NVIDIA Makes the Agent the Unit of Compute at GTC Taipei
Jensen Huang opened GTC Taipei this morning by reframing the data centre as an agent factory. The headline silicon is the Vera Rubin VR200, which NVIDIA put at roughly 50 PFLOPS of FP4 compute and 288 GB of HBM4 — a claimed 3.3x throughput jump over the B300 — and pitched explicitly for multi-node inference and agentic pipelines rather than training alone (ServeTheHome, BIZON). The second reveal was RTX Spark, the Windows-on-Arm consumer chip co-developed with MediaTek that was trailed for months as "N1X" — NVIDIA's first push into the laptop SoC market, with the full CUDA stack and RTX-class graphics aimed squarely at Apple, Intel, AMD and Qualcomm (TechRadar, VideoCardz).
The throughline was agents, not chips. NVIDIA announced an open agent stack — NemoClaw, a framework for building secure private agents, and OpenShell, a runtime for self-evolving ones — plus platform-level support for the OpenClaw harness, with Adobe, Atlassian, Salesforce and ServiceNow named as early adopters (NVIDIA Blog, ChatForest). For a European buyer the keynote cuts two ways: cheaper on-device inference via RTX Spark loosens the cloud dependency at the edge, but the compute layer that the whole agentic stack runs on stays a single-vendor, US-controlled platform. The hardware gets more sovereign-friendly; the supply chain underneath it does not.
Microsoft Build Opens Tomorrow With Windows as an Agent Platform
Build returns to San Francisco on 2–3 June with the same one-word framing as NVIDIA: agents. The expected throughline is the Azure AI Foundry Agent Service maturing past last year's GA into multi-agent orchestration, combined routing across OpenAI and open-weight models, and cost-governance tooling aimed at the runaway-token-spend complaint that dominates enterprise AI budgets (Notebookcheck, Windows News). On-device, DirectML 2.0 is set to abstract Intel, AMD and Qualcomm NPUs behind one API so Windows apps can run transcription, image generation and lightweight LLMs locally.
The most strategically loaded item is Project Polaris, Microsoft's own in-house coding model, expected to replace the default OpenAI model in GitHub Copilot from August (ChatForest). Read alongside the platform owners moving up the stack, the pattern is the same one European buyers have to plan around: the firms that own the runtime are reducing their dependence on any single frontier lab, and the deciding question for a regulated workload shifts from "which model" to "whose infrastructure, under which jurisdiction, with what audit trail."
Brussels' Sovereignty Package Lands Wednesday — With a Legal Definition Attached
After three slips — from 25 March to 15 April to 27 May — the EU Tech Sovereignty Package is now firmly scheduled for 3 June, and this time it arrives with teeth (Digital Watch, Euronews). The core is the Cloud and AI Development Act (CADA), which would establish four statutory levels of cloud sovereignty — graded on control of the service, the supply chain, where AI inference data is processed, infrastructure location, and cybersecurity — that public authorities must weigh in procurement depending on a use case's sensitivity (Implicator.ai, eu-cloud-ai-act.com). Alongside it sits a revised Chips Act with crisis powers to override supplier contracts and fines up to €300,000 for withholding capacity data — and, for the first time, the EU's own legal definition of "digital sovereignty."
This is the single regulatory lead in today's brief, and it earns the slot: a statutory definition of sovereignty plus a procurement-grading scheme is exactly the kind of load-bearing event the news cycle is built around this week. Editor's note: the other two leads are hardware and enterprise; the Sovereignty Package is carried as the one Regulation main, consistent with the standing one-per-day cap, because it is the defining European event of the week. The practical effect is that "processing of AI inference data" stops being a marketing line and becomes a procurement criterion that a buyer can be audited against — which is where jurisdiction and the ability to run a model on infrastructure you control move from compliance overhead to a scored requirement.
Quick Hits
- Computex opens behind the keynote — The show floor opens 2–5 June under the theme "AI Together," with Qualcomm's Cristiano Amon and Intel's Lip-Bu Tan giving their own keynotes and a dedicated zone for AI robots pushing physical AI from software into factories and logistics (NVIDIA).
- The open-weight cost counter holds — The recent Chinese open-weight cohort — GLM-5.1, Kimi K2.6, DeepSeek V4 and MiniMax M2.7 — continues to sit within a point or two of the Western frontier on agentic coding at under a third of the cost, the structural backdrop to every "which model" decision (LLM-Stats).
- Mistral's Paris data centre goes live — Mistral's first Nvidia-powered European facility near Paris comes online this month, funded by a €722M debt round from a French-led bank consortium (Bpifrance, BNP Paribas, HSBC, MUFG), as it targets 200 MW across Europe by end-2027 (Capacity).
- Anthropic's enterprise lead, and its strain — Claude now sits near 34.4% of US business adoption against OpenAI's declining ~36.5% peak, on a $30B+ run-rate, backed by a new $100M partner network — even as outages and rate-limit resets expose the cost of that growth (VentureBeat).
- The week's other dates — Apple's WWDC keynote on 8 June is expected to preview Siri 2.0 with third-party Extensions, and SpaceX is reported to price its IPO around 11 June with the Anthropic Colossus compute deal as a headline AI-revenue line (Apple).