The usual framing treats the person who shows up at a county meeting to oppose a data center as a technophobe standing in the way of progress. I think the opposite case is stronger. If you look at how these businesses make money, a skeptical neighbor is behaving more rationally than the boosters.
Start with the business model. The large AI companies are in the business of selling access to compute that most people cannot afford to own. That is only a good business if the compute stays scarce. Every GPU a hyperscaler buys is one an individual, a small company, or a local-model enthusiast cannot buy, and as prices rise, running a capable model on your own hardware becomes a luxury. People who cannot run models locally have to rent them, and the renter's position is the one the industry is trying to build. Its profits depend on vacuuming up all the supply, and the vast spending on new data centers is the mechanism for doing that.
You can see the incentive at work in how the industry treats efficiency. Quantization lets a model run in a fraction of the memory with modest loss in quality, which is exactly what makes local inference practical. An industry whose revenue depends on scarce hardware has little reason to push people toward it. I'm not claiming a conspiracy. The providers themselves use quantization internally to cut costs. But the marketing, the benchmarks, and the headline products all emphasize ever-larger models that need ever-larger clusters. Efficiency that reaches ordinary users is a threat to the model, and efficiency that stays inside the data center is just a margin improvement.
The financing makes this look fragile. Chip makers invest in AI companies that spend the money on chips, cloud providers lend to customers who then buy cloud services, and the revenue that justifies each step is partly the other steps. Circular financing like this holds up only while confidence holds up, and it exists largely to keep the buying going and the supply locked up.
Now consider how it ends. Sooner or later someone, very plausibly a Chinese manufacturer doing industrial espionage, figures out how to make GPUs as good as NVIDIA at scale and sell them cheaply. Prices fall on that supply alone. Then people discover that a modest local model on affordable hardware handles most of what they were paying a subscription for, demand for centralized inference drops, and the data centers become unprofitable. Their owners sell off GPUs to raise cash, which pushes prices down further. What's left is a very large windowless building in a cornfield. If you're a county commissioner, that is the asset you're being asked to give tax abatements for. Maybe you'd like that. Maybe you'd like to require them to install enough bathrooms and showers now that the building can be converted into a homeless shelter when it's abandoned.
The second argument is about the people who would use these services. A business that builds its operations on someone else's AI platform takes on a risk that is easy to underrate. The platform owner can cut you off because of something you said, a policy you supported, or a change in their terms. Maybe you were critical of Zionism, maybe a left-wing government came to power and you were problematic or gave contributions to the wrong party. Or maybe they just wait for vendor lock-in and raise prices knowing you can't leave. This matters more for AI than for most software, because companies are replacing customer service staff and back-office work with agents. Once your agents run on a provider, losing access is like being debanked: no payments, no customers served, nothing working. A small business that depends on a platform it doesn't control has handed its survival to someone else.
So opposing a data center is not hostility to technology. It's a reasonable response to a model that depends on scarcity, leans on circular money, and is likely to leave the community with an empty shell and a lot of lost leverage. Building a data center is feeding a machine that seeks to control you. It is far better if local models take over the market, and if these hyperscalers don't get built in America they'll still get built somewhere else, and you'll still be able to access them over the internet. There's no particular reason to bring that noise pollution into your community.
No comments:
Post a Comment
Please keep it civil