{"slug": "nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it", "title": "NVIDIA Revives Rubin CPX With HBM4, Ditching the Cheap GDDR7 That Justified It", "summary": "NVIDIA has revived its Rubin CPX GPU for production in the first quarter of 2027, according to supply-chain analyst Ming-Chi Kuo of TF International Securities, switching from 128 GB of GDDR7 to 168 GB of HBM4 and a standalone MGX ETL rack. The redesigned CPX, which pairs 1:1 with the Vera Rubin NVL72 for prefill and decode, now consumes 2,300 W per GPU, matching the standard Rubin, and its shift to HBM4 could tighten the already constrained memory market.", "body_md": "*Ming-Chi Kuo says Rubin CPX is back on the roadmap for 1Q27 production, with 168 GB of HBM4 and a rack of its own.*\n\nRubin CPX was supposed to be the cheap one. When NVIDIA [announced it](https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference) at the AI Infra Summit last September, the whole argument rested on a single observation: the prefill stage of long-context inference is compute-bound, not bandwidth-bound. Feed a model a million tokens of source code and the GPU spends its time doing math, not waiting on memory. So NVIDIA built a monolithic die good for 30 PFLOPS of NVFP4 and hung 128 GB of GDDR7 off it, on the grounds that GDDR7 costs less than half what HBM does per gigabyte and prefill would never notice the difference.\n\nThen CPX quietly stopped showing up in roadmap talk, and most of the industry wrote it off.\n\nIt is not dead. Ming-Chi Kuo of TF International Securities says his [latest supply-chain checks](https://x.com/mingchikuo/status/2094426162193993964) have NVIDIA restarting the program, with production slated for the first quarter of 2027 and a redesign extensive enough that calling it the same product is a stretch.\n\n## 168 GB of HBM4 and a 2,300 W budget\n\nThe GDDR7 is gone. Kuo puts the revived CPX at 168 GB of HBM4, against 288 GB on a standard Rubin part. Compute lands close to full Rubin, and so does the power envelope: 2,300 W per GPU, matching Rubin’s own ceiling. That is a very different chip from the one described a year ago, when CPX was the frugal sidekick that let you throw silicon at prefill without paying HBM prices for the privilege.\n\nThe packaging changed with it. The original plan folded CPX into shared racks alongside Rubin, in the [Vera Rubin NVL144 CPX](https://www.techpowerup.com/350947/nvidia-details-the-rubin-architecture-die-annotation-vera-cpu-hbm4-and-disaggregated-inference) configuration. Kuo says the new version gets a standalone MGX ETL rack instead, with customers choosing 64, 128, 192 or 256 CPX GPUs. It still has to be paired with a Vera Rubin NVL72, and the recommended ratio is reportedly 1:1: CPX handles prefill and builds the KV cache, then ships it to Rubin over Ethernet RDMA for the decode pass.\n\n## Why a PC builder should care about a datacenter part\n\nThat 1:1 recommendation is the number worth staring at. Every CPX that ships now eats HBM4 stacks the old design would have satisfied with GDDR7, in a deployment model that explicitly wants one CPX for every Rubin in the building.\n\nHBM4 is already the tightest thing on the memory market. SK hynix, Samsung and Micron are all steering leading-edge DRAM capacity toward it, and every wafer that becomes an HBM stack is a wafer that does not become GDDR7 or DDR5. That is the same squeeze that has been dragging consumer memory contract prices upward all year, and the reason a console barely a year into its life just got a price increase instead of a price cut.\n\nNone of this is official. Kuo is reporting supply-chain intelligence, not an NVIDIA announcement, and his record is good rather than perfect. Roadmaps at this end of the market get rewritten constantly too: CPX being shelved and then un-shelved inside twelve months is the proof. NVIDIA has said nothing publicly, and the numbers could move again before anything reaches a substrate. But if the 1Q27 date holds and the 1:1 pairing survives contact with real customers, the 2027 memory picture just got tighter than it already looked.", "url": "https://wpnews.pro/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it", "canonical_source": "https://hwbusters.com/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it/", "published_at": "2026-09-01 15:11:43+00:00", "updated_at": "2026-09-01 16:25:34.076888+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-chips", "ai-infrastructure"], "entities": ["NVIDIA", "Ming-Chi Kuo", "TF International Securities", "Rubin CPX", "Vera Rubin NVL72", "HBM4", "GDDR7", "SK hynix"], "alternates": {"html": "https://wpnews.pro/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it", "markdown": "https://wpnews.pro/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it.md", "text": "https://wpnews.pro/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it.txt", "jsonld": "https://wpnews.pro/news/nvidia-revives-rubin-cpx-with-hbm4-ditching-the-cheap-gddr7-that-justified-it.jsonld"}}