Two of the largest gatherings on the data and AI calendar landed within a couple of weeks of each other this May/June, and anyone who sat through both walked away with a strong sense of déjà vu. The headline messages from the Databricks Data and AI Summit 2026 echoed those from the Snowflake event that preceded it, almost word for word. Both companies declared that frontier models have grown smart enough that the real enterprise bottleneck is context, and that governance, cost control, and openness will decide who leads the agentic era. Databricks put on a very convincing show, delivering clear, strong messages about bringing data islands together into a single cohesive platform to enable enterprise AI.

Here are some of the top announcements from the event which I will dissect in detail below –

  • Genie One and Genie Ontology – A new enterprise context layer plus an agentic coworker, now generally available, with $10 of tokens per user per month included.
  • Unity AI Gateway – One control plane to govern models, agents, MCP services, and skills, with hard spend caps and smart routing across providers.
  • LTAP – A new architecture that unifies transactional and analytical data on a single copy in the lake with no pipelines. Lakebase already handles 12 million database launches per day.
  • Lakehouse RT and the Rayden engine – Real-time analytics clocking more than 12,000 queries per second at subsecond latency.
  • Zerobus Ingest – Serverless, up to 10 GB per second, a petabyte ingested in under 24 hours, and Kafka Producer API compatibility landing this July.
  • Omnigent – An open-source meta-harness, Apache 2.0, that sits above Claude Code, Codex, and Cursor.
  • Two market moves – The Panther acquisition pushes Databricks deeper into security, and CustomerLake takes it into the agentic customer data platform space.

What follows is my read of the announcements that matter, the architecture underneath them, and the places where enterprise buyers should keep their guard up.

Context Became the Center of Gravity

Ali Ghodsi opened the keynote with a deceptively simple argument. AI does not have an intelligence problem. It has a context problem. He walked the audience through the now familiar failure mode of an agent that spends ten to fifteen minutes wandering across documents, links, and MCP servers, burns tokens along the way, and still answers from a fraction of the available knowledge. The fix Databricks proposed is Genie Ontology, a continuously computed knowledge layer that connects to enterprise systems, extracts relationships and subject matter expertise, and ranks that knowledge with a permissions-aware algorithm the company calls OntoRank.

Databricks reported a thirty-plus percentage point accuracy improvement and roughly half the run time against leading generic coding agents, and it pointed to PepsiCo, where Genie inside the SpendWise procurement platform drew 30,000 query engagements within weeks and set up a consolidation of more than 30,000 reports into roughly 15 consoles. That PepsiCo outcome carries more weight than any benchmark, because it is a named customer with an operational result. Genie One, Genie Spaces, Genie Code, etc. all layer on top of Genie Ontology for context.

I have to say that given that Genie Ontology is still in preview mode, it still needs independent validation across messy, multi-source enterprise environments before anyone treats the performance numbers as settled. But, if executed well, agents running on Databricks can be trusted and that will boost the adoption across more enterprise use cases.

Unity AI Gateway Is the Keystone

Unity AI Gateway is the one product that seems to hold the entire strategy together. Databricks extended Unity Catalog from governing data assets to governing the runtime interactions between models, agents, MCP services, skills, and tools, and that single move makes the rest of the portfolio coherent.

The substantive advances sit in the shift from static access control to stateful, contextual control. A new On-Behalf-Of attribute lets policies distinguish an action a human takes directly from an action an agent takes on that human’s behalf. Contextual Service Policies, now in beta, let administrators allow, deny, or require approval at the granularity of a single tool. Column masking can trigger specifically when an agent, rather than a person, makes the call. 

Two design choices elevate the Gateway beyond a feature checklist. Hard spend caps paired with intelligent routing turn cost governance into an enforceable control. The open ecosystem play matters even more. By announcing runtime integrations with CrowdStrike, Palo Alto Networks, Zscaler, Okta, Ping Identity, and others, Databricks positioned the Gateway as a neutral control plane rather than a walled garden, which answers the lock-in objection head-on.

Buyers should still price in the structural reality. A control plane that governs every model call also meters every model call. The same architecture that gives the enterprise unified oversight gives Databricks unified visibility into, and consumption from, the customer’s full AI traffic. That serves good business and good governance at the same time, and both facts hold true.

Zerobus Quietly Contests the Ingestion Layer

Among the technical sessions, Zerobus Ingest earned more attention than its modest billing suggested. Zerobus is a serverless Rust engine that writes producer data directly into Unity Catalog Delta tables and skips the message bus entirely when the lakehouse is the only consumer. The session presenters stated plainly that Zerobus does not replace Kafka for microservice fan-out, only for the common case where teams run a bus purely as a pipe into the lake. Zerobus claims up to 10GB/sec to a single table, 5 seconds to queryability with a 200ms offload, and a benchmark that ingested more than a petabyte into one Delta table in under 24 hours, all backed by a reproducible public blog. 

A Kafka Producer API compatibility layer arriving in beta in July 2026 will let teams redirect a Debezium or Kafka Connect pipeline with a configuration change and no new code. Set against the broader movement around Kafka and the IBM and Confluent dynamic, Zerobus represents Databricks contesting an ingestion layer that streaming vendors have long treated as defensible. It also serves as the front door to a unified real-time stack that pairs Zerobus with Spark Real-Time Mode and Lakehouse RT, though several components of that stack remain on the roadmap rather than in production today.

LTAP and Reyden Show the Architecture Underneath the Story

The most defensible technical claims of the week live below the AI layer. LTAP, which Databricks describes as Lake Transactional and Analytical Processing, makes an architectural argument rather than a packaging one. Storage services run bound by I/O with spare CPU, and LTAP uses that idle capacity to transcode row-oriented writes into columnar Delta in place, with no separate pipeline and a net performance gain from better compression. The on-stage quip that CDC stands for “Continuous Data Corruption” evoked laughter, and the serious point underneath dismissed both HTAP, which collapses workload isolation, and zero-ETL, which hides the pipeline rather than removing it. Lakebase gives the claim a foundation with 12 million database launches per day and named customers that include Block, Zillow, and Superhuman.

Reyden, the engine that powers Lakehouse RT, reflects an equally sharp design philosophy. Rather than starting with database techniques, the Databricks team began with workloads and trained a machine learning model on the platform’s telemetry to select the right algorithm at runtime. The demonstrated figure of more than 12,000 queries/sec at sub-second tail latency stands as a vendor benchmark and deserves that label. That approach still answers a real and longstanding tradeoff that Michael Stonebraker documented decades ago. Design partner results include a 16x speedup at Evers and longest queries cut by nearly one hundred times.

Both LTAP and Reyden land partly as roadmap items. LTAP is coming soon as part of Lakebase, and the comparison tables ran inside controlled demos. But, the early performance numbers and the concept have already created a lot of noise (both positive and negative) in the industry.

Omnigent and the Bet on the Meta-Harness

Omnigent may prove to be the coolest yet left-field move Databricks made all week. It is an open-source meta-harness, released under the Apache 2.0 license, that sits above Claude Code, Codex, Cursor, and custom agents, making them interoperable. Databricks places itself as the abstraction layer above the agent harnesses that its competitors and partners build. Matei Zaharia, CTO and Co-founder of Databricks, framed it as a layer that remains constant while models and harnesses change underneath, applying the Kubernetes analogy to agents.

The commercial logic follows directly. If Omnigent becomes the place engineers actually work, Databricks owns the orchestration, policy enforcement, cost metering, and telemetry for agent work that runs largely on other vendors’ models, and the managed version channels all of it through Unity AI Gateway. Meta-harnesses only matter if the meta-layer gains adoption, and coding-agent vendors have every reason to keep developers within their own tools. This is a bet on a category that does not yet fully exist, which makes it the announcement I would watch most closely over the next year.

The Expansion Into Security and Marketing

Databricks widened its footprint into two adjacent markets, and the strategic logic holds, even as the framing deserves precision. These read more like acquisition-led category entries than organic product builds. The security play combines Lakewatch with three acquisitions, namely Panther alongside the earlier Antimatter and SiftD.ai. Panther carries real weight as an AI SOC platform with 100+ integrations, detection-as-code, and the credibility of defending Anthropic’s own environment. The wedge into legacy SIEM holds up, because incumbent economics force most organizations to analyze only a fraction of their security data, and an open lakehouse with agents attacks exactly that constraint. Read my past post on how I look at the disaggregation of the SIEM market.

CustomerLake invites the closest scrutiny. By entering the CDP (Customer Data Platform) and marketing activation space with its Infinity Campaigns and campaign agents, Databricks moves adjacent to, and partly into conflict with, the partners it relies on for activation. Leadership drew a deliberate boundary in the go-to-market session, stating an intent to own the data layer and cede personalization to Adobe and Salesforce Marketing Cloud. Boundaries of that kind tend to erode once a product team carries revenue targets, and Databricks itself acknowledged that overlap with ecosystem partners will grow as it productizes what customers used to build. I spoke with some of the CDP vendors (who were also Databricks partners) in the expo hall and they were fuzzy in explaining how they would work with Databricks going forward, given this competitive overlap.

The Cost Argument 

Ghodsi’s warning that AI spend will turn prohibitive for most organizations by roughly December 2026, reinforced by the story of a company which consumed an annual AI budget in a single quarter, framed cost as the urgent enterprise problem of the year. I wholeheartedly agree with the sentiment. “Tokenmaxxing” is the word on every CFO’s mind these days. While leveraging tools like Unity AI Gateway for cost governance and policy enforcement will be appropriate, it is still not a limiter on how much AI will be used by an enterprise. In fact, counterintuitively, with so many AI features in Databricks, the token usage will only keep increasing more. And this is not about just Databricks. Every vendor is adding to this problem.

Stratola’s Take

Databricks delivered the most internally coherent platform narrative in the data and AI space right now. Live demos across the keynotes and the sessions earned the confidence of the audience further. The framing of Context, Control, Cost, and Choice gave the portfolio a clean spine, Unity AI Gateway supplied the connective tissue, and LTAP and Reyden contributed architecture that genuinely runs ahead of the marketing. Omnigent represents the most ambitious and least certain bet. The move into security and marketing offers a smart expansion of the addressable market, even as it carries real channel risk beneath the enthusiasm. Overall, I am truly blown away with the way Databricks has delivered at this event and presented their vision for the future as well.

Beyond all that, the one thing that stands out as a key deciding factor in this battle of the data AI platform giants is who can create and deliver the right enterprise context to the agents to make them run accurately, cost effectively and efficiently. Each data platform vendor, including Databricks, wants to deliver the context themselves with the claim that they own and have access to all the enterprise data already. However, there are still several neutral players in the market that want to deliver a strong context layer that can work more seamlessly and openly across multiple platforms. This might be a more recommended way of designing a future-proof architecture for enabling enterprise AI.