May 7, 2026

The end of the pointer era: Chang She on rebuilding data infrastructure for AI

The end of the pointer era: Chang She on rebuilding data infrastructure for AI
Spotify podcast player badge
Apple Podcasts podcast player badge
Spotify podcast player iconApple Podcasts podcast player icon

For 50 years, the pattern has been the same: store the data in the database, keep the big files somewhere else, link them with a pointer. It's powered most production systems we've ever built. Chang She thinks AI is about to break it.

In this episode, Pete Soderling sits down with Chang She ahead of AI Council SF 2026 to talk about why the old data stack wasn't built for what's coming, what agents are doing to database throughput, and why anyone with a serious background in performance "starts to shake in their boots a little" when they think about agentic data access at scale.


About Chang

Chang She is CEO and co-founder of LanceDB, building modern data infrastructure for AI. Previously, he architected the ML and experimentation stack at TubiTV as VP of Engineering. In the mythical pre-pandemic epoch, Chang was the second major contributor to pandas, CTO/co-founder of DataPad, and a recovering financial quant.


Timestamps

  • 00:00 — Storing blobs inline vs. as pointers: the trade-offs
  • 02:34 — When you've blown past the bandwidth limit on object storage
  • 04:03 — Six months trying to make Spark on Parquet work, and why it didn't
  • 06:04 — The moment Chang decided to build something new from scratch
  • 07:35 — Why Chang wasn't worried about adding another tool to the AI ecosystem
  • 11:34 — Agents are firing 100,000 QPS, and most stacks weren't built for it
  • 13:32 — Latency, scale, and the new ceiling for production AI workloads
  • 14:54 — Pipelines written by agents, not humans
  • 16:12 — From co-authoring pandas to rebuilding the stack on top of it
  • 17:27 — Why Chang predicts "multimodal by default" within three to five years
  • 19:48 — What Chang is most looking forward to at AI Council


Mentioned in this episode

  • LanceDB Blob V2 API and multi-base feature
  • Apache Arrow and the future of database integration
  • The "hodgepodge tax" — what happens when one customer takes 24+ hours to process a single day of data
  • Claude Code, Codex, and agent-driven data pipelines

--

Mark your calendars for May 12-14, 2026 and join us for AI Council, the technical conference built by engineers for engineers. It's the only event where you'll be able to connect with 1,500+ practitioners for 3 days of technical deep dives, battle-tested architectures and production insights in San Francisco, the heart of the AI world.

Visit us at https://aicouncil.com/sf-2026 for the latest updates or subscribe to the newsletter as we release more details about most exclusive AI conference in San Francisco!