A reality check for data centers: challenging the hype narrative first

I compared Gavin Baker's claim that "data centers might be the best thing ever for the American working class" with my own actual records of running inference services on an open-source model over the past four weeks. I actually ran it through. Conclusion upfront: Data centers aren't castles in the air, but framing their value as a universal dividend is too smooth.

Data centers are places where servers, networks, and cooling equipment live; compute is the calculation resources models need to run; inference is that moment when a user asks and the model answers. Baker's logic goes roughly like this: AI needs massive compute, compute needs data centers, and data centers drive small-town taxes, construction, grid upgrades, and operations, allowing electricians, plumbers, HVAC technicians, etc., to earn money. He even said worrying about oversupply now is a misunderstanding; the real risk is under-construction.

Recently, I moved a small model from cloud GPUs to an environment closer to self-built nodes and ran it for four weeks. Not a large cluster, just single-card to dual-card, doing document Q&A and code snippet generation. At first, there was some surprise: after chunking company Feishu docs and a batch of PDFs, the model's responses to internal terminology were more stable than general chatbots. But later, I got stuck on long contexts. In the console logs, as soon as requests queued up, latency jittered; VRAM usage hit the limit, forcing me to shrink the context window and add another caching layer. The premise of "compute being like water and electricity" is cheap power, durable equipment, and stable loads.

Pros

Data centers aren't just pure financial stories. GPUs, switches, cooling, and power retrofits all translate into physical assets and engineering. Baker mentioned small-town taxes and blue-collar jobs; this direction exists. Construction, cooling, electrical work, network ops—these roles are more real than "everyone writing prompts." When I run code completion and long-document summaries, as soon as the context gets longer, latency and VRAM pressure spike immediately. Increased inference volume means node expansion isn't just a slogan. Low natural gas prices in the US offer cost advantages for high-energy facilities.

Cons

The first con is that "working-class benefits" are understated. Electricians and HVAC techs do make money during construction, but ordinary residents bear the costs of electricity, water, land, noise, transformer retrofits, and tax games. Baker himself admitted there were many reasonable objections to data centers over the past year and a half. These reasons don't disappear because of a phrase like "under-construction." Running my small node, what annoyed me most wasn't the model, but heat dissipation and noise. Putting the rack next to the office meant hearing fans like a plane taking off in summer.

The second con is employment structure. High-paying blue-collar jobs will grow, but won't fall evenly on every working-class family. Data centers often choose locations with low electricity prices, cheap land, and fast approvals. Profits are split among local finances, contractors, equipment vendors, cloud providers, and chipmakers. Those actually replaced or price-squeezed by AI might be customer service, basic translation, junior copywriters, and outsourced testers. Baker says learning to be an electrician has a higher net present value than college; it sounds great, but the premise is you can enter that construction and certification system.

The third con is the payback narrative. Materials say compute investment payback is less than a year. I remain skeptical. My tests show that if a single node serves only a low-traffic demo, it's nowhere near that fast; only if the load is stable, inference calls are dense, and the model is light enough does the cost curve look good. But "less than a year" easily becomes a fundraising PPT. Hardware generations, electricity prices, bandwidth, compliance, and ops incidents drag out the accounting. My worry is people using ideal loads to talk about average returns, then talking about average returns as a universal dividend.

This suits those who have calculated costs, have stable loads, and can secure power and cooling resources. It doesn't suit those treating it as a rags-to-riches story for the working class. Nor does it suit local governments looking only at taxes while ignoring resident electricity bills, water usage, and long-term maintenance responsibilities. For enterprises, ask three questions first: Is the load stable? Can electricity rates be locked? Does the model really need dedicated nodes? Otherwise, "self-building a data center" ends up being building a very expensive server room.

It depends. Data centers aren't a bubble, but "data centers are the best thing for the working class" is over-packaging. There's overheating in this direction, and actual implementation is early. Likely trend towards polarization: top cloud vendors continue stacking large clusters, small companies retreat to hybrid deployment and edge nodes; local opposition will strengthen, approvals will slow down, power contracts and water agreements will become core terms. Whoever clearly explains energy consumption, employment, taxes, and electricity prices is the one who truly lands.