Data & AI watch, week of 12 to 18 October 2026

The Databricks and Azure Databricks connectors in Power BI and Fabric are moving from the Simba Spark ODBC driver to the Databricks ADBC driver, now generally available. Existing connections keep working. You opt in per connection, by adding Implementation="2.0" to the source step of the Power Query expression, per workspace or for the whole tenant. The Power BI service will move from ODBC to ADBC on its own during the first quarter of 2027. ADBC supports user OAuth, machine-to-machine OAuth and Microsoft Entra ID.

The monthly summary brings, as generally available, a background colour for the chart plot area (bar, column, line, area, combo, ribbon, waterfall, scatter), with conditional formatting, and auto-expanding hierarchies on the axis, compatible with field parameters. Personal bookmarks now keep the width and order of table and matrix columns. The report authoring skill for Copilot can add custom visuals, imported ones included. In preview: a native Gantt visual with milestones and zoom, and a dropdown mode for the list slicer.

Announced last week, Fabric policies (preview) define who can create which item type on a capacity, and who can change a workspace's inbound and outbound access protection or customer-managed keys. They apply to the tenant or to a capacity, once the Policies (Preview) tenant setting is on; Power BI item creation and external sharing will follow. Generally available since 13 October, two admin APIs give a Fabric administrator 24-hour temporary access to a user's My workspace, and an orphaned workspace can be restored for 30 days after the Entra account is deleted.

A new version of the Capacity Metrics app (preview) installs alongside the current one. It adds a health page comparing capacities, for instance on time spent throttled, an hourly heat map of consumption, and a per-item page with fourteen days of history. It installs from the notification in the current app or from the Microsoft Marketplace. In Fabric Data Warehouse, custom SQL pools are generally available: reads and loads are split into two pools, each capped at 50% to start, and the split can be adjusted freely, including by application name.

A materialized lake view can now ingest CSV or Parquet files dropped into OneLake with a single USING OneLake_Files clause, with no separate ingestion step before the bronze layer. Subfolders are scanned, the schema can evolve (DYNAMIC mode) or stay pinned (FIXED mode, which rejects drift), and a __filepath__ column records each row's source file. Lineage starts at the source folder.

Google released Gemini 4 Argon, generally available on the Gemini API (model ID gemini-4-argon) and in the Gemini app for Google AI Pro and Ultra subscribers. Launch pricing is $2 input and $10 output per million tokens, rising to $4 and $20 after 30 November 2026. Google positions it for software engineering, enterprise knowledge work and cyber defence, and highlights benchmark results, including first place on the Vals Index, which are its own figures.

Meta made Muse Spark Open available for download: a multimodal model with 816 billion parameters, 42 billion of them active per request, on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry. Meta presents it as built for coding and agents, and compares it favourably with the best Chinese models, according to its own tests. Meta says it runs on two linked Mac Studios with M5 Ultra chips. The Muse Model License allows internal use and embedding in your products; an AI service above $20 million in revenue over twelve months needs a separate agreement, and the notice Built with Muse Spark Open is mandatory.

On 12 October, the European Commission opened a public consultation, until 9 November, on a draft implementing act setting out the AI Act's penalty procedures: access to the file, protection of business secrets, limitation periods. At an AI summit in Dublin, Executive Vice-President Henna Virkkunen said the Commission had started using its enforcement powers, with requests for information on model safety and monitoring, following the incidents covered last week. No formal proceedings have been opened at this stage.

According to Axios, reported by Gizmodo, the US FTC sent OpenAI and Anthropic civil investigative demands of more than fifty pages each on 14 October, covering security incidents, safeguards, marketing materials and privacy policies. This is evidence gathering, with no accusation at this stage. On 15 October, Anthropic announced it had brought in six people from METR, an independent evaluation organisation, with access comparable to its own risk teams and the right to publish their analyses without prior approval.

Anthropic announced the Anthropic Public Benefit Fund, a wholly owned company to which it commits 5% of its shares, subject to stockholder approval. The board can add 2% for each 50% increase in the company's value above a baseline. The fund may support nonprofits or give directly to specific populations, for example in response to AI's effects on employment; its equity will be used for neither lobbying nor political giving. Governance details will follow in the coming months.