ConferencesAI agentsAI hardwareLocal LLM

GTC Berlin 2026: open models, sandboxed agents and AI factories in a box

What I took away from NVIDIA GTC Berlin 2026: more interest in open models on your own hardware, OpenShell sandboxes for agents, full agent platforms like SUSE AI Factory, and hardware from 128GB laptops to Vera Rubin servers that need liquid cooling and almost twice the power of a B200. I end with Revolut's foundation model for banking events, and with analysts who say power, not chips, is Europe's real limit.

Siim Tolk · 4 November 2026 · 17 min read

GTC started for me with a full-day hands-on workshop, Deploying and Optimizing AI Inference at Scale. We worked on a lab Kubernetes cluster with 4x A100 GPUs, NIM, an Envoy AI Gateway and Dynamo. It was the most useful part of GTC for me, so it has its own article, a day at NVIDIA’s inference-at-scale workshop.

NVIDIA GTC Berlin ran 20-22 October 2026, with Jensen Huang’s keynote at the Tempodrom on 21 October. That was two weeks ago. I only got to writing it up now, but the conclusions have not shifted since. I took away four things:

  1. Growing interest in open-weight models on your own hardware.
  2. OpenShell for agent isolation.
  3. Full agent platforms.
  4. New hardware at every size, down to laptops. Huang held up an RTX Spark with 1 petaflop and 128GB of memory, the same as the DGX-1 he delivered to OpenAI in 2016, but in one hand (NVIDIA keynote blog).

Analysts and other attendees mostly agree, with two additions. Power is the real limit in Europe, and “sovereign” means less than the stage made it sound.

The conference at a glance

Workshops ran on 20 October, and talks and exhibits on 21-22 October at STATION-Berlin. The keynote was at the Tempodrom (NVIDIA). There were 153 sessions and six full-day workshops. Infrastructure had by far the most sessions. NVIDIA has not given an attendance figure, but pre-event listings expected about 10,000 (Baxtel), and both venues were packed.

Three hours before the keynote I walked past two queues, each about 500 metres long, and gave up on getting in. Instead, I followed the keynote comfortably from one of the STATION halls, as did, allegedly, about 22M people online.

Physical AI and AI for science got a lot of stage time. On stage were the Agile ONE SN humanoid and Hyperion in Volvo, Volkswagen, JLR and Mercedes-Benz cars. Robots will move into warehouses and factories as fast as certification allows. In AI for science, Lila Sciences showed agents designing a possible lymphoma therapy, and there was work on radiation therapy with Siemens Healthineers. Labs will have to prove their models are validated (NVIDIA keynote blog, Mind & Machine Weekly). Both were impressive, but they matter less for my undertakings, so below I focus on the four themes above, and end with one talk, Revolut’s.

1. Open models on your own hardware

A growing number of organisations want to run open-weight models on hardware they own, not only through someone else’s API. Privacy comes first, because the model may write most of your product’s code or clean all your data. Cost comes next, once the volume is high. A model you host also stays under your control. No one changes or retires it under you, and you can fine-tune it on your own data. The conference pointed the same way:

  • NVIDIA pitched open models (Nemotron, Cosmos, GR00T and others) fine-tuned on your own data as the route to what Huang called “sovereign intelligence”, plus a new Agent Knowledge Foundation for connecting agents to documents and databases (NVIDIA keynote blog).
  • New models: Nemotron 3.5 Super VL and Bolt, and Nemotron Retriever models for search (NVIDIA).
  • But the schedule tells the real story. Only 3 of the 94 talks and panels were in the Open Models track, and 29 were on AI infrastructure. The inference talks were about cost per token. NVIDIA spoke of “useful AI output per megawatt”, Vertiv of tokens per second per watt and per euro, and NVIDIA’s “Data Center Under Your Desk” of routing between a model on your desk and one in the cloud (semihub guide).
  • The courses say the same. There was a full-day workshop on serving models at scale, expert sessions on post-training open models on your own data, and a course on distilling, quantising and pruning a model to cut inference cost (NVIDIA training, Connect With Experts).
  • The pushback is that interest is not yet use. In Futurum’s September 2026 survey (n=511), 6.7% use Nemotron, against 74.2% for Claude and 67.5% for GPT; planned Nemotron adoption rose from 13.8% to 18.0% since March (Futurum).

Where is it heading? I do not think serving explains the gap in Futurum’s numbers. Serving models is becoming a well-understood problem. The best open-weight models trail the flagship closed ones by only about four months on average (Epoch AI), and the serving software (vLLM, SGLang, TensorRT-LLM, NVIDIA’s Dynamo) is mature. The open questions have moved on, to agent safety and to how easily agents connect to your data.

So this is how I read the gap between interest and use. In workflows where safety and privacy matter little, Claude and GPT were easy to adopt, and organisations have adopted them quickly. The core workflows, with proprietary data, are different. There you have to solve many problems besides serving, and that just takes time.

One of those problems is how much to let an agent touch. To do real work in your organisation, it needs your files, tools and data. Cut it off from them and it is useless. Give it everything and one mistake can do a lot of damage. So you want an agent with enough access to be useful, and a small blast radius if it misbehaves. That is the problem OpenShell, in the next section, tries to solve.

2. OpenShell: agents and agent fleets in a sandbox

The common thread in the agent talks was orchestration and isolation, not what agents can do. NVIDIA’s answer is OpenShell. It is an open-source (Apache 2.0) runtime that puts each agent in its own kernel-isolated sandbox between the agent and its files, credentials, tools and network. It first appeared at GTC in March. In Berlin NVIDIA added the OpenShell Reference Architecture. It uses Vera Rubin for inference, Vera CPUs for the agent harness, and Sentry on BlueField-4 DPUs to watch agents and stop them “within milliseconds” (NVIDIA keynote blog, NAND Research).

How it works, in short. You write a policy that says which files, processes and network endpoints the agent may use. OpenShell blocks everything no rule allows. The rules are enforced outside the agent, so they hold even when the agent ignores its instructions. Linux kernel controls limit files and processes, and a policy proxy decides which network requests go out. Every allow and block decision goes into an audit log. Before a policy is applied, a policy prover checks what the agent can reach under it. It is a deterministic check, not an LLM judge. It has two gaps for now. First, it checks one agent at a time. Say a policy bars an agent from reading your GitHub code and posting it outside. The agent could split the job between two sub-agents, one that reads and one that posts, and each alone passes the check. NVIDIA says it is working on checking several agents together. Second, it proves what the written policy allows, not that the running sandbox really enforces it. A bug in the sandbox would not show up in the proof (VentureBeat, Anthropic).

What others said:

  • Anthropic is adopting it. On 28 September 2026 Anthropic announced that Claude Managed Agents works with OpenShell. The agent loop stays on Anthropic’s servers, and the tool calls run in a sandbox you control, on your own infrastructure or with a managed sandbox provider, with OpenShell as its runtime. The harness, session state and credentials stay outside the sandbox. Anthropic’s advice is to start with narrow permissions, read the log, and use Claude to tighten the rules to the least access a task needs (Anthropic).
  • Broad backing. NVIDIA counts more than 100 partners. Salesforce has built it into Slack, where teams see agent activity and approve or reject permission requests. SAP builds it into its Joule Studio runtime, and Red Hat, SUSE and Canonical into their operating systems. CrowdStrike, Palo Alto Networks, Cisco and Zscaler are on the security side (VentureBeat, ABI Research).
  • Nscale backs it too. The GPU cloud Nscale says it will bring OpenShell and Sentry into its AI factories. Its CISO Latha Maripuri: “Security for agentic AI has to sit below the agent, in infrastructure it can’t reach or rewrite” (Nscale). At GTC, Tom Matthews gave Nscale’s talk on adapting OpenShell, “Engineering Agents on NVIDIA B200: Running a Software Factory on Open Models for NVIDIA Cluster Bring-Up” (Nscale).
  • A standard in the making, not yet a standard. The effort is tied to the Linux Foundation’s Open Secure AI Alliance (120+ organisations), but no formal standard or certification exists (VentureBeat).
  • One agent, not the fleet. OpenShell answers “what can this agent do on this box?” Agent-to-agent traffic, agent identity and cross-sandbox policy are out of scope, and the Kubernetes chart is experimental (Tigera).
  • The strongest part needs NVIDIA silicon. Sentry runs only on BlueField-4, which favours NVIDIA hardware. NVIDIA’s own advice is that most organisations do not need Sentry (Futurum, VentureBeat).
  • Other vendors say the same. n8n’s talk was titled “Your Agent Doesn’t Get Root” (Mind & Machine Weekly).
  • Developers are less sure. On Hacker News, around GTC Berlin, the loudest complaint was that telemetry is on by default (“Taking default telemetry is insane”). Others pointed out that OpenShell is not the sandbox itself: “It needs Docker or other virtualization.” One commenter just back from GTC said a significant part of the demos used it (Hacker News). Reviewers who read the code call it early. The README says alpha (Secure in Seconds, Starlog).

3. Full agent platforms: SUSE AI Factory

Another theme I noticed is that vendors now sell the whole agent platform, not only parts of it. SUSE is one example. At GTC Berlin it became a launch partner for the OpenShell Reference Architecture and built it into SUSE AI Factory with NVIDIA. One platform covers the OS (a new SLES for NVIDIA, due in November 2026), Kubernetes, agent sandboxes with proxies that block unauthorised network requests, and private model serving on NVIDIA NIM (Futurum, IT Brief). I found no other vendor at GTC Berlin selling the whole stack as one product. The other partners, such as Canonical, Red Hat, SAP and ServiceNow, supply only parts of it (NVIDIA). NVIDIA also launched a new Agent Toolkit, next to the Agent Knowledge Foundation (section 1). The direction is agents you own, running on open models, with audit trails.

What others said:

  • Isolation is no longer the difference. Red Hat and Canonical ship the same OpenShell runtime. Futurum says proof of what an agent did should decide instead. Full event export is off by default, and in their survey only 45% of organisations log agent actions (Futurum).
  • Savings need proof too. Keeping models in-house does not stop agents from sending data out through approved tools, and the saved API fees have to be set against the infrastructure cost (Futurum).

Where is it heading? I think buyers will ask for controls they can check, not a “sovereign” label.

4. Capable hardware at every size, and what Vera Rubin asks of a server room

NVIDIA now sells hardware at four sizes, from a container down to a laptop:

  • Data centre. The DSX POD is a Vera Rubin AI factory in a container, built by partners. It holds 7 NVL72 racks, 504 GPUs in all, and draws about 1.2 MW. It installs on site in 14 days instead of the usual 6 months or so. It ships from early 2027.
  • Rack/server level. For a few dozen GPUs in a handful of servers, there are HGX Rubin NVL8 servers with eight GPUs each. Fully liquid-cooled, all eight GPUs fit in a 2U server (ASUS). More on these below. Partners ship them in the second half of 2026. There was no new launch at this size in Berlin.
  • Desk. DGX Spark now comes as a new, smaller model with 64GB of memory. You can join up to 8 of them in one cluster.
  • Laptop. RTX Spark Windows laptops have been on sale since 16 October. They have up to 128GB of unified memory and 1 petaflop of FP4 compute.

Sources: ServeTheHome, ABI Research, Futurum, Tech Insider, NVIDIA HGX, 2CRSi.

The laptop. The 128GB laptop is the one that matters for running open models locally, and the one Huang held up in the keynote. The chip, developed as N1X with MediaTek, has a 20-core Arm CPU and a Blackwell GPU with 6,144 CUDA cores. They share up to 128GB of LPDDR5X at about 300GB/s, and NVIDIA rates typical draw at 45-80W (Tech Insider). Most of that memory can go to a model. Reviewers set it against Apple’s high-memory Macs, with CUDA as the difference. The first Best Buy listings ran from $2,599.99 for a Surface Laptop Ultra with 24GB to $3,999.99, and the most memory on offer there was 64GB. The cheaper 18-core chip tops out at 32GB (HotHardware). So the 128GB laptop is the headline, not yet the shop shelf.

A laptop runs a model for one person. A team that wants to keep its data in house needs a server room, and there the step from Hopper and Blackwell to Vera Rubin is less about speed than about power and cooling.

From Hopper and Blackwell to Vera Rubin. Vera Rubin is the generation after Blackwell. It pairs Rubin GPUs with NVIDIA’s own Vera CPUs. NVIDIA says it is ramping into full production (NVIDIA HGX).

The direct successor of an 8x B200 server is HGX Rubin NVL8: still eight GPUs on one board, now with 288GB of HBM4 each. That is 2.3TB in total, against 1.4TB on HGX B200 and 640GB on a DGX H100 (NVIDIA DGX H100). The host can be a Vera CPU or a normal x86 board. NVIDIA claims up to 10x the inference throughput of HGX B200, and the same training speed with 4x fewer GPUs. Both are projections, and the 10x compares sparse FP4 on Rubin with dense FP4 on B200 (NVIDIA HGX).

Power and cooling are the bigger change, as the table shows. There is no air-cooled Rubin. Even ASUS’s hybrid 3U server, which cools the CPUs with air, still needs a liquid loop for the GPUs. The Rubin power figures are industry estimates, not NVIDIA specs.

System GPUs Power Cooling
DGX H100 8 ~10.2 kW, 8U Air
DGX B200 (today) 8 ~14.3 kW, 10U Air
HGX Rubin NVL8 node 8 ~24 kW, 2U (estimate) Direct-to-chip liquid only
HGX Rubin NVL8 node, hybrid (ASUS XA NR1I-E12L) 8 ~24 kW, 3U (estimate: same GPU board and CPUs as the 2U node) Liquid for the GPUs, air for the CPUs
GB300 NVL72 rack 72 ~135-140 kW Mostly liquid
Vera Rubin NVL72 rack 72 ~190 kW (Max Q) to ~230 kW (Max P) 100% liquid, up to 45°C inlet water

Sources: NVIDIA DGX H100, NVIDIA DGX B200, 2CRSi, ASUS, Hashrate Index.

The good news is the warm water. At 45°C inlet, most climates can cool without chillers, with dry coolers alone, as in the DSX POD (Hashrate Index). The bad news is that a server room built for air-cooled 8-GPU boxes cannot take a Rubin server as it is. It needs a liquid loop to the rack and almost twice the power per server of a B200, more than twice that of an H100. So if you plan to replace a Hopper or Blackwell server, plan the plumbing and the power feed first.

The same limit shows at the top end. Analysts read the DSX POD as a story about speed to revenue, not chips. Huang put a 1 GW AI factory at about $50 billion to build and close to $40 billion a year in revenue, so a three-month delay costs about $10 billion (ABI Research). The DSX POD was the big launch, and I expect more of these ready-made AI factories. In a year or two, power, not chips, will limit GPU clouds. A provider that cannot get a grid connection for a new site cannot add GPUs, however fast the racks install (see “Power is the bottleneck” below).

5. Foundation models beyond text: Revolut

A foundation model is a large model trained once on broad data, then adapted to many tasks (Wikipedia). Most are trained on text. Revolut’s, PRAGMA, is trained on structured data. On Thursday, 22 October, Anton Repushko and Amey Baokar told its story in “From research to real time: Scaling a foundation model for structured data in production”, also written up as a blog post (Revolut).

The trick is to treat a bank’s event log like text. Each event becomes tokens, so a user’s history reads like a long document. PRAGMA learns patterns the way a language model does: hide parts of the history and guess what is missing. It saw 24 billion events from 26 million users (arXiv). One model then serves many tasks. A simple linear model on its output already does well on credit scoring, fraud and recommendations, and light fine-tuning does better: fraud recall rose 64.7% against the paper’s baselines. It failed at anti-money laundering (F0.5 down 47.1%), because the model reads each user alone and laundering is about links between users.

Running it on huge volumes was a data problem more than a model problem. Tokenising events from half a dozen systems at request time would be too slow inside a card payment. So a data platform tokenises in the background and stores each history ready to use. Six months of history turned out to work as well as several years. Scoring then means: fetch the tokens, run the model.

One backbone model stays on the GPU, and each task adds a small LoRA adapter that updates only 2-4% of the parameters (NVIDIA case study). So one GPU serves many tasks instead of one model per task. They serve the 100M-parameter model on B300 GPUs, with Triton batching about 40 users at a time; the average response is 1.16 s. It is in production for 80 million customers (NVIDIA), in real-time credit decisions, subscription detection and product ranking.

Revolut’s lessons: build the data platform first and the model second. Tokenise once and serve everywhere. Keep the tokeniser fixed, so the same input always gives the same answer a regulator can trust. Let adapters make the economics work.

What other reviews said

I read the analyst and attendee write-ups published in the days after the event. They mostly agree with the direction and question the details:

  • Power is the bottleneck. The 14-day install assumes power is already on site. Grid connections in Frankfurt, London, Dublin and Amsterdam can take years (Futurum, ABI Research). Counterpoint’s Neil Shah says the keynote “addressed the real constraints”, but grid connections and permits stay the real bottleneck in Europe, and the DSX POD does not solve that (Counterpoint). Energy vendors were on the main floor (Mind & Machine Weekly).
  • The hardware comes later, the software now. Rajesh Beri sums up GTC Berlin as “the hardware story is large and arrives later”. The DSX POD ships in early 2027 with no published price. OpenShell is open source and can be installed today, but it is a 0.1.x runtime, and “the honest weakness is maturity” (The Daily Brief).
  • Sovereignty with a catch. Europe gains control of data, models and apps, but the chips are US-designed and Taiwan-made. Open models do not mean independence if the tools tie you to one platform (NAND Research, Counterpoint). “Don’t buy ‘sovereign’; buy controls you can check” (Mind & Machine Weekly). Beri sees a strong case for a container AI factory only where a workload cannot go to a hyperscaler region at all; for everyone else “the case is much weaker” (The Daily Brief).
  • Vendor numbers. The claim that AI factories can draw up to 30% less grid power, the millisecond Sentry response and revenue figures are NVIDIA’s own claims, not yet checked by anyone else (ABI Research). The same goes for the three customer cases where post-trained Nemotron models beat OpenAI models by 30 to 41 points on the customer’s own task (Counterpoint).

Sources