20VC: Five Predictions for a World of Agents | The Ads Business Model Will Die | Biggest Lessons from Working with Elon Musk at Twitter with Parag Agrawal, Parallel

Agents will use the web 1000 times more than humans

Parallel is building web search for software rather than people. Agents will use the web a thousand times more than humans, so search has to be far cheaper and faster, and the ad-funded model that paid for the human web stops working. Agrawal's answer is to pay content owners for the value agents actually draw from their content, in a market he expects to grow rather than shrink.

Takeaways

• Agents will use the web 1000 times more. Parallel's first idea is that agents will search far more than humans, so search must become 10 to 100 times more efficient. Agents need answers in 100 milliseconds for voice or minutes for background work, so Parallel offers Turbo and advanced processors.
• Ads break with agents, so pay content owners. Ads do not work with agents in their current form. Parallel plans variable payments each time an agent benefits from content, an AdSense for agents. In his example, a 90-cent run matches a one dollar run without one owner, so that owner earns the 10 cents.
• Web search is mispriced and will get cheaper. Parallel charges about 1 dollar per thousand searches at its quality, against 7 to 14 dollars from rivals. The roughly 10 dollar price traces back to Google's ads CPMs. Agrawal expects prices to fall at least tenfold.
• Frontier models grow while small models catch up. Agrawal expects frontier models to keep growing, while a much smaller model matches any fixed performance level within about six months. A useful model and the frontier model can differ in price by 100 to 1,000 times. He rejects the view that open models take 90% of tokens.
• Every site must eventually let agents in. Amazon has blocked agents while Shopify and Expedia welcome them. Agrawal says everyone will have to open up, with terms unsettled. Nobody yet knows how to pay for unique data like PitchBook's at inference time, beyond per-seat pricing.
• The Monitor API moves the web from pull to push. Instead of agents repeatedly polling the web, the Monitor API triggers on events, like Google alerts built on LLM search. Agrawal says it can save 10 to 50 times the compute for the same answer, as persistent agents wait for external events.
• Parallel's growth tracks inference, with agent risk. Agrawal puts inference revenue near 200 to 300bn dollars and says Parallel grows at the same rate while holding share. A 100bn dollar valuation is possible within years. The main risks are agents failing to deliver value, or Parallel losing on technology, content or customers.

Agrawal's cautions on safety and society

• Alignment is totally unsolved, and models are not adversary proof.