Skip to content
Source · Daily Brief

AI Daily Brief — 18 March 2025

The biggest keynote of the AI year so far. Jensen Huang framed AI as a $1 trillion computing inflection point and proceeded to ship — or pre-announce — almost every layer of the stack at once: Blackwell Ultra now, Vera Rubin in 2026, Rubin Ultra and Feynman beyond, plus open humanoid foundation models, an open inference framework, a joint physics engine with DeepMind and Disney, desktop AI computers, silicon-photonics switches, and partnership rollouts across GM, T-Mobile, Yum Brands and Foxconn.

Top stories

  • Blackwell Ultra (B300 / GB300 NVL72) AI factory platform. 15 petaFLOPS dense FP4, 288 GB HBM3e, 8 TB/s memory bandwidth, 1,400 W TDP, NVLink 5, PCIe Gen6. The GB300 NVL72 rack pairs 72 Blackwell Ultra GPUs with 36 Grace CPUs, delivering 1.1 ExaFLOPS FP4 — 1.5x the GB200 NVL72 — and ships in H2 2025. via NVIDIA
  • Vera Rubin roadmap to H2 2026 — HBM4, 88-core Vera CPU. Rubin shifts from HBM3e to HBM4 with 13 TB/s bandwidth at 288 GB capacity. NVLink doubled to 260 TB/s; new CX9 inter-rack link at 28.8 TB/s. Vera CPU is a custom Arm chip with 88 cores / 176 threads and 1.8 TB/s NVLink C2C to the Rubin GPU. via Tom’s Hardware
  • Rubin Ultra NVL576 ‘Kyber Rack’ — 600 kW per rack, H2 2027. 576 Rubin Ultra GPUs across four pods of 18 vertical blades, 15 EFLOPS FP4 inference per rack, 5x the 120 kW of current Blackwell B200 racks. After Rubin: ‘Feynman’. via Data Center Dynamics
  • Isaac GR00T N1 — world’s first open humanoid foundation model. Vision-Language-Action model with dual-system architecture (VLM + diffusion transformer). Trained on 780,000 synthetic trajectories (~6,500 hours of human demos) generated in 11 hours. Early access to Agility Robotics, Boston Dynamics, Mentee Robotics, NEURA. Released openly on GitHub. via NVIDIA
  • Dynamo — open-source distributed inference framework for reasoning models. Disaggregated prefill/decode, LLM-aware request routing, dynamic GPU scheduling, KV-cache transfer. Supports SGLang, TensorRT-LLM and vLLM. NVIDIA claims up to 30x request throughput running DeepSeek-R1 on Blackwell. via NVIDIA
  • Newton — open physics engine from NVIDIA + Google DeepMind + Disney. GPU-accelerated robotics simulation built on NVIDIA’s Warp framework, interoperating with DeepMind’s MuJoCo and NVIDIA Isaac Lab. Disney’s BDX droid ‘Blue’ joined Jensen onstage. Open-source release planned later in 2025. via NVIDIA Developer
  • DGX Spark and DGX Station personal AI computers. DGX Spark (formerly Project DIGITS): 150x150mm, 1.2 kg, GB10 Grace Blackwell Superchip, 1 petaFLOP FP4, 128 GB unified memory. DGX Station: GB300 Grace Blackwell Ultra, runs >1 trillion-parameter models on one desktop. Built with ASUS, Dell, HP, Lenovo. via NVIDIA
  • Spectrum-X and Quantum-X silicon photonics switches. Industry’s first 3D-stacked silicon-photonics CPO switches — Spectrum-X (Ethernet) at 1.6 Tb/s per port, Quantum-X (InfiniBand). NVIDIA claims 3.5x power efficiency, 10x network resilience, 1.3x faster deployment vs traditional pluggable optics — saving ~40 MW per large datacenter. Built on TSMC 3D hybrid bonding. via NVIDIA

Who shipped

The day was almost entirely NVIDIA. Co-stars in the keynote: Google DeepMind and Disney on Newton; GM on next-gen AVs and Halos safety stack; T-Mobile, Cisco, MITRE, ODC and Booz Allen on AI-RAN 6G; Yum! Brands on AI drive-thrus across 500 Taco Bell, KFC, Pizza Hut and Habit locations; Foxconn debuting FoxBrain on the GTC stage.

Open-source pulse

Open releases inside Jensen’s keynote: Isaac GR00T N1 (humanoid foundation model), Dynamo (inference framework), Cosmos Predict and Cosmos Transfer (world models, on Hugging Face), Llama Nemotron Nano/Super/Ultra reasoning models, and the upcoming Newton physics engine — making this NVIDIA’s most open-weights-heavy GTC ever.

Money, infra & hardware

NVIDIA also unveiled the AI Data Platform reference design with DDN, Dell, HPE, Hitachi Vantara and IBM as launch storage partners. Yum’s 500-location drive-thru AI rollout (planned for Q2) is the most concrete enterprise commitment from the day. via CNBC

Quiet corners

Non-NVIDIA frontier-lab news was essentially zero — the entire AI press cycle deferred to the keynote. OpenAI’s o1-pro API came a week later on Mar 25; Anthropic and xAI stayed quiet for the day.

By the numbers

  • 1.1 ExaFLOPS FP4 — GB300 NVL72 rack (1.5x GB200 NVL72)
  • 288 GB HBM3e per Blackwell Ultra GPU; 13 TB/s HBM4 per Rubin
  • 600 kW — Rubin Ultra NVL576 ‘Kyber’ rack power draw
  • 780,000 — synthetic trajectories training Isaac GR00T N1
  • ~40 MW — datacenter power savings from Spectrum-X / Quantum-X photonics
  • 500 — Yum! Brands locations getting NVIDIA-powered drive-thru AI
  • Most-mentioned company: NVIDIA

Compiled by AI Feed’s editor from verified web sources for 18 March 2025.