Datacentre Dive: Do AI Datacentre Physics Make On-Premise Unviable?

Why It Matters

The physics‑driven redesign raises capital costs and operational complexity, pushing most firms toward outsourced hyperscale or colocation solutions for training and hybrid models for inference. This reshapes the datacenter market and influences investment decisions across the tech sector.

Key Takeaways

• GPUs >700 W require mandatory direct‑to‑chip liquid cooling
• AI racks now consume 140‑200 kW, outpacing air‑cooling limits
• 800 V DC power reduces cable bulk, enabling megawatt racks
• On‑premise AI facilities risk multi‑million‑dollar obsolescence
• Hybrid strategy: outsource training, keep liquid‑cooled inference onsite

Pulse Analysis

The AI boom is redefining datacenter physics. Modern GPUs such as Nvidia’s H100 operate above the 700‑watt threshold, a point where air‑based cooling can no longer prevent thermal throttling. Direct‑to‑chip liquid cooling removes heat at higher fluid temperatures, allowing the use of dry coolers instead of energy‑hungry chillers. This efficiency gain is evident at TeraWulf’s under‑construction 750‑megawatt AI factory on Lake Ontario, where liquid‑cooled loops will power thousands of high‑density racks while dramatically cutting water consumption.

Beyond cooling, power delivery is undergoing a parallel revolution. Traditional low‑voltage AC distribution struggles with the amperage required for 200‑kilowatt server configurations, prompting a move to 800‑volt DC architectures. Higher voltage means fewer amps, slimmer cables, and the elimination of the eight‑cable bundles that would otherwise crowd a megawatt‑scale rack. However, DC systems lack natural zero‑crossing points, necessitating solid‑state circuit breakers to isolate faults without collapsing entire clusters. The capital outlay for such upgrades can run into several million dollars—roughly $3‑5 million when converted from pounds—making retrofits a high‑stakes decision for enterprises.

For CIOs, the strategic implication is clear: building a traditional on‑premise datacenter for AI training is increasingly uneconomic. Instead, a hybrid model is emerging. Companies outsource the training phase to hyperscale or specialized colocation providers that already host 800‑V DC, liquid‑cooled infrastructure, then bring the inference workload in‑house using modular, plug‑and‑play liquid‑cooled enclosures that fit existing footprints. This approach balances the need for massive compute power with data‑locality, latency, and regulatory concerns, while preserving capital efficiency. As liquid‑cooling technology matures, standardised solutions will further lower entry barriers, allowing more enterprises to adopt high‑density AI compute without the risk of rapid obsolescence.