Summarize this blog post with:
Most enterprises attending the 2025 Data + AI Summit in San Francisco arrived with the same question: why isn't our AI working the way it should? Ali Ghodsi, Databricks CEO, gave a direct answer from the opening keynote — the bottleneck isn't the model. It's the data infrastructure underneath it.
That framing set the tone for the entire event. With 81% of Databricks customers already running generative AI and 95% on classical ML, adoption has stopped being the measure of progress. Reliability in production has. That shift in focus was reflected in the announcements that followed — less about introducing new capabilities, and more about making existing ones work reliably at scale.
- Lakebase: A New Database Built for AI Workloads. Ghodsi put it simply: "once you've picked a transactional database and put your data in it, it's nearly impossible to move off of it." Tightly coupled compute-storage systems have enforced that for decades. Lakebase — built on Neon — separates the two: sub-10ms latency, 10,000+ QPS, Postgres-compatible, on open formats inside the Lakehouse. With 80% of new databases now created by AI agents, that separation has become a practical requirement.
- Agent Bricks and Mosaic Gateway: The summit confirmed that hype and production reliability are different things. Agent Bricks is built for that gap: describe the task, and it handles data generation, benchmarking, LLM-judge evaluation, and deployment. Mosaic AI Gateway (now GA) adds Fallback — automatic provider switching when one model fails or costs too much. For teams running multiple AI applications, that control layer matters.
- How Virgin Atlantic Made the Case for Governance: The governance argument came from Virgin Atlantic, not Databricks. Their session explained why human oversight had to be in the Medallion architecture from day one. Unity Catalog's 2025 upgrades follow: a Metric Catalog for consistent KPI definitions, attribute-based access control, and lineage tracking that shows where any number originated. Governance isn't the obstacle to AI. It's what stops AI from creating bigger ones.
- Expanding Access: AI/BI Genie hit general availability with 500% user growth in a year. Databricks One extended the platform to business users — role-based dashboards, natural language querying, no SQL required. Databricks committed $100 million to global education and launched a free tier.
Every person of every skill level should have equal access to work with data and use AI.
- LakeFlow Designer and Lakebridge: Reducing the Data Engineering Backlog. Migration debt is the line item most AI budgets undercount. LakeFlow Designer gives analysts a visual, plain-language way to build production-grade pipelines — backed by Spark, with Enzyme detecting schema changes and optimizing ETL automatically. Porsche Holding reported 85% faster development time. For legacy systems, Lakebridge automates up to 80% of warehouse migration work — relevant for any team still on Teradata or older infrastructure.
Look past the individual announcements and a bigger move becomes visible. Databricks isn't presenting itself as a data platform anymore — it wants to own the full chain, from the point data arrives to the moment an agent acts on it. As Microsoft CEO Satya Nadella observed at the summit, the compounding effects across AI's development stages — pre-training, test-time compute, application maturity — are stacking. June 2026 is when that ambition gets weighed against what enterprises have actually built.
- Agentic AI shift from Pilot to Production: The honest question for 2026 isn't whether agentic AI works — most organizations have already proven it does in controlled settings. The harder question is what breaks when you scale it. The difficulties change to governance, cost control, and failure handling when agents are managing fundamental business activities instead of demos: how do you keep an agent within predetermined parameters, how do you assess whether it's providing genuine value, and how do you step in when it doesn't? The 2026 sessions are centred on those discussions.
- Real-Time Data Engineering for AI-Ready Infrastructure: Most enterprise data pipelines were designed for batch processing — collect, transform, load, analyse. That model doesn't support the latency requirements of production AI. The 2026 sessions reflect a broader shift toward streaming and declarative pipelines that keep data current as AI workloads run on top of it. Toyota's session on real-time data streaming for connected vehicles is one example of where this is already operating at scale.
- AI/BI: From Dashboard to Decision: After 500% user growth in 2025, the Genie sessions at 2026 move past adoption into harder questions — how do you ensure natural language outputs are accurate enough for real decisions, and how do you govern them at scale? Building analytics layers that business teams rely on daily—rather than only for exploratory queries—is the main goal. Atlassian is presenting on exactly that journey, from deployment through production use.
- AI Applications and Agents Beyond Experimentation: A shift from basic models to whole AI systems from applications, agents, and workflows directly linked to business outcomes is suggested by the message surrounding the 2026 meeting. There seems to be less emphasis on "which model wins" and more on creating enterprise-scale systems that are dependable, controlled, and functional.
- What AI Has Actually Delivered: The 2026 sessions include quantified outcomes, not just capability showcases. BP cut cloud spend by 43% and reduced critical job runtimes from days to under 8 hours. Ensemble Health Partners eliminated their pipeline infrastructure on Lakebase across 800TB. The broader expectation for 2026 is clear: AI investments need to show up in revenue growth or cost reduction, not platform adoption metrics.
The tools from 2025 — Lakebase, Agent Bricks, Unity Catalog upgrades, LakeFlow Designer, and Lakebridge — are in production. They were built for organizations that have already moved past the pilot phase and are running AI on core business operations.
The 2026 summit will show how many organizations actually got there — and what the ones that did have in common. The question for 2026 is no longer whether enterprises have adopted AI. Adoption already happened. The question is whether AI systems have become reliable enough to run core operations — and which organizations have learned how to build that foundation. That appears to be the real theme emerging around this year's summit.
See what Polestar Analytics is bringing to the 2026 Data + AI Summit. From Lakebase deployments to agentic AI in production — get a preview of what we're presenting this year.
Explore our 2026 Summit lineup
Kshitij GuptaData Strategist
Manish Sharma
