GTC Berlin 2026: Simulation Is the Rehearsal. The Ground Is the Test.

Notes on Jensen Huang's GTC Berlin keynote. Almost every physical demonstration rehearsed in simulation before touching the world. That makes the evidence of what happens on real ground the next thing to build.

Jerson Boyd Milan

Founder, Neicrone

The Tempodrom in Berlin at dusk, with the green GTC26 letters in front of the entrance

On 21 October, Jensen Huang took the stage at the Tempodrom in Berlin for the NVIDIA GTC Berlin 2026 keynote. Ninety minutes covered AI factories, agents, telecom, healthcare, cars and robots, framed for Europe: nearly $10 billion committed to the region's ecosystem and more than 12,000 partners.

This is not a summary of everything announced; NVIDIA's own recap does that well. It is what we took from the keynote as a company working on one narrow part of the problem: what happens when intelligence leaves the screen and has to perform on real ground.

The GTC26 sign in front of the Tempodrom entrance in Berlin — GTC26 outside the Tempodrom, Berlin.

Infrastructure first

Huang opened on what he called the five-layer cake of AI: energy, AI factories, tokens and intelligence, models, and applications. The order was the point. Infrastructure comes before models. The DSX POD, a reference design for complete Vera Rubin AI factories built in modular containers, aims to bring on-site deployment from six months to 14 days once site utilities are ready.

Jensen Huang on stage beside AI factory compute hardware — Frame from the NVIDIA GTC Berlin 2026 keynote. Source: NVIDIA.

The same framing applies to physical AI. A capable model is the top of the stack, not the whole of it. Underneath sit the sites, the conditions and the measurements that decide whether a system can be deployed at all.

Simulation before reality, in every demo

The thread that ran through the physical demonstrations was consistent: rehearse in simulation, then act in the world.

At Siemens' Erlangen factory, agents prepared new products for production by testing plans in a digital twin before changes reached the factory floor.
Dassault Systèmes showed engineering agents using simulation to find and fix a vehicle design problem.
NVIDIA and Siemens Healthineers described adaptive radiotherapy in which a digital twin of the device and the patient's anatomy simulates tumour targeting, with live imaging adjusting treatment as the patient moves.
For automotive, the keynote pointed to new tools to build, validate and deploy autonomous vehicles at scale on NVIDIA Hyperion, with Volvo Cars adopting the platform.

Jensen Huang presenting in front of a Siemens slide during the keynote — Frame from the NVIDIA GTC Berlin 2026 keynote. Source: NVIDIA.

From simulation to a physical robot

The clearest example came near the end. An agent helped train Hugging Face's Microduck: it set up a course, simulated the robot in Isaac Sim, sent training to the cloud, evaluated the learned policy on an RTX Spark, and deployed it to the physical robot. Then four Microducks walked out on stage.

Rendered Microduck robots training on a simulated course in Isaac Sim — Simulated, then real: Microduck in Isaac Sim (a rendered scene) and on stage. Frame from the NVIDIA GTC Berlin 2026 keynote. Source: NVIDIA.

Three physical Microduck robots on the keynote stage — Simulated, then real: Microduck in Isaac Sim (a rendered scene) and on stage. Frame from the NVIDIA GTC Berlin 2026 keynote. Source: NVIDIA.

Alongside it, NVIDIA introduced the Agile ONE SN reference humanoid, built on Isaac GR00T and designed to help researchers develop, train and validate humanoid skills, together with Isaac Lab 3.0, the Isaac OS runtime and an open physical AI dataset.

The Neicrone perspective

Simulation is the rehearsal. The tooling for it is getting remarkably good, and the keynote made the loop from simulation to deployment look almost routine. That is exactly why the step after it matters more.

A policy that passes in simulation has passed a rehearsal. The ground is the test.

Every one of those demonstrations ends at the same handover: a system trained and evaluated in a model of the world meets a specific site, with its own surfaces, slopes, weather, traffic and wear. What a team needs at that moment is not more simulation. It is evidence:

Where should this system be tested, so the test covers the conditions it will actually meet?
How does it behave there, measured independently of the system under test?
Is there a record of that, traceable from the site and its conditions to the result, that a buyer, an insurer or a regulator can trust?

That is the part Neicrone works on. Andromeda ranks candidate sites against terrain, hazard, access and operational constraints, with the reasoning visible. Validation engagements, now being scoped, turn those sites into structured tests with separate reference measurements, and each one ends in a traceable Neicrone Evidence Record.

The keynote showed how quickly intelligence is being built and rehearsed. The constraint is moving downstream, to the moment it meets real ground, and to the evidence of what happened when it did.

Neicrone is a member of NVIDIA Inception. The full keynote is on NVIDIA's YouTube channel.

— Neicrone