NVIDIA unveils Rubin CPX (built for million-token inference)

NVIDIA unveils Rubin CPX (built for million-token inference)

NVIDIA announced Rubin CPX, a new class of GPU optimized for long-context AI like full-codebase reasoning and hour-long video understanding/generation. It slots into the new Vera Rubin NVL144 CPX platform aimed at massive-context inference at data-center scale.

Key specs & platform
• Up to 30 PFLOPS NVFP4 per GPU, 128GB GDDR7, and 3× faster attention vs GB300 NVL72.
NVL144 CPX rack: 8 exaflops of AI, 100TB fast memory, 1.7 PB/s bandwidth, and ~7.5× the AI performance of GB300 NVL72 systems.
• Integrated video decode/encode + long-context inference on one chip; works with Quantum-X800 InfiniBand or Spectrum-X Ethernet (Spectrum-XGS, ConnectX-9).

Ecosystem & software
• Backed by the full NVIDIA stack: Dynamo for scaled inference throughput, Nemotron multimodal models, AI Enterprise with NIM microservices, and CUDA-X libraries.

Early partners
Cursor (coding agents), Runway (generative video), and Magic (agentic software engineering) are exploring Rubin CPX for long-context workloads.

Availability
• Expected end of 2026. NVIDIA also floats an aggressive ROI claim tied to “token revenue” at scale; treat as forward-looking.


Other News

Anthropic Claude can now create & edit files.
GitHub OpenAI GPT-5 and GPT-5 mini are now generally available in Copilot.
OpenAI ChatGPT Voice update (Release Notes).
Adobe Premiere Pro 25.5 adds 90+ new effects, transitions, and animations.
NVIDIA Blackwell Ultra sets new MLPerf inference records.
NVIDIA Rubin CPX GPU designed for massive-context inference.
AWS Managed tiered checkpointing on Amazon SageMaker HyperPod.
GitHub New JFrog integration for secure, traceable builds.


AI Tools of the Day
Murf AI — Create AI voiceovers for videos, ads, and training.
Beeyond AI — 50+ AI tools for content, art, and marketing.
HeadshotPro — Generate LinkedIn-ready headshots from a few selfies.
SEOpital — AI SEO writer that researches, outlines, and drafts ranking-focused content.
guidde — Record once; get step-by-step video docs and how-tos auto-generated.