The announcement feels like a tombstone on a debate that has been inflaming the tech community for years. A researcher among the founding fathers of artificial intelligence at Nvidia — the very company that is the main hardware supplier of this race — has stated that AGI, artificial general intelligence, simply will not come. And to explain the sector's direction, he used an analogy that few expected: the closed models of OpenAI and Anthropic are like the walled-garden internets of AOL and Prodigy in the 1990s, destined to be swept away by openness and customization.

The thesis is sharp. We won't pursue a single omniscient artificial mind; instead, we will build thousands of specialized models, trained on the proprietary data of each business and kept under their own control. A vision that redefines the priorities of those evaluating Large Language Model adoption today, shifting the focus from chasing the largest model to the ability to orchestrate bespoke inference.

The parallel with AOL and Prodigy: closed is dead

The walled internets were protected gardens, with curated content and no real interoperability with the open web. Today OpenAI and Anthropic offer powerful APIs, but the models remain black boxes hosted in their clouds, with recurring costs, unpredictable latencies, and virtually no control over prompt engineering, versions, or data. For many enterprises, especially in regulated sectors or with digital sovereignty needs, this schema is a straitjacket. The answer, according to the Nvidia executive, is customized open source: each organization starts from an open base (such as Llama, Mistral, or similar) and refines it with its own data, keeping the model on its own infrastructure. This is not a return to artisanal do-it-yourself; it is the logical evolution of enterprise IT, which has already lived through the same arc with Linux, databases, and now AI.

What it means for on-premise deployments

For teams already operating in air-gapped contexts or needing to meet strict GDPR requirements, this statement is more than an endorsement: it's a market signal. The real bottleneck is no longer the availability of open models but the ability to perform fine-tuning and inference locally with sustainable TCO (Total Cost of Ownership). Here the maturation of stacks like vLLM, llama.cpp, and quantization tools comes into play, allowing LLMs to run on consumer GPUs or mid-range servers, bypassing the VRAM hunger that until yesterday relegated these tasks to the cloud. The promise is a tailor-made model, trained on internal manuals, customer history, and operational procedures, that can remain confined to the corporate network without ever passing through an external endpoint. This is not science fiction: several industrial and financial organizations are already conducting proof-of-concepts along these lines, coupling compact models with vector databases for retrieval-augmented generation.

The expertise hurdle remains. Customizing an LLM is not plug-and-play. It requires fine-tuning skills, dataset management, and evaluation of safety and performance benchmarks. However, the proliferation of frameworks and the lowering of the technical barrier are turning this competence from an academic niche into an increasingly sought-after skill in IT departments. The message from the Nvidia source, echoed on specialized forums, should also be read as a call to prepare: the competitive advantage will not belong to those who adopt a generalist model first, but to those who can integrate AI as an organic component of their own stack, with full control over data and costs.

In the meantime, the AGI debate can wait. The urgency now is elsewhere.