Mistral ships Vibe agent + Nvidia pledges $150 billion a year to Taiwan + Cognition hits $26 billion

Mistral replaced Le Chat with Vibe, a single agent platform with a Work mode that plans tasks across Google Workspace, Outlook, SharePoint, Slack, and GitHub before execution, and a Code mode that runs coding agents in isolated sandboxes from a web UI, a CLI, or a new VS Code extension. Pricing starts at a free tier with Pro at $14.99 per user per month and Team at $24.99 per user per month, and every prior Le Chat conversation auto-migrates. The launch consolidates Mistral's chat, agent, and coding surfaces into one license at a moment when OpenAI, Anthropic, and Google are each fragmenting their own offerings across separate apps.

At the groundbreaking for a new Taipei headquarters, Jensen Huang announced that Nvidia will spend $150 billion a year in Taiwan, up from $10 to $15 billion annually four to five years ago. The headquarters will employ 4,000 people and target a 2030 opening, anchoring Nvidia closer to TSMC, which fabricates its chips, and Foxconn, which assembles them into server racks. The pledge lands as Washington has spent two years pushing chipmakers to build US capacity, and it cements Taiwan rather than the United States as the structural center of advanced AI manufacturing for at least the rest of the decade.

Cognition, the maker of the Devin autonomous coding agent, closed a round of more than $1 billion at a $25 billion pre-money valuation, led by Lux Capital, General Catalyst, and 8VC with Ribbit Capital and Atreides joining. The company disclosed $492 million in annualized revenue and named Mercedes-Benz, NASA, Goldman Sachs, and Santander among its enterprise customers. The valuation more than doubles from a $10.2 billion post-money mark eight months ago, a vote of confidence that standalone coding-agent vendors can hold ground even as Anthropic, OpenAI, and Google push their own coding products and Microsoft pulls Claude Code from its own employees.

Robinhood launched an agentic trading beta that lets users create a separate AI-agent account funded by a dedicated wallet. Agents connect over the Model Context Protocol and can analyze concentration risk, look up analyst notes, and place stock trades, with some orders requiring user approval. The agent can only spend pre-loaded balances, and all activity is mirrored to the user's app for review. Options, crypto, event contracts, futures, and prediction markets are slated to follow, putting Robinhood alongside Stripe, Amazon, and Google as the first retail-facing platforms to wire agents into live financial transactions.

Mistral acquired Vienna-based Emmi AI and used the deal to launch Physics AI, a class of data-driven models that learn from physics solver outputs and predict the behavior of physical systems in seconds on a single GPU, replacing simulations that traditionally take hours or weeks. Mistral named Airbus, ASML, Safran, and Siemens Energy as launch partners, targeting aerospace, semiconductors, energy, and industrial equipment. The offering pitches three concrete use cases: exploring thousands of design variants for new products, optimizing factory tooling, and running real-time digital twins on live sensor data — engineering acceleration with named industrial deployments rather than future-tense lab promises.

YouTube announced it will automatically label videos that are AI-generated or meaningfully altered with AI, replacing the current self-disclosure-only system. The platform will combine creator disclosures with detection signals from partners and its own classifiers, and labels will appear in the description by default and in the player for content depicting realistic people or events. The shift covers the largest video platform on the open web at roughly the same moment that the US Federal Trade Commission has begun enforcing the Take It Down Act and that Meta is testing AI-tier paid subscriptions in Singapore, Guatemala, and Bolivia.

Google Research deployed a new privacy-preserving analytics technique into Android SafetyCore that lets safety-model developers measure the true-positive rate of on-device classifiers across a global fleet without ever seeing the private content that triggered each alert. The technique combines a lattice-based cryptographic protocol that allows a single encrypted submission per device, instead of the multi-round handshakes prior secure-aggregation methods required, with trusted execution environments that attest the server code. It is now used to detect model drift, surface hidden classifier biases, and measure Smart Reply error rates at production scale.