AI Is Training People While People Think They Are Training AI When people talk about artificial intelligence learning from humans, the usual picture is simple: people teach, AI learns. A human writes a prompt. The AI responds. The human corrects the answer, chooses a better response, gives a rating, or explains what went wrong. It sounds like a one-way relationship in which the person is the teacher and the machine is the student. But that picture leaves out something important. While people are teaching AI how to respond, AI systems are also changing how people write, think, search, evaluate, and solve problems. That does not mean today's AI is secretly training every person who uses it. Nor does it mean every conversation automatically becomes training data. The reality is more specific: AI systems are designed to learn from human feedback during some training and post-training processes, while their repeated use can also influence human behavior and skills. That crea...
A modern AI server rack can draw over 100 kilowatts of power — more than ten times a traditional server rack's 5-10 kilowatt draw — which is why data center design, cooling, and even site location are all being rebuilt specifically for AI workloads.
Here's a number that explains most of what's happening in AI infrastructure right now: a traditional server rack draws somewhere between 5 and 10 kilowatts of power. A modern AI cluster rack can draw over 100 kilowatts — more than ten times as much, in the same physical footprint. That gap is why data centers, chip design, and even where AI companies choose to build are all being rewritten at once.
Why AI Workloads Are Just Different
A website or a database processes fairly predictable, modest workloads. Training a modern AI model is a different category of problem entirely — models with billions or trillions of parameters, trained on datasets that require thousands of processors running in parallel for weeks or months. That's not a scaled-up version of normal computing; it needs a different kind of infrastructure built specifically for it.
| Infrastructure Type | Typical Power Draw |
|---|---|
| Traditional server rack | 5–10 kW |
| High-density computing rack | 10–20 kW |
| AI infrastructure rack | 20–80 kW |
| Advanced AI clusters | 80–100+ kW |
The Cooling Problem This Creates
More power in the same space means more heat, and traditional air cooling simply doesn't scale to these densities. Three approaches are actually solving this in production right now:
- Direct liquid cooling — coolant circulated directly across the processor rather than around the whole server, which is now standard in new AI-optimized facilities
- Immersion cooling — submerging hardware entirely in non-conductive fluid, which pushes density even higher than direct liquid cooling alone
- Microfluidic cooling — channels built directly into the chip itself, still emerging but aimed at the next generation of even denser processors
This is a genuinely underrated competitive axis. Two facilities with identical raw compute can have very different economics if one packs twice the density into the same building because its cooling actually works — which matters more as land, construction timelines, and grid capacity all become constraints on how fast a company can add compute.
The Custom Silicon Race, Named Specifically
Rather than relying entirely on external GPU suppliers, the largest AI companies are building their own chips, each with a different focus: Google's TPUs (mature enough that they now run a meaningful share of external customers' workloads, not just Google's own); Amazon's Trainium for training and Inferentia for inference, built as separate chips because those two workloads have genuinely different hardware demands; Microsoft's Maia and Cobalt, focused as much on cooling density as raw speed; and Meta's MTIA line, narrowly optimized for its own enormous recommendation-engine inference volume rather than built to be sold externally.
In October 2025, OpenAI and Broadcom also announced a partnership to co-design custom AI accelerators specifically for OpenAI's own infrastructure — a sign that even AI labs without decades of chip experience are now entering this race rather than relying purely on GPU purchases.
GPU-as-a-Service: The On-Ramp for Everyone Else
Building AI-grade infrastructure from scratch is out of reach for most organizations, which is why renting GPU capacity through cloud platforms — rather than owning hardware — has become the default path for startups and mid-sized companies. It trades some cost efficiency at scale for the ability to start immediately without a capital outlay, which is the right trade for almost anyone who isn't training frontier-scale models themselves.
The Part That Doesn't Get Enough Attention: Networking
Raw compute is useless if data can't move between processors fast enough. Large-scale AI training clusters depend on ultra-low-latency, high-bandwidth networking to keep thousands of chips synchronized — a bottleneck that's easy to overlook next to flashier GPU announcements, but one that determines whether a cluster's theoretical compute actually translates into real training speed.
The Sustainability Trade-off Nobody's Fully Solved
None of this is free environmentally. Electricity demand, water use for cooling, and embodied carbon in constantly-refreshed hardware are all rising alongside AI adoption, and the industry's answer so far is a mix of renewable energy procurement, more efficient chip designs, and better cooling — genuine progress, but not yet a fully solved problem, and worth treating with some skepticism when a company markets a data center as simply "carbon neutral."
Frequently Asked Questions
Why do AI workloads need so much more power?
Training a modern AI model involves thousands of processors running in parallel continuously for weeks, a fundamentally different load than a website or database.
What cooling methods handle this power density?
Direct liquid cooling and immersion cooling, since air cooling can't move enough heat out of the small footprint AI racks require.
Why did OpenAI partner with Broadcom on chips?
To co-design custom AI accelerators specifically for OpenAI's own infrastructure, announced in October 2025 — a sign that even AI labs without chip-design history are entering the custom silicon race.
Conclusion
The AI infrastructure boom isn't just "more data centers" — it's a fundamentally different physical problem than the computing industry has solved before, at ten times the power density of what came before it. The companies solving the density and cooling problem, not just the raw chip-speed problem, are the ones actually positioned to keep scaling as this continues.
Comments
Post a Comment