Six years after AlphaGo, Google DeepMind is once again marking the debate on AI evolution, this time by publishing a map of four paths that could lead from what we call AGI — artificial general intelligence, able to match human cognitive abilities — to ASI, artificial superintelligence, capable of systematically surpassing humans in every field. The work, relaunched yesterday with only an AFP credit, is not yet fully available, but the title alone raises heavy questions about how organizations currently investing in LLMs and on-premise training will gear up.

No longer science fiction: why the mapping matters

The transition from AGI to ASI is often depicted as a quantum leap that is hard to govern. DeepMind, always attentive to safety and alignment, is instead trying to chart evolutionary paths, probably distinct in modular architecture, self-improvement capability, learning methods, and integration with external data. For those in the enterprise world, this announcement is not pure theory: it signals that the race toward more powerful systems will be punctuated by non-trivial infrastructural choices.

For instance, if one of the hypothesized paths involved rapid surpassing of human performance through massive-scale reinforcement learning, companies currently fine-tuning LLMs locally, on self-hosted stacks powered by high-VRAM GPUs, would need to quickly rethink their assets. It would no longer be about serving models with hundreds of billions of parameters, but about managing computational loads where the boundary between training and inference blurs, with memory and bandwidth requirements that today belong only to a few data centers.

Implications for on-premise deployment

AI-RADAR has long tracked the tension between cloud and on-premise for the most demanding AI workloads. DeepMind’s mapping adds a strategic piece: if ASI is on the horizon of research roadmaps, today’s decisions about hardware, cooling, and networking could soon become obsolete — or, conversely, turn out to be the anchor for maintaining sovereignty over data and models.

Those who are evaluating servers with multi-GPU configurations (think systems with 512GB of total VRAM, interconnected via NVLink 4.0) already know that TCO is a key variable. The prospect of having to support even larger models, with extended context windows and possibly recursive reasoning capabilities, puts pressure on three axes: CapEx for hardware purchase, OpEx for energy and maintenance, and hidden costs related to security and compliance (GDPR and similar). It’s no coincidence that several banks and industrial groups are moving sensitive AI workloads to private infrastructures, away from public clouds.

What DeepMind’s framework can teach us

Even without knowing the paper’s details, we can contextualize the announcement within an already familiar path: the growth in model performance is forcing developers to rethink pipelines and serving frameworks. If today tools like vLLM or Ollama allow optimization of inference on consumer or prosumer hardware, tomorrow, with systems oriented toward superintelligence, architectures designed for continuous training and quantization pushed to yet-unstandardized levels will be needed.

The debate opened by DeepMind, in short, reminds us that the real bottleneck will not be only algorithmic. The ability to locally manage an ASI — assuming it arrives — will depend on choices made now in terms of orchestration, storage, and data isolation. For this reason, those following the evolution of on-premise LLMs would do well to consider the roadmap to ASI not as a speculative exercise, but as a compass for investments that take years to implement.