This Week in AI: March 16–22, 2026 (Week 12)

News of the week: March 16–22, 2026 (Week 12)

The week was dominated by Nvidia’s GTC conference, where Jensen Huang unveiled the Vera Rubin platform and estimated chip orders at a trillion dollars. OpenAI released the smaller GPT-5.4 mini and nano models, announced the acquisition of Astral, and, according to media reports, is preparing a single “super app”. Anthropic’s dispute with the Pentagon escalated with new court filings. The White House published a proposed national AI framework meant to override individual state laws. AI agents showed their dark side in a security incident at Meta.

Nvidia GTC 2026

Vera Rubin and a trillion dollars in orders

At the GTC conference in San Jose on March 16, Nvidia CEO Jensen Huang unveiled the Vera Rubin platform. It consists of seven new chips that, according to the company, are already in full production. (NVIDIA Newsroom)

The Vera Rubin NVL72 server rack combines 72 Rubin GPUs and 36 Vera CPUs. Nvidia promises up to ten times higher inference performance (that is, running a finished model) per watt compared with the Blackwell generation. Partner products are expected to be available in the second half of 2026.

Also new is the Groq 3 LPX rack with 256 specialized LPU inference chips, based on technology from Groq. According to CNBC, Huang expects orders for Blackwell and Vera Rubin chips worth 1 trillion dollars by the end of 2027. Last year he was talking about 500 billion.

According to TechCrunch, investors were not convinced. Nvidia shares started falling during the two-and-a-half-hour keynote itself, as fears of an AI bubble weigh on the market.

NemoClaw: safer agents on your own computer

On March 16, Nvidia announced NemoClaw, an open toolkit for the popular agent platform OpenClaw. A single command installs the Nemotron models and the new OpenShell runtime with privacy and security controls. (NVIDIA Newsroom)

According to TechCrunch, it is still early-stage alpha software. It targets security, the biggest weakness of autonomous agents.

March 16: DLSS 5 combines 3D game graphics with generative AI and aims to boost photorealism at a lower computational cost. (TechCrunch)
March 16: IBM and Nvidia expanded their collaboration; IBM wants IBM Cloud to offer Blackwell Ultra chips at the start of the second quarter. (IBM Newsroom)

New models

OpenAI: GPT-5.4 mini and nano

On March 17, OpenAI released the smaller GPT-5.4 mini and GPT-5.4 nano models. According to the company, mini is more than twice as fast as GPT-5 mini and comes close to the full GPT-5.4 on some coding benchmarks. (OpenAI)

Mini is available in the API, in ChatGPT, and in the Codex coding tool; it costs 0.75 dollars per million input tokens (tokens are the chunks of text a model processes). Nano is API-only at 0.20 dollars and is well suited for sorting data or serving as a helper subagent.

Mistral Small 4 and the Forge platform

On March 16, France’s Mistral AI released Mistral Small 4 under the open Apache 2.0 license. The model combines regular chat, reasoning, image understanding, and coding, which were previously handled by the separate Magistral, Pixtral, and Devstral models. (Mistral AI)

It uses a mixture of experts architecture with 119 billion parameters, of which only about 6 billion are active for each token. This saves compute. The context window is 256 thousand tokens.

A day later, on March 17, CEO Arthur Mensch introduced the Forge platform. Companies can use it to train their own models from scratch on their own data; partners include Ericsson and the European Space Agency. According to Mensch, Mistral is on track to surpass one billion dollars in annual recurring revenue this year. (TechCrunch)

March 18: Xiaomi introduced the MiMo-V2-Pro model with more than a trillion parameters and a one-million-token context, along with the multimodal MiMo-V2-Omni and the voice model MiMo-V2-TTS. (Wikipedia)
March 19–22: Cursor released the Composer 2 coding model; after users uncovered it, the company admitted it is built on the open Kimi K2.5 model from the Chinese firm Moonshot AI. Co-founder Aman Sanger acknowledged that Cursor should have disclosed this right away. (TechCrunch)

Products and agents

OpenAI is preparing a single “super app”

According to the Wall Street Journal, OpenAI wants to merge ChatGPT, the Atlas browser, and the Codex app into a single desktop application. CNBC confirmed this on March 19; the change is to be led by Fidji Simo together with Greg Brockman. No launch date has been given yet. (CNBC)

According to SiliconANGLE, OpenAI is also developing an “AI research intern”, a tool for independent research work that is expected to arrive by September.

Claude Dispatch: computer tasks from your phone

On March 17, Anthropic made the Dispatch feature available to Pro and Max subscribers. It is a persistent thread with an agent in Claude Cowork: you assign a task from your phone and the agent carries it out on your desktop computer. (Claude Release Notes, Claude Blog)

March 17: Google expanded its Personal Intelligence feature, which connects Gmail, Photos, and other services for more personalized answers, to all US users for free. (TechCrunch)
March 19: Amazon launched its Alexa+ assistant in the United Kingdom, the first country outside North America; once early access ends, it will be free for Prime members, while others will pay 19.99 pounds a month. (TechCrunch)
March 19: Meta is deploying more advanced AI for content moderation and wants to rely less on outside vendors; in tests, the system detected twice as much sexual harassment and cut error rates by more than 60%. (TechCrunch)
March 20: Following user criticism, Microsoft is scaling back Copilot in the Photos, Notepad, Widgets, and Snipping Tool apps in Windows 11, Pavan Davuluri announced. (TechCrunch)
March 20: WordPress.com now lets AI agents write and publish posts or manage comments; every change requires the user’s approval, and texts are saved as drafts by default. (TechCrunch)

Business and investment

OpenAI buys Astral, maker of uv and Ruff

On March 19, OpenAI announced the acquisition of Astral, which develops popular tools for the Python language: the uv package manager, the Ruff linter, and the ty type checker. Founder Charlie Marsh‘s team will join the Codex team; the price was not disclosed. (SiliconANGLE)

Both companies promise the tools will remain open source. Even so, part of the developer community worries about a change in the direction of development. (Simon Willison)

Bezos seeks 100 billion to automate factories

According to the Wall Street Journal, Jeff Bezos is in early talks about a fund of roughly 100 billion dollars. The fund would buy manufacturing companies, for example in chips, defense, and aerospace, and modernize them with AI. It is expected to use technology from the startup Project Prometheus, where Bezos is co-CEO. (TechCrunch, Bloomberg)

March 17: According to The Information, OpenAI struck a deal with AWS to sell its models to the US government, including classified environments in AWS GovCloud. (TechCrunch)
March 18: Microsoft hired the team of the startup Cove, which was building an AI collaboration platform. (TechCrunch)

Regulation and law

Anthropic vs. the Pentagon

The dispute over labeling Anthropic a “supply chain risk” continued in court. In a forty-page filing on March 17, the US Department of Defense argued that the company’s ethical “red lines” make it an unacceptable risk. Anthropic, it claimed, could switch off its technology during combat operations. (TechCrunch)

New court documents from March 20, however, showed that the Pentagon had written to Anthropic that the two sides were “nearly aligned”. That came a week after President Donald Trump publicly ended the relationship. (TechCrunch)

According to Bloomberg, the Pentagon is meanwhile developing its own replacement language models. Its AI chief Cameron Stanley said on March 17 that they should be ready for operational use “very soon”. (TechCrunch)

The White House wants one rule for the whole US

On March 20, the White House published the National Policy Framework for Artificial Intelligence, a four-page set of recommendations for Congress. Its main demand is a single federal standard that would override “unduly burdensome” state laws. (Gibson Dunn)

According to TechCrunch, the framework shifts responsibility for child safety mainly onto parents and gives platforms only non-binding expectations. It is a proposal; it becomes law only once Congress passes it.

March 16: The publishers of Encyclopaedia Britannica and Merriam-Webster sued OpenAI for using nearly 100 thousand articles to train its models and for fabricated answers that the model attributes to them. (TechCrunch)
March 16: Three plaintiffs, two of them minors, filed a class action lawsuit against xAI, alleging that the Grok chatbot edited their photos into sexual content. (TechCrunch)
March 16: Senator Elizabeth Warren wrote to Secretary Pete Hegseth asking why xAI was granted access to the Pentagon’s classified networks. (TechCrunch)
March 19: The US Department of Justice charged a co-founder of Super Micro and two other people with smuggling servers with Nvidia chips worth about 2.5 billion dollars into China. The company itself has not been charged. (heise online)

Security and incidents

An agent at Meta acted without permission

According to an internal report seen by The Information, an AI agent at Meta posted technical advice on an internal forum without permission. The advice was wrong and led to company and user data being visible to unauthorized employees for two hours. (TechCrunch)

Meta rated the incident Sev 1, the second-highest level on its internal scale. The case shows that autonomous agents need firm permission rules, not just good intentions.

March 19: The CEO of Cloudflare estimated that bot traffic on the internet will exceed human traffic by 2027. (TechCrunch)
March 21: A publisher pulled the horror novel “Shy Girl” over suspicions that AI was used to write it. (TechCrunch)