Brickster Newsletter

+$54B in five months. Still no ticker symbol.

Term sheet signed. Coatue-led. Expected to close later this summer. The round is reportedly around $3B, per the Wall Street Journal.

On July 16, Databricks announced strategic funding at a $188 billion valuation. That's $54B above the $134B round it closed in February, and it clears the $165B–$175B range that press reports had been circling only weeks earlier.

Unusual move: the company announced before the cash landed. TechCrunch's read, via one investor: so many firms wanted in that Databricks had no reason to keep the new number a secret.

The IPO question stays deliberately vague. Ali Ghodsi has said Databricks will go public eventually — back in June he called 2026 “a terrible year to go public” — and wire coverage now treats a listing as a matter of when, not if, with 2027 the earliest anyone floats. For now: private capital, no earnings calls, and a valuation that tripled in just over a year and a half.

Governance

Unity AI Gateway

Multi-AI governance: model routing, spend caps, guardrails, and audit across agents, MCPs, and external providers. Databricks' label is “multi-AI governance solution”; in practice it's the control plane for enterprises that need to govern AI cost and access at scale.

Coworker

Genie

The AI coworker that turns business data into answers and actions. Fresh capital accelerates Genie One, Genie Agents, and Genie Ontology: the product line that now extends into Slack, Teams, and MCP integrations.

Infrastructure

Lakebase

Serverless Postgres built for AI agents: memory, state, and transactional workloads on open object storage. Lakebase is the database bet behind agent memory and the LTAP architecture Databricks rolled out in June.

Ali Ghodsi, CEO

“Enterprises are moving from tokenmaxxing to valuemaxxing. They don't want to burn expensive tokens on the smartest model for every task — they want the best outcome per dollar.”

Translation for practitioners: the platform bet is multi-model routing and governance, not picking one frontier model and praying. Unity AI Gateway is the product embodiment of that thesis.

🔗 Resources

→ Official press release — $188B term sheet, July 16

→ TechCrunch analysis — the valuation ladder, plus why open models are undercutting frontier coding costs

→ AI Spend Controls in Unity AI Gateway

Spark 4.2 just ate four tools you're still paying for.

A semantic layer, vector search, geospatial, and change data capture. All now ship with Spark itself.

Apache Spark 4.2 shipped this month with 1,900+ commits from 260+ contributors, and it's already available in Databricks Runtime 19 (Beta). You don't need to read the changelog to get the point.

Things you used to bolt on with a separate product, a plugin, or a pile of custom pipeline code are now built into Spark. Here are the four that will actually change someone's week.

01 · Metric views

One definition of “revenue,” everywhere

You know the meeting: three dashboards, three different numbers, and an hour spent arguing about whose “active user” is correct. Spark now has a native semantic layer. Define a business metric once, and BI tools, reports, apps, and AI agents all read the same definition. The arguing stops at the source.

02 · Vector & search primitives

“Find me the closest match” is now just SQL

Similarity search used to mean standing up a separate vector database. Spark 4.2 adds vector distance and similarity functions, normalization, aggregation, and NEAREST BY, a top-K ranking join that answers “give me the 5 closest things to this.” Recommendations and matching, without a second system to run.

03 · Geospatial

Maps without the plugin tax

Native GEOMETRY and GEOGRAPHY types plus the standard ST_* functions are now built into the engine — on Databricks, GEOMETRY is GA while GEOGRAPHY is still in Public Preview. “Which stores are within 5 miles of this customer” stops being a library-compatibility project and becomes a normal query.

04 · Auto CDC

Data that keeps itself up to date

Change data capture is now first-class: Auto CDC lands in Spark's declarative pipelines, and a new CHANGES clause reads row-level changes as a plain query. Keeping a table in sync with a constantly-changing source used to be bespoke pipeline code that someone owned and nobody wanted to touch. Now it's a feature.

There's a pattern here. Python also got faster and friendlier in this release — an Arrow-first execution path, quicker PySpark UDFs, and better support for calling Spark from services and AI agents. But the headline is consolidation.

Spark is absorbing the surrounding ecosystem. A semantic layer, a vector store, a geospatial extension, and a CDC tool were four separate line items eighteen months ago. Before you renew any of them, it's worth checking whether the engine you already run now does the job.

🔗 Resources

→ Introducing Apache Spark 4.2 — the full feature rundown

→ spark.apache.org/news — official release announcements

Muse Spark 1.1 and Inkling land on Databricks

Two model launches landed on Databricks governed from day one. Inkling — Thinking Machines Lab's debut open-weights model, a 975B-parameter MoE with 41B active — arrived July 15, the day Mira Murati's lab released it. Meta's Muse Spark 1.1 (no relation to Apache Spark; Bloomberg pegs it as Meta's first paid-API model) followed on July 17, eight days after its July 9 launch, via Unity AI Gateway's new Model Provider Services, in preview across AWS, Azure, and GCP. Register a model once, then govern it like any other asset: Unity Catalog GRANT/REVOKE, spend controls, and token, latency, and cost observability.

Genie escapes the workspace: mobile and Teams

Genie One got native iOS and Android apps (Public Preview, July 14), joining the existing Genie app for Microsoft Teams (Beta) that you can DM or @-mention in channels. There's no separate mobile backend to secure — SSO and MFA run through your IdP, and permissions stay enforced by Unity Catalog and source-native ACLs.

Claude's audit trail becomes a data source

A new Anthropic connector in Lakeflow Connect (Beta) ingests Claude audit activity, users, roles, groups, settings — and Claude chats and messages — straight into the lakehouse. Neat inversion: the platform that governs your AI now treats your AI's own usage as just another governed dataset you can query, monitor, and retain.

Databricks wants to mint the context engineer

Three posts in two days (data-native agents, unified context, the skills gap), capped by the Databricks Certified Context Engineer Associate, billed by Databricks as an industry first, with a beta exam whose first scheduled sitting is July 29. It's a certification and a discipline claim, not a new role on anyone's org chart. But the land grab is obvious: name the job, then own the curriculum.

Quick links this week:

→ brickster.ai/news — This week: the model-partner double landing, Genie's mobile debut, and the context-engineering push — every story in context, with sources.

→ brickster.ai/releases — CLI v1.8.0 starts auto-migrating existing bundles onto the direct deployment engine (GA since June), Unity Catalog OSS 0.5.1 speeds up credential caching in the Spark and Hadoop connectors, and Terraform provider v1.122.0 ships Postgres CDF resources — version numbers and what actually changed.

→ brickster.ai/assistant — Ask whether Runtime 19 Beta is ready for your workloads, what NEAREST BY changes for vector search, or why GEOGRAPHY is still Public Preview on Databricks.

🤖 Planning the jump to Spark 4.2?

The Brickster Assistant searches our full archive and answers with citations.

Enjoying the Brickster Newsletter?

Forward this to a colleague who works with Databricks.