The Sustainability Question Around AI Models
There’s something about this AI wave that feels different from past tech cycles. The speed, the hype, the reach — it’s all unprecedented. But as everyone rushes to build, release, and integrate, I keep wondering: is any of this truly sustainable?
When we talk about “sustainability” in tech, the first thing that comes to mind is usually environmental impact — energy use, carbon footprint, resource consumption. But with AI, sustainability goes deeper. It’s not just about how much power these models consume, but about the ecosystem that keeps them alive: compute, capital, and control. How much longer will AI tools be free/affordable?
The cost of intelligence
Someone in a discussion I had recently said something that’s really true: “These models are energy-intensive, and the current pricing structure isn’t sustainable.”
That line stuck with me. Because when you look at the sheer scale of what it takes to run systems like GPT, Claude, or Gemini, it’s staggering. These aren’t static models sitting quietly in a data center. Every query, every API call, every chat or image generation spins up thousands of GPUs burning power at an industrial scale.
It’s not just a matter of hardware — it’s the economics behind it. AI companies are in an arms race to make models smarter, faster, and more accessible, but each step up comes with exponential cost. Someone, somewhere, is paying that bill. Right now, it’s venture-backed firms and hyperscalers. But for how long?
The capitalist engine behind “open” AI
Others in that same conversation pointed to what’s maybe the less visible issue — the capitalist structure of modern tech. The people building the most powerful AI tools also own the platforms that distribute them, the servers that host them, and the companies that profit from their use.
It’s a closed loop — and it’s hard to imagine true openness inside that.
Even the supposedly “open” frontier models depend on infrastructure that’s owned by the same few companies. You can fork the code, train a small version, or deploy a fine-tuned variant — but in the end, you’re still renting your compute from the same giants who built the original models.
It makes you wonder whether this ecosystem is designed for democratization or for deeper dependency.
A possible shift: self-hosted and smaller
That’s why I find the rise of self-hosted models so interesting.
Developers and organizations are beginning to realize they don’t always need a massive, general-purpose model. For many tasks, a small, domain-specific model is enough — and it can run locally, without cloud APIs or recurring costs.
The tools are improving too. Frameworks like Ollama, vLLM, and Hugging Face’s inference tools are making it surprisingly easy to spin up your own LLM. You can fine-tune, host, and serve it privately.
This shift feels like a quiet rebellion — not against AI itself, but against its centralization. It’s a movement back toward autonomy.
And in that sense, it reminds me of the early days of open-source software — when people built because they could, not because it was profitable.
The rise of open and Chinese models
Another recurring point in these conversations is the growing presence of Chinese models and open alternatives.
It’s not just diversity for diversity’s sake. It’s competition. Models like Yi, Qwen, and DeepSeek are advancing rapidly, and they’re breaking the Western monopoly over AI capability.
That shift could be healthy for the industry. It introduces pressure — on pricing, access, and innovation. It forces transparency and keeps the power balance in check.
In the long run, the end users — the developers, creators, and small startups — benefit from that competition. When there’s more choice, there’s more room for sustainability.
The quiet dependency no one’s talking about
But there’s another kind of sustainability we don’t talk about enough — workflow dependence.
We’re entering a phase where an entire generation of developers is being shaped by AI-assisted tools. Things like Copilot, Replit AI, or N8N make development smoother, faster, and more enjoyable. They reduce friction and lower the barrier to entry.
But they also shift ownership.
When your entire workflow — your automation, your code generation, your pipelines — lives inside a closed ecosystem, you’re not working independently. You’re working inside a subscription model that can change, limit, or lock your access at any time.
Take N8N for example. It’s one of the most exciting no-code automation tools out there, but it’s slowly drifting toward paywall-based limitations. You can self-host it, yes, but the message is clear: freedom is available, but convenience costs money.
That’s the new dependency — not on a single model, but on the infrastructure of convenience.
Power, cost, and ownership
When we talk about sustainability in AI, we often focus on energy usage or financial viability. But maybe the real sustainability question is about ownership.
Who owns the models?
Who owns the pipelines?
Who owns the tools we use to create?
Because if the answer to all three is “someone else,” then what we’re building isn’t a sustainable ecosystem. It’s a rental economy disguised as innovation. And if the majority of students, developers and other professionals are hooked on these tools that have suddenly been made expensive, we might be looking at the new “white gold”.
The future we seem to be moving toward is one where access replaces ownership, convenience replaces control, and automation replaces understanding. And maybe that’s fine for some — but it shouldn’t be the only path forward.
Where we go from here
There’s still hope in the small things. The open-source communities building local models. The engineers writing papers outside corporate labs. The creators pushing for smaller, decentralized systems that can thrive independently.
That’s where sustainability might actually live — not in the biggest data centers or the most powerful models, but in the ability to build, host, and maintain tools on our own terms.
AI doesn’t have to be unsustainable. But it does need to evolve beyond this cycle of scale and dependency. Otherwise, we’ll wake up one day surrounded by incredible tools — none of which we actually own or understand.
And that’s not intelligence. That’s just another form of control.