NVIDIA Ships Nemotron 3 — an Open-Weight Reasoning Model Built for Long-Running Agents
NVIDIA put its own open weights on the table this week. Nemotron 3 Ultra, unveiled by Jensen Huang at Computex and published to the public on 4 June, is a 550-billion-parameter Mixture-of-Experts model (55B active) using a hybrid Mamba-Transformer architecture, tuned for frontier reasoning and orchestration inside agentic systems. NVIDIA claims roughly up to ~6× higher inference throughput than comparable open models (NVIDIA's own figure) on long-running agent workloads, and the weights are downloadable today from Hugging Face, OpenRouter, ModelScope and NVIDIA NIM (NVIDIA Newsroom, NVIDIA Technical Blog, MarkTechPost). Thirteen launch adopters — among them Accenture, CrowdStrike, Palantir and Perplexity — are already building on the family, and NVIDIA shipped an agent toolkit (NemoClaw and the OpenShell runtime, which Canonical and Red Hat are packaging as a policy-controlled open-source agent runtime) alongside it (Cryptobriefing).
The honest framing matters more than the throughput number. Several day-one write-ups call Nemotron 3 Ultra the strongest US open-weight model while noting the open frontier overall is still set by Chinese labs — DeepSeek, Qwen, GLM and the freshly-launched MiniMax M3 (The Decoder). That is the point worth holding onto: the open-weight tier is no longer one lab's project but a crowded, multi-jurisdiction field where the best choice shifts model to model and task to task. For a European operator, a downloadable agentic reasoning model that runs on infrastructure you control is exactly the substrate a curated-catalogue-plus-router platform like Stav is built to exploit — the value is no longer in betting on one model but in routing across the best open weights as they leapfrog each other.
EU's Cloud and AI Development Act Puts Sovereignty Tiers — and Open Source — Into Operative Law
With the Tech Sovereignty Package now published (3 June), the operative text is finally readable, and two details stand out for European buyers. The Cloud and AI Development Act (CADA) proposes an EU Cloud Sovereignty Framework built on four assurance levels, scored on control over the service, control over the supply chain, treatment of data, infrastructure location and cybersecurity (European Commission, Shaping Europe's digital future). This is the mechanism prior briefs have been tracking in draft form — it turns "EU residency" from a contractual nicety into a graded category of public-procurement eligibility.
The second detail is quieter but structural: open source appears as a named principle in the proposal's operative text, not buried in a recital — what one analysis calls the most ambitious open-source commitment in EU law to date (SUSE). US commentary is already reading the package as disguised industrial protectionism aimed at AWS, Microsoft and Google, which is the predictable reaction to a framework that scores supply-chain and infrastructure control as first-class procurement criteria (techUK). For regulated-sector teams, the takeaway is concrete: the assurance level a provider can clear now bounds which workloads it is even allowed to bid for.
Big Law Goes All-In on AI — and the Deal Term Is Data Control
The clearest enterprise signal of the week came from a sector Stav counts as a core market: legal. On 4 June, Kirkland & Ellis — the world's highest-grossing law firm — announced a multiyear partnership with Palantir to build an exclusive "Fund Enterprise Platform" on Palantir's AIP, aimed at private-equity fundraising: drafting fund documentation, preparing side letters, tracking investor agreements and monitoring compliance for PE managers and their LPs (Kirkland & Ellis, Bloomberg Law). More than 1,000 lawyers are expected to use it, against a firm-wide $500M AI commitment over three to four years (CityAM).
The detail that should travel is the deal term, not the spend: Kirkland's framing stresses that confidential client data stays under the firm's control (the precise vendor-access terms are not public). That is the sovereignty-and-confidentiality argument expressed commercially — in a regulated profession, the data-control boundary is now part of the contract, not an afterthought. It lands the same week the Financial Times framed Harvey, Legora and Anthropic's Claude for Legal (launched 12 May) as squeezing incumbents Thomson Reuters and LexisNexis, who are racing to upgrade (Artificial Lawyer). Legal AI has moved from pilots to platform commitments — and the firms making those commitments are negotiating for exactly the data-isolation guarantees that sovereign deployment is designed to provide.
Quick Hits
- Altman and Amodei sign a biosecurity letter — Sam Altman, Dario Amodei and other lab leaders signed an open letter urging stronger tracking of synthetic DNA that could be used in AI-designed bioweapons, the latest sign that frontier labs want governance scaffolding around their own capability gains (Fortune).
- WWDC is T-3 — Apple's 8 June keynote is expected to centre on a rebuilt Siri reportedly running on a custom ~1.2-trillion-parameter Gemini model (~$1B/year to Google), with an Extensions system letting users route queries to Claude, Gemini or ChatGPT — leaving open questions about where EU users' data is processed (TechRadar).
- MiniMax M3 weights watch — M3's full open weights and technical report are still pending, due roughly ten days from the 1 June launch (~11 June); the actual download remains the next real reproducibility test for the open-weight cohort (The Decoder).
- Capital rotates to the AI supply chain — Taiwan's and South Korea's stock-market capitalisations have overtaken India's as investors favour economies whose firms sit in the AI hardware supply chain — a reminder that, for now, the durable AI margins still accrue closer to the silicon than the application layer (Financial Times via Techmeme).
Editor's note: regulation gets exactly one main today — the CADA sovereignty framework — and only because the operative text just became public, making the four-assurance-level structure and the open-source-first principle genuinely new detail rather than countdown coverage. The lead and the enterprise story were both stronger out of the Models/Open-Source and Enterprise buckets, in line with the standing guidance to push regulation toward quick hits when those buckets have credible news. They did.