The Hidden Cost of Intelligence: The Trust Story Hiding in Plain Sight

AI runs on more water, energy, and trust than anyone budgeted for. The real cost isn't the price tag. It's the foundation underneath it.

Jun 11, 2026

This is Part 1 of The AI Reckoning: A Future of Trust Series

There is a trust story hiding in plain sight, and there’s nothing artificial about it.

I am writing this as a builder, not a critic. I am inside this industry, developing what we see as the answer. That context matters because what follows isn’t pessimism or alarm. It’s the view from within, and it’s the same view that led us to build something fundamentally different.

If you’re an investor, an enterprise leader, or someone trying to understand where this is actually going, I’d encourage you to read this carefully. The reckoning is already here.

There is much to be excited about. What we are experiencing is extraordinary. AI is delivering on promises that would have seemed implausible a decade ago. The capability is real. The momentum is real. The investment flowing into it is the largest concentration of capital in the history of technology.

So is the cost. And it’s hiding in plain sight.

Water. Energy. Infrastructure. Quantities that the business case almost never includes. The bill is real, and it is growing. What is also real is that it is being paid by people who were never given the invoice.

That gap is where trust breaks down.

Trust rarely collapses because people disagree with a decision. More often, it erodes when people discover they were never given the full picture in the first place.

The issue is not that AI consumes water, energy, and infrastructure. Every transformative technology consumes resources. The issue is whether those costs are visible, understood, and included in the decisions being made. When costs remain hidden, confidence becomes fragile. Not because anyone intended deception, but because transparency and understanding failed to keep pace with adoption.

The Physical Reality Nobody Budgeted For

Let’s start with water, because water makes this tangible in a way that terawatts and gigabytes don’t.

AI data centers are currently withdrawing 550 million gallons of water every single day. That’s roughly the same rate as the entire global bottled water industry.ⁱ In 2025 alone, AI data centers broke 264 billion gallons of water withdrawal for the year, the equivalent of the annual water usage of 1.8 million Americans. This was happening while 63% of the United States was experiencing drought conditions as of early 2025.ⁱⁱ

A single Google data center in Council Bluffs, Iowa consumed 1 billion gallons of water in 2024. One data center. One year. That’s enough water to supply all of the state’s residential users for five days.ⁱⁱⁱ^iv Microsoft increased its water consumption by 34% in a single year.^[v]The training run for GPT-4 consumed 11.5 million gallons of water in July 2022 alone, in one location, in one month.^viAccelerated AI adoption alone could result in an additional 4.2 to 6.6 billion cubic meters of water withdrawal by 2027. That is four to six times the annual water withdrawal of Denmark. ^vii

Why so much water? Because of heat. Servers generate extraordinary amounts of it, and water is the most efficient way to absorb and dissipate it. In evaporative cooling systems, water absorbs heat from the servers and converts to steam that is vented into the atmosphere, not returned to the source. It doesn’t disappear from the water cycle, but it leaves the local watershed as humidity, unavailable to the communities and ecosystems that depended on it. The compressor-based alternative uses refrigerant instead of water but requires roughly one watt of cooling for every watt of compute. At the scale of modern data centers spanning hundreds of thousands or millions of square feet, it simply isn’t physically viable.

Water cooling isn’t a greedy choice. At this scale, it is the only viable one.

Here is what that means in practice. Data centers are predominantly sited where water is cheapest and most accessible, which puts them in direct competition with agriculture, municipalities, and the people who live nearby. In much of the American West and Midwest, groundwater aquifers are already being drawn down faster than they recharge. A data center doesn’t just use water, it joins a queue that was already oversubscribed. The communities downstream don’t lose access because of malice, they lose it because the data center got there first, and the permitting process never asked who else was depending on it.

Some will point to desalination as an answer. It is worth examining. Desalination is extraordinarily energy-intensive, which increases electricity demand, which requires more cooling, which loops back to water. It also produces concentrated brine discharge with its own environmental consequences. It is not a solution to this problem. It is a relocation of it.

The more compute, the more heat. The more heat, the more water drawn from watersheds that were never part of the business case. It is a compounding loop, and it scales with every query, every conversation, every token generated. That is not a side effect. That is the deal. The costs are environmental. The consequences are human. And the people absorbing both were never included in the conversation that created them.

Now add energy.

AI’s appetite for electricity is growing even faster than its appetite for water. In 2025, electricity demand from data centers grew by 17%, compared to just 3% growth in global electricity demand. AI-focused data centers climbed even faster. By 2030, electricity consumption from data centers is set to double, and power use from AI-specific facilities is poised to triple.^viii The implications extend far beyond utility bills. New transmission infrastructure, grid upgrades, and power generation capacity are required to support that demand. Those costs don’t disappear. They are distributed across communities, ratepayers, and public infrastructure long before they appear in an AI business case.

The communities surrounding these facilities will feel those demands first. Grid strain, infrastructure upgrades, and rising energy costs are not abstract future concerns. They are the practical consequences of scaling a technology whose resource requirements continue to grow. And we are deploying it at extraordinary speed while still learning what those second-order consequences are.

The Decision That Started All of This

So what’s driving the scale of these demands?

The answer lies in a decision made in 2017. That year, a team of researchers at Google published a paper introducing what they called the transformer. It was a genuine breakthrough. A new way of processing language that outperformed everything that came before it. It became the foundation of every major AI language model that followed. GPT. Claude. Gemini. All of them.ix

That architecture might have remained a research breakthrough, powerful but contained, were it not for a parallel development. Nvidia’s GPU advances, particularly the V100 released that same year and the A100 that followed in 2020, provided the computational infrastructure to train and run these models at previously unimaginable scale.

The transformer gave AI its engine. The GPU gave it a highway. What neither provided was a map of where that highway was going, or who was paying for the road.

The moment those two forces converged in a product the world could touch, everything accelerated. ChatGPT launched publicly in November 2022 and became the fastest-growing consumer application in history within weeks. Investors who had been funding research could suddenly see the destination. Capital followed at a scale the industry had never seen. The largest concentration of investment in the history of technology didn’t begin with a boardroom strategy. It began with a public demo that made the future feel immediate.

Here is the issue. The transformer was designed for research. It was not designed for the economics of running billions of conversations a day at civilization scale. The standard transformer design is dense, meaning every single parameter in the model is activated for every single token it processes. Every word. Every punctuation mark. Every invisible reasoning step happening behind the scenes before a visible response appears. The entire network fires every time, for everything. At research scale, this is manageable. At the scale we are now asking it to operate, the compounding costs of that density are the bill nobody budgeted for.

The Token Economy Nobody Explained to the CFO

Talk is cheap. Tokens are not.

Inference, the process of running a model to generate a response, now accounts for more than 60% of total AI computation among major providers.^x Training a model is a significant but bounded cost. You can plan for it, budget it, and measure it against a defined outcome. Running it at scale is different. Inference costs are ongoing, variable, and increasingly difficult to predict as usage patterns shift, reasoning demands grow, and agentic workflows multiply the token consumption of every interaction. Every word generated costs real energy, real water, real money. That unpredictability is where enterprise budgets are breaking down.

Now factor in reasoning mode.

The transformer was designed for research. It was not designed for the economics of running billions of conversations a day at civilization scale.

When AI models engage extended chain-of-thought reasoning, the kind that makes them genuinely better at complex tasks, token consumption increases dramatically. The model thinks through a problem step by step before producing a visible response, and those internal thinking steps generate tokens you pay for even though you never see them. You send 50 tokens, receive 100 tokens back, and are charged for 650 tokens total because 500 reasoning tokens ran silently in between.^xi Agentic workflows involving planning, tool use, and multi-step reasoning have caused token consumption per task to increase 10 to 100 times since late 2023.^xii^xiii

The hidden multiplier that most enterprise buyers never see until the bill arrives.

Here is the counterintuitive detail that the next piece examines more closely: token prices are actually falling. Dramatically. And enterprise bills are still going up. That paradox is the economic story underneath this one.

For enterprises deploying these systems, the reckoning is already underway. Average enterprise AI spend hit roughly $7 million in 2025 and is projected to jump 65% to $11.6 million in 2026.^xiv Yet only 29% of organizations report seeing significant ROI from generative AI. Forrester predicts a market correction, with enterprises deferring 25% of planned 2026 AI spend into 2027.^xv The bills arrived before the returns did. This is where the trust conversation begins to shift from environmental costs to organizational costs.

When expectations are set higher than outcomes can realistically support, people start searching for explanations. Employees question leadership. Leaders question vendors. Investors question assumptions. Everyone begins looking for someone to blame when the deeper issue may be that the underlying economics were never fully understood. Trust is often lost in the space between promise and reality.

This is not a technology problem in isolation but rather a structural one, rooted in the mismatch between what the underlying model was built for and what we are asking it to do. Structural problems require structural thinking, not more spending on the same foundation. A different foundation entirely.

The Proposed Solutions, and What They’re Missing

The industry is not sitting still, and some of what is being tried is genuinely promising.

On water, the most significant development is the shift to closed-loop cooling. Microsoft’s Fairwater facility in Wisconsin, one of its most advanced AI campuses, has filled its cooling system with water once during construction and now recycles it continuously. Microsoft CEO Satya Nadella claimed at Microsoft Build 2026 that the facility’s annual water consumption is comparable to that of a single local restaurant. Google has pledged to replenish 120% of the water it consumes by 2030.^xviThese are meaningful moves in the right direction.

But here is the context that can be easy to miss: these innovations apply to new facilities. The vast majority of existing data center infrastructure uses evaporative cooling and will continue doing so for years. And even the most water-efficient new designs still require enormous energy, which brings us to the bigger question.

On energy, nuclear is the most credible long-term answer being pursued at scale. Microsoft is reviving Three Mile Island. Amazon has secured a 17-year nuclear power purchase agreement. Meta has partnered with Oklo for a 1.2 gigawatt campus in Ohio with 16 small modular reactors.^xvii The logic is sound: nuclear provides reliable, around-the-clock, low-carbon power that renewables alone cannot guarantee at the scale AI demands.

But here is what is not being said alongside the enthusiasm.

Nuclear power plants are among the most water-intensive energy sources that exist. They consume 20 to 83% more water than coal-fired plants of the same capacity. A typical 1,000 megawatt reactor requires roughly 35 to 65 million liters of water per day for cooling, depending on design.^xviii Nuclear plants have already been forced to reduce output or shut down during drought conditions in Europe and the United States because river temperatures rose too high or water levels fell too low.^xix

We are proposing to solve AI’s water problem with an energy source that has its own significant water dependency. That is not a solution to the water crisis.

Nuclear also carries vulnerabilities that deserve acknowledgment: unresolved waste disposal, physical security and cyber risk when reactors are concentrated near high-value data campuses, and regulatory safeguards being quietly dismantled to accelerate deployment. Those safeguards were built from the hard lessons of Three Mile Island, Chernobyl, and Fukushima. Removing them in the name of speed is precisely the kind of second-order consequence we should be paying attention to.^xx

Nuclear is not the wrong answer. It is an incomplete one.

Then there is the space proposal. In early 2026, Elon Musk merged SpaceX and xAI and announced a vision for one million orbital, solar-powered data centers circling the Earth. Others have suggested similar futures. At first glance, it sounds elegant. Move computation into space where solar energy is abundant and many of the terrestrial constraints disappear. The physics of low Earth orbit communication are more viable than science fiction suggests, and the solar energy argument has genuine logic behind it.

But I find myself wondering whether this is another example of a question we are not asking. The challenges facing AI are not answered by simply changing where we place the infrastructure. Moving data centers into orbit does not change the amount of computation required. It does not change the underlying architecture. It does not change the growing demand for tokens, reasoning, and increasingly complex AI workloads. What it does is relocate the consequences.

Low Earth orbit currently contains over 14,000 satellites and an estimated 120 million debris fragments, all traveling at roughly 27,000 kilometers per hour.^xxi A January 2026 study calculated that a complete loss of satellite command for just 24 hours carries a 30% chance of triggering Kessler syndrome, the cascade of collisions that renders entire orbital shells permanently unusable.^xxii We do not have a solution to space debris at scale. This is not for lack of effort. Serious work is underway on active debris removal, controlled de-orbit and collision avoidance, and some of it is promising. But the question that matters is one of sequence. Does that mitigation mature and deploy at the speed orbital compute is being proposed, or does the infrastructure go up first while the solutions are still being built? Confidence that the answer will arrive in time is not the same as the answer having arrived.

Removing a single 95 kilogram satellite currently costs 86 million euros.^xxiii Introducing one million orbital data centers into an already critically congested environment is not a bold engineering vision. It is a proposal that carries risks the global space community is already warning about. A cascade failure in low Earth orbit would not just affect AI infrastructure, it would threaten GPS, weather systems, global communications, and human spaceflight for generations.

It is a bold vision. It is still not the answer the future requires. And depending on how it is pursued, it could make several other problems significantly worse.

Second Order Consequences

History is filled with examples of technologies that worked exactly as designed while producing consequences nobody anticipated. The most important decisions are often not about whether something can scale, but whether the foundations beneath that scale remain sustainable once second-order effects begin to emerge.

Can we create systems capable of extraordinary intelligence while remaining accountable to the people, communities, and resources they depend upon?

We’re at an inflection point. Eyes wide open is not pessimism, it is the only responsible way to build.

The conversation about what comes next is happening. It is just not happening loudly enough given the scale of investment already committed to the current path. And that is partly because the people most invested in the transformer model have the least incentive to ask the question seriously.

This is what it looks like to deploy technology before we understand its second-order consequences. Not malice, institutional inertia dressed up as progress. When capital commits at this scale, it doesn’t just fund a direction, it forecloses questions. The people closest to the problem can develop a vested interest in preserving the assumptions that brought them this far because the weight of what has already been built makes certain questions feel too costly to ask. That is something more dangerous than deception. It is the point at which bias and incentive become indistinguishable from each other. It becomes a reaction when what is needed is a pause and a deliberate response that is oriented toward the future. Not scaling a broken system but considering a redesign.

Why do we keep responding to consequences by scaling the system that created them, rather than pausing to consider a better path forward? That is ultimately a trust and leadership question.

What the Future Requires

AI will continue to advance. The upsides are too promising. The deeper question is whether we are willing to challenge the assumptions that got us here.

The rethinking is already happening, quietly, among researchers and builders like us, asking a different question than the one that dominated the last decade. Not how do we power the transformer at greater scale, but how do we build something that inherently makes compute more efficient and responsible as we scale.

Yann LeCun, former Chief AI Scientist at Meta, has argued publicly for years that the transformer is the wrong foundation for the next generation of AI and left Meta in late 2025 to pursue his Joint Embedding Predictive Architecture full time. ^xxiv State Space Models like Mamba offer near-linear computational complexity compared to the transformer’s quadratic demands, reducing the resource requirements of processing long sequences significantly. ^xxv As of late 2025, several of these alternative approaches have transitioned from theoretical research to production deployment and are matching or surpassing transformer performance on key benchmarks. ^xxvi

I mentioned at the beginning of this article that I am a builder. My co-founder and I did not start with the transformer and try to make it cheaper. We started with the problem: how do you build AI that is relational, explainable, and architecturally efficient from the ground up? That question led us to build a fundamentally different architecture. Not a wrapper. Not an optimization. Designed for the cost and resource realities of the world we are actually in, not the research environment of 2017. I am not ready to say we have solved everything. But I am ready to say, with confidence, that the path forward exists and that it will not look like a more expensive version of what we have now.

In the next piece, we go deeper into AI economics. Because the cost of inference is not just an environmental story. It is a business model story. And the math is stranger than most people realize.

AI is here to stay. And so are we. The only question that matters for the future is whether we build it in a way that remains sustainable, accountable, and worthy of the trust being placed in it.