Summarize this blog post with:
The most revealing moment of this year's Databricks Data + AI Summit wasn't a product launch. It was a candid admission. Speaking to a room full of customers, CEO Ali Ghodsi said the part most vendors prefer to leave unspoken: running agents is "going to get extremely expensive. We are just scratching the surface."
It's a striking thing for a consumption-priced company to tell its customers, and it captures exactly where enterprise AI stands in 2026. The excitement around agents has run into the reality of the invoice, and for most organizations, the invoice is now driving the conversation.
That tension framed all four days at Moscone, where roughly 30,000 people from 150 countries packed into 800+ sessions. It has quietly become the place enterprise infrastructure bets get previewed before they get made, and this year the guest list proved it: Microsoft's Satya Nadella in a fireside chat, OpenAI's Greg Brockman on stage in person, and Reliance's Mukesh Ambani describing a plan to democratize intelligence for 1.5 billion people.
Here's what you need to know about all 4 of the following topics during each keynote: Context, Control, Cost and Choice. They were presented as foundational ideas by Databricks throughout the event, creating a common theme amongst announcements of new products. The underlying message is clear - many enterprise AI products do not function properly yet.
Ghodsi gave that admission a sharper edge in his keynote. AI, he argued, "does not have an intelligence problem, it has a context problem. If a CFO can't get a model to explain why margins moved, the problem isn't the model's intelligence. It's that the model doesn't know the business. Almost every release below is an attempt to close that gap.
1. Genie One
The biggest product story was Genie, which graduated from a feature into a full family. Genie One, now generally available, is pitched as an AI coworker for finance, sales, and marketing teams that works across structured and unstructured data and actually produces things, documents, reports, scheduled tasks, rather than just answering questions.
Why It Matters - With no seat licensing and a small free monthly allowance per user, the barrier that usually kills adoption, buying seats for people who might use it, disappears. A pilot can start with a handful of real users and scale only where value actually shows up, which is how tools get embedded rather than mandated. For most organizations the constraint on analytics was never talent, it was throughput, and Genie One is aimed squarely at throughput.
2. Genie Ontology
It is more than just a chatbot; it is an endless improving context layer for retrieving the meaning of your business from your tables, dashboards, documents, and connected applications. Databricks provided data showing Genie correctly answered 84.5 percent of actual employee inquiries on the first attempt compared to the general-purpose coding agent who answered only 52.4 percent of all employee inquiries correctly on the first attempt. The 32-point difference between the results from these examples demonstrates the difference between something you see in a demonstration to something that will give a controller the confidence to trust it.
Why It Matters - By learning meaning continuously from your own tables, documents, and usage rather than from a one-time manual mapping, Genie Ontology moves the reliability question away from the model and onto the context around it. The 84.5 versus 52.4 percent gap is the concrete version of that shift. It is the entire difference between a tool they double-check and a tool they rely on, and that reliance is the precondition for AI ever leaving the pilot stage.
The pricing was the plot twist. There's no seat licensing, with up to $10 of free usage per user each month, and you pay only for what you consume. For anyone who has tried to budget a 5,000-seat rollout, that's a deliberately low bar to a real pilot.
1. LTAP
LTAP, or Lake Transactional/Analytical Processing was another step forward for databricks. It unifies transactional, analytical, and streaming data on one copy of storage, with no ETL pipelines and no replicas, by storing operational Postgres data in Delta and Iceberg format from the moment it's written.
Why It Matters - The impact is structural, not incremental. Every pipeline between an operational and an analytical system is both a cost center and a failure point. Collapsing that to a single copy removes the plumbing, the duplicate storage, and the sync lag in one move. Engineering time shifts from maintaining pipelines to building products, and analytics start reflecting the business now rather than last night's batch. If it holds up under real load, it's the bet most likely to reshape how stacks get designed over the next 12 to 18 months. Smart teams will pressure-test the consistency model before betting the quarter on it, but the direction is hard to ignore.
2. Lakebase
Sitting underneath is Lakebase, the serverless Postgres engine now running 12 million database launches per day. It gained git-style branching, cross-region disaster recovery, and, most usefully, native hybrid vector plus full-text search. If that search delivers, it quietly removes the separate vector database most agent architectures bolt on today. One less system to run, one less thing to govern.
Discover how Lakebase combines transactional speed with the Lakehouse architecture to power the next generation of real-time AI applications.
Databricks Lakebase Explained Guide
3. Lakehouse//RT
The third piece is Lakehouse//RT, a real-time layer that delivers sub-100ms latency directly on governed Delta and Iceberg tables. It runs on a purpose-built new engine called Reyden, designed from the ground up for the fast, concurrent queries that live dashboards and customer-facing apps demand.
Why It Matters - This lands hardest for any team that stood up a second database just to make something feel live. That serving tier carries its own cost, its own sync, and its own tendency to drift from the source of truth. Running sub-100ms queries directly on governed tables removes it entirely. The payoff is fewer systems to run, one less place for numbers to disagree, and governance that carries over automatically.
1. Agent Bricks
Agent bricks has matured into a full developer platform, and the usage numbers explain why Databricks keeps talking about cost. The platform now has 100,000+ agents built on it, processing over one quadrillion tokens a year. A quadrillion. That's the scale that turns Ghodsi's "extremely expensive" warning from a soundbite into a line item.
2. Omnigent
Databricks also used the moment to make a bet against lock-in. Omnigent, a new open-source meta-harness, lets teams combine and switch between agent frameworks, a "harness of harnesses" that means you aren't trapped in one vendor's approach as the tooling keeps shifting. It's a quietly confident move: own the platform, and you can afford to be open about the frameworks running on top.
Why It Matters - Omnigent earns its place because the agent tooling underneath everyone is still moving fast. Committing to one framework today risks an expensive rebuild the moment a better approach arrives. An open meta-harness that lets you combine and swap frameworks makes that commitment reversible: adopt what works now without stranding the work later. Rebuilds are among the most expensive and least visible costs in enterprise software, so that flexibility has a real financial payoff, and it signals Databricks believes the platform, not the harness, is where value sits.
3. Unity AI Gateway
This might be the most quietly important release of the week. It brings spend visibility across providers, hard spend caps, smart routing to balance quality against cost, and PII and prompt-injection guardrails, all governed through Unity Catalog and extended to models, agents, and MCP services. The message from every governance session was blunt and worth repeating: set this up before you scale agents, not after. It sounds obvious. Most teams still don't do it, and retrofitting governance onto agents already in production is the expensive way to learn the lesson.
4. Unity Catalog
Underpinning all of this is Unity Catalog, which expanded from a governance catalog into the semantic layer that grounds every agent. The headline addition is Unity Catalog Metrics: define a KPI like revenue or churn once, and every dashboard, API, and agent answers it the same way, instead of three teams reporting three different numbers. With a new business glossary and domains for shared meaning, it's what lets an agent's answer actually be trusted, which is the whole point.
Databricks also widened the on-ramp. Its Free Edition, now past 500,000 users, added Genie Code, serverless GPUs, Lakebase, and Agent Bricks at no cost.
The summit was really an argument about where enterprise AI goes next, and most of it applies well past the Databricks ecosystem.
The first shift is that context, not model quality, is now the constraint. For two years the industry chased more capable models. The Genie Ontology results, where a grounded system answered correctly far more often than a strong general-purpose agent, point to a different bottleneck: a model that does not understand your business will fail no matter how clever it is. That reframes the work. The advantage moves to whoever models their meaning, their metrics, definitions, and relationships, most rigorously, and away from whoever simply buys the biggest model.
The second shift is that cost has become an architecture decision, not a procurement one. A quadrillion tokens a year is the number that turns agent spend from a footnote into a design input. When running an agent well is expensive, choices like which model handles which task, where spend caps sit, and how much redundant infrastructure you maintain stop being afterthoughts and start shaping the system from day one. This is why governance and cost control, the least glamorous parts of the stack, were treated as headline features.
The third shift is consolidation. LTAP collapsing the pipeline between operational and analytical data, Lakehouse//RT removing the separate real-time tier, and Lakebase folding vector search into the database all point the same way: fewer moving parts. For years the default answer to a new requirement was to bolt on another specialized system. The summit's bet is that the sprawl itself has become the problem, and the systems that win will be the ones that reduce the number of things a team has to run and govern.
None of this is unique to Databricks. Every major data and AI platform is converging on the same three ideas, which is what makes them worth watching regardless of which vendor you have standardized on. The lesson of DAIS 2026 is that the next phase of enterprise AI will be won on foundations, governed context, controlled cost, and consolidated infrastructure, rather than on model horsepower alone.