By Angelo Fienga,  Senior Business Advisor – AI, Sustainability and Cloud – Deloitte Italy

Last week, Nvidia announced a partnership with Span and homebuilder PulteGroup to install mini data centers – small, liquid-cooled GPU nodes called XFRA units – directly onto residential homes. Span claims it can deploy 8,000 of these units six times faster and at one-fifth the cost of building a comparable 100MW centralized data center. They’ve already tested prototypes with paying customers, and plan to deploy 100 units in new homes in the southwestern US this fall.

The logic is compelling: instead of concentrating hundreds of megawatts in a single location (with all the grid strain, water usage, and community pushback that follows), distribute the compute across existing infrastructure. Use untapped residential grid capacity. Put the processing closer to where it’s needed.

This is not just an energy efficiency play. It’s an architectural shift.

Because once you have compute nodes sitting in homes, the natural next step is edge inference. Your AI services – voice assistants, home automation agents, local search, personalized recommendations – running locally, with lower latency, no cloud dependency, and no data leaving your premises.

And here’s where the timing gets interesting. In the past few months, the open-weight SLM (Small Language Model) ecosystem has reached a tipping point. Google released Gemma 4 in April 2026 with a model (26B MoE) that activates only 3.8 billion parameters per token – delivering strong reasoning capabilities while running on a single consumer GPU. Quantized versions run on laptops and mobile devices. Apache 2.0 license, full commercial freedom. And Gemma 4 is not alone: Qwen 3.6, Llama 4, Nemotron 3, GPT-OSS – the space is crowded with capable, efficient, fine-tunable models.

The critical word here is fine-tunable. These models can be adapted to specific domains and use cases. A home automation SLM trained on your energy patterns, your routines, your preferences. A personal health assistant fine-tuned on your medical history. A financial advisor that knows your portfolio, your risk profile, your goals.

So we have three converging trends: distributed compute infrastructure at the edge, highly efficient open-weight models that run on affordable hardware, and the ability to customize those models for hyper-personalized services. Lower latency, lower cost, better energy efficiency, full personalization.

It looks like the perfect storm.

But it raises a question that few are asking: what happens when your local LLM becomes a container of your most personal information?

A fine-tuned model doesn’t just access your data – it absorbs it. Your health records, financial behavior, daily routines, communication patterns: all encoded in the model weights. This is fundamentally different from a database you can encrypt at rest. The model is the knowledge. And if someone exfiltrates it, they don’t get encrypted files – they get a functioning AI that can answer questions about your life.

The attention points are at least three. First, governance: who controls the model, who can update it, who decides what data it ingests. Second, cognitive delegation: the risk of blindly trusting a local AI without understanding its limitations, its biases, its potential for error – especially in critical domains like health or finance (see my previous post here). Third, and perhaps most fundamentally: protecting the model itself.

Traditional encryption protects data at rest and in transit. But once data is loaded into a model for inference, it’s in the clear. This is the gap that homomorphic encryption addresses – the ability to perform computation on encrypted data without ever decrypting it.

This is exactly what DataKrypto’s FHEnom for AI™  does. It combines Confidential AI with Trusted Execution Environments (TEEs) to keep models and data encrypted even during inference. The model only understands encrypted inputs and produces encrypted outputs. The decryption key never leaves the secure enclave. If someone steals the model, they get something that is computationally useless without the key.

In a world where AI models are moving from centralized clouds to distributed edge nodes – sitting in homes, in offices, in vehicles – this kind of protection is not optional. It’s structural. The model is the sensitive asset. Protecting it from the inside out is not a feature. It’s a requirement.

The edge AI revolution is coming. The infrastructure is being built. The models are ready. The real question is: are we ready to protect what they’ll contain?