NVIDIA GTC Berlin 2026: what companies need to know about the next wave of AI

NVIDIA GTC Berlin 2026 will bring together developers, researchers, technology leaders, and companies from October 20 to 22 around a question that is gaining increasing importance: how to take Artificial Intelligence from experimentation to systems capable of operating continuously within real organizations.

The event's agenda allows us to observe where that conversation is shifting. NVIDIA has organized more than 100 sessions around areas such as Agentic AI, infrastructure, accelerated computing, open models, inference, robotics, and Physical AI. Jensen Huang's keynote is scheduled for October 21 at the Tempodrom in Berlin.

For companies, the interest goes beyond new models or GPUs. The next stage of AI adoption requires addressing issues related to infrastructure, data, inference costs, security, integration, and the ability to maintain these systems in production.

The signal left by the GTC Berlin programming is clear: building enterprise AI increasingly requires a complete stack.

Enterprise AI is entering a more operational phase

The first generative AI projects allowed many organizations to experiment with assistants, content generation, semantic search, or automation of certain tasks.

The next challenge arises when those trials must turn into applications used daily.

A system that works correctly during a demonstration may face thousands of users in production, different permissions, changing business information, security requirements, and execution costs that increase with volume.

The architecture then becomes as important as the model.

NVIDIA is framing GTC Berlin around the complete AI stack, from accelerated computing and infrastructure to models and applications. The agenda explicitly includes AI infrastructure, agentic systems, inference, open models, and Physical AI.

For a company looking to advance in Artificial Intelligence, this means jointly evaluating software, data, and infrastructure rather than treating the model as an isolated component.

AI specialists can work precisely on that layer of integration between models, applications, and business processes.

Agentic AI takes automation to systems capable of acting

One of the major focuses of GTC Berlin is Agentic AI.

NVIDIA describes these systems as specialized agents capable of reasoning, planning, and acting to transform business data into digital workflows ready for production.

This evolution changes the role of AI within an application.

A traditional chatbot receives a question and generates a response. An agent may need to interpret a goal, decide which steps to execute, consult different tools, access business information, and take actions before returning a result.

The difference is significant because each additional capability introduces new responsibilities.

An agent that only consults documentation has a different level of risk than one capable of modifying information in a CRM, executing a financial operation, or interacting with infrastructure.

Therefore, the development of Agentic AI needs to define aspects such as:

• what tools each agent can use;

• what information they can consult or modify;

• what operations require human approval;

• how their actions are recorded and audited;

• what happens when the system does not have enough information to decide;

• how their behavior is evaluated before and after deployment.

Autonomy can generate efficiency, but it also amplifies the consequences of an incorrect decision. Governance, observability, and permissions thus become part of the system's technical architecture.

Data continues to determine how much AI can really do

Business agents need context.

An organization may have advanced models and still achieve limited results if its information is fragmented, outdated, or difficult to consult.

Documents, databases, CRM, ERP, data warehouses, APIs, and internal tools are part of the environment with which an AI application may need to interact.

This explains why the evolution of models does not reduce the importance of data engineering.

Data Engineers can build pipelines, integrations, and access layers that provide reliable information to intelligent applications.

The quality of the context also affects autonomy. If an agent can execute actions, it becomes even more important to know where the information used to make a decision comes from.

The business challenge is starting to shift from “what model do we use” to “what information can the system use and under what conditions.”

AI factories turn AI capability into infrastructure

Another central concept within the GTC programming is AI factories.

NVIDIA uses this term to describe infrastructure specifically designed to produce, manage, and deploy AI capabilities at scale. At GTC Berlin, the infrastructure sessions cover everything from large compute facilities to networking, training, and inference.

From a business perspective, the concept helps to understand an important shift.

An organization that uses AI occasionally can consume external services without needing specialized infrastructure. When numerous products, teams, and processes start to run models continuously, capacity, availability, and costs become operational issues.

The infrastructure must then respond to different questions.

How much capacity does each workload need? Where will the models run? What resources need to be reserved? How are loads distributed? What happens when demand increases? How much does it cost to keep each intelligent application running?

Cloud Engineers can intervene in these decisions when the organization needs to combine Cloud services, accelerators, networking, storage, and scalable capacity.

Inference can become one of the major economic variables of AI

Training models receives a considerable amount of public attention, but for many companies, the recurring cost will appear later.

Every time a model responds, classifies information, processes an image, executes reasoning, or participates in an agent, there is an inference process.

Multiplying that process by thousands or millions of operations turns performance and efficiency into business variables.

NVIDIA dedicates specific sessions at GTC to the deployment, optimization, and scaling of inference, with an emphasis on performance and efficiency for modern AI applications.

This has direct consequences on architecture.

A company may use a very powerful model for a task where a smaller one provides sufficient results with lower latency and cost. It may also distribute different models based on the complexity of the work.

The goal shifts from always using the model with the highest capacity to selecting the appropriate capacity for each workload.

Machine Learning Engineers can participate in this optimization, working on models, pipelines, and behavior in production.

Open models expand architectural options

GTC Berlin also dedicates a specific area to Open Models and Datasets, including customization, deployment, and real applications.

For companies, this expands the available alternatives.

Using exclusively APIs of commercial models can accelerate initial development. However, certain projects may require greater control over deployment, customization, data, or costs.

Open models allow for consideration of other architectures.

An organization can deploy models within controlled infrastructure, adapt them to specific domains, or combine different models based on the characteristics of the workload.

The choice may depend on multiple factors: necessary quality, latency, privacy, available infrastructure, operational cost, and the team's ability to maintain the system.

Therefore, the future of enterprise AI will likely not be dominated by a single model used for all tasks.

The next AI architecture may use multiple models simultaneously

An enterprise application can combine a general model for certain tasks with other specialized ones for classification, extraction, vision, speech, or specific domain processing.

Agents amplify this possibility because they can decide which tools or services to use during a workflow.

This creates an architecture where different capabilities work together.

The challenge then becomes operational: maintaining versions, controlling costs, measuring results, and knowing which component produced each decision.

The IT infrastructure must be prepared to support these loads without each new AI project ending up building a completely independent environment.

From digital models to Physical AI

GTC Berlin also dedicates a significant block to Physical AI and Robotics.

NVIDIA presents this area around models, libraries, and simulation frameworks used to build factories, robots, and autonomous vehicles. The industrial programming also includes sessions on digital twins applied to manufacturing.

Here, AI leaves the exclusively digital environment.

A Physical AI system needs to perceive the world through sensors, interpret that environment, and make decisions that subsequently produce physical consequences.

This introduces different requirements than those of a business assistant.

Latency, simulation, operational security, perception, and real-time processing become more important.

Digital twins also allow for simulating facilities, processes, or machines before transferring certain changes to the real environment. In sectors such as manufacturing, logistics, and robotics, this combination of simulation and AI can accelerate tests that would be costly or risky directly on physical infrastructure.

GTC Berlin shows that AI needs a complete stack

The main areas of the event can translate into concrete business decisions.

The connection between these trends is more important than analyzing them individually.

An agent needs models. Models need data. Running them requires infrastructure. Scaling them introduces costs and observability. Allowing them to act introduces governance and security.

The company ultimately needs capabilities distributed among different technical profiles.

Bringing AI to production also changes the work of DevOps

When an organization incorporates numerous AI systems, deployment ceases to be a one-time event.

Models, services, agents, and pipelines can change independently. Each modification needs validation and must subsequently be monitored for behavior in production.

DevOps Engineers can work alongside AI teams to automate deployments, monitoring, and infrastructure management.

Observability also needs to evolve.

In a traditional application, it is important to know availability, errors, and latency. An AI application adds questions about consumption, model utilization, agent behavior, and the quality of certain responses.

This combination of software operation and model operation explains why practices like MLOps are gaining relevance as AI enters productive systems.

What a company should review after GTC Berlin 2026

The business utility of an event like GTC appears when trends turn into decisions.

An organization preparing its AI strategy for 2027 should review at least five areas: what processes truly justify automation through agents, what data will be available for those systems, what infrastructure will support the loads, how inference costs will be controlled, and what security and governance mechanisms will accompany the deployment.

It should also analyze its internal capacity.

Experimenting with an AI API requires relatively little infrastructure. Building agents connected to business data, deploying specialized models, and keeping them available for multiple applications requires a much broader set of competencies.

The roadmap should grow following the actual maturity of use cases and not exclusively the speed at which new technologies emerge.

The advantage is shifting from the model to the architecture

GTC Berlin 2026 places Agentic AI, infrastructure, inference, open models, and Physical AI on the same agenda. This combination reflects a significant transformation in how intelligent systems are built.

Accessing advanced models is becoming increasingly easier.

Turning them into systems capable of using business data, executing actions securely, operating with sustainable costs, and remaining available at scale continues to be considerably more difficult.

There lies probably one of the main technological decisions for companies preparing for 2027.

The next wave of Artificial Intelligence will be determined both by the evolution of models and by the infrastructure that exists around them. Competitive advantage will begin to emerge in the ability to integrate, operate, govern, and scale AI within real business processes.