At NVIDIA’s GTC Berlin keynote, the conversation moved beyond faster chips to AI factories, secure autonomous agents, energy efficiency and the economics of putting intelligence to work. The next phase of AI competition may be decided by how efficiently businesses turn computing power into useful outcomes.
- The AI factory changes the economics of intelligence
- Autonomous agents change what AI infrastructure must do
- Enterprise AI security must move beyond the model
- Power, cooling and deployment speed are now part of the AI advantage
- The opportunity extends beyond NVIDIA and foundation-model companies
- What enterprise leaders should do next
- The next competitive advantage is operational intelligence
By FutureIsNow Editorial | October 22, 2026
For years, the artificial intelligence race has been measured through familiar questions: Which model is smarter? Which company has the most powerful GPUs? How quickly can a system generate an answer?
At NVIDIA’s GTC Berlin keynote on October 21, CEO Jensen Huang offered a different way to understand the industry’s direction. AI, he argued, is becoming a manufacturing industry—a system of infrastructure, software and computing capacity designed to produce intelligence at scale.
The shift matters because the commercial value of AI is no longer determined by model capability alone. It increasingly depends on whether organisations can deliver computing capacity, control costs, protect sensitive data and put AI to work inside real business processes.
Autonomous AI agents make this transition more consequential. Unlike a conventional chatbot that answers a question, an agent can reason through a task, consult information, call tools, take actions and repeat the process until it reaches an outcome.
That creates a new workload for the computing industry—and a new set of problems for businesses.
The next AI race is not simply about who can build the most capable model. It is about who can turn intelligence into reliable, secure and economically useful work.
The AI factory changes the economics of intelligence
Huang presented AI as a five-layer system: energy, advanced computing, AI factories, models and applications. Each layer depends on the infrastructure beneath it.
The analogy is revealing. A factory does not create value merely because its machines exist. It must have power, operate reliably, use its capacity effectively and produce something customers will pay for.
AI infrastructure follows a similar logic.
The computing hardware matters, but it is only one part of a larger operating system that includes networking, storage, software, cooling, electricity and the models running on top. A facility that has expensive accelerators but cannot operate them at useful capacity is not delivering the intended economic return.
This is the context for NVIDIA’s announcement of DSX POD, a reference design for factory-built AI infrastructure based on its Vera Rubin systems. The company says these integrated, tested systems can reduce deployment at a prepared site from approximately six months to as little as 14 days. Worldwide partner deployments are planned to begin in early 2027.
The important qualification is that this is not a promise to build an entire data centre in two weeks. Site readiness—including power, cooling, networking and supporting infrastructure—remains essential.
Nevertheless, the commercial logic is straightforward. Moving integration and testing into the manufacturing process could reduce on-site complexity and bring productive computing capacity online sooner.
NVIDIA’s DSX POD announcement makes the intended economic outcome explicit: reduce the time between installing infrastructure and putting it to work.
That interval could become an increasingly important competitive variable as AI infrastructure investments grow.
The critical question for an enterprise is therefore changing from How much computing capacity can we acquire? to How quickly can we turn that capacity into useful production, and at what cost?
Autonomous agents change what AI infrastructure must do
The growing importance of AI agents explains why the infrastructure conversation is expanding.
A conventional generative AI interaction may involve a prompt, model processing and a response. An agentic workflow can involve multiple model calls, planning, retrieval, tool execution, verification and additional reasoning before a task is complete.
Each step can consume computing resources. Longer workflows, repeated reasoning and greater use of tools can increase inference demand.
This does not mean every agent is inherently more expensive than every chatbot. Costs depend on the model, task, architecture, implementation and degree of human intervention. But it does mean that a simple count of user prompts becomes less useful as a measure of AI workloads.
An agent is better understood as a system with several interdependent components:
- Model: The reasoning and generation capability.
- Harness: The software that manages the agent’s workflow, memory and interactions.
- Sandbox: The environment that limits the agent’s access to the wider system.
- Knowledge: The information the agent uses to interpret its task.
- Tools: The systems and services through which it performs actions.
Huang outlined these elements during the keynote as the basis of an autonomous agent architecture.
The distinction is important. A powerful model can generate a convincing response, but an enterprise agent must do considerably more. It must access the correct information, select appropriate tools, execute authorised actions and produce a result that can be evaluated.
Consider an agent tasked with preparing a procurement recommendation. It might retrieve approved vendor records, compare contract terms, check prices, identify discrepancies and draft a recommendation for a manager.
The economic value does not come from the volume of text generated. It comes from completing a business task more accurately, quickly or cheaply than the existing process.
This is why businesses need to measure the cost per completed task, not merely the cost per token or the number of agent interactions.
FutureIsNow has examined the practical side of this shift in its analysis of agentic AI use cases, adoption and risks. As agents move from experimentation into business processes, their reliability and economics become as important as their intelligence.
Enterprise AI security must move beyond the model
When an AI system can take actions, the security problem changes.
A chatbot that produces an incorrect answer may create a knowledge or decision-making problem. An agent with access to company files, business applications, credentials or external networks could potentially turn an incorrect decision into an unauthorised action.
The distinction between generating an answer and executing a task makes security architecture central to enterprise adoption.
NVIDIA addressed this challenge at GTC Berlin with its OpenShell Reference Architecture. According to the company’s announcement of the architecture, the approach combines isolated execution environments, policy enforcement and additional monitoring outside the agent itself.
The underlying principle is that an agent should receive only the access required for its task. Restrictions must be enforced by the surrounding infrastructure rather than left entirely to the model’s willingness to follow instructions.
This has practical implications for enterprise deployment.
An agent handling customer-service enquiries should not automatically have the same permissions as an agent processing refunds. An agent analysing financial records may need read access without the ability to alter those records. A software-development agent may need a controlled environment in which to execute code without unrestricted access to production systems or credentials.
Permissions also need to be monitored as tasks change. The access required to summarise a document is not necessarily appropriate for sending an email, approving a payment or modifying a database.
Security architecture can reduce these risks, but it does not eliminate them. Enterprises must still test their systems, monitor behaviour, audit actions and maintain appropriate human approval for consequential decisions.
The larger signal is that agent security is becoming an infrastructure and governance problem, not just a model-safety problem.
As businesses delegate more work to autonomous systems, the ability to control what those systems can access and do could become a prerequisite for adoption at scale.
Power, cooling and deployment speed are now part of the AI advantage
The AI infrastructure conversation is often dominated by chips and model performance. But available electricity, cooling capacity and deployment constraints determine how much computing infrastructure can actually be operated.
NVIDIA’s GTC announcements underline this point.
The company says its DSX MaxLPS software can produce the same token output using up to 30% less power under its stated operating conditions. It also describes cooling configurations intended to reduce water consumption, including estimated savings of 54 million litres per day for a two-gigawatt deployment.
These figures are NVIDIA’s claims, not independently established outcomes for every facility. Actual performance will depend on workloads, local climate, system design and operating conditions.
Even so, they highlight a genuine strategic issue: the economics of AI depend on more than purchasing computing hardware.
FutureIsNow explored the same constraint in its article on AI infrastructure, power, cooling and grid access. The ability to connect, power and cool computing capacity can influence where infrastructure is built and how quickly it becomes productive.
This also explains why NVIDIA’s deployment claims deserve careful interpretation. A factory-integrated computing system may shorten installation once the site is ready, but it cannot independently solve every obstacle involving electricity supply, land, permitting or network access.
An AI infrastructure strategy must therefore consider the whole system: capital expenditure, operational expenditure, utilisation, energy, water, reliability and time to production.
For businesses buying AI services, these constraints may eventually influence availability, pricing and the choice between centralised cloud services, regional infrastructure and local deployment.
For infrastructure providers, they create opportunities around power management, cooling, integration and operational efficiency—not just accelerator supply.
The opportunity extends beyond NVIDIA and foundation-model companies
NVIDIA’s strategy is significant not simply because it expands the company’s technology stack. It illustrates how value can be distributed across the wider AI economy.
The opportunity can be understood across three layers.
Infrastructure and capacity. Companies provide computing hardware, cloud platforms, networking, energy, cooling and the systems required to operate AI facilities. This layer demands substantial capital and technical capabilities.
Agent platforms and control. Businesses develop orchestration software, secure execution environments, knowledge systems, monitoring and evaluation tools. These capabilities help enterprises deploy autonomous systems within defined operational boundaries.
Applied intelligence. Organisations build specialised solutions around industry workflows, proprietary information and measurable business outcomes. Examples could include procurement analysis, customer-service operations, manufacturing inspection, financial reconciliation or internal knowledge management.
These layers are connected, but they do not require every company to compete in all three.
Most enterprises will not build their own foundation models or operate gigawatt-scale AI facilities. Their opportunity may instead lie in identifying where intelligence can improve a valuable workflow and assembling the right models, data, tools and controls to make it work.
That could create room for specialised software companies, enterprise integrators, industry-focused AI providers and businesses with valuable proprietary knowledge.
Model choice will also become more strategic. In his keynote, Huang argued that enterprises will use a combination of closed, open and customised models rather than relying on a single model for every task.
A business might use one model for general reasoning, another for a specialised workflow and a locally deployed model where data-control requirements justify the additional complexity.
This is an architectural possibility, not a guarantee that using more models will always produce better results. The economics depend on the task, integration overhead, performance requirements and governance obligations.
The durable commercial advantage is likely to come from understanding a particular business problem better than a general-purpose provider does—and delivering an outcome that customers can measure.
What enterprise leaders should do next
The shift from AI experimentation to operational deployment requires a more disciplined approach than simply giving employees access to increasingly capable tools.
Five priorities stand out.
Start with the workflow, not the model. Identify a process with measurable costs, delays or quality problems. Define the outcome before selecting the technology.
Establish an economic baseline. Measure the existing process, including employee time, error rates, turnaround time and direct costs. Compare these figures with the full cost of the proposed AI workflow, including model usage, integration and supervision.
Control access from the beginning. Give agents narrowly defined permissions. Separate read-only activities from actions that change records, move money or affect customers. Require human approval where the consequences justify it.
Evaluate systems against real tasks. Test accuracy, completion rates, failure modes, latency and cost under realistic conditions. A compelling demonstration is not evidence of production reliability.
Scale only when performance is repeatable. Expand access and autonomy after the workflow has demonstrated its value and the organisation can monitor, audit and recover from failures.
These principles apply whether an organisation uses a closed commercial model, an open model, a cloud service or a locally hosted system.
They also help prevent a common mistake: measuring AI adoption by the number of pilots launched rather than the value delivered by systems in production.
The next competitive advantage is operational intelligence
NVIDIA’s GTC Berlin keynote points to a broader transition in the AI industry.
The first phase concentrated attention on model breakthroughs and access to computing power. The next phase is putting greater emphasis on building the infrastructure, software and controls needed to turn those capabilities into dependable production systems.
AI factories address the physical economics of supplying intelligence. Autonomous agents expand the range of work that AI systems can attempt. Security architectures address the risks created when those systems can take action. Enterprise applications connect the underlying technology to business value.
Each part matters, but none is sufficient by itself.
More tokens do not automatically mean more productivity. More capable agents do not automatically mean safer automation. More computing capacity does not automatically mean profitable utilisation.
The winners will need to connect all three: capability, control and commercial value.
For enterprise leaders, the strategic question is no longer just where AI can generate a better answer. It is where an organisation can delegate work safely, demonstrate a measurable improvement and scale that improvement economically.
Intelligence is being industrialised. The next moat will not simply be access to AI. It will be the ability to put AI to work—reliably, securely and profitably.
FutureIsNow — Signals over Noise.
