Bots devour the web: ai traffic will drown humans by 2027
Matthew Prince, the man who routes one-fifth of the planet’s HTTP requests through Cloudflare, just dropped a stopwatch on the human era of the internet. 2027 is the year bot traffic overtakes human clicks, he told a packed room at SXSW. The countdown has already started.
The 5-site shopper vs. the 5,000-site agent
Prince’s maths is brutal. Ask a human to pick a mirrorless camera and she’ll open five tabs, maybe six. Ask an AI concierge and the same task spawns 5,000 server calls—price scrapes, review aggregates, stock-check loops, spec-sheet comparisons—before it dares to answer. Multiply that by the 100 million daily queries lobbed at ChatGPT, Gemini, Claude and the clones, and the fabric of the web begins to fray.
Cloudflare’s edge sensors see the spike in real time. Pre-ChatGPT, bots hovered around 20 % of global traffic, mostly Google’s crawler doing its librarian routine. Today the share is already pushing 47 %. At the current compound growth rate it flips the majority line in 36 months, Prince warns, ‘and that’s before we factor in agent swarms negotiating with other agent swarms.’
Data centres choke first, wallets second
The strain isn’t theoretical. Every extra hop burns CPU cycles, bandwidth, joules. Prince showed a heat-map of Cloudflare’s Madrid PoP: orange zones where AI scrapers hit 200 requests per second, cooling fans whining like jet turbines. ‘We’re retrofitting racks for 80 kW per cabinet,’ he said. ‘Five years ago we designed for 15 kW.’
Site owners feel it in latency spikes and cloud bills. A mid-size electronics retailer told me their AWS egress costs jumped $38 k last quarter, 70 % traced to AI crawlers that never buy anything. ‘We’re paying to educate someone else’s model,’ the CTO spat over tacos in Austin.
Dead-internet theorists say the flip already happened
Imperva, a security vendor that guards 500 million browser sessions, claims bots crossed the 51 % threshold in mid-2025. Their definition is looser—it includes credential-stuffers, scalpers, SEO farms—but the trend line is identical. Sam Altman nodded to the same timeline when he mused that self-training loops could render the public web useless for humans by 2028. ‘We’ll be eating our own data tail,’ he told a podcast backstage.
Prince refuses to surrender the open web so easily. Cloudflare is shipping two tools this summer: Agent Audit, a signed-token protocol that forces crawlers to identify purpose and budget, and Cache-for-AI, edge storage that serves pre-chewed answers so the crawler doesn’t re-render the same page 4,000 times. ‘If we don’t architect friction now, we’ll inherit collapse later,’ he said, voice hoarse from three days of investor pitches.
The irony? The faster AI devours the web, the more expensive it becomes to feed. Power grids in Virginia and Ireland are already denying new data-hall connections. Space-based solar farms—once sci-fi gossip—are now serious RFPs. The next frontier isn’t software; it’s literal megawatts in orbit.
Prince closed his slide deck with a single line: ‘The human click was the scarce resource. We spent 30 years optimising for it. Now we have to optimise for the machine query that doesn’t even blink.’ The room, mostly engineers, laughed. Then they opened their laptops and started coding, because the bot clock is ticking louder every second.