← Tech
💻Tech

NVIDIA GTC Berlin 2026 Keynote Recap: Rubin, DSX POD, OpenShell and RTX Spark

GTC Berlin (Oct 21, 2026) recap: Vera Rubin in production, DSX POD, OpenShell agent sandbox, open models, RTX Spark tests and what you can try today.

T Terapep ·

TL;DR: The most recent NVIDIA GTC keynote is GTC Berlin, October 21, 2026, at the Tempodrom. Jensen Huang didn't announce a new GeForce GPU or a new DLSS version. The keynote was about AI factories and agents. The headliners were Vera Rubin "in full production", the containerized DSX POD (NVIDIA says it can be deployed in as little as 14 days), the OpenShell reference architecture for sandboxed AI agents, and a large batch of open models: Nemotron, Cosmos, GR00T, Alpamayo and Kumo. For PC users, the news is local AI on RTX Spark laptops and DGX Spark/Station desktops, not gaming hardware. (NVIDIA keynote recap)

Which keynote this is

GTC Berlin was the last stop on Huang's autumn Europe tour, after Paris, London, Stockholm and Dublin. It is the most recent GTC keynote as of November 2, 2026. The next one is GTC DC on December 1. NVIDIA used the event to pitch "sovereign AI" to Europe. It says it is committing nearly $10 billion to the European ecosystem and has 12,000+ partners there (vendor claims, NVIDIA). AP's coverage summed up the speech as Huang pushing Europe's AI industry to move faster, with modular AI data centers and a "sandbox" for agents.

How to read the labels: "Vendor claim" means NVIDIA or a partner said it, and nobody has tested it independently. "Independent" means a third-party reviewer measured it. Some items below were released on GTC Berlin day but were not shown on stage. Those are marked (GTC-week release).

The key announcements

1. Data center: Vera Rubin ships, and DSX POD packages it into containers

  • Vera Rubin is "in full production", according to Huang on stage. The Rubin/Feynman roadmap itself was laid out at GTC San Jose in March. Berlin did not change it.
  • DSX POD is a reference design for a complete Vera Rubin "AI factory." Partners build it inside modular containers. NVIDIA claims deployment at a prepared site drops from six months to as little as 14 days, with partner deliveries worldwide in early 2027 (vendor claims). Partners include Dell, HPE, Lenovo, Foxconn (Ingrasys), GIGABYTE, QCT, Wistron, Pegatron, Schneider Electric, Vertiv, Eaton, Delta, Bull, Crusoe, DayOne, Armada and Anduril (NVIDIA Newsroom).
  • The DSX AI Factory Platform comes with two efficiency claims: up to 100% less cooling water, and the same AI output on up to 30% less power (vendor claims).
  • Google Cloud A5X (GTC-week release) runs on Vera Rubin NVL72 racks. General availability is planned for November 2026, with Europe to follow. Google says its network can scale to 960,000 Rubin GPUs across sites. The first customer is Ineffable Intelligence, David Silver's lab (Google Cloud, NVIDIA).
  • Smaller GTC-week releases: Confidential Bare Metal, cuDSS reaching general availability, the Factory Operations Blueprint (FOX) in early access on GitHub, and Minerva AIOps.

2. Agents: OpenShell, Agent Toolkit and Kumo

  • The OpenShell Reference Architecture for Safe AI Agents is built on OpenShell. OpenShell is NVIDIA's open-source (Apache-2.0) runtime. It runs agents in kernel-level sandboxes and controls their filesystem, process, network and credential access through policies. This is the "sandbox" in AP's write-up.
  • The Agent Knowledge Foundation expands the NVIDIA Agent Toolkit with Nemotron parse, embed and rerank tools, plus new NVIDIA Kumo models for structured (tabular) data.
  • Independent: The Register reported the new Nemotron 3.5 retrieval models: 1B and 8B multimodal embedding models, an 8B reranker, 34 languages, and an NVFP4 version of the 1B embedder. The accuracy figures it quoted came from NVIDIA (The Register).

3. Open models

The keynote covered six model families: Nemotron, Cosmos, GR00T, Alpamayo, Ising and Earth-2. New in the GTC week were Cosmos 3.5, Alpamayo 2 Edge (for DRIVE AGX Thor), RFD4-Proteina (BioNeMo) and new Ising releases (NVIDIA).

Nemotron 3.5 Super VL (120B) and Bolt 74B-A9B were showcased, but both actually shipped on October 15, six days before the keynote. NVIDIA calls Super VL "5x faster" than the next-best model in its class (vendor claim).

4. Telecom, healthcare, cars and robots

  • Telecom: Huang highlighted the first commercial phone call on T-Mobile's network with Nokia ("CUDA makes phone calls"). Huang framed cell sites as distributed data centers: about 100,000 sites and about 100 GW (his figures).
  • Healthcare: Siemens Healthineers is working on AI-enabled adaptive radiotherapy.
  • Cars: Volvo Cars is adopting DRIVE Hyperion and the Alpamayo stack. NVIDIA says next-generation vehicles from Mercedes-Benz, JLR, Volvo and Volkswagen's vans unit will be Hyperion-compatible. Bolt (with Lucid), Mercedes-Benz, VW Commercial Vehicles and Einride showed L3/L4 and robotaxi use cases.
  • Robots: Agile ONE SN is a reference humanoid built on Isaac GR00T, with 31 degrees of freedom, expected next year. NVIDIA says it will cut researchers' setup time "from weeks to days" (vendor claim). Agility, Figure, KUKA and NEURA build on the same platform.

5. Local AI on the desk and in a laptop

  • RTX Spark is NVIDIA's Arm "superchip" laptop platform. It launched on October 16, and Huang featured it on stage: "1 PFLOP" and 128GB in a laptop, compared with the 2016 DGX-1 (vendor claim). His "Microduck" robot demo used Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt.
  • DGX Spark 64GB became available on Oct 23. NVIDIA Sync can now pool up to 8 DGX Sparks (about 1TB unified memory), and NVIDIA said a Model Launcher would follow later in October (GTC-week release).
  • DGX Station offers 748GB of coherent memory and "up to 20 PFLOPS" (vendor). It gets 7 new IT-management integrations, early access to Cloud Functions, and a Windows version in December (NVIDIA local AI blog).

Spec and roadmap table

Item Key spec (source type) When
Vera Rubin (NVL72) "In full production" (vendor, on stage) Now
DSX POD Full Rubin AI factory in containers; "6 months to as little as 14 days" (vendor) Partner deliveries early 2027
Google Cloud A5X Vera Rubin NVL72; up to 960,000 Rubin GPUs across sites (Google) GA November 2026; Europe planned
OpenShell reference architecture Open-source sandbox runtime, Apache-2.0 Available now (GitHub)
RTX Spark (N1X) 6,144 CUDA cores, 20-core Arm CPU, 300 GB/s, 45–80W (PC Gamer spec table); up to 128GB, "1 PFLOP" (vendor) Shipping since Oct 16
DGX Spark 64GB 64GB unified memory (vendor) Oct 23, 2026
DGX Spark cluster via Sync Up to 8 units, ~1TB (vendor) Now; Model Launcher to follow
DGX Station 748GB coherent, up to 20 PFLOPS (vendor) Now; Windows version December
Agile ONE SN humanoid GR00T-based, 31 DoF (vendor) 2027
Nemotron 3.5 Bolt 74B-A9B 74B total / 9B active, native 2-bit, fits 24GB (model card) Released Oct 15
NVIDIA local AI memory ladder: DGX Spark 64GB, RTX Spark 128GB, DGX Station 748GB, 8x DGX Spark 1TB (vendor claims)DSX POD deployment time claim: about 6 months vs as little as 14 days (vendor claim)

What it means for PC gamers

No new GeForce card, no DLSS 5.x update and no driver feature came out of this keynote. If you were waiting for an RTX 50 Super or an RTX 60 tease, Berlin wasn't it. NVIDIA's own keynote recap lists none. Recent gaming news came through GeForce News instead. See our DLSS 5 supported games list and the GeForce hotfix 617.70 guide.

The gaming-relevant product on stage was RTX Spark. NVIDIA's gaming claim is "1440p at 100 fps" (vendor). Independent testing is more mixed:

  • Micro Center (independent), Surface Laptop Ultra with N1X, Cyberpunk 2077, Ultra preset at 1440p: 48.6 fps native, 65.3 fps with DLSS Balanced, and 151.2 fps with DLSS + 4x frame generation. With RT Ultra at 1440p, DLSS Balanced gives 37.6 fps (Micro Center). The 100 fps-class figures depend on Multi Frame Generation.
  • PC Gamer (independent), ASUS ProArt P16: RTX Spark ran close to an RTX 5070 laptop in Black Myth: Wukong. It fell well behind in Baldur's Gate 3 (42 vs 65–69 fps). It drew about 43W of GPU power against 77–97W, but lasted only 86 minutes on battery while gaming, versus 131–134. EA Javelin anti-cheat titles such as Battlefield and FC 27 don't run yet. PC Gamer's verdict: impressive, but not one to buy purely for gaming (PC Gamer).
PC Gamer independent 1080p FPS: RTX Spark N1X vs RTX 5070 laptops

Pricing: NVIDIA says RTX Spark laptops start at $2,399. PC Gamer put the cheapest one it could find at about $2,700. Check the retailer before you trust either number. If you mainly game and want to try DLSS 4.5 frame generation on an existing RTX 50 card, our Cyberpunk 2077 Dynamic MFG guide is more useful than anything from this keynote.

What it means for AI developers

  • Agents are moving into sandboxes. OpenShell plus the reference architecture is NVIDIA's answer to the question of how to let an agent run code without handing it the whole machine. It's open source and runs on Docker or Podman today.
  • Local models are getting large enough to be useful. Nemotron 3.5 Bolt runs a 74B-parameter model in about 24GB. With 128GB of RTX Spark memory or a 1TB DGX Spark cluster, frontier-adjacent models can run on a desk. Independent: Micro Center measured 35.2 tokens/s for Qwen3.5 122B-A10B and 55.6 tokens/s for GLM 4.7 Flash on RTX Spark.
  • Inference speedups (GTC-week releases, vendor claims): llama.cpp MTP speculative decoding is up to 1.4x faster on RTX PRO 6000 (Qwen3.8 Flash Next). vLLM is up to 1.3x faster on DGX Spark (Qwen3.8 27B) and 1.2x faster on DGX Station (GLM 5.3 Flash). NVIDIA kernels make Recursive Flint, a 40B MoE with 2.5B active parameters under Apache 2.0, generate "+55%" faster on RTX Spark.
  • Retrieval/RAG: the Nemotron 3.5 embed, rerank and parse models plus Kumo for tables are open weights. They are worth benchmarking against your current embedder. See our LLM leaderboards roundup for context.
  • The cloud gets Rubin first. Google Cloud A5X general availability in November is the first realistic way to rent Vera Rubin. DSX POD matters to whoever builds your provider's next region, not to you directly.

What to do now

  1. Run Nemotron 3.5 Bolt locally.
    • Ollama: ollama run nemotron-3.5-bolt:74b-a9b. If you're on a smaller GPU or a Mac, use ollama run nemotron-3.5-bolt:e4b instead (an 8B model with 4B effective). The 74B model isn't supported on Apple Silicon.
    • llama.cpp: download nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF with hf download, then run llama-server with the DFlash drafter. NVIDIA's model card recommends 32GB+ of VRAM (the 28GB recipe) or 48GB+ (full 256K context) (Hugging Face).
    • Not sure which app to use? Read Ollama vs LM Studio.
  2. Sandbox an agent with OpenShell. On Linux or macOS, run curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh, then openshell status and openshell sandbox create --name demo. You need Docker Engine 28+ or Podman 5 with cgroups v2 (docs).
  3. Try Recursive Flint from Hugging Face (Apache 2.0) in llama.cpp, LM Studio or vLLM. Its 2.5B active parameters make it fast even on mid-range RTX cards.
  4. Generate images with FLUX.3 Klein. The 5.6B model runs in 8GB of VRAM. NVIDIA claims about 4 seconds per 1-megapixel image on an RTX 5090 (vendor).
  5. Update llama.cpp and vLLM to current builds to pick up the MTP and speculative-decoding paths.
  6. For RAG stacks, pull the Nemotron 3.5 embed and rerank models from build.nvidia.com or Hugging Face.
  7. For factory, ops and HPC work, request FOX early access on GitHub. cuDSS (the sparse direct solver) is now generally available in the CUDA math libraries.
  8. If you own DGX Sparks, update NVIDIA Sync to pool them. Wait for DGX Station for Windows in December if you're Windows-only.

Every speedup above is NVIDIA's own number until someone independent reproduces it on your model and your hardware.

Sources

#nvidia#gtc#jensen huang#vera rubin#dsx pod#openshell#nemotron#rtx spark#dgx spark#local ai
← Back to all posts