Construction AI Monday: floating data centres land at Posidonia, the UK gets its public-sentiment data, and a London ConTech worth watching

The most material new construction story of the past week wasn't on a UK stage at all. At Posidonia 2026 in Athens between 2 and 3 June, Samsung Heavy Industries signed two memoranda of understanding to commercialise floating data-centre designs. The first is a three-party MOU with Capital Clean Energy Carriers Corp and Lloyd's Register: Samsung Heavy handles vessel technology and construction, Greek shipowner Capital leads site origination and investment, and Lloyd's Register — the British classification society — covers regulatory standards and certification. The second is a bilateral MOU with Lloyd's Register Advisory covering economic feasibility and North American market analysis. One day earlier, at the Innovate APAC 2026 exhibition in Taipei on 1 June, Samsung had also signed a joint development agreement with Supermicro, the US AI server specialist, covering operational verification of AI server infrastructure in river and marine environments. Samsung is developing positioning control and salt-and-humidity barrier technologies for the marine platform; Supermicro is validating the AI server stack against the environmental envelope.

Why this matters for UK construction. A floating data centre is a new typology that sits at the intersection of marine engineering, classification, power infrastructure, and AI hardware integration — and the British classification society is in the middle of all three Samsung agreements. The likely sequencing is that early commercial floating data centres will get classed and certified out of the UK and northern European maritime cluster, drawing in UK marine engineering firms, electrical infrastructure specialists and offshore-construction-management capability. The wider context is the data-centre boom now eating into pipelines across UK contractors and consultants. McKinsey's projection of roughly $7tn into data infrastructure by 2030 has been circulating widely; the National Grid AI Growth Zones work commissioned earlier in the year is now starting to influence planning decisions. None of this guarantees that floating designs win commercially over greenfield, but the Posidonia agreements have moved the conversation from "could you?" to "who will the first commercial customer be?"

The procurement angle: If your firm has marine, offshore, or classification capability, treat this as a pipeline signal worth building a small business-development effort around now. The big offshore wind contractors and the marine consultancies have an obvious overlap with the skills these projects will need.

Worth knowing if any part of your work touches data-centre planning. Public First — the strategy and polling firm — surveyed 2,023 UK adults between 30 March and 1 April 2026, and benchmarked against 2,093 US adults polled in January. The picture is more positive than the loud headlines might suggest, but more nuanced than vendor optimism would have you believe. Nationally, more British people support new data centres than oppose them, and they can see both the national case (UK AI capability, jobs) and the local case (employment). When asked about a council plan to build a data centre within three miles of where the respondent lives, support stays roughly in the same place (35%) but opposition rises from 19% to 25%. The politics, in Public First's own words, are not settled — there's a large middle group that's undecided, neutral or unsure, which is exactly the audience consultation work needs to address well.

The concerns to plan for are concrete. National worry is led by cost to the taxpayer (31%), environmental damage (26%) and job losses as AI improves (23%). The local picture shifts toward disruption during construction (27%) alongside the taxpayer (28%) and environmental (25%) concerns. None of these are insurmountable, but they're not the concerns the industry usually leads its public-affairs messaging with. The relevant counter-evidence — renewables-led power strategies, structured community-benefit packages, transparency on local employment numbers — needs to land before opposition crystallises.

For anyone in development, planning or consultation: Read the Public First survey before your next data-centre planning conversation. The numbers give you a defensible baseline and a clearer view of which arguments actually move the undecided middle.

A UK ConTech AI worth knowing about. Depixen is a London-based technology company building ArchiPick — a modular digital ecosystem for construction designed around semantic-data infrastructure rather than LLM-first design. The technical approach matters because it's genuinely different from most of the ConTech AI category. Where the dominant tooling now stacks an LLM on top of project documents and trades off against hallucinations and audit-trail confidence, Depixen builds on W3C Linked Data principles, RDF, knowledge graphs and sector-specific ontologies. The pitch in essence: deterministic structured data with AI components on top, rather than probabilistic language models with structured-data wrappers retrofitted afterwards.

The proof points are early and the marketing voice is high-flown, but the architecture is the right shape for a real problem. The biggest unsolved issue in construction-document AI is reliability — making sure that a specification reference, a product code or a compliance fact comes back the same way every time, and traces cleanly to a source. Semantic graphs and Linked Data approaches are credible answers to that. For firms that have hit reliability ceilings on LLM-only document-AI pilots — and "the model said something different the second time we asked" is the most common failure mode I hear — Depixen's structured-first approach is worth at least a procurement-options conversation. The caveat is the one that applies to any UK ConTech still scaling: the engineering claims need to be tested on your own data and your own evaluation harness, not the vendor's demo.

Worth doing: If you're stuck on the reliability or audit-trail problem in document-AI, add Depixen to your next evaluation alongside the LLM-first incumbents. Different architecture, different trade-offs, useful to test both.

A shorter note on the wider-AI side. The shift from chat-based assistants to autonomous browser agents that complete multi-step work is now a clear category. Perplexity launched its "Computer" feature — a browser agent that carries out multi-step tasks on the user's behalf — and OpenAI is pushing Codex as an always-on background coding agent. Both are early; both have visible failure modes; both point to the same direction of travel that DCW's vendor pitches were also describing. The Procore CDE relaunch, Bluebeam Max's Claude-via-MCP integration, and the Anthropic Claude Platform on AWS launch I covered Friday all sit on the same axis: agents that complete a task end-to-end rather than draft a paragraph for a human to copy. The pragmatic posture for construction-AI buyers stays the same — keep the human approval gate, demand the audit trail, route the workload to the cheapest model that clears your quality bar.

One last restatement, because Monday morning is the day this gets decided or doesn't. The deliverable for the month is one workflow. Picked deliberately, specified to a quality bar you can defend, run with a human approval gate, written up on LinkedIn in your own voice. The momentum from DCW is the asset; that asset depreciates by Wednesday afternoon if it doesn't translate into a specific decision.

Two practical mechanics. First, get the LinkedIn DCW write-up out by Monday lunchtime. Even a tight 250 words with two specific named stand or session moments and one honest opinion outperforms a polished post on Wednesday. The slop-throttling point from a fortnight ago still matters — drop "delve", "leverage", "harness", "robust", "seamless", "comprehensive", "navigate" if you find them. Second, block thirty minutes today to decide the workflow. Compliance-document search across a live project, an RFI triage agent inside Procore or your CDE, a reality-capture-to-BIM pilot on a refurb, a takeoff trial on Togal or Kreo — any of these are defensible. Whichever has the strongest sponsor in your business is the right answer; the worst answer is "three of them at once".

Today's action: Decide the workflow before lunch. Post the DCW reflection before the day ends. The week after that is for delivery.