Parallel, which builds APIs for AI agents to search and access the web, closed a $100m Series B at a $2bn valuation last month.
Parag Agrawal, Parallel's co-founder and former Twitter CEO, says agent web use is 10 to 20 percent of human use, implying at least 5,000x to 10,000x growth ahead of his 1,000x forecast.
Agrawal warns that the web's ad, per-seat and subscription models are at risk, and that Parallel pays content owners each time an agent reads their content, comparing it to AdSense.
Parag Agrawal Says AI Agents Will Dwarf Human Use of the Web
Parallel is betting that AI agents will become the web's main users, far beyond today's levels. Agrawal says agent use is only 10 to 20 percent of human use and should grow at least 5000x. Agents need faster, cheaper search than humans do, and the web's ad and subscription models will have to change to pay content owners when agents read their pages.
Takeaways
Agent web use may be 10,000x human use. Agrawal predicted three years ago that agents would use the web 1000 times more than humans ever have. He now thinks that was low. Agent use is 10 to 20 percent of human use, which leaves at least 5000x to 10000x growth.
Agent traffic is already passing human traffic. Asked about a claim that coding agents make up 60 percent of traffic on Vercel's indices, Agrawal says the more relevant fact is that agent traffic on the web now exceeds human traffic.
Agent search needs different infrastructure from human search. Parallel's search must be far more efficient than old-school search and runs on GPUs. Voice agents need results in 150 to 200 milliseconds, versus 600 to 700 for humans. Parallel expects to grow with inference, roughly 10x.
Better search lowers customers' model costs. Customers such as Harvey and Notion buy Parallel alongside model tokens when the search bundled with a model falls short. Parallel's search lets models run half the searches and use fewer tokens.
The web's business models depend on human browsing. Agrawal says ads, per-seat subscriptions and affiliate models will not last once agents read the web. Parallel built a system that works out what to pay each content owner whenever an agent reads their content, comparable to AdSense.
Agent security needs many layers of defence. Agrawal expects consequential cyber events every few weeks. Models alone cannot stop prompt injection, so defence must sit at several layers. Good agents should declare themselves so they can be treated differently from adversarial ones.
GPU supply is messy but has not blocked customers. Parallel has raised a reported $230m and buys and rents GPUs. Availability is worked out over iMessage and finance finalizes the paperwork. It has had several fire drills but never turned away a customer.
Tap a name to go there · tap a line to see why they are linked