Databricks says it opened One Sansome last week at 201,000 square feet, growing to 235,000 square feet by next fall.
The detail that matters is Adaptive Instructed Retriever reserves sequential search for only complex, multi-step prompts.
That default decides how agents find tables, notebooks, dashboards and documents without brute-force exploration on every request.
MemCurrent adds the other half by keeping relevant memory consistent and fresh, so Genie and Agent Bricks Memory do not start from zero each turn, according to Databricks.
KARL is pitched as custom enterprise models that beat frontier models on inference cost, latency and quality at once, according to Databricks.
You feel that tradeoff in production when an agent hammers a warehouse with repeated retrieval calls and slow context rebuilds.
Lakebase, Genie and Unity Gateway are the product surface where that work lands, with Lakebase as serverless Postgres for agents, Genie as an always-on work agent and Unity Gateway for governance and cost controls, according to Databricks.
One Sansome now houses that AI research and product work alongside customers Databricks names as OpenAI, Anthropic, Cognition, Cursor and Decagon.
Databricks says it employs more than 12,000 people, serves more than 20,000 customers and crossed a $7 billion revenue run rate in August with greater than 80 percent year-over-year growth.
Forge lists Databricks stock at $288.40 and cites strategic funding at a $188 billion valuation, a touch below the $190 billion figure Databricks cites, according to Forge.
Databricks ties the building to a $1 billion-plus San Francisco investment by early 2028 and a Data plus AI Summit commitment through 2030 that drew more than 30,000 attendees this year, according to Databricks.
If you build on this stack, retrieval strategy and persistent memory will move your bill more than model choice.
