Rubin CPX by NVIDIA: status and timeline

Milestones

Next · Vera Rubin NVL144 CPX system production

Rubin CPX announcedOct 2026Complete.
Vera Rubin NVL144 CPX system productionTarget Q4 2026Not yet reached.

Most important, last 3 months

NVIDIA unveils Rubin CPX, a GPU optimized for million-token contextsOct 2026

Upcoming

Q4 2026Vera Rubin NVL144 CPX system production (next)

Current obstacles

Integration with decodingDecoding still requires high-throughput GPUs like Blackwell. Scheduling prefill (Rubin) and decode (Blackwell or LPU) across multiple GPU types adds complexity.

Physics limits

Memory latency vs parallelism trade-offGDDR7 per-access latency is higher than HBM, but parallel fetches across a large context matrix hide this latency.

How it works

3 parts

GDDR7 memory

Cheaper than HBM, fast enough for prefill

GDDR7 is lower cost and faster to produce at scale. Suitable for prefill where throughput matters more than per-access latency.

Attention optimization

Kernels for long sequences

Custom attention algorithms handle million-token contexts. Achieves 3x speedup on attention over prior generation.

NVFP4 precision

4-bit floating-point

NVIDIA's 4-bit float format reduces memory bandwidth and power. Suitable for prefill where numerical precision is less critical than throughput.

Spec sheet

Spec	Rubin CPX
NVFP4 compute per GPU	30 petaflops^R (reported)
Memory per GPU	128 GB GDDR7^R (reported)
Attention speedup vs prior generation	3x faster^R (reported)
Vera Rubin NVL144 CPX per-rack	8 exaflops^R (reported)
NVL144 CPX memory bandwidth	1.7 PB/s^R (reported)

^R reported by the company

Update log

1 update

Oct 2026

Major: NVIDIA unveils Rubin CPX, a GPU optimized for million-token contextsPressHardware

About NVIDIA

The team behind Rubin CPX

NVIDIA

AI GPUs, rack systems, CUDA, networking

NVIDIA designs the GPUs, networking and software behind most AI training and inference. Its CUDA software made GPUs the default for AI. In October 2025 it became the first company worth $5 trillion.

Vera Rubin in production in 2026; cloud partners get it from the second half of the year.
Licensed Groq's inference technology and hired its leaders for about $20B (Dec 2025).
Plans to invest up to $100B in OpenAI, tied to at least 10 GW of NVIDIA systems (Sep 2025).

nvidia.com (opens nvidia.com)Company profile →

Founded: 199333 yrs
Headquarters: United States
Status: Public
Listed: NVDA (opens google.com)NASDAQ
Valuation: $5.4Tmarket capSep 2026
Staff: ~36,000Jan 2025
Works in: ComputingAI accelerators & CPUsAIRobotics
Coverage: 10 programs · 30 updateslatest 2d agochecked 25 Sep
Customers: OpenAICoreWeave
Partners: IntelGroq
Suppliers: TSMCSK hynix
People: Jensen Huang (opens en.wikipedia.org)Co-founder, President and CEOColette Kress (opens en.wikipedia.org)CFOBill Dally (opens en.wikipedia.org)Chief Scientist

Spec

Rubin CPX

NVFP4 compute per GPU

30 petaflops^R (reported)

Memory per GPU

128 GB GDDR7^R (reported)

Attention speedup vs prior generation

3x faster^R (reported)

Vera Rubin NVL144 CPX per-rack

8 exaflops^R (reported)

NVL144 CPX memory bandwidth

1.7 PB/s^R (reported)