Construction AI: Google I/O reset, a £1bn west London data centre, and the agent plumbing settles

Google I/O 2026 opened last night at Shoreline Amphitheatre with the main keynote livestreamed worldwide. Headlines from Day 1: Gemini Omni, a single-architecture model that folds text, image and video generation into one system; Gemini 3.5 Flash, the new low-cost workhorse; Gemini Spark, a redesigned creative surface inside the Gemini desktop; Stream to Cursor, deeper IDE integration for developers; and a substantive preview of Android XR for smart glasses, ahead of Samsung's Galaxy Glasses at Unpacked in July. The wider narrative was clear: from chat surfaces that answer questions to proactive agents that do work across Workspace, Android and the browser.

For construction AI buyers, the immediate practical advice is unchanged from yesterday: hold non-urgent tooling decisions until end of week. Pricing for Gemini 3.5 Flash and Omni will affect multiple AI-video and document-automation roadmaps in your existing pipeline, and Day 2 announcements may further reshape the answer.

SEGRO and Pure Data Centres have secured planning approval from the Old Oak and Park Royal Development Corporation for a £1bn, 72MW hyperscale data centre at Premier Park, west London. The 22,365 sq m scheme - designed by Scott Brownrigg, replacing a redundant warehouse - will deliver three storeys, nine data halls, office space, plant, storage and a dedicated substation. The sustainability brief is solid for a hyperscale build: BREEAM Excellent target, A-rated EPC for offices, PV panels, rainwater harvesting and potential heat recovery. Construction is set to start in 2026.

This continues the now-familiar pattern. The UK AI infrastructure build-out keeps landing in the south-east, with London-adjacent industrial-zoned land the dominant location. For UK contractors and consultants with MEP, civils, ground investigation, grid-connection or commissioning capability, this remains the long-cycle pipeline to invest in - and it is now well-correlated with PlanOps-style compliance and digital-thread tooling demand on the build itself.

IBM Technology's Day-1 video from this week ("MCP vs ADK: How Modern AI Agents Connect and Work Together") has crystallised what was a confused buying question. The framing: Anthropic's Model Context Protocol (MCP) is the agent-to-tool layer - how agents connect to data sources, APIs, file systems, GitHub, Slack, Drive and so on. Google's Agent Development Kit (ADK) and the Agent-to-Agent (A2A) protocol are the agent-to-agent layer - how multiple agents coordinate, hand off tasks and share state. They are complementary, not competing.

For construction-tech teams designing agentic workflows - submittal review across procurement and commercial, RFI orchestration between design and site, compliance evidence chains spanning design, fabrication and handover - the practical implication is straightforward. Build for MCP for connectivity, design for A2A/ADK for multi-agent coordination. Avoid frameworks that lock you into one or the other.

A practitioner proposal circulating this week from AI LABS introduces ADLC - Agent Development Life Cycle - as a replacement for SDLC in teams where most code is produced by AI. The seven phases run from hypothesis through to deployment and continuous learning, with explicit slots for simulation and proof-of-value before implementation. It is still a discussion paper rather than a standard, but it is a useful conceptual scaffold for engineering leaders working out how to govern AI-heavy teams without throwing away the discipline that SDLC provided.

Two recent releases worth a fresh look together. Google's Gemma 4 (released April under Apache 2.0) ships in four variants (2B, 4B, 26B MoE, 31B dense) with native multimodal support for video, images and - for the smaller variants - audio. The 31B model ranks #3 on Arena AI as of April with 89.2 per cent on AIME 2026 and 80 per cent on LiveCodeBench. Microsoft's MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 (released April) are Microsoft's first production models built independently of OpenAI - MAI-Transcribe-1 achieves 3.8 per cent average WER across 25 languages at roughly 50 per cent lower GPU cost than leading alternatives; MAI-Voice-1 produces 60 seconds of expressive audio in under one second on a single GPU.

For UK construction firms with data-residency or PI concerns about cloud-only AI, the open-weight quality story has materially improved. Gemma 4's Apache 2.0 licence (no MAU caps, no acceptable-use enforcement) is the critical detail - it removes the procurement objection that has blocked many "local AI" trials.