UK energy gap, Erlangen factory, Omnibus today — and two new features

DSIT has published a forecast that the UK will need at least 6GW of AI-capable data centre capacity by 2030 — a threefold increase on today. Its Compute Evidence Annex, updated April 23, estimates UK greenhouse gas emissions from AI compute over 2025–2035 could reach 34 to 123 MtCO₂, or 0.9–3.4% of projected UK totals. DSIT’s goal: nationally significant sites capable of 500MW+ each, at least one AI Growth Zone exceeding 1GW by 2030.

DESNZ, reported by The Guardian on Sunday, projects far lower demand growth for the broader commercial services sector — a figure incompatible with DSIT’s AI-specific projection. Two branches of the same government, incompatible models. The practical consequence: at DSIT’s forecast, the UK cannot meet demand without dramatically expanding grid connectivity, cutting planning timelines, and addressing electricity prices running at roughly four times US levels. The same pricing barrier paused OpenAI’s UK Stargate project. West London data centre clusters including Slough and Heathrow are already near saturation. DSIT is routing new capacity north via AI Growth Zones — but only where power genuinely exists.

The Science, Innovation and Technology Committee has opened an inquiry into whether low-energy compute architectures can change the trajectory. On current numbers, the UK builds more data centre capacity than its grid can serve.

At CES in January 2026, Siemens and NVIDIA announced an expanded partnership to build the world’s first fully AI-driven, adaptive manufacturing sites globally, with the Siemens Electronics Factory in Erlangen, Germany as the first blueprint site. The factory operates on what Siemens calls an “AI Brain” — integrating software-defined automation, industrial operations software, and NVIDIA Omniverse digital twin libraries. Operational changes are tested virtually then pushed across the site in near real time.

At Hannover Messe last week, NVIDIA confirmed the programme is live. Humanoid’s HMND 01 wheeled robot — running NVIDIA’s Jetson Thor edge AI module, developed in seven months using Isaac Sim and Isaac Lab simulation tools — has completed autonomous logistics operations inside the Erlangen production environment. This is the first proof of concept inside an operational factory rather than a controlled test cell.

Under the partnership, NVIDIA provides AI infrastructure, simulation libraries, models, and frameworks. Siemens deploys hundreds of industrial AI experts alongside its hardware and software portfolio, including the Eigen Engineering Agent announced at Hannover, which writes and troubleshoots the PLC automation code the factory runs on. The goal is to replicate the Erlangen blueprint at Siemens facilities globally, then license the operating system to external manufacturers.

What is agreed: Both Parliament (voted 569-45 on March 26) and Council (position adopted March 13) are aligned on the core postponement. Annex III stand-alone high-risk AI systems move to December 2, 2027. Annex I AI embedded in regulated products moves to August 2, 2028. A targeted ban on AI nudifiers is agreed. Registration obligations for self-assessed non-high-risk systems are restored by both.

What is contested: Sectoral integration. Parliament wants Annex I-A categories (machinery, medical devices, toys) moved to Annex I-B with AI requirements integrated into existing sectoral law. The Council has not accepted this. The Standing Committee of European Doctors has called for medical devices to remain under the AI Act directly. Synthetic content watermarking: Parliament proposes November 2, 2026, Council proposes February 2, 2027. Germany is also pressing for an industrial AI carve-out following Merz’s Hannover statement.

If no deal today: The original August 2, 2026 deadline for Annex III high-risk AI systems applies. Every organisation planning around December 2027 must revert immediately. Formal adoption after a deal is expected May–June; Official Journal publication targeted July 2026.

01 — DeepSeek V4 benchmarksDeepSeek published V4-Pro on Friday with self-reported benchmarks placing it within 3–6 months of GPT-5.4 and above all open-source alternatives in coding and mathematics. Independent verification from LMSYS, Hugging Face, and the UK AI Security Institute is expected this week. DeepSeek’s R1 self-reported claims held up in January 2025. If V4 holds too, it confirms that Huawei Ascend chip-based training produces frontier-class open-source models — a direct challenge to US chip export control policy and EU semiconductor strategy.

02 — Cohere + Aleph Alpha regulatory reviewThe April 24 merger triggers mandatory notification in Canada and the EU. In Germany, the Bundeskartellamt will review the deal given Aleph Alpha’s German public sector market position. The German federal government’s anchor customer role gives this review political weight beyond competition law. Clearance is expected in three to six months. Whether Germany formalises its anchor commitment before or after clearance affects how the deal is structured and valued.

03 — Mythos breach investigationAnthropic’s investigation into the Discord group that accessed Claude Mythos Preview through a third-party vendor environment is continuing. The breach is acutely sensitive: UK financial institutions began deploying Mythos this week, and ESMA chair Verena Ross gave her Reuters interview warning about AI accelerating financial cyberattacks the same day. Anthropic has not disclosed which vendor was breached, the group size, or whether any outputs related to financial system vulnerabilities.