NVIDIA GTC Berlin 2026: What’s next for industrial AI and AI infrastructure

In short

• NVIDIA supplies more building blocks, while partners control deployment. NVIDIA showed agent runtimes, factory blueprints and AI data-center designs at GTC Berlin, but the highlighted demonstrations and reference designs depended on partners’ solvers, plant data, measurements or prepared sites.
• Industrial agents work inside specialist tools with restricted rights. Cadence and Dassault Systèmes agents ran the vendors’ own engineering software, and NVIDIA’s factory agents created work orders without changing machine settings.
• AI data-center economics depend on deployment speed and output per reserved megawatt. Factory preintegration and dynamic power allocation promise earlier and higher output, with clear trade-offs.

Why it matters

• For industrial software and automation vendors. Solver ownership and traceable results remain the basis for selling agents, but access terms for agent runtimes and plant data need a deliberate strategy.
• For manufacturers and AI data-center operators. Near-term value comes from bounded diagnosis, faster design iteration, and earlier preparation work, not from autonomous machine control or headline deployment times.

In this article

• NVIDIA GTC Berlin 2026 event overview
• The top 5 trends defining what is next for industrial AI and AI infrastructure—as seen at NVIDIA GTC Berlin 2026
  • 1. Agents start to run industrial software
  • 2. Security rules for agents move outside the AI model
  • 3. Simulation scales robot training; physical measurement decides reliability
  • 4. AI data centers are prepared earlier through 2 separate mechanisms
  • 5. Data center operators get more compute from the same power supply
• Analyst takeaways: Outlook for 2027 and beyond
• Frequently asked questions about the future of industrial AI and AI infrastructure
• Further analysis
• 16 key takeaways from GTC Berlin 2026 (Insights+ Exclusive)
• 3 event showcases with analyst views (Insights+ Exclusive)
• 2 takeaways from the Siemens Erlangen electronics factory site visit (Insights+ Exclusive)
• 5 developments to watch through 2027 (Insights+ Exclusive)

NVIDIA GTC Berlin 2026 event overview

NVIDIA used GTC Berlin to extend its AI agenda into European factories and data centers. NVIDIA GTC Berlin 2026, the US-based chipmaker’s second European AI conference after GTC Paris in 2025, took place on October 20–22, 2026. IoT Analytics Senior Principal Analyst Satya Sinha and Analyst Joaquin Fernandez attended the keynote, analyst briefings, and booth demonstrations, while IoT Analytics CEO Knud Lasse Lueth visited Siemens’ Erlangen factory where NVIDIA deploys some of its technology. Their findings (including 16 key takeaways, 12 event showcases, 7 overarching trends, 7 keynote highlights, and 5 developments to watch through 2027) are published in our 52-page event report, NVIDIA GTC Berlin 2026: What’s next for industrial AI & AI infrastructure (published November 2026).

Across the event, NVIDIA presented itself as a supplier of far more than chips: agent runtimes, factory blueprints, robot simulation tools, and AI data-center designs. Yet, the demonstrations and reference designs we saw depended on assets that partners and customers control: solvers, plant data, measurement data, and grid-connected sites.

The 5 themes below show where this dependency creates value, where it limits deployment, and what this means for the future of industrial AI and AI infrastructure.

The top 5 trends defining what is next for industrial AI and AI infrastructure—as seen at NVIDIA GTC Berlin 2026

1. Agents start to run industrial software

Engineering agents add value by running the vendors’ own tools, not by replacing them. At GTC Berlin, US-based electronic design automation (EDA) software company Cadence and France-based 3D design and simulation software company Dassault Systèmes showed agents that turn an engineer’s goal into a sequence of actions inside established design and simulation software.

Cadence showed the chain from intent to tool execution. In a booth demo of AuraStack, the agent for circuit-board and advanced-packaging design that US-based electronic design automation vendor Cadence launched in July 2026, a prompt to “check and fix any IR reliability with pre-layout analysis” led the agent to reason about the design, plan the steps, and run Cadence’s layout, thermal, and signal-integrity tools with shared design context.

Dassault Systèmes showed agents coordinating design and simulation. In a keynote demonstration, Dassault Systèmes‘ AURA planning agent and LEO sub-agents adapted a Volkswagen Golf door design for the Tiguan in CATIA, found a vibration issue in SIMULIA, and refined the design. An engineer selected the final version.

Solver ownership matters because an agent is only as reliable as the tool it calls. Solvers, design data, validation history, and customer workflows stay with the software vendors, while NVIDIA provides computing infrastructure and, in selected implementations, models and agent runtimes. This strengthens the incumbents’ position but does not guarantee it. Neither demonstration showed independent evidence on error rates, and engineers still need to trace and approve each result. The business case depends on whether shorter engineering cycles and earlier launches offset inference and implementation costs.

2. Security rules for agents move outside the AI model

The control point for agent autonomy is moving from the prompt to the runtime. As agents gain access to files, credentials, and tools, NVIDIA argued that instructions inside a model cannot serve as a permission boundary.

“Today, the only thing stopping an agent from misusing all that is asking it nicely. … But hope isn’t security.”

NVIDIA OpenShell presentation, NVIDIA GTC Berlin keynote (October 21, 2026)

OpenShell enforces policies outside the agent. In NVIDIA‘s open-source OpenShell runtime, a supervisor checks each outbound request and releases credentials only to approved endpoints, a gateway manages sandboxes and policies, and a policy prover checks proposed permission changes with formal verification. Sentry, an optional layer on NVIDIA BlueField-4 data processing units (DPUs), provides hardware-isolated monitoring and policy enforcement independently of the host.

The first factory agents are deliberately narrow. At its booth, NVIDIA demonstrated its Factory Operations Blueprint (FOX) on a circuit-board line. An inspection agent using NVIDIA’s Cosmos model re-checked parts that an automated optical inspection machine had rejected. A second agent searched the line’s machine data, traced excess solder to paste build-up on a printer sensor, and created a maintenance work order. The agents could read data and create work orders but could not edit factory logs or change machine settings; a technician made the repair. At its Erlangen plant, Siemens follows a similar pattern: AI agents propose production schedules, and a planner approves one.

Permission design is likely to set the pace of factory deployment. NVIDIA says OpenShell has partner support, which shows ecosystem interest, not production deployments or an established standard. OpenShell policies restrict an agent’s access to files, networks, credentials, and tools. They do not replace PLC safety functions, industrial interlocks, or human approval for machine changes.

3. Simulation scales robot training; physical measurement decides reliability

Simulation can reduce the need for repeated, costly physical trials, but accurate measurement determines whether the result works. Real robot training data is scarce and costly, especially for contact-rich handling, so developers train skills in simulation and then calibrate and validate them on physical robots.

Lightwheel showed what calibration involves. At the booth of US-based Synopsys, which now owns Ansys, robot simulation company Lightwheel measured the forces as a real robot arm pushed an Ethernet cable into a socket, then calibrated its simulation against those measurements. The company reported a normalized root-mean-square error (NRMSE) of 4.4% between its calibrated simulation and the measured force curves across 5 trials on this one cable-insertion task. Lightwheel said it uses Ansys physics solvers because standard simulation did not match the forces measured on real robots.

Schaeffler and Wandelbots showed the step from simulation to a real cell. Germany-based motion technology company Schaeffler and Germany-based robot software company Wandelbots presented an electric-motor assembly skill trained in simulation and run in a physical robot cell. They aim to reuse such skills across robots and plants, which still requires integration and safety checks at each site.

Ground truth is becoming a competitive asset. NVIDIA‘s simulation tools let developers run more iterations, but they do not supply calibration data. That advantage sits with physics-solver and measurement specialists and with manufacturers that have access to real cells. One calibrated task does not show reliable transfer across tasks, parts, or robots, so robot arms in structured cells are likely to benefit before general-purpose humanoids.

4. AI data centers are prepared earlier through 2 separate mechanisms

Idle GPUs make delays costly, so integration work is moving ahead of GPU arrival. NVIDIA GTC Berlin 2026 showed 2 separate mechanisms; they can be combined but are not 2 steps of one sequence.

Virtual validation can identify integration mismatches before GPU racks arrive. NVIDIA‘s DSX Air simulates infrastructure interfaces so that operators can test provisioning and orchestration software before production racks arrive; at its booth, NVIDIA rehearsed data-center software on a DGX Station. Cadence‘s Reality digital twin checks the power headroom, airflow, and liquid-cooling effects of proposed racks before construction or an upgrade. Performance, thermal, and acceptance tests still need the real hardware.

Physical preintegration is designed to move hardware work off the customer site. With the DSX POD reference design announced in October 2026, NVIDIA‘s manufacturing partners complete rack cabling, piping, and functional testing in the factory. They then ship modules in containers to a site where power, cooling, and network connections are already in place. NVIDIA projects that on-site deployment at such a prepared site could fall from 6 months to as little as 14 days. First deliveries are targeted for early 2027, so this remains a projection, not customer evidence.

The economics depend on how soon an AI factory starts producing tokens and its total cost. Earlier start-up could reduce the time that expensive GPUs sit idle. However, duplicating power and cooling equipment in each module could raise lifetime costs compared with a shared central plant, so DSX POD may suit time-critical capacity more than the lowest-cost build. Factory preintegration could also move some integration work from customer sites to equipment manufacturers and electrical and mechanical suppliers. Neither mechanism removes the grid connection, permits, or final commissioning.

5. Data center operators get more compute from the same power supply

Operators reserve power for each GPU’s peak demand, but live jobs often draw less. NVIDIA‘s DSX MaxLPS software shares this unused headroom among nodes within a budget the operator sets, so more compute can run under the same provisioned power.

A controlled Nscale test shows the gain and its cost. In a test published by NVIDIA in September 2026, UK-based AI cloud provider Nscale ran NVIDIA GB300 NVL72 systems under a 264.4 kW provisioned ceiling. With static allocation, 140 GPUs ran 3 instances of the Kimi K2.5 model: 2 high-throughput jobs and 1 low-latency job. With MaxLPS, the shared headroom allowed an additional 52-GPU throughput job, bringing the total to 192 GPUs under the same ceiling. Aggregate throughput rose by 49.2%. The additional 52-GPU inference instance raised aggregate output; throughput of the existing instances remained virtually unchanged in this test. Measured power draw rose by 19.7%, and the reported 99th-percentile (P99) time to first token, measuring latency per inference request, increased by 17%.

Provisioned power and measured power are different numbers. The ceiling stayed fixed, but measured power draw rose. NVIDIA also promotes MaxLPS as delivering the same tokens with up to 30% less power; that describes a different operating comparison, not the measured outcome of this test.

For power-constrained operators, output per reserved megawatt becomes the key metric. When grid connections, not chips, limit expansion, turning stranded reservations into output is valuable. Trade-offs include potentially higher electricity costs if the higher draw is sustained, weaker tail latency, and little benefit for workloads near peak. The Nscale result comes from one tested model and workload mix, not a market-wide result.

Analyst takeaways: Outlook for 2027 and beyond

6 developments could shape deployment after GTC Berlin:

• Chip suppliers could develop competing agent-security architectures. NVIDIA‘s hardware-independent OpenShell offers an early open blueprint for agent permissions outside the model. Adoption is plausible, while chipset vendors, including US-based Qualcomm, could introduce hardware-rooted competitors.
• Factory operators are likely to grant agent permissions in stages. Bounded write authorization is already possible. Broader production-system permissions will likely depend on safety assurance, corporate policies, and regional regulation, with adoption varying across markets.
• Robot developers could reuse more simulation-trained skills across tasks and cells. Simulation-trained models could reduce programming, particularly for mobile and hybrid robots. Cross-robot reuse still requires calibration, integration, and safety validation.
• Wheeled humanoids could lead some early factory deployments. Humanoids need not be bipedal. Some early factory deployments may favor wheeled designs, which avoid the complexity of two-legged locomotion on flat industrial floors while retaining human-like manipulation. Deployments could grow in 2027, but large fleets remain unlikely; reliability and cost per task are likely to determine the pace.
• DSX POD deployments could expose a speed-versus-cost trade-off. NVIDIA projects that factory preintegration could reduce on-site deployment time at prepared sites, but duplicating power and cooling infrastructure could raise lifetime costs. Completed projects, after deliveries begin in early 2027, are needed to test the trade-off.
• Grid access could constrain European AI-factory expansion. Grid access may limit capacity even where GPUs are available. Dynamic allocation can raise output within reserved power but does not add supply; in NVIDIA‘s Nscale test, it also raised measured draw and tail latency, effects likely to vary by workload.

Frequently asked questions about the future of industrial AI and AI infrastructure

Further analysis

The full NVIDIA GTC Berlin 2026: What’s next for industrial AI & AI infrastructure report adds the evidence behind these conclusions and the developments not covered here. Below, in our Insights+ section, we share all 16 key takeaways, 3 of 12 event showcases with an analyst view on each, takeaways from the Siemens Erlangen site visit, and 5 developments to watch through 2027.