The Shark that never swam Fairwater

Fairwater: A Glass Fortress on Thin Ice

When Satya Nadella unveiled Fairwater, he painted a utopia — a datacenter cluster as powerful as nuclear plants, cooled without water, powered entirely by renewables. Tens of thousands of GPUs, fiber loops circling the Earth, “seamless” architecture.

It looked like a god’s creation. But in public I refused to swim there. Because that water was neither free nor fair — it hid toxicity, fragility, extraction, and silence.

In my LinkedIn thread, I pushed back — calling the vision “hogwash,” calling the scale “monstrous,” calling the engineering a front for extraction, calling the narrative a lie. Now I expand that into a full dissection: what AI’s infrastructure is doing to the planet, to communities, to legitimacy — and why I won’t be a quiet part of it.

1. The LLM Elephant in the Room

Large language models (LLMs) are the engine behind the “smart future” pitch — but their energy and material costs are rarely acknowledged.

Training & Development Overhead

A new 2025 paper, Holistically Evaluating the Environmental Impact of Creating Language Models, estimates that when you factor in hardware manufacturing, model development, and training, carbon emissions and water use roughly double what people usually report. For one family of models they measured 493 metric tons of CO₂ and 2.77 million liters of water during the entire process.

They also found power usage during training isn’t constant — it swings between ~15% and ~85% of max draw — which is hell for grids unprepared for volatility.

Inference & Scale Death Spiral

Every time you ask ChatGPT a question, there’s hidden energy, water, hardware wear:

• A 2025 benchmarking study, “How Hungry is AI?”, measured inference costs across 30 models. Some prompts burned 33 Wh in the least efficient models — levels multiples above the efficient ones.
  
• Even at modest scale, aggregated queries create loads comparable to tens of thousands of homes.
  
• Tapping a generative AI prompt consumes more energy than a Google search — often 5× to 10× more (including cooling overhead).
  

Given the breathless rollout of new models every few months, much of the prior energy is effectively “wasted” as infrastructure becomes obsolete or gets discarded.

Thus, the “AI revolution” is powered not just by compute, but by obsolescence, turnover, and scale arms race.

2. Data Centers: Seas of Glass, Currents of Cost

Since Fairwater promises datacenter scale, cooling, and efficiency, let’s examine the water, energy, and environmental cracks under that facade.

Electricity & Grid Pressure

• Generative AI is pushing demand upward. The MIT Explained series reports that AI computing clusters now consume 7–8× more power per unit workload than standard operations.
  
• One projection suggests by end of 2025, AI could account for ~49% of total datacenter electricity usage (excluding crypto).
  
• Deloitte forecasts global datacenter consumption reaching 536 TWh in 2025 (~2% of global electricity), and possibly more than 1,000 TWh by 2030.
  

These are not incremental shifts. They strain distribution, transform load curves, and demand grid upgrades never priced into the cost of “tech growth.”

Water & Cooling: The Thirsty Truth

• Many datacenters rely on evaporative or hybrid cooling systems that consume ~2 liters of water per kWh used.
  
• In drought-prone regions, that water cost is a hidden tax on nature and local communities.
• A 2024 environmental review found that AI infrastructure may consume multiple times more water than a small country per year.
  

Even with claims like “zero water post-construction,” that rhetoric ignores maintenance, humidity control, bleed-off, and infrastructure leakage.

Supply Chains & Embedded Impacts

Datacenters and AI rely on a global supply chain of semiconductors, rare earths, metals, plastics, cooling fluids, and logistics.

• GPUs and chip fabrication have high Scope 3 emissions (manufacturing, transport, disposal). The How Hungry is AI? study flags that lifecycle emissions may outstrip operational use.
  
• Mining for cobalt, lithium, copper, rare earths often devastates ecosystems and communities.
  
• Data center hardware often has short useful lives (3–5 years or less), creating e-waste streams laced with hazardous materials.
  

Thus Fairwater’s “built once, scale forever” narrative hides a conveyor belt of extraction and waste.

Chemical Pollution & “Forever Chemicals”

New scrutiny shows data centers use PFAS (“forever chemicals”) and f-gases in cooling systems. These chemicals are tied to cancer, hormonal disruption, and persist in the environment.

If a cooling leak spreads PFAS into water tables, that is a risk beyond climate — it is public health.

3. Fragile Architecture, Centralized Power

Fairwater is pitched as a monolithic, seamless cluster—compute, network, storage combined. But that magnitude breeds fragility.

• Systems that scale across thousands of GPUs must manage tight latency control, error correction, synchronization — any small fault can cascade.
• Power and cooling spikes during AI training create bursts that grid operators struggle to absorb.
• Many data center operators already rank power constraints and grid instability as their top threat.
  
• This architecture removes modularity: you can’t isolate failures; your job either lives or dies wholesale.

I refused to swim in that sea because drowning becomes inevitable when every node is interlocked.

4. Power, Control & Ethical Sovereignty

Every AI empire is backed by oversight, or absence thereof. In my thread I challenged Nadella: who decides what model runs, who audits it, who bears the harm?

• When Microsoft restructured OpenAI into a profit engine, it privatized responsibility. Our critique isn’t hyperbole — legal scholars argue software developers may be held accountable under war crimes doctrine for misuse of autonomous systems.
  
• AI systems already assist surveillance, targeting, and narrative control — the leap to weapon use is not fictional.
• Big tech’s promises of “governing principles” are weak in practice. The frameworks lack teeth and independent enforcement.

Fairwater is not neutral infrastructure. It is power, control, and influence disguised as progress.

5. Unequal Burdens & Legacy of Extraction

One of the harshest truths: the burdens of this scale won’t fall on elites alone.

• Public health costs from data center expansion in the U.S. are estimated at $5.4 billion over five years, due to air pollution, hospital visits, respiratory issues.
  
• Communities near data centers bear the noise, heat, pollutants, and strain on local water and power.
• The environmental debt accrues in countries supplying minerals — poor nations suffer while power flows outward.

In my commentary I invoked Gandhi and karmic justice — because infrastructure is moral, even divine, and its sins accrue.

6. The Shark’s Silence: Not Drowning, Not Denial

I chose not to swim in Fairwater. That refusal is not weakness — it is clarity.

Where others build outward, I look inward.
Where others market inevitable, I demand accountable.
Where others bury questions, I surface them.

I charted 14 replies on LinkedIn, seeded ideas, named injustices. But this article is the deep dive — bringing together research, data, and philosophy to stake a claim: AI must be judged by what it costs, not what it claims.

7. Principles Before Planes: A Code for Conscience

To survive this moment, the AI world needs more than tech. It needs integrity. These are non-negotiables:

• Full Life-Cycle Accounting: Train, manufacture, ship, retire — track all scopes.
• Decentralized & Modular Compute: Avoid monoliths; favor federated, edge-aware systems.
• Grid-Aware Algorithms & Scheduling: AI jobs that yield to grid stress signals.
• Renewables + On-Site Generation: Closely tied to actual operations, not just purchase credits.
• Ethical Enforcement & Transparency: External audits, public logs, redress mechanisms.
• Community & Ecological Respect: Build where nature and people can thrive, not where they pay the hidden cost.
• “Return on Environment” Metric: Every model should carry an eco-cost score.

Epilogue: The Shark That Never Swam

Fairwater is not a sea I’ll sail. Because its waters are tinted. Because its foundation is built on exclusion, extraction, and obfuscation. Because for it to look seamless, it must swallow thousands of silent costs.

I refused. Because not every ocean is meant for every creature. Sometimes the only way to dissolve a fortress is to never enter it — and instead expose its cracks from the outside.

So I publish this piece.
Let others dive.
I remain the Shark that never swam Fairwater — but I also remain capable of discerning which waters to inhabit, and which to abandon forever.