AI News Roundup: NVIDIA Rewired the Whole Stack in a Day
From chips to open models to a kernel sandbox, GTC Berlin was a vertical integration masterclass.
The October 21 AI news cycle belonged to NVIDIA. At GTC Berlin, Jensen Huang put the Vera Rubin platform into full production, showed a DSX POD that drops an AI factory into modular containers deployable in about 14 days, and wrapped agents in OpenShell, a kernel-level sandbox with least-privilege access. A reference humanoid, Agile ONE SN, rounded out the robotics pitch.
On the model side, NVIDIA expanded its open family the same day: Nemotron 3.5 Super VL and Bolt for language, Nemotron Retriever for search, Cosmos 3.5 world models, Alpamayo 2 Edge for in-vehicle reasoning, and BioNeMo RFD4-Proteina for protein design, built with the Baker Lab and reported about 13x faster than its predecessor on a 200-amino-acid benchmark. Recursive Flint, a 40B open coding model, runs above 220 tokens per second on local RTX Spark hardware.
Local hardware got a refresh too: eight DGX Sparks clustered give 1TB of unified memory for frontier-class models, and the 64GB DGX Spark starts shipping October 23. NVIDIA also committed nearly 10 billion dollars across Europe and expanded substrate capacity with Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron -- a 2.5x rise over three years.
The throughline is vertical integration: NVIDIA is selling the model, the sandbox to run it safely, the chip to run it locally, and the factory to build it at scale. Competitors are racing one layer at a time.
Updated October 28, 2026: hardware ship dates below were confirmed at announcement; the 64GB DGX Spark started shipping October 23 as planned.
Frequently asked questions:
Q: What did NVIDIA announce at GTC Berlin?
A: Vera Rubin platform full production, a DSX POD that deploys an AI factory in modular containers in about 14 days, OpenShell kernel-level agent sandboxing, and a reference humanoid called Agile ONE SN.
Q: Which open models did NVIDIA release?
A: Nemotron 3.5 Super VL and Bolt, Nemotron Retriever, Cosmos 3.5 world models, Alpamayo 2 Edge, BioNeMo RFD4-Proteina, and Recursive Flint - a 40B open coding model running above 220 tokens per second on RTX Spark.
Q: What about local hardware?
A: Eight clustered DGX Sparks give 1TB of unified memory for frontier-class models, and the 64GB DGX Spark began shipping October 23.