Sadiq Jaffer – Energy Optimisation Strategies for HPC Data Centres

At BTC Prague, I made the case that Bitcoin mining’s most underrated superpower isn’t decentralisation — it’s flexibility. As Cloud Innovation Lead at KPMG UK, I’ve been exploring how Bitcoin mining can act as a buyer of last resort inside next-generation HPC data centres, turning wasted energy into a strategic asset.

The Bridge That Barely Stands

I opened with an analogy that engineers will recognise immediately: anyone can build a bridge that stands, but it takes a real engineer to build a bridge that barely stands. The point is efficiency — using exactly the resources you need and not a single watt more. That same principle applies directly to how we design and operate data centres today.

What a Data Centre Actually Has to Do

The core design principles are performance, reliability, scalability, energy efficiency, security, and maintainability. With the explosion of generative AI workloads, data centres that couldn’t adapt have been completely overwhelmed. The ones that thrive are those built with modularity and flexibility at their core — able to scale rack by rack, workload by workload.

Firm Load vs Flexible Load

There are two fundamental energy consumption profiles. A firm load requires consistent, predictable electricity — traditional data centres fall here because critical workloads can never go down. A flexible load can be shifted and interrupted — Bitcoin mining is a perfect example. You can shut miners down instantly. That adaptability is exactly what makes them so valuable in this context.

AI Workloads Have Very Different Energy Profiles

AI models run in two modes: training and inference. Training is consistent and intensive — you cannot interrupt it. Inference, on the other hand, has a lower baseline but is highly spiky and asynchronous. Every time you send a query to ChatGPT or Gemini, that’s an inference request. This natural variability creates real opportunities for dynamic energy optimisation.

By monitoring GPU and CPU utilisation in real time, you can apply frequency scaling, voltage scaling, and deactivation — spinning devices down when idle and back up on demand. Scaling GPUs this way reduces energy consumption, extends hardware lifespan, and drives meaningful cost savings. The thresholds for switching must be adaptive, never static, because workloads constantly evolve.

What Do You Do With the Energy You Save?

Here’s where it gets interesting. When you design a data centre, you commit to a power purchase agreement — a fixed allocation of watts. If your containment strategies free up energy, you can’t just let it sit there. Your options are on-premise battery storage, selling back to the grid under a sellback clause, or — and this is the key insight — using Bitcoin mining as a buyer of last resort.

Bitcoin miners create a price floor for your surplus energy. When GPU demand drops and miners spin up, you maintain a constant energy consumption profile from the grid’s perspective, even while what’s happening inside your data centre fluctuates wildly. Add fast-access capacitors for rapid discharge and slower UPS batteries for stability, and you have a genuinely elegant system.

The Four-Way Equilibrium

The model I’ve been working through creates a four-way equilibrium: your GPUs, your Bitcoin miners, your on-premise batteries, and your grid sellback. A resource manager monitors all of this in real time — if the grid’s price signal exceeds your Bitcoin mining profitability, you sell back. If not, you mine. Workloads are orchestrated across a containerisation layer like Kubernetes, with device-level and workload-level controls operating simultaneously.

The Data Centre That Barely Runs

The end goal is a data centre that is never overprovisioned — one that uses as little energy as possible while keeping critical workloads running reliably. Combine on-site energy production, GPU containment strategies, intelligent energy management