Nvidia’s engineers are writing less code than ever. And Jensen Huang is perfectly fine with it. “These agentic systems are…,” the CEO began, suggesting that his teams’ preference for building AI agents over Python coding is now well established. Not a step back, but a promotion.

The remark, reported by The Next Web, is more than a campus curiosity: it contains a precise thesis about where software design is heading. Huang isn’t talking about coding assistants suggesting snippets, but about a paradigm shift in which the engineer no longer produces code firsthand, but instead assembles and orchestrates agents based on LLMs that generate, test, and integrate code autonomously. The craft moves from syntax to supervising probabilistic behaviors.

Who wins and who loses in the agentic factory

This transformation already has clear winners and losers. Nvidia is the most obvious beneficiary: if the unit of work is no longer the line of code but the inference of an agent, compute consumption multiplies. Every agent requires one or more models running continuously, and every LLM call translates into compute demands that must be served with GPUs capable of high throughput and low latency. For a company that produces the hardware those models run on, having its own engineers embrace this approach also means building internally the demand that the market will later replicate.

At risk are cloud service providers whose value relies on simply hosting repositories and traditional CI/CD pipelines. When code is generated by agents, intellectual property resides not only in final artifacts but also in prompts, context datasets, and the agent’s decision chain. The organization developing software therefore has a strong incentive to keep the entire cycle within its own boundaries—not just for control, but to protect know-how. On-premise or self-hosted deployment becomes the natural choice for anyone wanting to avoid having their agent logic end up under third-party control.

The sovereignty and latency knot

There’s a flip side: orchestrating agents locally means having GPU clusters sized to sustain continuous inference. A nightly build server is no longer enough; you need infrastructure capable of handling dozens of simultaneous real-time requests, with memory and bandwidth requirements that grow as agent complexity increases. For IT teams, the shift to agentic pipelines isn’t a routine software upgrade—it’s a rethinking of supporting hardware.

This is where AI-RADAR’s perspective fits for those evaluating on-premise deployment of LLMs: the analytical frameworks available on the platform help compare the variables at play—TCO, energy consumption, latency constraints, and data sovereignty—without stopping at a superficial cloud-vs-bare-metal comparison.

A promotion, not a threat

Huang chooses the word “promotion” carefully. It’s not a sugarcoating tactic, but the recognition that the ability to orchestrate agents will be a rarer and more valuable asset than pure coding skill. The implicit message is that the profession isn’t disappearing; it’s moving to a higher plane: those who can design feedback loops, evaluate model outputs, and align agents with business constraints will be the new protagonists of the development cycle.

Open questions remain. If agents become so central, doesn’t the concentration of hardware suppliers risk becoming a structural bottleneck? And what skills should we teach newcomers to the workforce today? For now, Nvidia seems to have already taken its position: push demand where supply is ready to respond.