Artificial intelligence has an appetite, and it’s physical. Behind every chatbot reply and every model training run sit data centers, power grids, and fiber networks that were never built for this kind of load. The result is a quiet rebuild of the internet’s plumbing.
And it’s happening faster than most infrastructure can adapt. The companies racing to build bigger models are also, almost by accident, redrawing the map of where compute lives and how data moves.
AI Runs on Other People’s Data
Modern models are only as good as the data they’re trained on, and that data lives on the open web. So AI firms send out crawlers to collect it at a scale the internet has never seen.
Businesses outside the AI labs are doing the same thing, just for different reasons. They scrape competitor prices, track product availability, and gather market signals across thousands of sites a day.
Doing this reliably depends on infrastructure most people never see: proxy networks that route requests through different IP addresses so a single server doesn’t get blocked. The type of proxy matters as much as the volume.
Datacenter proxies are fast and cheap but easier for sites to flag, while residential addresses look like ordinary users and slip through. IPRoyal’s article on residential proxies vs data center breaks down that tradeoff in detail.
The volume is staggering. Fastly’s threat research reported that AI crawlers made up almost 80% of all AI bot traffic in mid-2025, with some fetcher bots firing over 39,000 requests a minute at unprotected servers.
The Power Bill Nobody Budgeted For
Data centers used to be a rounding error on the grid. Not anymore.
The International Energy Agency expects global data center electricity use to roughly double by 2030, reaching around 945 TWh, with AI-driven servers responsible for nearly half of the net increase. That’s close to 3% of all electricity worldwide.
Engineers are squeezing out efficiency where they can; AI chips now do roughly 100 times more computation per watt than they did in 2008. But model size is growing faster than those gains.
The money backing this is enormous. Roughly $580 billion went into AI-focused data center infrastructure in 2025 alone, and Deloitte expects AI data center power demand to grow more than thirtyfold by 2035.
The strain is already local. In 2023, data centers ate about 26% of Virginia’s electricity, according to Pew Research, plus double-digit shares in states like Nebraska and Iowa. Some new housing projects near London now wait years for a grid connection because nearby data centers got there first.
An Arms Race Over Access
Site owners are fighting back. Roughly one in five visits to a typical site was a scraping attempt in 2025, nearly double the rate three years earlier, and that pressure has pushed operators toward rate limiting, CAPTCHAs, and bot-management tools.
Even the open-source world feels it. The Wikimedia Foundation saw the bandwidth it uses to serve images jump 50% in little over a year as bots ingested its freely licensed content.
And there’s a deeper worry. When an AI assistant answers a question directly, it often doesn’t send anyone back to the source that trained it, which threatens the traffic-and-ad model that funded the web in the first place.
Networks Built for Humans, Now Serving Machines
There’s an architectural shift underway too. Centralized server farms are giving way to edge computing, where smaller facilities sit closer to users to cut latency. Some designs already target response times under 10 milliseconds for regional traffic.
Location is becoming a competitive factor in its own right. A request routed through a nearby server beats a distant one every time, which is why companies now pick infrastructure by geography first and raw speed second.
Why does this matter beyond the engineers building it? Because where the computer physically sits decides who gets fast service and whose data gets harvested along the way.
The same buildout that powers better models is also concentrating control in a handful of companies and regions. That tension will define internet policy for years.
What Comes Next
The next few years won’t be about smarter chatbots so much as the wires and watts behind them. IPv6 rollouts, post-quantum encryption, and distributed micro-data-centers are already moving from research papers into procurement plans.
Whoever controls that physical layer will shape how everyone else experiences the network. The firms thinking hard about the physical layer today, rather than only model size, are the ones likely to still be standing when the dust settles. That’s the real race.