NVIDIA GTC Berlin 2026 announcements

The headline product at GTC Berlin was not a chip - it was a shipping container. DSX POD packages seven liquid-cooled Vera Rubin NVL72 racks, roughly 17 kilometers of Spectrum-X cabling, and all associated networking, cooling, and management hardware into a single factory-tested, transportable unit delivering about 25 EFLOPs of AI inference ^[2]. NVIDIA says that packaging shift - moving integration, cabling, and liquid-cooling work off-site into factory manufacturing - can compress deployment at a prepared site from roughly six months to as little as 14 days ^[1]. An early proof-of-concept from Lambda, using DSX MaxLPS on NVIDIA HGX B200 systems, reported 24% more cluster token throughput and up to 30% less power for the same token output, numbers NVIDIA is using to sell the economics of speed, not just raw compute ^[1]. Worldwide container delivery is slated for early 2027, with a partner roster - Crusoe, Dell Technologies, HPE, Lenovo, Foxconn's Ingrasys, and more - that reads like NVIDIA assembling its own supply chain for prefabricated data centers rather than leaving that work to hyperscalers alone ^[1]. Crusoe is folding the Vera Rubin DSX Reference Design and Omniverse DSX Blueprint directly into its own gigawatt-scale AI-factory buildout ^[3]. Taken together, DSX POD reads less like a product launch than a statement: NVIDIA wants to be judged on how fast an AI factory goes from empty lot to revenue-generating tokens, not just on FLOPs per chip.

NVIDIA's other big bet at GTC Berlin was that autonomous AI agents need to be policed from outside their own process, not trusted to report on themselves. The OpenShell Reference Architecture treats every agent workload as untrusted throughout execution, granting access to data and tools only when needed ^[4]. Enforcement runs on NVIDIA Sentry, a watchdog on BlueField-4 DPUs that enforces policy in silicon and can quarantine a misbehaving agent within milliseconds of detecting a threat or violation ^[4]. That hardware-first posture matches Jensen Huang's own framing that the only safe starting point for an autonomous agent is zero trust, and that AI safety itself needs to be accelerated rather than bolted on later. The architecture runs across Linux and Kubernetes distributions from Red Hat, SUSE, and Canonical, with Red Hat Enterprise Linux now supporting Vera CPUs and the Vera Rubin platform ^[5], and SUSE positioning its own integration as bringing sovereign control to enterprise agentic AI ^[6]. The adoption list is the real signal: agents from eight separate model builders - including Anthropic, Google, Meta, OpenAI, and Perplexity - can run inside OpenShell, ServiceNow and SAP are building on it, and banks including BBVA and Santander are reportedly evaluating it ^[7]. ServiceNow didn't wait to ship - it took its Arc personal enterprise agent to general availability running inside the OpenShell runtime and governed through its own Super Intelligence Control Tower, with BlueField-4 policy revocation as the enforcement backstop ^[8]. If that adoption holds, NVIDIA has positioned itself as the default trust layer underneath a wave of enterprise agents it does not itself build.

Huang's pitch to European businesses was expansive: every company in Europe will become an AI company, built on a five-layer stack of energy, compute, AI factories, models, and applications, backed by close to 10 billion dollars NVIDIA has committed to the European ecosystem and more than 12,000 regional partners ^[9]. The framing drew pushback from multiple directions at once. Forrester analyst Paul Miller argued that anything with critical ongoing dependencies on NVIDIA - a non-European firm - is unlikely to meet the purest definitions of European tech sovereignty, cutting against the keynote's own branding ^[10]. Protesters interrupted the keynote itself over allegations that NVIDIA chips have turned up in Russian drones, a reminder that the company's global hardware footprint carries dual-use and reputational risk a sovereignty narrative does not resolve ^[11]. That skepticism echoed outside the venue too: community discussion elsewhere reported that NVIDIA had denied several neoclouds access to GTC Berlin after they added rival AMD and TPU chips to their cloud offerings, and a separate thread discussed Huang's own blunt public criticism of Germany's technology standing. None of this derailed the keynote's reception among investors - analyst Gene Munster called NVIDIA still underappreciated and pointed to DSX POD's deployment speed as a real return-on-investment story ^[12]- but it shows the sovereign-AI narrative is being contested in real time, not accepted at face value.

Beneath the keynote's biggest headlines, NVIDIA quietly expanded the retrieval infrastructure that agentic systems actually run on. The Nemotron 3.5 family adds open, commercially available Embed, Rerank, and Parse models aimed at agentic document, chart, and table retrieval ^[13]. Nemotron 3.5 Embed 8B BF16 posted 61.17 nDCG@10 on the ViDoRe v3.1 benchmark, ranking first among dense retrieval models as of October 20, 2026, while a distilled 1B version retained 95.4% of that score - a meaningful efficiency trade for teams that cannot run an 8B model at inference scale ^[13]. Nemotron 3.5 Rerank is pitched on cost rather than raw accuracy, cutting agentic token cost per query by roughly 14%, and Nemotron 3.5 Parse processes pages at up to 16.1 pages per second on a single H100 while still ranking first among sub-1B-parameter parsing models ^[13]. The more consumer-facing half of the family is Nemotron 3.5 Super VL, a multimodal model that combines text and visual understanding so agents can extract and act on information from webpages, PDFs, tables, and charts ^[15], and its compressed sibling Nemotron 3.5 Bolt, a mixture-of-experts hybrid Mamba-Transformer model with 74B total and 9B active parameters distilled down from the 120B-parameter Super VL to fit in 24GB for local agentic use on RTX and DGX Spark hardware ^[14]. Together these are less flashy than DSX POD or OpenShell, but they are the models that make NVIDIA's agent-security and agent-infrastructure story actually runnable.

The quieter story at GTC Berlin was how fast NVIDIA's cloud and telecom partners are locking in Vera Rubin. Google Cloud brought its first Vera Rubin NVL72-powered A5X instances live for an initial customer on the same day as the keynote, with general availability planned for November 2026 and a roadmap to scale toward 960,000 Rubin GPUs across multiple sites ^[16]. The first named customer is Ineffable Intelligence, the London-based reinforcement-learning lab founded by AlphaGo creator David Silver, who said Vera Rubin's compute performance and connectivity let his team push the scale of its research further ^[16]. On the telecom side, NVIDIA and Nokia, working with T-Mobile, completed what they describe as the world's first AI-RAN call on a live commercial network, roughly eight years into the companies' collaboration ^[17]. Nearly 20 telecom operators globally have now committed to AI-RAN work with NVIDIA and Nokia a year after NVIDIA's one-billion-dollar investment in Nokia, with Nokia integrating Nemotron into its own Autonomous Networks Suite ^[17]. Neither announcement generated the stage time of DSX POD or OpenShell, but both show the same pattern: NVIDIA's partners are being wired into Rubin-era infrastructure well before most enterprises have even seen the hardware.