The Heat Is On - part 2: Powering the AI Revolution, and the Fallout
In part 1 of this 2-part, we considered the march of liquid immersion cooling to help. Here, we review a hardware aceleration that is alreayd putting 1Mw racks on the horizon.
With DCs already using more energy than south australia, the stakes could not be higher!
2025: Blackwell, GB300, and the Liquid-First Era
NVIDIA’s GTC 2025 dropped the GB300 NVL72: 72 Blackwell Ultra GPUs, 36 Grace CPUs, and direct-to-chip liquid cooling as standard. With 800 Gb/s ConnectX-8 SuperNICs and 20 petaflops of FP4 AI compute per rack, air cooling’s a relic. This is huge for Australia, where AI demand is scaling up but power grids are wobbling. ResetData’s clients are jumping on liquid-cooled racks, aligning with our sustainability-first ethos to deliver high-impact compute. ([NVIDIA GB300](https://www.nvidia.com/en-us/data-center/gb300-nvl72/) & [more reading](https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-aims-to-solve-ais-water-consumption-problems-with-direct-to-chip-cooling-claims-300x-improvement-with-closed-loop-systems))
x86 vs. ARM: The AI Workload Shake-Up
AI’s rise is shaking the CPU world. x86 (Intel, AMD) ruled data centres for decades, built for high-performance computing. But ARM’s energy-efficient architecture, king of mobiles, is storming servers. AWS’s Graviton chips fired the first shot in 2018, followed by Microsoft and Google’s custom ARM silicon. ARM’s lower power draw suits AI’s parallel workloads, especially inference, which needs less scalar grunt than x86’s enterprise apps and databases. Will x86 die? I highly doubt it, but it’s on notice. Legacy workloads still love x86’s compatibility and power. AI training and inference, though? ARM’s stealing the show, with chips like NVIDIA’s Grace CPU optimised for GPU-heavy clusters. Hybrid data centres are emerging, blending x86 for traditional tasks and ARM for AI. Equinix’s Toronto facility, using cold water from Lake Ontario, supports hybrid ARM/x86 setups. Europe’s push to repatriate data, driven by GDPR and geopolitical tensions, accelerates this - firms want local control, and ARM’s efficiency aligns with energy regs. Australia’s data sovereignty and power costs are nudging operators toward ARM and hybrids, especially for AI.
Density, Footprint, and the Cost of Dirt
AI racks hit 200–300 kW, dwarfing the 30 kW max of older Aussie sites. Land in Sydney or Melbourne isn’t cheap, so liquid cooling’s smaller footprint saves millions. Globally, data centres eat 4% of energy and spit out 1% of CO2. Water use and refrigerant bans are under fire, pushing hyperscalers toward liquid cooling to hit net-zero goals. Australia’s Safeguard Mechanism adds local pressure. Meanwhile, waste heat reuse is gaining traction - Equinix’s Helsinki site heats homes, and Aussie urban data centres could follow, turning AI’s thermal byproduct into community wins, a sustainability play ResetData champions. (fuelling DC’s)
Hyperscalers Go Nuclear: Powering the AI Beast
AI’s power demand is feral - 100 MW per data centre, doubling by 2030. Hyperscalers like AWS and Microsoft are buying nuclear plants or betting on small modular reactors (SMRs). Why? Grids are buckling, and Australia’s east coast is no exception, with coal plants fading and renewables lagging. Nuclear’s stable and low-carbon, but SMRs won’t hit Aussie soil before 2035 - if ever. Regulatory hurdles, a mining magnate stranglehold, and a government that’s kicked the nuclear can down the road for decades have left us in a bind. The industry’s paying for past inaction, and local utilities are bristling as hyperscalers hog the grid. (heating up)
Where Do We Dive Next?
Cooling’s evolution - fans to pipes to tanks - fuelled every compute leap. AI’s raised the stakes. Five years ago, sustainability was lip service - carbon credits or dodgy offsets. Today, in 2025, we’ve got true sustainable cloud: liquid immersion with zero waste water. Lead times and conversions are hurdles, but the proof’s in - customers can choose real sustainability, not greenwashing.
Will immersion be the standard rack by 2030? Can Aussie air-cooled sites retrofit affordably? Could underwater pods become edge AI reefs off Perth, cooling compact AI clusters for IoT or mining? Will ARM and DPUs redefine hybrid data centres?
I’m all about being all-in on sustainable innovation to ensure AI doesn’t fry the planet.
