Skip to content
Source · Daily Brief

AI Daily Brief — 19 March 2025

GTC Day 3 turned into NVIDIA partner day. The lab unspooled the open Nemotron reasoning family, Dynamo as the inference substrate, Spectrum-X / Quantum-X co-packaged optics, and a Boston quantum research center — while Oracle, Microsoft, Cisco, HPE and IBM all locked in deeper enterprise stacks. OpenAI counter-programmed with o1-pro at a record API price tag.

Top stories

  • OpenAI ships o1-pro in the API at $150/M input + $600/M output. Its most expensive model ever — 2x GPT-4.5 input cost and 10x regular o1 output. Limited initially to API customers who’d spent at least $5; supports vision, function calling, structured outputs, Batch API. Uses more compute per query than o1 for harder problems. via TechCrunch
  • NVIDIA Llama Nemotron open reasoning models go live. Nano (edge), Super (single-GPU) and Ultra (multi-GPU) tiers with toggleable reasoning. Post-training reportedly lifted accuracy up to 20% over base Llama and delivered ~5x faster inference than other leading open reasoning models. Released on build.nvidia.com and Hugging Face — NVIDIA’s reasoning answer to DeepSeek-R1. via NVIDIA Developer
  • NVIDIA Accelerated Quantum Research Center (NVAQC) — Boston. NVIDIA committed to a Boston-based research center integrating quantum hardware with AI supercomputing, with a GB200 NVL72-based system of 576 Blackwell GPUs over Quantum-2 InfiniBand. Partners: Harvard (HQI), MIT (EQuS), Quantinuum, QuEra and Quantum Machines. Notable because Jensen had said weeks earlier that useful quantum was decades away. via NVIDIA
  • Oracle + NVIDIA — 160+ AI tools and 100+ NIM microservices into OCI Console. No-code AI Blueprints deployment, vector-search acceleration in Oracle Database 23ai via cuVS, and NVIDIA AI Enterprise availability across OCI public regions, Government/sovereign clouds, OCI Dedicated Region, Alloy, Compute Cloud@Customer and Roving Edge. via NVIDIA
  • GM, Gatik and Torc deepen NVIDIA AV partnerships. GM plans NVIDIA-powered ADAS in consumer cars (not robotaxis) plus Omniverse + Cosmos for factory planning and robotics simulation. Gatik (autonomous trucking) and PACCAR-owned Torc Robotics deepen use of DRIVE AGX/Thor. via TechCrunch
  • Cisco Secure AI Factory with NVIDIA. Cisco announced a Secure AI Factory architecture co-designed with NVIDIA — Cisco Hypershield workload protection, Cisco AI Defense for model/app security, Spectrum-X Ethernet networking, with storage from Pure, Hitachi Vantara, NetApp and VAST Data. via Cisco

Who shipped

NVIDIA ran the entire enterprise stack — open reasoning models, inference framework, optics switches, quantum lab — under the GTC banner. OpenAI chose this day to put a $600/M output price tag on its frontier-tier reasoning, sharpening the price-performance argument. Microsoft, Oracle, Cisco, HPE and IBM announced deeper integrations across the day.

Open-source pulse

Llama Nemotron (open weights, on build.nvidia.com + Hugging Face) and Dynamo (open inference framework supporting SGLang, TensorRT-LLM, vLLM) made NVIDIA’s open-weights footprint as wide as it’s ever been.

Money, infra & hardware

Microsoft Azure made Container Apps serverless GPUs with NVIDIA NIM generally available (per-second billing, optimized cold start), and added Nemotron and Cosmos to Azure AI Foundry. HPE rolled out Blackwell Ultra servers and a Private Cloud AI developer kit.

Quiet corners

No major China-labs release dated Mar 19 — Tencent’s Hunyuan T1 was still two days out. No notable arxiv landmark.

By the numbers

  • $150 / $600 per million input/output tokens — o1-pro pricing
  • 20% — accuracy lift Llama Nemotron over base Llama
  • 5x — faster inference vs other open reasoning models
  • 576 — Blackwell GPUs in the NVAQC quantum supercomputer
  • 160+ / 100+ — NVIDIA AI tools / NIM microservices into Oracle OCI
  • Most-mentioned company: NVIDIA

Compiled by AI Feed’s editor from verified web sources for 19 March 2025.