Company@X — 2026-05-18#

Signal of the Day#

Cursor is graduating from an API consumer to a foundational model builder, announcing a joint training run with SpaceXAI on a “million H100-equivalent” Colossus 2 cluster. This marks a major strategic shift in the AI ecosystem, as leading application layer companies increasingly move to vertically integrate their models to reduce reliance on external APIs.

Key Announcements#

[Cursor] · Source Cursor launched Composer 2.5, built on Moonshot’s open-source Kimi K2.5 base, boasting significant intelligence gains and up to 10x efficiency improvements over similar models. More importantly, the company revealed an ambitious partnership with SpaceXAI to train a massively larger model from scratch. Leveraging Colossus 2’s compute and combined RL techniques, Cursor expects this to be a major leap in its underlying coding capabilities.

[Anthropic] · Source Anthropic announced the acquisition of Stainless, an SDK and Model Context Protocol (MCP) server platform. Stainless has powered Anthropic’s SDKs since the earliest days of its API, marking a clear move by the AI lab to bring critical developer ecosystem tooling and infrastructure in-house.

[Hugging Face / Dell] · Source At Dell Technologies World, Hugging Face and Dell announced a major enterprise push for on-prem AI. High-performing open-source models—including Kimi K2.6, DeepSeek V4 Pro, GLM 5.1, and MiniMax M2.7—are now available via the Dell Enterprise Hub, optimized to run on Dell’s PowerEdge XE9780 servers equipped with NVIDIA’s new B300 accelerators.

[NVIDIA] · Source NVIDIA confirmed that SpaceX is deploying its new Vera CPU. NVIDIA explicitly positioned the Vera silicon as being “purpose-built for agentic AI,” signaling a growing hardware specialization and demand for architectures optimized beyond standard GPU workloads for autonomous agents.

[Google] · Source Ahead of its I/O conference, Google globally launched “Preferred Sources” in Search, allowing users to heavily weight specific websites in their search results, resulting in a 2x increase in click-through rates for those domains. The company also teased broader product updates slated for tomorrow across Search, Gemini, DeepMind, GoogleAIStudio, and “Antigravity”.

[OpenClaw] · Source The open-source AI agent framework released version 2026.5.18, shipping a massive suite of reliability upgrades. The release introduces real-time voice sessions for Android via Gateway relays, robust xAI/Grok OAuth sidecar authentication, and better modal awareness for browser automation tools.

Also Noted#

• [Georgi Gerganov] (Source): llama.cpp added MTP support for the Qwen3.6 family, pushing local inference speeds from 25 to 45 tokens per second on commodity hardware.
• [xAI] (Source): Grok introduced “Skills,” enabling users to automate complex workflows using both prebuilt and custom integrations.
• [Y Combinator / InsForge] (Source): YC-backed InsForge crossed 10k GitHub stars for its platform that turns underlying infrastructure into actionable primitives for AI coding agents.
• [Andreessen Horowitz] (Source): Gaby Goldberg joined a16z as a Partner on the New Media team, focusing on the intersection of communications, networks, and investing.