What mattered last week in AI

OpenAI builds its own chip

OpenAI and Broadcom unveiled Jalapeño — OpenAI's first custom chip, a reticle-sized inference processor taken from design to tape-out in just nine months, likely the fastest advanced-silicon cycle ever. OpenAI used its own models to speed the design. Early testing shows performance per watt substantially better than today's best, with first deployment by end of 2026. OpenAI is now a hardware company. The race is to own the whole stack — model, chip, and data center.

Anthropic accuses Alibaba

Anthropic told U.S. senators and the White House that operators tied to Alibaba's Qwen lab used nearly 25,000 fake accounts to run ~29 million exchanges with Claude — harvesting its software-engineering and agentic-reasoning skills. Anthropic calls it the largest distillation campaign yet against a U.S. AI company. Senators are already drafting sanctions. The new war isn't about who has the best model. It's about who owns the thinking — and who can copy it.

Vibe coding cost $81,267

An employee at fintech startup Slash set out to build a meme shooter with AI. The bill came to $81,267 in credits — "a genuine accident; I underestimated my own ability." Slash is now reviewing its AI-coding policy (and promoting the game to write off the cost). Vibe coding is free to start. The bill is not. Budgets and guardrails aren't optional.

Figma makes animation easy

At Config 2026, Figma launched Motion — native, timeline-based animation right on the canvas, with keyframes, spring physics, and an agent that builds the animation from a description. Export straight to CSS, React, MP4, or GIF. No After Effects round-trip. The tools keep collapsing hours into seconds. The barrier to shipping polished product keeps dropping.

Qwen builds a world of agents

Alibaba's Qwen team released Qwen-AgentWorld — one model that simulates seven environments (MCP, terminal, browser, desktop OS, Android, search, and software dev). Instead of choosing actions, it learns to predict what the environment does next — a "language world model" that tops GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro on its own benchmark. The agent is no longer just a worker. It's learning to model the world it works in.

Enjoyed this? Get it every week.

Sugata AI Weekly — the AI news that actually matters, in 5 minutes. Free, no spam.

Want to grow with AI?

Sugata AI turns this stuff into shipped software — autonomous agents, MCP tools, workflow automation, and full-stack AI apps. Let's build what's next for your business.