.png "Global Insight")
Ask a simple question — can the internet handle the age of AI? — and you get an unsettling answer: nobody is sure, and the people who run the internet are starting to say so out loud. On a recent episode of TechCrunch's Equity podcast, hosts described the net's foundations as "held together by glue and dreams": routing protocols designed decades ago, a power grid being asked to feed data centres at a scale it never planned for, and a disclosure system for software flaws that now moves faster than any human team can patch. AI did not create these weaknesses. But AI is exposing them, all at once, and the strain is starting to show.
The old protocols underneath everything
The internet's core plumbing — systems like DNS, which turns names into addresses, and BGP, which routes traffic between networks — was built for a world of human-scale traffic: people clicking links, sending emails, streaming videos. AI workloads break those assumptions. Training runs generate sustained, massive bandwidth demands; inference at scale creates unpredictable spikes that stress load balancers designed for predictable human usage. Content delivery networks, certificate systems and authentication layers all sit on assumptions from an earlier era, and AI's concentrated, machine-to-machine traffic pushes against every one of them.
The warning turned concrete this month with the PHP flaws known as PHP2Shell. PHP powers a huge share of the world's websites, and when emergency patches were issued, exploitation attempts surged from thousands to millions within hours, according to Cloudflare's chief executive. Our own earlier report covered the emergency patching; the deeper point, made on the TechCrunch podcast, is structural. These specific flaws were found by human researchers, not AI — but with AI tools now able to audit old code at scale, the next flaw may be found and weaponised faster than defenders can respond. The disclosure clock has shrunk from weeks to hours.
The power queue nobody can build fast enough
Even if the software held, the electricity might not. Data-centre developers have reportedly requested about 1,066 gigawatts of US grid connections, according to figures attributed to Wood Mackenzie — but only around 28 percent of that is expected to be committed by utilities. A request is not a power plant: developers ask several utilities at once while deciding where to build, and the queue overstates what will actually be energised. Still, the direction is clear. AI computing needs generation, transformers, substations, cooling and transmission, and even the richest companies cannot build physical infrastructure as fast as AI adoption grows.
Investors have noticed. Shares in Australian data-centre builder NextDC fell about 40 percent after delays at its flagship Melbourne site, with local opposition to giant power poles part of the story. A London developer, Nscale, has struggled to attract buyers for a listing, and CoreWeave dropped sharply on concerns about its financial strength. Meanwhile, communities are pushing back: New York State has imposed a one-year permit moratorium on large new data centres, San Francisco has passed its own, and Amazon and Microsoft have both pledged to stop using non-disclosure agreements when negotiating with local governments. The AI boom has a neighbourhood problem as well as an engineering one.
So will the internet break?
Probably not all at once — and that is both reassuring and misleading. The internet is designed to route around damage, and its failures tend to arrive as local breakdowns: a region's sites forced offline, a government agency scrambling to patch, a neighbourhood fighting a substation. What AI changes is the frequency and the coupling. More flaws found faster, more traffic concentrated in fewer places, more power demanded from grids that plan in years while models scale in months.
The fix, if there is one, is unglamorous: replace or harden the old protocols, build power with honest timelines instead of headline queues, and give defenders disclosure systems that work at machine speed. None of that is as exciting as the next model launch. But the next decade of AI will be decided as much by cables, substations and forty-year-old code as by chips — and the companies that forget the foundations may discover, too late, that the ground underneath them was never as solid as it looked.
For another test of new technology in the real world, see our report on the iPhone Duo durability test failure. Sources include the TechCrunch Equity discussion and TechXplore's data-centre report.