{
  "video": {
    "id": "1UmZHb_E_SM",
    "title": "How Web Data Infrastructure Powers the Next Generation of AI — Patricija Žemaitytė, Oxylabs",
    "duration": 1143,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "1UmZHb_E_SM",
    "title": "How Web Data Infrastructure Powers the Next Generation of AI — Patricija Žemaitytė, Oxylabs",
    "one_liner": "An Oxylabs product manager tells three war stories — a 5-petabyte/month video API built in two weeks, a sub-second SERP API rebuilt from a 4-second baseline, and scaling a web unblocker from 10k to 60k requests/second — to argue that the next generation of AI will be powered by better infrastructure around models, not better models.",
    "summary": "Patricija Žemaitytė argues that training alone is no longer enough: models need fresh, live external data, and the infrastructure layer that fetches it is where the hard work now lives. She walks through three escalating episodes at Oxylabs — adapting product scope (a video-download request that grew into transcripts, subtitles, search and metadata, becoming a full video API suite in ~3 months), adapting architecture for latency (rebuilding search delivery from ~4s average to 550ms average, 650ms P90 on the first rebuild), and adapting systems for scale (10,000 → 60,000 requests/second in under two months, now becoming 'project 150'). The recurring lesson is that clients don't buy the first iteration, they buy your ability to adapt — 'this is not a build-once business, this is an adapt-forever business.'",
    "key_points": [
      "Oxylabs (founded 2015) positions itself as a web intelligence platform and premium proxy provider; the speaker's claim is that the harder job is the infrastructure layer for extracting public web data at scale, not the proxies themselves.",
      "A sales trip to San Francisco produced a video-API-for-AI-training request: two-week deadline, at least 5 petabytes per month. The team had never built anything like it and shipped on time — but that was only version one.",
      "Requirements kept mutating: transcripts were built and every request failed, because the client actually needed subtitles; then multi-language video search; then metadata and channel information. In roughly 3 months it became a whole video API suite with its own internal glue library. By 2026 the client had gathered 30 petabytes — and payment was still outstanding.",
      "AI infrastructure is going multimodal: pipelines now need video, metadata, transcripts, subtitles and the structural context around content, not just text.",
      "SERP data changed role under AI — from SEO/analytics/market intelligence to feeding retrieval pipelines, grounding answers and letting agents use live information instead of stale training memory; the speaker cites Google's grounding documentation positioning Google Search as the connection to current public knowledge.",
      "A 2024 request for sub-second SERP delivery (against a ~4-second average baseline) was built, never tested by the client, and shelved because the market wasn't ready. The insight was that the fix wasn't making the old scraper faster: the regular scraper maximizes coverage (ads, widgets, rich results, AI-generated results, varied layouts), while a fast search API narrows to what AI systems need — organic results, top stories, news — and cuts the heavy layout.",
      "A 2025 client demanded zero data retention, sub-800ms latency, geolocation and query parameter support, in under two weeks. A from-scratch rebuild hit ~650ms P90 — then got badly blocked live on the client call, forcing a second, harder iteration that had to lean on browsers, which are 'slow, expensive, complex, and deeply incompatible with dreams about low latency.'",
      "Speed came from hunting seconds across layouts, parsers, sessions and proxies rather than one breakthrough. Today: fast search API at 550ms average latency, and scale from 400 million to almost 6 billion daily requests — described as a change in operating model, not just growth, affecting cost, observability and failure domains.",
      "'Project 60' scaled the web unblocker from ~10,000 to 60,000 requests/second in under two months, where a request means an end-to-end scraping job (routing, rendering, proxy handling, browser execution, parsing, retries, normalization, delivery) — so adding 2,000 more servers doesn't solve it.",
      "The bottleneck surfaced in load testing, not an outage: synthetic traffic was easy, organic-like traffic was hard, and they hit a wall around 20,000 requests/second where the real problem was uncertainty about whether the system could go further. Telemetry itself became part of the load; they scaled gradually and accepted that real testing happens with production traffic. Project 60 is already becoming project 150 (~100,000 requests/second)."
    ],
    "takeaways": [
      "Design retrieval APIs around what AI systems actually consume — organic results, top stories, news — rather than trying to speed up a maximal-coverage scraper; scope reduction is the latency lever.",
      "Treat sub-second delivery as a product constraint, not an optimization target: at 4 seconds you have a slow pipeline, under a second you have something that can sit inside an AI workflow.",
      "Get latency from accumulated small cuts across layouts, parsers, sessions and proxies; expect browsers to be the thing fighting your latency budget.",
      "Don't trust dev or synthetic-traffic test results — build load tests from organic, real-usage-like traffic, scale gradually, and accept that final validation comes from production traffic; budget for telemetry becoming part of the load at scale.",
      "Expect requirements to move (transcripts vs subtitles), and build for repeated adaptation — clients buy your ability to adapt, not the first iteration."
    ],
    "topics": [
      "web-scraping",
      "infrastructure",
      "latency",
      "scaling",
      "serp",
      "grounding",
      "multimodal-data",
      "observability"
    ],
    "tools": [
      "Oxylabs",
      "Google Search",
      "Google grounding documentation",
      "Oxylabs fast search API",
      "Oxylabs Web Unblocker",
      "Oxylabs Video API suite"
    ],
    "quotes": [
      {
        "text": "The next generation of AI will not be powered by better models. It will be powered by better infrastructure around it.",
        "at": "17:47",
        "url": "https://www.youtube.com/watch?v=1UmZHb_E_SM&t=1067s"
      },
      {
        "text": "This is not a build-once business. This is an adapt-forever business.",
        "at": "17:21",
        "url": "https://www.youtube.com/watch?v=1UmZHb_E_SM&t=1041s"
      },
      {
        "text": "There is a difference between system that works in development, system that works in a test, and system that actually survives reality.",
        "at": "09:59",
        "url": "https://www.youtube.com/watch?v=1UmZHb_E_SM&t=599s"
      },
      {
        "text": "And this is how systems becomes fast, not by giant breakthroughs as we thought at first, but by small decision that adds up.",
        "at": "11:05",
        "url": "https://www.youtube.com/watch?v=1UmZHb_E_SM&t=665s"
      }
    ],
    "words": 3059
  },
  "summary_url": "/#1UmZHb_E_SM",
  "transcript": {
    "html": "/transcripts/1UmZHb_E_SM.html",
    "txt": "/transcripts/1UmZHb_E_SM.txt",
    "vtt": "/transcripts/1UmZHb_E_SM.vtt"
  }
}