Welcome back. This will likely be the last newsletter for the year as I take a short break and enjoy the festive season so lets get into it. It's been another week of musical chairs at the top of the AI leaderboard. Google releases Gemini 3.0, OpenAI then counters with GPT-5.2 and Anthropic's Claude is still inching forward on reasoning benchmarks, particularly in coding. The flagship models are locked in a fascinating game of leapfrog where the "best" AI changes every few weeks. To be honest however, for most engineering work, the differences are starting to matter less than the hype suggests.

And the sustainability narrative? It's far more nuanced than many headlines suggest.

Here's what will be covered in today's newsletter:

If you've been following the AI sustainability debate, you've probably heard the apocalyptic projections: AI will consume 10-20% of global electricity by 2030, ChatGPT queries are boiling the oceans, and we're building a technology that will cook the planet before it saves it.

Let's not sugarcoat this: training large AI models is an requires an insane amount of energy. Training GPT-4 consumed an estimated 50 gigawatt-hours (GWh) of electricity, enough to power roughly 5,000 Australian homes for an entire year. Google's Gemini 3.0? Likely around double that. These training runs require thousands of GPUs running at full throttle for weeks, generating so much heat that cooling systems consume 40% of the data center's total power draw.

The math compounds when you consider frequency. OpenAI, Google, Anthropic, and Meta are each training multiple frontier models per year, with each generation requiring exponentially more compute. Throw in the thousands of smaller companies fine-tuning specialized models, and you're looking at a genuine infrastructure crisis.

Here's where it gets really problematic: most training happens in regions with dirty grids. Northern Virginia, the world's largest data center hub (consuming 25% of the state's electricity), sources much of its power from natural gas. Hyperscalers are signing deals to restart retired coal plants and extend the life of aging nuclear reactors just to keep the silicon fed.

The cooling challenge is equally brutal. A single NVIDIA H100 GPU rack generates 10.2 kW of thermal load. Pack 50 racks into a training cluster, and you're dealing with half a megawatt of heat in a space smaller than a tennis court. Traditional air cooling can't handle it and as such Liquid cooling is becoming mandatory, but it comes with water consumption issues—1 to 5 liters of water per kilowatt-hour in evaporative systems. In water-stressed regions like Arizona or Australia's proposed inland data centers, this is highly significant.

Here's the counterintuitive part that the headlines miss: actually using AI, what's called inference, is nowhere near as bad as the training story suggests.

A groundbreaking study from Google and UC Berkeley revealed that a single ChatGPT query consumes approximately 0.3 watt-hours (Wh) of electricity. That's equivalent to leaving an LED bulb on for 30 seconds. For perspective:

The study found that inference accounts for only 10-20% of a model's lifetime energy consumption, with training dominating the carbon footprint. Once a model is deployed, serving billions of queries is remarkably efficient because inference doesn't require the massive parallelization that training demands. You're running calculations through a frozen network, not updating billions of parameters simultaneously.

The hardware evolution is accelerating this gap. NVIDIA's latest Blackwell architecture delivers 2.5× better inference performance per watt compared to Hopper. Google's TPU v5e achieves even better efficiency for specific workloads. More radically, startups like Extropic are developing thermodynamic computing architectures that claim 10,000× efficiency improvements by computing with probability distributions rather than deterministic logic (though real-world validation is pending).

1. Train Smart, Not Often
Rather than training from scratch, techniques like Low-Rank Adaptation (LoRA), quantization, and model pruning can reduce training energy by 90% while maintaining performance. For engineering firms deploying specialized models (predictive maintenance, process optimization, compliance checking), you should never be training from scratch, instead fine-tuning existing models or augmenting models with external knowledge bases is both faster and vastly more efficient.

2. Engineer the Cooling
This is where mechanical and process engineers have genuine opportunities. Waste heat recovery from data centers (40-80°C) is perfect for district heating, industrial preheating, or absorption cooling. The challenge is co-locating compute with thermal loads, which requires infrastructure planning, not just IT procurement.

3. Measure What Matters
The AI-is-killing-the-planet narrative conflates training (energy-intensive, infrequent) with inference (efficient, continuous). A more useful metric is energy payback: does the AI system optimize operations enough to offset its training and operational footprint?

The Aramco-Yokogawa deployment at Fadhili demonstrates this perfectly. The autonomous control agents were trained once using simulation and historical data, then deployed to run continuously on existing control hardware. The 5% power reduction delivered across the entire gas plant dwarfs the one-time training cost by orders of magnitude. That's the sustainability win: using AI to optimize industrial processes at scale delivers net energy reductions that far exceed inference costs.

The AI energy narrative needs a reset, with it being critical to understand the real sustainability issues that exist with AI, and applying the same rigorous, physics-based engineering principles to its deployment as any other industrial system.

This newsletter seeks to engage and challenge the way engineers see AI and its potential for application in industry. Any thoughts, questions or arguments are welcome! Finally, if you enjoy the content, please refer it on to your friends and colleagues, I would love the support.