Imagine you hired an architect to design your dream house, and before pouring a single foundation, they handed you a video game version of it. You could walk the hallways, flip the light switches, run the shower, and watch the electricity bill tick up in real time. Move the kitchen three feet left, and the model instantly tells you how much that decision costs you over ten years.
That is roughly what NVIDIA is now offering to the people who build the buildings your AI agents live in. The platform is called DSX, and it launched in 2026 as a way to design AI factories using digital twins — virtual replicas that let engineers test power and cooling setups before construction starts.
If you’ve never thought about where your AI chatbot physically lives, that’s fair. Most of us haven’t. But that gap in understanding is exactly why this matters.
What an AI Factory Actually Is
When you ask an AI agent a document or book a flight, the thinking doesn’t happen on your laptop. It happens in a warehouse-sized building packed with specialized chips, all running hot enough that keeping them cool becomes one of the largest engineering challenges of the whole operation.
These buildings are what the industry now calls AI factories. The name is deliberate. A traditional factory takes in raw materials and outputs products. An AI factory takes in electricity and data, and outputs answers — what the industry measures as tokens, the small chunks of text and code that AI models produce.
Which reframes the whole thing. Your AI assistant isn’t software in the way a spreadsheet is software. It’s closer to a service delivered by industrial infrastructure, like tap water or electricity. And like both of those, the pipes matter enormously.
Why Design-Before-You-Build Is Such a Big Deal
Here is the practical problem DSX addresses. Building an AI factory is expensive and slow. Once the concrete is set and the cooling pipes are run, changing your mind is painful. Traditionally, you find out whether your power and cooling design was any good after you’ve already committed to it.
Digital twins flip that sequence. You simulate the building first, spot the inefficiencies in software, and fix them where fixing things is cheap. NVIDIA’s stated aim with DSX is improving energy efficiency and reducing costs, and the platform includes pieces for simulation and operations alongside the design work.
The partnership side gives a sense of how seriously the industry is taking it. On August 17, Trane Technologies and Eaton unveiled an integrated power-and-cooling reference design aligned with DSX. Their claim is up to 15% better energy efficiency, plus reduced installation effort. Trane makes cooling systems. Eaton makes power management gear. Neither is a chip company, which tells you something about where the hard problems now sit.
Tokens Per Watt
At the AI Infra Summit in September 2026, NVIDIA showcased DSX advances alongside its Vera Rubin architecture, framed around optimizing tokens per watt. That phrase is worth sitting with for a second, because it’s a quietly useful way to think about AI.
Tokens per watt means: how much useful AI output do you get per unit of electricity? It’s a fuel-economy number for intelligence. And it explains why a cooling company’s reference design is news. If you can’t cool the chips efficiently, you burn power on air conditioning instead of answers. The plumbing directly determines how much thinking you get per dollar.
What This Means If You’re Not an Engineer
You will probably never touch DSX. But the reason to pay attention is that it reveals what actually constrains AI right now. The public conversation focuses on models getting smarter. The industry’s own investments point somewhere else entirely — toward electricity, heat, and real estate.
- AI agents run on physical infrastructure with physical limits, not in an abstract cloud.
- Energy efficiency is becoming a competitive feature, not just an environmental talking point.
- The companies shaping AI’s near future include ones that make air handlers and switchgear.
- When AI services get cheaper or more available, infrastructure efficiency is often the reason.
There’s something clarifying about this. AI can feel like magic precisely because the machinery is hidden. Learning that a chip designer is now shipping tools to plan cooling loops makes the whole field feel less mystical and more like what it is — a very large, very hot engineering project that a lot of people are trying to make cheaper to run.
Next time an AI agent answers you in a second flat, you can picture the building. Somebody simulated its air flow before they built it.
🕒 Published: