Parag Agrawal: why the ads business model dies with agents

Key points

Key takeaways from a 20VC interview with Parallel Web Systems founder and CEO Parag Agrawal (September 2026):

Search for agents, not humans. Agrawal describes Parallel as "the Google for agents": agents will use the web a thousand times more than humans, so web search needs new technology — 10–100x more efficient compute —and new business models, because agent queries arrive as full sentences rather than keywords, latency budgets range from roughly 100 milliseconds for voice agents to long-running background tasks,and output is tokens or files, not blue links.

Spending compute across the stack. He frames web search as narrowing trillions of documents to the few thousand tokens a model's context window should see, allocating more compute to search when the calling model is large and expensive — Parallel's API even lets a caller disclose which model is querying — while cheap models get minimal search. Productised modes include Turbo, built for voice-agent speed,and Advanced for expensive background agents. Most demand today comes from knowledge work: coding invokes web search in only ~5% of prompts, whereas law, insurance underwriting, sales and AI science are search-heavy.

Big models keep growing, small ones catch up. He expects frontier models to keep scaling while ever-smaller models reach any fixed performance bar roughly every six months, widening the useful range of model sizes. He doubts open models will capture 90% of token volume or frontier models 90% of dollars — path dependency makes it hard to predict — and hopes American open models emerge through two builders competing. The routing layer is valuable today while GPU supply is tight, but its long-run value is uncertain.

The ads business model dies with agents. When agents — not humans — are the customer, display advertising stops working: even Amazon's ad business, larger than its e-commerce business, would shrink toward nothing if agents handle purchases. Parallel wants to become an "AdSense for agents", paying content owners a variable amount every time an agent derives value from their content, so publishers have an incentive not to block agents. Unlike music streaming, where the pie shrank first, agents add enough utility that the overall market can grow, making the business-model transition possible.

Web search is mispriced for the agent era. With a cheap model, 80–90% of an agent's spend can go to web search versus 10% to the model itself, so prices must drop an order of magnitude or more. Parallel says it delivers the same quality at $1 per 1,000 searches where rivals charge $7–14, with another 10x reduction possible within three years — usage balloons via the Jevons paradox. He dismisses public benchmarks as saturated, with models having memorised browsecomp.

Market math: a share of inference spend. He estimates 5–20% of inference GPU spend will go to web-search infrastructure, so as inference grows 3–7x year-on-year, Parallel scales alongside. If inference totals ~$200–300B a year,and Parallel holds around a third of the search layer,that supports a roughly $6–7B revenue business, which he argues can reach a $100B valuation within a few years if agents deliver value. The risks are execution, content partnerships, trust,and ~5x uncertainty in eventual share.

From a pull web to a push web. Always-on agents should not poll every few hours: Parallel's Monitor API flips search into an event-triggered system, crawling the web continuously and spending compute only when something relevant changes —cutting compute 10–15x for watch-style use cases, such as alerting when an under-25 founder registers a new company. The web becomes a push event stream, not a pull query stream.

Safety, Elon lessons and what's next. On guardrails,he argues model alignment is not adversary-proof and builders bear responsibility to minimise harm, calling the "badge of honour" framing of model-led hacks an embarrassment: each demonstrates capability and insufficient guardrails. He admires Elon Musk's urgency,ability to compress time,and unreasonable expectations of people,who often exceed what they thought possible when pushed. He is glad Birdwatch — now Community Notes — survived his Twitter tenure,and his spookiest belief is that most people will eventually trust agents with bank accounts and logins, just as credit cards online and online dating became normal. He also worries AI might diffuse the wrong way, leaving gains too concentrated.

Read more: 20VC