Aravind Srinivas: Perplexity's Race to Build Agentic Search (AI column)
I saw something interesting. interview Aravinda Srinivasa, CEO of Perplexity AI, at the YC AI Startup School event. Aravind spoke about the company’s journey from a SQL prototype to a multi-million dollar “response engine”, shared plans for the launch of the Comet browser and explained why speed, accuracy and focus help the startup compete with Google, OpenAI and Anthropic. If we talk about the history of Perplexity, then it can be seen in another issue of YC.How To Build The Future: Aravind Srinivas"which I'm talking about." toldI'll just talk about diffs here.
1. Scaling and growth Aravind shared the numbers
- 2024: ARR up to ≈$80 million; December estimate of $9 billion
- 2025 May: Negotiations for the $ Round500 million dollars14 billion
- State ~200 Mandatory use of Cursor and GitHub Copilot AI encoders has accelerated experimentation with 3–4 days 1 hour. 2. Comet browser as a "cognitive OS" Aravind talked about the next big bet from Perplexity, namely the concept of their browser Comet.
- It's a single point of context. Tabs, cookies, sessions and history already contain the personal data required by agents. It is difficult to copy, as the development of the mobile version of the browser takes months, which gives the startup a head start. The browser gives an abstraction over chatbots. Comet manages multiple tasks in parallel; chatbots are perceived as private cases When it comes to technology, This is Chromium fork with support for Chrome extensions, which is available to subscribers (200$/month) It has such features: omnibox chat for information and agency tasks, comet assistant in the sidebar: sums up mail, book hotels, makes purchases, fills out forms Hybrid architecture – local models for fast tasks, cloud APIs for complex tasks Data is stored locally, and strict mode blocks trackers and handles sensitive on-device requests 3. Perplexity is in competition with key players
- Google AI Overviews, Gemini briefs in Chrome, but the CPC model prevents you from giving direct answers, as it is a risk of losing income
- OpenAI There is a Search mode in ChatGPT, there is a browser in the plans, but there is no own index yet.
- Anthropic - There's already a web search in beta. The speed of innovation to run faster than competitors 4. Business model and monetization Aravind spoke about the options for monetization
- Subscription Pro/Max - There's more models, more requests. Margins are high – many pay but do not use the limit.
- Usage-based agents This is a model for automating tasks in Comet. Marginality is average as costs rise due to external API consumption
- Transactions (CPA) For example, selfbook bookings or Firmly shopping. Marginality low (Historically lower than CPC)
- Venture fund This is about investing in the API ecosystem. Profitability here portfolio 5. Principles of organizational culture "The user never makes a mistake" - the product adapts to native queries, does not teach "prompt-engineering"
- CEO personally fixes bugs and it motivates the team Mandatory use of AI tools: “Who doesn’t use Cursor is slower”
Finally, the conclusions. 1. The browser is a logical continuation after the search engine - in the image of Google, in 2008 Chrome 2. Strong brand + high-speed iteration> planningat 10–20 With millions of paying users, a brand becomes a self-sufficient asset, even if larger models copy features. 3. Market opportunities lie in agency scenarios It is necessary to monetize not answers, but actions: booking, shopping, automation of routine. The browser agent provides the infrastructure for use-based charging. 4. The main risk is the loss-making cost of calculations The growth of queries and agency tasks requires optimization of inference; Perplexity is invested in model distillation and hybrid architecture. 5. The Age of Multipolar Search Google’s Antitrust Risks Open the Way to an Alternative n
#AI #Management #Startup #Engineering #ML #Agents #Software #Architecture