September 2026 Summaries

Agentic AI is increasingly enabling nontechnical users to query enterprise data through natural language, but its iterative model and tool calls can generate heavy, unpredictable workloads that require centralized governance, observability, permissions, and cost controls. Onehouse introduces its AI Gateway as a Kubernetes-native layer combining inference routing and MCP-based tool access, designed to let organizations use their own model-provider credentials, connect to open table formats and multiple catalogs, and deploy in cloud, on-premises, or neocloud environments. The company contrasts its approach with Snowflake Cortex and Databricks Unity Gateway, arguing that those platforms impose varying limitations involving provider choice, proprietary infrastructure, logging costs, serverless dependencies, or support for open formats. Onehouse says its gateway supports OpenAI and Anthropic currently, with AWS Bedrock and Baseten planned, and provides model aliases, failover routing, centralized access policies, credential brokering, traces, latency metrics, and estimated spending data. Future plans include semantic layers for unstructured information, infrastructure optimization based on agent behavior, SQL and tool-call optimization, and broader support for open models and neocloud inference services.

Databricks Photon is a native C++ execution engine that accelerates supported Spark SQL and DataFrame workloads through vectorized, columnar processing and SIMD, requiring no code changes while retaining Spark’s planning layer. It is enabled and included by default on Databricks SQL warehouses and serverless compute, but on job and all-purpose clusters it adds a substantial DBU premium, commonly up to about 2.9 times the standard rate. Photon is most effective for large, CPU-bound scans, joins, aggregations, and Delta or Parquet operations, while short, I/O-bound, UDF-heavy, driver-side, model-training, and stateful-streaming workloads receive little benefit or may silently fall back to standard Spark. Its financial value therefore depends on measured runtime improvement: savings occur only when faster execution offsets the higher DBU rate and infrastructure costs, making workload-specific benchmarks and Spark UI verification important. The text recommends enabling Photon selectively rather than globally, checking supported query stages, addressing general compute waste first, and comparing results against alternatives such as Quanton, a Spark execution engine presented as running in customer cloud environments with per-GiB billing rather than compute-hour charges.

Onehouse introduces a lakehouse-based workflow for creating reproducible AI training datasets from operational data, agent traces, feedback, and business outcomes, addressing the limitations of ad hoc exports that lose lineage, version history, and correction paths. Its platform uses OneFlow to ingest data into open Apache Hudi and Iceberg tables, Quanton to curate, validate, version, and export training examples, and optional Apache Airflow orchestration, while Baseten, Fireworks, and Together AI perform the actual model training and Lakegres provides inference-time context. The approach emphasizes that post-training, fine-tuning, distillation, and agent improvement require carefully selected examples tied to explicit evaluations and outcomes rather than unreviewed production logs. It highlights evidence from legal and coding-agent post-training efforts showing that measurable gains depend on well-defined tasks, realistic environments, expert evaluation criteria, and reliable data selection. A contract-analysis example illustrates how labeled table data can be transformed into supervised fine-tuning examples, while noting the need for balanced negative cases, stable source identifiers, validation, and immutable dataset versions. By retaining source snapshots, lineage, training configurations, and provider handoff records, the system aims to let teams inspect past training inputs, identify effects of changed labels or data, rebuild datasets as feedback arrives, and connect the continuous cycle of training, evaluation, inference, and operational outcomes.