has already equipped its GeForce RTX 5090 gaming cards with the memory and plans to use GDDR7 in the China-focused B40 chip. In that case, lower-bandwidth GDDR7 will replace HBM to comply with U.S. export controls. Analysts expect B40 shipments to reach at least 1 million units this year and up to 5 million next year.

The trend is spreading beyond Nvidia. Canadian AI startup Tenstorrent used 24GB of GDDR6 in its Wormhole accelerator last year and 32GB in its newly unveiled Blackhole. While a generation behind GDDR7, these moves highlight how new players are deploying GDDR memory for inference tasks, where it complements HBM’s strengths.

Industry experts note that “demand for HBM is soaring, but supply remains far short,” adding that “GDDR7 could carve out a new niche.” Since inference workloads require less bandwidth than training, mass-producible GDDR7 could quickly gain ground. Some caution that GDDR7 alone cannot stably power high-end AI servers, but Nvidia’s adoption may push competitors toward similar hybrid strategies.

Samsung’s Rise as a “Dual Supplier”

The shift carries direct implications for suppliers, particularly Samsung. The company recently expanded its graphics DRAM production lines at its Pyeongtaek campus, with operations expected to begin as early as this month. Initially prepared for next-gen products, the lines have been repurposed to meet surging GDDR7 demand. This expansion positions Samsung to handle Rubin CPX volumes and future demand for gaming and data center GPUs.

Samsung also leads in technology. It was the first to develop GDDR6 in 2022, unveiled GDDR7 in July the following year, and achieved mass production of 24GB GDDR7 on a 12nm process last October. With speeds up to 48Gbps—double the prior generation—Samsung’s GDDR7 has demonstrated sufficient competitiveness for AI inference chips. These advances suggest Samsung’s role could extend beyond fulfilling orders to driving new market demand.

According to U.S. research firm Industry Growth Insights, the graphics DRAM market will expand from $3.2 billion in 2018 to $4.8 billion by 2030. GDDR7 is expected to anchor this growth by serving both AI chips and high-performance graphics cards. For suppliers, securing stable GDDR7 capacity means tapping into a lucrative new revenue stream.

Samsung stands out as the only company supplying both HBM and GDDR7. While HBM remains essential for ultra-high-performance training chips, GDDR7 is poised for rapid adoption in inference and general-purpose GPUs, where cost and efficiency dominate. With its Pyeongtaek expansion, Samsung is not only meeting immediate demand but also accelerating a structural realignment of the global memory ecosystem.