NVIDIA Rubin CPX: A Full Guide to the Context GPU, GDDR7, NVL144 CPX, the HBM4 Redesign and Limits

NVIDIA announced Rubin CPX in September 2025 as a new class of GPU purpose-built for massive-context processing, and said its flagship rack would return $5 billion in token revenue for every $100 million invested. Six months later, the chip had vanished from the GTC 2026 keynote, and NVIDIA's Ian Buck told reporters "we've pulled CPX".

Then, in late August 2026, analyst Ming-Chi Kuo reported that NVIDIA had revived the program with HBM4 and production expected in 1Q27. NVIDIA hasn't confirmed it. So which Rubin CPX is real: the 2025 GDDR7 chip, the shelved one, or the rumored redesign?

This full guide covers all three. It explains why NVIDIA split inference into a context phase and a generation phase, the original Rubin CPX specs, the Vera Rubin NVL144 CPX rack, why Groq 3 LPX took its slot, what the HBM4 redesign report says, and the honest limits, all as of October 30, 2026. For the rack generation that follows, see Inside Deep Tech's NVIDIA Rubin Ultra and Kyber full guide.

Last updated: October 30, 2026.

Key Takeaways

• NVIDIA Rubin CPX was announced on September 9, 2025 as a prefill GPU with up to 30 petaflops of NVFP4 compute and 128GB of GDDR7, expected at the end of 2026.
• Its showcase system, Vera Rubin NVL144 CPX, promised 8 exaflops, 100TB of fast memory and 7.5x the AI performance of GB300 NVL72 in one rack.
• NVIDIA dropped CPX from its GTC 2026 roadmap, and Groq 3 LPX took its place as the specialized inference rack beside Vera Rubin NVL72.
• A redesign reported by Kuo would swap GDDR7 for 168GB of HBM4 at about 2,300 W in standalone racks of 64 to 256 GPUs. NVIDIA hasn't confirmed it.
• Honest limit: NVIDIA's Q2 FY2027 earnings call didn't mention CPX, so nobody outside NVIDIA can say if, when or in what form it ships.

What NVIDIA Rubin CPX actually is

Rubin CPX is a Rubin-architecture GPU designed for one job: the context phase of inference, also called prefill. That's the step where a model reads the whole prompt, document set, codebase or video and builds the key-value (KV) cache it needs before it writes a single output token.

NVIDIA built it as a cost-efficient monolithic die with dense NVFP4 math and integrated video decoders and encoders. Unlike the flagship Rubin GPU, which uses two reticle-sized dies and HBM4, CPX was meant to be cheaper per unit of compute, with ordinary graphics memory instead of stacked memory.

Jensen Huang pitched it as "the first CUDA GPU purpose-built for massive-context AI". The target workloads were million-token coding agents and long video, where NVIDIA said a model can take up to 1 million tokens for an hour of content.

CPX was never meant to work alone. It sat beside regular Rubin GPUs and Vera CPUs, handing finished KV caches to the Rubin GPUs that run decode. That pairing is the whole idea, so it helps to understand why the two phases are so different.

Why prefill and decode want different chips

NVIDIA's own Rubin CPX technical blog frames the split clearly. The context phase is compute-bound: it processes huge amounts of input in parallel and keeps math units busy. The generation phase is memory-bandwidth-bound: each new token reads the model weights and KV cache again, so speed depends on how fast memory can feed the chip.

NVIDIA's Dynamo disaggregated serving docs describe the three steps in production terms: prefill workers compute the KV cache, the cache moves to decode workers, and decode workers generate tokens. Each pool can then scale on its own.

On a general-purpose GPU, prefill leaves expensive HBM bandwidth idle, and decode leaves expensive compute idle. CPX attacks the first waste. It keeps lots of FP4 compute and drops HBM, the most expensive and supply-constrained part of a modern AI GPU. Inside Deep Tech's GDDR7 full guide explains why GDDR7 is far cheaper to build into a board than stacked memory.

Rubin CPX specs at a glance

NVIDIA published only a handful of official CPX numbers. Independent analysts filled in more, so the table labels each source. Treat the SemiAnalysis column as estimates, not NVIDIA specs.

Source: NVIDIA Rubin CPX announcement; SemiAnalysis Rubin CPX analysis; NVIDIA Vera Rubin NVL72 page and NVIDIA CES 2026 Rubin release for the Rubin GPU; die count per Inside Deep Tech's Vera Rubin NVL72 guide.

The 3x attention claim matters most. Attention cost grows with context length, so a chip that runs attention faster keeps time-to-first-token (TTFT) manageable as prompts grow toward a million tokens. NVIDIA made that claim against GB300 NVL72 systems, not against Rubin.

SemiAnalysis estimated that GDDR7 costs more than 50% less per GB than HBM, and that skipping CoWoS packaging frees CPX from the same advanced-packaging bottleneck that limits flagship GPUs. Those cost points were the commercial case for the chip.

Inside the Vera Rubin NVL144 CPX rack

NVIDIA's showcase system packed CPX into the same MGX rack as the standard Vera Rubin design. Per NVIDIA's technical blog, the Vera Rubin NVL144 CPX rack combined 144 Rubin CPX GPUs, 144 Rubin GPUs and 36 Vera CPUs.

NVIDIA quoted the rack at 8 exaflops of AI compute, 7.5x the AI performance of GB300 NVL72, 100TB of fast memory and 1.7 petabytes per second of memory bandwidth. It also said a dedicated Rubin CPX compute tray would be offered so customers could add CPX to existing Vera Rubin racks.

The rack counted GPU dies, which is why it carried the NVL144 name. NVIDIA later switched to counting GPU packages, so the standard rack is now sold as Vera Rubin NVL72. Inside Deep Tech's Vera Rubin NVL72 full guide covers that naming change and the base rack in detail.

Networking options included Quantum-X800 InfiniBand or Spectrum-X Ethernet with ConnectX-9 SuperNICs. SemiAnalysis noted the CPX trays relied on PCIe Gen6 rather than NVLink, at roughly 1 Tbit/s per CPX versus 14.4 Tbit/s of NVLink, which was fine for handing off KV caches but not for tight GPU-to-GPU work.

Power was the other trade. SemiAnalysis estimated the NVL144 CPX rack at about 370 kW versus about 190 kW for the standard rack, and said the fixed ratio of prefill to decode GPUs in one rack was a real downside, because real workloads don't hold one ratio.

The economics NVIDIA promised

NVIDIA sold CPX on token economics rather than raw speed. The headline was $5 billion in token revenue for every $100 million invested in the NVL144 CPX platform, a 50x figure. The developer blog put the range at 30x to 50x return on investment.

Those are NVIDIA marketing numbers built on assumed token prices and utilization. They aren't measured customer results. Still, the logic was sound: if prefill runs on cheaper GDDR7 silicon, the scarce HBM4 GPUs spend more of their time on decode, the phase that most directly drives output tokens per second.

Three named partners gave the pitch weight. Cursor, Runway and Magic all said they were exploring Rubin CPX, and Magic cited its 100-million-token context window as a fit for a context GPU.

Why NVIDIA pulled Rubin CPX

At GTC 2026 in March, NVIDIA presented the Vera Rubin platform as seven chips and five rack-scale systems. Rubin CPX wasn't one of them. Tom's Hardware reported that CPX was removed from the roadmap and absent from the keynote slides, with Groq 3 LPX taking center stage.

Buck explained the decision in a press Q&A. Per the Tom's Hardware transcript, he said "we've pulled CPX. It's still a good idea," and added that NVIDIA would be "thinking about CPX more in the next generation." Asked whether CPX could still launch in 2026, he declined to rule it out.

The reason was where the bottleneck moved. Buck described the high-value workload as trillion-parameter models serving long contexts at around 1,000 tokens per second per user. At that speed, decode latency becomes the hard problem, and CPX only sped up prefill.

Groq 3 LPX attacks decode instead. NVIDIA says an LPX rack holds 256 LPUs with 128GB of on-chip SRAM and 640 TB/s of bandwidth, and pairs with Vera Rubin NVL72 through Dynamo. In that design, Rubin GPUs handle large-scale context processing and the LPUs handle fast token generation. The prefill job CPX was built for went back to the Rubin GPU.

The Groq deal raised the stakes. Tom's Hardware reports NVIDIA's $20 billion licensing agreement with Groq, and NVIDIA said in August 2026 that Groq 3 LPX was in full production. Two specialized inference racks in one generation would have split engineering, supply and customer attention. Inside Deep Tech's Groq LPU and Groq 3 LPX full guide covers that rack in depth.

Rubin CPX timeline: announced, pulled and reportedly revived

Source: NVIDIA CPX announcement; NVIDIA CES 2026 release; NVIDIA GTC 2026 release; Tom's Hardware Q&A transcript; NVIDIA LPX production release; Q2 FY2027 call transcript; Kuo's post; NVIDIA Q3 call notice.

One detail is easy to miss. NVIDIA told ComputerBase in March that the program was on ice but not abandoned, and that the idea could return with the Feynman generation at the latest. That fits Buck's comments and keeps the door open without promising anything.

The HBM4 redesign report

Kuo's August 31 report describes a very different chip. As summarized by TechPowerUp and RCRTech, the revived Rubin CPX would carry 168GB of HBM4 instead of 128GB of GDDR7, draw about 2,300 W, and enter production in 1Q27.

• Standalone racks. CPX would move into its own MGX racks of 64, 128, 192 or 256 GPUs, built from 64-GPU modules of eight trays of eight GPUs plus a switch tray.
• Smaller NVLink domain. NVLink would link GPUs only inside a tray, at about 1 to 1.5 TB/s per CPX versus 3.6 TB/s for Rubin.
• Ethernet between trays. Spectrum-6 Ethernet would connect modules, with copper inside a module and optics between modules, per RCRTech.
• One-to-one pairing. Each CPX would pair with a Vera Rubin NVL72 GPU and hand off KV caches over Ethernet RDMA, per TechPowerUp.

The stated reason is workload mix. Kuo's report says context processing and KV-cache building now make up more than 50% of inference workloads, which would justify a dedicated prefill rack even with LPX handling decode. In that picture, a full NVIDIA inference stack has three tiers: CPX for prefill, Rubin GPUs for the main model, and LPX for the fastest decode.

The redesign also undercuts the original pitch. CPX was supposed to avoid HBM. Moving to HBM4 puts it back in the most contested part of the supply chain. Inside Deep Tech's HBM full guide explains why every stack of HBM4 is already spoken for by flagship GPUs.

NVIDIA Rubin Ultra and Kyber: A Full Guide

The rack generation after Vera Rubin: Kyber vertical blades, 600 kW racks, 800 VDC power, HBM4E and the NVL576 naming change.

Inside Deep Tech

Rubin CPX vs Groq 3 LPX vs Rubin GPU

The three NVIDIA inference options each target a different bottleneck. The table compares them on the facts that are public.

Source: NVIDIA CPX announcement; TechPowerUp on the redesign report; NVIDIA GTC 2026 release; NVIDIA LPX production release; NVIDIA Vera Rubin NVL72 page.

For most buyers, the decision is already made by availability. Vera Rubin NVL72 is shipping. LPX is shipping, with Nebius named as the first cloud to deploy it. CPX isn't available in any form.

Who should plan around Rubin CPX, and who shouldn't

Plan around it if you run very long-context inference at scale, such as repository-scale coding agents, long-video search, or document review over millions of tokens, and you already disaggregate prefill and decode with Dynamo or a similar stack. A cheaper prefill tier would change your cost per query more than anyone else's.

Don't wait for it if you need capacity in 2026 or early 2027. Vera Rubin NVL72 already handles prefill, and NVIDIA's own LPX material assigns context work to Rubin GPUs. A purchase plan that depends on CPX depends on a product NVIDIA hasn't re-announced.

Watch it if you design data centers. The reported redesign at about 2,300 W per GPU would push rack power and liquid cooling needs well beyond the original CPX tray, which changes facility plans.

Honest limits

• No official current status. NVIDIA's Q2 FY2027 call transcript has no CPX mention, and the original NVIDIA developer blog post now returns an error; only an archived copy remains.
• The redesign is one analyst's supply-chain check. Every outlet covering it, including TechPowerUp and RCRTech, traces back to Kuo.
• The 2025 economics are vendor math. The $5 billion per $100 million figure assumes token prices and utilization NVIDIA didn't publish in detail.
• Memory bandwidth and power were never official. The 2 TB/s and 800 W figures are SemiAnalysis estimates.
• Fixed ratios are a real risk. SemiAnalysis flagged that a rack with a set prefill-to-decode ratio can waste silicon when workloads shift.

Inside Deep Tech's take

Inside Deep Tech's take: NVIDIA was right about the problem and early on the product. Disaggregated inference is now standard in Dynamo, and NVIDIA's own GTC 2026 platform splits work across specialized racks. CPX lost its slot because decode speed, not prefill cost, became the scarcer resource in 2026.

If the HBM4 redesign is real, it says something important. A context chip with HBM4 isn't a cheap chip anymore, so NVIDIA would be betting that prefill volume, not prefill cost, is the new constraint. That's a plausible bet in a market where agents read far more tokens than they write. It's still a bet NVIDIA hasn't made in public.

Rubin CPX racks still rely on the Vera CPU memory pool. For that layer, see the SOCAMM guide, and for the generation after Rubin, the NVIDIA Feynman guide.

What to watch next

• NVIDIA's Q3 FY2027 call on Tuesday, November 17, 2026: any CPX mention, or a three-tier inference story, would be the first official signal since March.
• Supply-chain reports on HBM4 allocation for 2027, since a 168GB part would compete with Rubin and Rubin Ultra for the same stacks.
• GTC 2027 roadmap slides, where NVIDIA said the CPX idea could return in the Feynman generation.
• Dynamo releases that add first-class support for a third, prefill-only hardware pool.

FAQ

The bottom line

Rubin CPX is the clearest public statement yet that prefill and decode want different chips. The 2025 GDDR7 version is off the roadmap, Groq 3 LPX shipped instead, and an HBM4 redesign is reported but unconfirmed. Until NVIDIA says more, plan around Vera Rubin NVL72 and LPX, and treat CPX as a 2027 option at best.

For what comes after Vera Rubin, including Kyber racks, 800 VDC power and HBM4E, read Inside Deep Tech's NVIDIA Rubin Ultra and Kyber full guide.