Tech Digest – March 25, 2026
The OpenAI Reconfiguration
OpenAI Kills Sora, Renames Its Product Org “AGI Deployment,” and Assembles a $120 Billion War Chest
OpenAI has finished pretraining its next flagship model, codenamed “Spud,” and expects it to accelerate the economy within weeks. To clear the runway, the company is shutting down Sora — its video generation platform — and has renamed its product organisation to “AGI Deployment.” Sam Altman is handing off direct control of safety and security teams to focus on raising capital, supply chains, and building data centres at planetary scale. The collateral damage includes Disney, which has ended its partnership with OpenAI — including plans for a $1 billion stake — as part of a broader pivot toward business and coding ahead of a potential IPO as early as Q4. An additional $10 billion brings the record funding round to $120 billion.
Note: Yesterday in this digest, Jensen Huang declared on camera that AGI has been achieved. Today, OpenAI put the word in an org chart and Arm put it on a chip (see below). The label has left the debate stage and entered the product catalogue. When the largest AI lab renames its product division “AGI Deployment” — not “AI Products,” not “Platform” — that’s not a technical claim. It’s a market signal that reshapes how every vendor, partner, and procurement officer frames what they’re buying.
Sources: The Information, Wall Street Journal, Variety, CNBC
Compute Architecture & Efficiency
Arm Unveils Its First “AGI CPU” — Meta Co-Develops, $25 Billion Revenue Target by 2031
Arm has made a dramatic departure from its three-decade role as a neutral IP licensor, unveiling a CPU designed from the instruction set up for AI workloads — and branding it the “AGI CPU.” The chip claims twice the efficiency of x86 on the most demanding AI tasks. Meta is the lead development partner, co-designing for gigawatt-scale infrastructure alongside its custom MTIA accelerators, with Cerebras, Cloudflare, and OpenAI among the launch partners. Arm projects $25 billion in revenue by 2031, with $15 billion from AGI CPU sales alone — against just $4 billion in total revenue in 2025.
Note: “AGI” on silicon. The label that Jensen Huang put into headlines yesterday is now a product category. For EU infrastructure planners, the more concrete signal: the x86 duopoly that has defined institutional IT procurement for decades just got a serious challenger purpose-built for a different workload — and the partner list reads like a who’s who of the organisations building the next generation of compute infrastructure.
Sources: Financial Times, Arm Newsroom, CNBC
Google’s TurboQuant Squeezes 8x More Performance from Existing H100 GPUs
Google Research introduced TurboQuant, a technique that quantizes the key-value cache to just 3 bits — without retraining or accuracy loss — achieving up to 8x throughput improvement on Nvidia H100 GPUs. The method works on hardware already deployed, meaning organisations running or procuring AI compute may need significantly fewer chips than recent projections assumed.
Note: The procurement maths just shifted. An 8x efficiency gain on installed hardware means some of last quarter’s capacity projections were written in a different cost regime. This is why locking into long-term compute contracts carries more risk than it appears — efficiency breakthroughs can rewrite the unit economics between one budget cycle and the next.
Sources: Google Research
The Physical Buildout
Microsoft Leases a 700-Megawatt Texas Data Centre as Crusoe Scales Renewable AI Compute 7x
Microsoft has agreed to rent a 700-megawatt data centre in Texas — originally developed for Oracle and OpenAI, adjacent to the Stargate campus. Separately, Crusoe Energy and Redwood Materials are scaling their renewable-powered compute partnership to nearly seven times the original deployment density in Nevada. The physical plant of AI continues expanding, but the Crusoe model signals that not all of it is grid-dependent: purpose-built renewable infrastructure is scaling alongside conventional capacity.
Note: Seven hundred megawatts is roughly the electricity consumption of a mid-sized European city — from one facility. The Crusoe model is the counterpoint worth watching: renewable power purpose-built for compute, sited independently of the grid. As energy conflicts stall data centre projects across Europe, the question of who builds power alongside the servers — rather than drawing from shared infrastructure — may determine which projects get approved.
Robotics Cross the Commercial Line
Amazon Enters Consumer Humanoids — Germany’s Agile Robots Integrates DeepMind Across 20,000 Installations
Amazon has acquired Fauna Robotics, a startup building a 42-inch humanoid capable of walking, gripping objects, and dancing — marking Amazon’s entry into the consumer humanoid market. In parallel, Munich-based Agile Robots has partnered with Google DeepMind to integrate Gemini Robotics models into its 20,000 installed robotic solutions worldwide. Two different market segments — consumer and industrial — crossing the same threshold from demonstration to deployment at scale.
Note: Twenty thousand installed robots gaining foundation-model intelligence isn’t a product launch — it’s a retrofit. The upgrade path for existing industrial robotics just shortened dramatically. And the EU has a front-row seat: Agile Robots is headquartered in Munich, making this one of the clearest examples of European industrial capability meeting frontier AI.
Sources: Bloomberg, Agile Robots
AI Meets the Real World
OpenAI Foundation Targets Alzheimer’s with $1 Billion Annual Budget — MIT Agents Run Physics Experiments Autonomously
The newly organised OpenAI Foundation, funded at $1 billion per year, is prioritising AI to map Alzheimer’s disease pathways and accelerate treatment personalisation. Separately, MIT researchers demonstrated that LLM agents can now autonomously execute high-energy physics analysis pipelines, with Claude Code handling everything from event selection to statistical analysis to paper drafting. Two signals, one direction: AI labs are pointing their most capable models at problems that were previously the exclusive domain of human researchers.
Note: The physics result is especially striking — an LLM agent didn’t assist a physicist, it ran the pipeline. The institutional question isn’t whether AI will change research. It’s whether funding structures, peer review processes, and university staffing models are ready for what “conducting research” means when the researcher is software.
Sources: OpenAI Foundation, arXiv
AI Reads Radar Imagery to Warn Communities of Imminent Landslide Risk
Researchers have developed an AI system that analyses radar satellite imagery to detect millimetre-scale ground movement, identifying communities at imminent risk of landslides and avalanches. The system processes shifts invisible to the human eye across geologically active and mountainous regions, offering early warnings that could transform disaster preparedness for settlements in the Alps, the Pyrenees, and beyond.
Sources: BBC Future
Meta Withholds New Display Glasses from the EU — Regulation Has a Price Tag
Meta’s latest Ray-Ban glasses — featuring a built-in display and expanded AI capabilities — will not launch in the EU. Bloomberg reports the decision stems from a combination of AI Act compliance concerns and battery regulations, compounded by supply constraints. It is one of the most visible cases of a major consumer tech product being withheld from the European market specifically because of the regulatory environment.
Note: The EU’s regulatory framework is designed to set global standards and protect consumers. It also means 450 million people won’t get this product. Every product withheld is a data point in the competitiveness debate that sits alongside the Digital Single Market — and for institutional leaders, a reminder that regulatory ambition and market access exist in tension, not in sequence.
Sources: Bloomberg
