When it comes to AI data center deployment, maybe it would help to analyze it at a more familiar scale.
Imagine you have an old laptop. It has 16GB of RAM, a six-core cpu, and an integrated graphics card with 4GB of VRAM. It would barely play Roblox to the satisfaction of a 12 year old. Yet this computer, with Ubuntu Server installed and a little finagling, could serve a modest social media platform to 100 users. It might need a bigger hard drive to store the data if people are uploading pics and videos, but that's it. The database, the reverse proxy, the entire platform, could all technically be run off of this device. I know this because I currently serve a Mastodon instance (denton.social) to 30-40 monthly active users using way fewer resources.
Now, take that same computer, and try to run an LLM on it. The model weights will first have to move from storage into VRAM and system RAM in order to run the mathematical operations on them that actually generate tokens. That is usually the first limit, and it's why GPUs and RAM sticks cost so much right now. On our example laptop, we could serve a four-billion parameter model to answer exactly one prompt at a time, and it would hallucinate so much that nobody would want to use it. That's it. Dozens of people able to talk to each other and the larger network, vs. one lonely conversation with a computer that can't even get the cast of Deep Space Nine correct.
Now scale up to one billion users, trillion-parameter models, etc. There are economies of scale at play, but they are overstated. You might be asking, "But why is it so cheap or sometimes even free to use these services, then?" That's because you are not paying the actual price per token. Remember when an Uber ride cost you $5 and the driver was somehow able to feed their whole family off it? It's like that times ten. We're still in the VC honeymoon phase, but it is quickly ending. What's worse, these GPUs last a couple years at the rate of use they get in the data centers, meaning without a near repeat investment, they're just buildings full of e-waste.
That's why the entire focus of OpenAI, Anthropic, Google, and Meta right now is getting China out of the market and jumping the shark with the data-hoovering plan of "OpenClaw, but make it kawaii." The game plan of "build the parachute on the way down" has reached 1000 meters of altitude, and the parachute is mostly looking like the military industrial complex. That means the data center you thought was being built to tell you whether that mole looks suspicious will actually be used to target and arrest "all threats foreign and domestic." And right now they are getting suckers to PAY to run around-the-clock agents to gather all that information. The business model is surveillance, plain and simple. So no, I don't want another data center.
#AI #DataCenter #DataCenterBoom
