Weekly AI News — 17 Oct 2025

This week brought big infrastructure deals, new enterprise integrations, and a notable UK-focused skills push. Here are the updates that matter most to non‑technical leaders running SMEs and charities in the UK.

OpenAI

• OpenAI and Broadcom to co‑develop 10GW of custom AI accelerators (2026–2029). OpenAI will design the chips and systems; Broadcom will develop and deploy racks, with Ethernet networking across scale‑up and scale‑out clusters. First deployments are targeted for H2 2026. Official announcement. What it means for UK teams: competition in the AI chip supply chain should improve capacity and, over time, pricing predictability for cloud AI workloads. Expect better availability for heavy research/analysis jobs in 2026+. ([openai.com](https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration//))
• AMD partnership confirmed (6GW) — context. Last week’s AMD deal locks in an additional 6GW of GPUs, with first MI450 deployments planned for H2 2026. OpenAI post. What it means: OpenAI’s multi‑vendor strategy reduces single‑supplier risk; enterprise buyers may benefit from steadier quota and potentially more stable pricing next year. ([openai.com](https://openai.com/index/openai-amd-strategic-partnership/))

Actions to consider

• When planning 2026 budgets, assume greater capacity for long‑running “deep research” and batch analysis, but keep 2025 quotas conservative.
• Ask vendors about their 2026–2027 compute roadmaps and whether they expect lower wait times or better SLAs as OpenAI’s new capacity comes online.

Microsoft

• “Hey Copilot” hands‑free voice and Copilot Vision on Windows 11. Microsoft is promoting voice‑activated Copilot and on‑screen understanding (“Copilot Vision”), now rolling out with support pages and product content updated 16 Oct. Windows page, support note. What it means: employees can trigger Copilot by voice and get context‑aware help on what’s visible on screen — useful for accessibility and time‑saving. ([microsoft.com](https://www.microsoft.com/en-us/windows/learning-center/copilot-on-windows-11))
• Model choice in Microsoft 365 Copilot (Anthropic models) and Researcher agent enhancements rolling through October. Microsoft flagged Outlook workflow bridges, Library, and “Open in Word” for Researcher; Anthropic models are supported in Copilot experiences. Community Hub, Microsoft 365 blog. What it means: more flexibility to choose models for reasoning and document work, potentially better results and cost control. ([techcommunity.microsoft.com](https://techcommunity.microsoft.com/blog/microsoft365copilotblog/what%E2%80%99s-new-in-microsoft-365-copilot--september-2025/4457317))

Actions to consider

• Ask IT to enable “Hey Copilot” where appropriate and update device policies. Confirm privacy settings for voice activation and Copilot Vision. ([support.microsoft.com](https://support.microsoft.com/en-gb/topic/copilot-wake-word-hey-copilot-82c2259e-8ca9-4518-97cf-a83b0f3a643e))
• Agree rules for when teams should use the Researcher agent versus standard Copilot prompts; set a monthly cap or approval step for long‑running Researcher jobs. ([support.microsoft.com](https://support.microsoft.com/en-us/topic/get-started-with-researcher-in-microsoft-365-copilot-e63ab760-f3de-4c47-ae87-dad601b0e9c4))

Anthropic (Claude)

• New Microsoft 365 connector and enterprise search. Claude now integrates with SharePoint, OneDrive, Outlook, Teams and more via the Model Context Protocol (MCP), bringing work content directly into Claude. Anthropic announcement (16 Oct). What it means: UK organisations standardised on Microsoft 365 can trial Claude with first‑party context, without manual file uploads. Check admin controls and data access scopes. ([anthropic.com](https://www.anthropic.com/news/productivity-platforms))

Actions to consider

• Ask admins to pilot the Microsoft 365 connector with a small group; review permission scopes and retention in line with your AI policy. If you don’t have one, see our AI Policy Pack templates.

Google (Gemini)

• Gemini Enterprise announced (9 Oct). Google is positioning Gemini Enterprise as the “front door” for AI at work, tying together company information and agent workflows. Google blog. What it means: clearer packaging for enterprise buyers; watch for UK pricing and Workspace admin controls. ([blog.google](https://blog.google/products/google-cloud/gemini-enterprise-sundar-pichai/))
• Gemini 2.5 Computer Use model (7 Oct) in preview. A specialised model for agents that operate UIs via the API, available through Google AI Studio and Vertex AI. Google DeepMind post. What it means: “desktop‑automation‑by‑AI” capabilities are maturing; expect more reliable UI agents in service desks and back‑office tools. ([blog.google](https://blog.google/technology/google-deepmind/gemini-computer-use-model/))

Actions to consider

• Run a 2–3 week pilot for a narrow back‑office flow (e.g., expense validation) using Computer Use; measure accuracy, exception rates and time saved.

NVIDIA

• DGX Spark is shipping (13 Oct). NVIDIA’s compact “personal AI supercomputer” offers up to 1 PFLOP and 128GB unified memory for local fine‑tuning and inference, available from NVIDIA and partners. Press release. What it means: for R&D‑heavy teams with data‑sovereignty needs, on‑prem experimentation becomes more feasible without a data‑centre footprint. ([investor.nvidia.com](https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-DGX-Spark-Arrives-for-Worlds-AI-Developers/default.aspx))

Actions to consider

• If you handle sensitive client data, compare 12‑month total cost of ownership of a pilot DGX Spark vs. reserved cloud capacity; include power, cooling and support.

Databricks

• Databricks to train 100,000 people in the UK & Ireland by 2028. Announced 16 Oct at the Data + AI World Tour (London), with >$10m invested in Free Edition learning. Press release. What it means: a practical route to upskill analysts and engineers on data and AI foundations. ([databricks.com](https://www.databricks.com/company/newsroom/press-releases/databricks-train-100000-people-data-and-ai-across-uk-and-ireland))
• Model Serving adds support for GPT‑5 family (preview). For teams already on Databricks, serving GPT‑5 series models is now in public preview. Release notes. What it means: simplifies hosting of frontier models alongside Lakehouse data; confirm licensing and usage caps. ([docs.databricks.com](https://docs.databricks.com/aws/en/release-notes/product/2025/october))

Actions to consider

• Nominate 5–10 staff for the free training pathways and set measurable goals (e.g., pass one Databricks certification by Q1). Align with your practical AI stack.

Snowflake

• Tri‑Secret Secure now supports private connectivity. New capability in early October release improves options for privately connecting your KMS. Release notes (9.31). What it means: stronger security posture for regulated data when using customer‑managed keys. ([docs.snowflake.com](https://docs.snowflake.com/en/release-notes/2025/9_31.html))
• Heads‑up: billing for Iceberg data compaction starts 20 Oct. A new view helps you monitor credits used for compaction; billing begins Monday 20 Oct 2025. Release notes. What it means: expect a small cost uptick if you rely on Snowflake‑managed Iceberg tables; update FinOps alerts. ([docs.snowflake.com](https://docs.snowflake.com/en/release-notes/2025/9_31.html))

Actions to consider

• Enable cost dashboards for ICEBERG_STORAGE_OPTIMIZATION_HISTORY and set thresholds to flag unexpected compaction spend from 20 Oct. ([docs.snowflake.com](https://docs.snowflake.com/en/release-notes/2025/9_31.html))

Meta (Llama/Meta AI)

• Recommendation personalisation will use AI interactions. Meta says it will use people’s interactions with Meta AI to personalise content and ads, with notifications starting 7 Oct; policy goes live 16 Dec. Meta newsroom. What it means: organisations using Meta platforms should review employee and customer communications about AI chat use and data sharing. ([about.fb.com](https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/))
• Ray‑Ban Meta glasses in India add Meta AI features (14 Oct). A regional update, but indicative of rapid voice‑assistant expansion across devices. Meta newsroom. What it means: expect more staff using voice assistants on wearables; update acceptable‑use policies accordingly. ([about.fb.com](https://about.fb.com/news/2025/10/new-meta-ai-features-makes-ray-ban-glasses-in-india-more-useful-fun/))

Actions to consider

• Ask your DPO to assess whether your privacy notices or staff social media guidance need updates ahead of 16 Dec changes. ([about.fb.com](https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/))

xAI (Grok)

• Service incident on 16 Oct (resolved). xAI reported an availability issue affecting registered users; resolved after ~90 minutes. Status page. What it means: reliability still varies across newer AI services; use status alerts and avoid single‑vendor dependence for critical workflows. ([status.x.ai](https://status.x.ai/grok-com/INC73c92399))

Policy and governance — UK

• ICO consultation: “Recognised legitimate interest”. Consultation open until 30 Oct on a new lawful basis introduced by the Data (Use and Access) Act 2025 (not yet in force). ICO page. What it means: DPOs should review implications for AI projects that rely on legitimate interests and consider responding. ([ico.org.uk](https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2025/08/ico-consultation-on-draft-recognised-legitimate-interest-guidance/))

Actions to consider

• Map AI use‑cases using legitimate interests, and draft a consultation response. If you’re tightening AI governance, see our UK AI governance guide and 5‑Day AI Evaluation Sprint.

What to watch next week

• Snowflake: Iceberg compaction billing goes live Mon 20 Oct — check alerts and budgets. ([docs.snowflake.com](https://docs.snowflake.com/en/release-notes/2025/9_31.html))
• OpenAI infrastructure: Further details on the Broadcom collaboration and potential ecosystem partners may drip through; watch OpenAI’s newsroom. ([openai.com](https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration//))
• Windows 11 Copilot: Staged rollouts — expect minor policy and support articles to update as features reach more UK tenants. ([microsoft.com](https://www.microsoft.com/en-us/windows/learning-center/copilot-on-windows-11))

Need a second pair of eyes?

If you’d like a concise plan to pilot these features (with guardrails, procurement questions, and a success dashboard), we can help — from a 30‑day cutover plan to a 6‑week RAG blueprint.