AI factories, energy limits, and the challenge of building responsibly

Imagine a world where every factory lives in duplicate: one to manufacture goods, one for AI. This is the world Jensen Huang, NVIDIA CEO, unveiled for us during his keynote at GTC 2025. His message was clear: we are not just building more powerful computers; we are building entire factories dedicated to generating tokens—the core currency of AI.

These AI factories, as Huang described them, won’t be an anomaly. They’re poised to become standard infrastructure for industries across the board: cars will have a factory, and then a factory for AI for the cars. Manufacturing will have AI factories, as will finance, healthcare, and education. The vision is broad, ambitious, and rapidly approaching reality.

But this wasn’t just a product reveal; it was a moment to pause and ask deeper questions. As we scale AI infrastructure globally, are we considering the full impact on our one Earth? What happens when every factory has a twin that consumes massive amounts of energy, materials, and attention?

The energy-intensive path of scaling up

Huang emphasized that the compute demands of AI are far greater than expected, especially for reasoning models, which require 150x more computation than traditional tasks and 20x more tokens to achieve more nuanced outputs. The inference phase—where AI models generate responses—turns out to be far more energy-intensive than many anticipated. This means that building and running AI systems at scale is no longer just a technical challenge; it’s a resource one.

Scaling up to meet these demands means constructing massive data centers—a buildout projected to reach a trillion dollars. He described the shift as a move from general-purpose computing to purpose-built AI factories, designed solely for token generation. These AI factories are poised to become as common as traditional factories in every industry.

But behind this vision is a tension Huang himself acknowledged: "We're a power-limited industry." He described the ultimate constraint not in terms of chip speed, but energy availability—what he called the "Ultimate Moore's Law" where performance progress is now determined by how much more can be computed within the same power envelope. In his words, the focus is now on "ISO power, not just ISO chips"—meaning it’s not just about advancing chip technology but about achieving more computation within fixed energy limits.

This shift forces us to reckon with the environmental consequences of AI’s rapid growth. The physical footprint of data centers, the energy they consume, and the resources they require are no longer side considerations. They are central to the conversation about AI’s future.

A humane tech lens: Doughnut Economics

This is where Doughnut Economics offers a timely lens. The framework, developed by Kate Raworth, centers on the idea that humanity must live within an ecological ceiling while ensuring a social foundation for all. It presents a framework for balancing human prosperity with planetary boundaries.

Building and powering AI factories requires enormous amounts of energy, water, and rare earth minerals—resources that strain local ecosystems and global climate goals. As we design infrastructure that can serve enterprise IT, autonomous vehicles, and more, we have to ask: can we do this while staying within planetary boundaries?

And beyond the environmental toll, there’s the issue of distribution. How are the benefits of this technology being shared? Are we reinforcing inequalities, or are we designing systems that uplift more people into that safe and just space?

Equitable scaling: Who benefits?

Huang mentioned a projected shortage of 50 million workers and positioned AI as a tool to address this gap, from digital workers to autonomous vehicles. Yet it’s worth asking: Who benefits from this acceleration, and what are the costs? Will those communities hosting massive AI factories see improved livelihoods, or will they bear the burdens—higher energy costs, environmental strain—without a proportional share of the gains? One way to go about this is to use the externality framework to minimize these harms.

We have to look at who is at the table when these infrastructures are planned, built, and governed. Are we considering the voices of those most affected by environmental extraction and energy usage? Are the long-term externalities being factored into these trillion-dollar investments?

There are promising developments. NVIDIA is investing in digital twin simulations to optimize data centers, and partnerships with companies like Perplexity AI suggest an openness to collaboration. Yet scaling responsibly requires more than technical efficiency—it calls for intentional design that centers both people and planet.

Designing with care

Huang’s keynote ended with a reflection on accelerating discovery so that researchers can accomplish their life’s work in their lifetime. The potential of AI to support scientific breakthroughs, address societal challenges, and fuel creativity is immense.

Yet realizing this potential means more than faster chips and better models. It requires a rethinking of how we build, for whom, and at what cost. As we stand at this inflection point, the conversation needs to expand: not just about compute power and token throughput, but about resilience, accountability, and care.

As we look ahead, the question isn’t just how much more we can compute, but rather how wisely we choose to build. Can we reimagine AI infrastructure that aligns with planetary health and human well-being? Can we shift from racing to scale to committing to stewardship? These are the choices before us—not just for engineers and executives, but for all of us invested in shaping a future that balances innovation with responsibility.

The challenge is real, but so is the opportunity. It’s not often that we get to help define the terms of a new technological era. Let’s ensure that as we duplicate factories for AI, we don’t duplicate the mistakes of extractive growth. Let’s choose to build with intention, humility, and care for the world we share.

Want to learn more about how you can build humane technology? Visit our website: https://www.buildinghumanetech.com/.