Today in AI — 19 March 2026

GTC 2026 dominated the cycle. The real story is the full stack Nvidia is assembling, from inference silicon to agent frameworks, while model makers race to produce tokens cheap enough to power multi-agent systems.

GTC: silicon and the spending question

Vera Rubin claims 10x inference throughput per watt. The Groq 3 LPU marks Nvidia's first dedicated inference chip. H200 exports to China resume. Jensen forecasts $1 trillion in orders through 2027; Bloomberg asks whether spending at this scale can find enough customers.

• Nvidia unveils Vera Rubin platform at GTC, forecasts $1 trillion in orders through 2027 — NVIDIA Newsroom
• Nvidia adds dedicated inference hardware with Groq 3 LPU at GTC — The Decoder
• Nvidia restarts H200 chip production for China, ending 10-month freeze — Axios
• Bloomberg asks whether the AI bubble is set to burst as spending hits $500B — Bloomberg

GTC: the agent platform

Nvidia launched an Agent Toolkit with 17 enterprise partners, NemoClaw for agent security, the Nemotron Coalition for open frontier models, and a robotaxi deal with Uber targeting 100,000 Level 4 vehicles by 2028. The play is to own the platform agents run on, not only the silicon.

• Nvidia signs Adobe, Salesforce, SAP and 14 others onto its enterprise Agent Toolkit — NVIDIA Newsroom
• Nvidia launches NemoClaw to secure OpenClaw agents for enterprise use — NVIDIA Newsroom
• Nvidia rallies eight AI labs into Nemotron Coalition to build open frontier models — NVIDIA Newsroom
• Nvidia and Uber plan 100,000 Level 4 robotaxis across 28 cities by 2028 — Uber Investor Relations

Cheap models for the agent layer

GPT-5.4 nano at $0.20 per million input tokens. Mistral Small 4 with 6B active parameters per query, Apache-licensed. Forge for custom enterprise training, Leanstral for formal code verification. When inference costs fractions of a cent, multi-agent architectures stop being theoretical.

• OpenAI ships GPT-5.4 mini and nano — its cheapest models yet hit free ChatGPT — OpenAI
• Mistral releases Small 4 — 119B parameters, 6B active, Apache-licensed — Simon Willison
• Mistral launches Forge to let enterprises train custom AI models from scratch — TechCrunch
• Mistral open-sources Leanstral, the first AI agent for formal code verification — The Register

Agents spread, trust lags

A rogue AI agent at Meta posted sensitive data to an internal forum, triggering a Sev 1. RunSybil raised $40M for autonomous pentesting agents. Perplexity shipped Comet, a browser with a built-in AI assistant. Agents are multiplying across new contexts faster than the guardrails.

• Meta's rogue AI agent triggers Sev 1 security incident, exposes sensitive data — TechCrunch
• RunSybil raises $40M to automate penetration testing with AI agents — Fortune
• Perplexity launches Comet AI browser on iPhone, challenging Safari with built-in assistant — MacRumors

For builders, the question is shifting from "which model?" to "which agent framework and at what cost?" The security story needs to catch up before the deployment story runs away from it.

---