Databricks

Nvidia: The Harness, Not Model, Drives Agent Success

Nvidia published research showing that the software 'harness' around an AI model — handling memory, runtime, tools and supervisory control — can be more important than the underlying model for long-horizon agent tasks. Using a custom harness with a supervising agent, researchers reported Claude Opus 5 achieved a 100% score on the interactive reasoning benchmark ARC-AGI-3, versus 30% without the harness. Nvidia released a harness design called Agentic Variation Operators (AVO) and argued that open harness components give users more control. The story situates Nvidia's findings alongside other research: OpenAI improved scores by adjusting harness settings but did not reach 100%, Microsoft published a study showing models struggle on long-horizon editing tasks, and Databricks noted harness choice can materially affect AI costs.

Two Huge Exits Boost a16z’s Megafund Model

Andreessen Horowitz (a16z) saw two major outcomes this week: SpaceX closed its acquisition of Cursor for $60 billion in stock, and Stripe agreed to acquire OpenRouter for around $8 billion, according to reporting. The combined value of a16z’s stakes in the two companies is reported to be north of $8 billion on roughly $320 million of investment. The deals were driven by a16z’s infrastructure practice led by general partner Martin Casado and involved other a16z personnel including Matt Bornstein and Chris Dixon. The story also highlights large reported revenue run rates at frontier AI firms (Anthropic, OpenAI), ongoing DOJ scrutiny of a16z over board seats, and broader implications for mega-funds, IPOs, and private-market liquidity.

Enterprise AI: Alliances, Ontologies and Lock‑in

This analysis argues the decisive battleground in the AI supercycle is enterprise context — proprietary, localized knowledge trapped inside companies — and that competition is happening at the level of alliances and the layers they open or hold. Palantir has positioned its Ontology (a typed operating model and decision surface) as a junction that it opens beneath but holds above the model, while major model labs (OpenAI, Anthropic) and hyperscalers (Microsoft, Amazon) have shifted their architectures and acquisition strategies this year to capture the layer above models (human implementation, deployment, and business-context harness). Nvidia convened an open-weight/security coalition; the roster of signatories and absences signal whose economics depend on closed versus commoditized models. The piece highlights product launches, acquisitions, and new services (OpenAI Frontier and Deployment Company, Anthropic’s Ode, Nvidia’s open-weight efforts), and warns buyers to score which layer an alliance opens and which it retains — and whether the retained layer is portable.