Callista AI Weekly (March 16-21, 2026)

That was NVIDIA's week. The GTC 2026 dominated the news cycle with new chips, a new rack architecture and an open platform for AI agents, which brought 17 major enterprise partners on board on day one. But the actual story of the week goes beyond a single company. From the White House's first national AI legislative framework to Xiaomi's secret placement of a trillion-parameter model on public leaderboards under a false name — the pace of change in AI accelerated on all fronts.

New AI use cases

The most important enterprise deployment news of the week came from NVIDIA yourself. At GTC 2026 in San Jose (March 16-19), Jensen Huang presented NemoClaw before, an enterprise-grade security and data protection layer based on the viral open-source agent framework OpenClaw. Seventeen major software companies joined as launch partners, including adobe, Atlassian, box, cisco, CrowdStrike, Dassault Systems, Red Hat, SAP, salesforce, siemens, ServiceNow and Synopsys. NemoClaw's OpenShell runtime enforces sandboxing, least-privilege access controls, and policy-based privacy guardrails — turning an extremely popular but security-related open source project into something that companies can actually use.

Alibaba entered the race for agent-based AI on March 17 with the launch of Wukong. Wukong is an enterprise AI agent platform that enables companies to manage multiple agents from a single interface. The platform performs tasks such as document processing, approvals, meeting transcription, and research. It is available as a standalone desktop app and as an embedded agent in DingTalk, Alibaba's collaboration platform with over 20 million enterprise users. Integrations with Slack, Microsoft Teams, and WeChat are planned. Alibaba wants to connect Wukong with Taobao, Tmall, 1688, Alipay and Alibaba Cloud as modular agent skills, positioning it as a hub for everything from e-commerce storefront design to cloud infrastructure orchestration.

microsoft used the RSAC 2026 security conference (from March 19) to announce Zero Trust for AI — an extension of the Zero Trust security architecture to cover the entire AI life cycle. This includes a new reference architecture, updated assessment tools, and practical patterns for validating AI on a large scale. The company also confirmed that Agent 365, its central control plane for monitoring, governing, and securing agents in organizations, will be generally available on May 1.

In a separate step, Microsoft launched on March 19 May image 2, his new text-to-image model. The model debuted in third place in the arena.ai rankings, behind Google's Gemini 3.1 Flash and OpenAIS GPT Image 1.5. MAI Image 2 is being rolled out via Copilot and Bing Image Creator, API access is available to selected customers such as WPP available. The model uses photorealistic results with natural lighting and precise skin tones, but currently only supports square editions and provides a 30-second break between generations.

Marvell Technology and Lumentum Holdings At OFC 2026 this week, they demonstrated a technology for optical circuit switching that is intended to support autonomous AI agents on a large scale through improved network performance in data centers.

Major vendor updates

NVIDIA's GTC Offensive

GTC 2026 was the main event of the week. Huang's two-hour keynote on March 16 covered more than most tech conferences can do in three days.

The headline hardware announcement was the Vera Rubinplatform, NVIDIA's next-generation AI infrastructure. It combines seven chips: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet Switch and the newly integrated Groq 3 LPU. NVIDIA requires up to 10 times more inference throughput per watt and a tenth of the cost per token for the platform compared to the Blackwell generation.

The Groq 3 Language Processing Unit is NVIDIA's first chip from the Groq acquisition, a $20 billion deal on Christmas Eve 2025. The LP30 chip is powered by samsung manufactured using the 4nm process and is to be delivered in the third quarter of 2026. It carries 512 MB of on-chip SRAM per die with 150 TB/s of memory bandwidth. A full LPX rack accommodates 256 LPUs. In the Vera Rubin architecture, the Ruby GPUs take over the computationally intensive prefill phase of requests, while the GROQ LPUs take over the decode phase for low-latency token generation.

Huang also showed a prototype of Kyber, the next generation of NVIDIA's server rack design. It integrates 144 GPUs that are arranged vertically instead of horizontally, increasing density and reducing latency. The cable-free design is intended to significantly reduce installation time. Kyber is coming with Vera Rubin Ultra, scheduled for 2027.

On the consumer side, NVIDIA announced DLSS 5 , which is out this fall for the RTX-50 series. Unlike previous DLSS versions that focused on upscaling and frame generation, DLSS 5 is an AI-powered lighting layer that adds photorealistic light, shadow, and material behavior to game scenes.

Huang predicted that orders between Blackwell and Vera Rubin will reach a volume of 1 trillion dollars by 2027 — more than previous estimates.

OpenAI's super app plans

OpenAI builds a desktop application that combines ChatGPT, Codex, and ChatGPT Atlas (its AI-powered web browser) into a single product. Bloomberg reported on March 20 that Fidji Simo, CEO of Applications, will co-lead the project with President Greg Brockman. The existing standalone ChatGPT app should continue to exist alongside the new product. A start date has not yet been announced.

Separately, OpenAI made GPT-5.4 mini available to free users and Go users via the “Thinking” feature, thus removing the payment barrier for its latest smaller model.

Google's Gemini push on the Mac

google began privately testing a dedicated Gemini app for Mac this week, Bloomberg reported on March 19. The app, which is internally codenamed Janus, can search the web, analyze uploaded documents and maintain the conversation history. Testers generate images, videos, music, tables, and charts. A feature called Desktop Intelligence allows Gemini to pull context from other active Mac apps, such as the calendar, to provide more personalized answers. A release date was not announced.

Xiaomi's stealth launch

xiaomi One of the most unusual model launches in the recent past was achieved. On March 18, a model called Hunter Alpha appeared on OpenRouter without developer assignment. The specifications were exceptional: over 1 trillion parameters in total, 42 billion of them active during inference, a context window of 1 million tokens, and an agent-oriented architecture. It topped the daily charts and processed over 1 trillion tokens before anyone found out that it was MiMo-v2-Pro — developed by Xiaomi's AI division under the former DeepSeek-Researcher Luo Fuli. Xiaomi offered developers a week of free access via agent frameworks such as OpenClaw, Cline, and Blackbox.

Anthropic Institute

Anthropic At the beginning of March, formally launched the Anthropic Institute, a research branch led by co-founder Jack Clark, which deals with the economic, social and security effects of AI. The institute brings together three existing research teams — the Frontier Red Team, Societal Impacts, and Economic Research — and has hired experts such as Matt Botvinick from Google DeepMind, Zoe Hitzig from OpenAI, and economics professor Anton Korinek. Anthropic is also expanding its public policy team and opening an office in Washington, D.C.

AI governance

The White House AI framework

The biggest governance announcement of the week came on March 20. Die Trump administration published their National Policy Framework for Artificial Intelligence, a six-part legislative draft for Congress. The framework includes protecting children and empowering parents, protecting American communities, upholding intellectual property rights, preventing censorship and protecting freedom of speech, promoting innovation and securing American AI dominance, and training Americans for an AI-ready working world.

The framework's most far-reaching proposal is federal preemption over national AI laws. The administration wants Congress to establish a uniform federal approach instead of allowing what it calls a “patchwork” of national regulations. This follows Trump's Executive Order of December 2025, which questions the enforceability of national AI laws. The administration urged Congress to act “this year,” although Bloomberg pointed out that the path through Congress will be difficult even with Republican control.

State level activities

While the White House pushed for federal preemption, the states continued to follow suit. Washington State passed five AI laws ahead of the adjournment on March 12: HB 1170 (disclosure requirements for AI), HB 2225 (chatbot security for minors), SB 5395 (AI for health insurance decisions), SB 5105 (AI deepfakes involving minors) and SB 5886 (fake digital images). Oregon passed a similar chatbot security bill the previous week. Colorado's AI Act, which came into force on January 1, continues to set the standard for regulating high-risk AI systems with its requirements for documentation, transparency, and risk mitigation.

EU regulatory adjustment

The European Council agreed on a position to simplify AI rules and proposed to extend the deadline for applying the rules for high-risk AI systems by up to 16 months. The delay gives the Commission time to ensure that the required standards and tools are in place before enforcement starts.

Enforcing export controls

On March 19, the US Department of Justice filed charges against three men for conspiring to smuggle from Supermicro-2.5 billion dollars worth of advanced NVIDIA chips to China, in violation of US export controls. Wally Liaw, co-founder of Supermicro, and Willy Sun were arrested. Steven Chang is still fleeting. The scheme allegedly used a Southeast Asian broker to create counterfeit papers and repackage servers before they were shipped to China. Supermicro released employees and is cooperating with the authorities. The company's shares fell 33%.

Switzerland focus

Exa opens first European office in Zurich

The San Francisco-based AI search startup Exa opened its first European office in Zurich this month. The company, which has raised 85 million dollars, is developing AI-optimized search infrastructure. The office is run by former Google employees Max Buckley and Felix Kroner. Exa cited Zurich's growing status as an AI hub and the presence of Anthropic, OpenAI, NVIDIA and Google DeepMind in the region as decisive factors for choosing a location.

Swiss AI Initiative is approaching the deadline

Die Swiss AI Initiative, the world's largest open science and open source project for basic AI models, has set its full application deadline for March 31. The initiative is the first project of the Swiss National AI Institute, a partnership between the AI Center of ETH Zurich and the AI Center of EPFL. The Swiss multilingual model Apertus, developed by EPFL, ETH Zurich and Switzerland's National High Performance Computing Center (CSCS), covers over 1,000 languages, including Swiss German and Romansh. The initiative will shift the focus to specialized models in 2026, particularly in medicine. that Lausanne University Hospital (CHUV) will start testing Meditron, a Swiss medical language model, in the emergency room from May.

ETH-Zurich-Stanford partnership

ETH Zurich and EPFL confirmed a strategic collaboration with Stanford Institute for Human-Centered Artificial Intelligence (HAI), formalized by the Swiss National AI Institute. The partnership signals Switzerland's intention to remain a globally networked AI research location.

Groundbreaking research

OpenClaw's explosive growth raises concerns about commoditization

OpenClaw, Peter Steinberger's open-source framework for autonomous AI agents, reached a critical turning point this week. CNBC reported on March 21 that the rapid distribution — over 250,000 GitHub Stars and thus the fastest-growing open-source project in history — raised concerns that AI models are becoming mass-produced. The project's ability to perform tasks autonomously across messaging platforms, combined with NVIDIA's enterprise layer NemoClaw, has created a complete stack of agent-based AI that is accessible to everyone.

Chinese authorities prohibited state-owned companies and government agencies from using OpenClaw on service computers due to security risks. At the same time, the city of Wuxi offered up to 5 million yuan (730,000 dollars) for OpenClaw-based breakthroughs in robotics and industrial applications. The tension between enthusiastic adoption and security concerns defines where agent-based AI stands right now.

NVIDIA's AI-Q blueprint

At GTC, NVIDIA released its AI-Q Blueprint for agent-based search, an open-source system based on LangChain, which tops the DeepResearch Bench accuracy rankings while cutting query costs in half. The blueprint provides a reference architecture for building research agents that can search, synthesize, and evaluate large collections of documents.

GPT 5.4 performance milestone

OpenAI's GPT-5.4 “thinking” model scored 83.0% on the GDPval benchmark and is therefore at or above the level of human experts in economically valuable tasks. On the Artificial Analysis Intelligence Index, GPT-5.4 Pro is on a par with Google's Gemini 3.1 Pro at 57 points — the highest values ever measured. These figures suggest that Frontier models are reaching expert levels in a wide range of real-world tasks.

conclusion

NVIDIA's GTC has set the hardware direction for the next two years. The Vera Rubin platform with its integrated Groq LPU and the promise of a 10-fold increase in inference efficiency gives companies a specific goal to which they can base their planning. But the more immediate change is taking place in software. The launch of NemoClaw with 17 enterprise partners shows that agent-based AI is moving from demos to deployments faster than many expected.

The governance picture is dividing. Washington State passed five AI laws in a single session, while the White House pushed Congress to oust state laws entirely. The EU is buying more time. These three approaches — aggressive national regulation, federal preemption, and delayed enforcement — cannot coexist in the long term. At some point, a decision must be made.

And then there's the quiet signal underneath all the noise. Xiaomi launched a trillion-parameter model anonymously and it topped public rankings before anyone knew who built it. OpenAI made GPT-5.4 mini free of charge. OpenClaw surpassed 250,000 stars. The costs of building and operating powerful AI systems are falling rapidly. The competitive advantage is shifting from who can build the best model to who can use it most effectively in the real world.

Ready to explore how Agentic AI can transform your organization? Visit us at https://www.callista.ch/agentic-ai to discover how we can guide your journey into this exciting new era of AI-powered productivity.

sources

• NVIDIA Newsroom - “NVIDIA Ignites the Next Industrial Revolution in Knowledge Work with Open Agent Development Platform” - March 16, 2026
• NVIDIA Newsroom - “NVIDIA Vera Rubin Opens Agentic AI Frontier” - March 16, 2026
• NVIDIA Newsroom - “NVIDIA Announces NemoClaw for the OpenClaw Community” - March 16, 2026
• CNBC - “Nvidia GTC 2026: CEO Jensen Huang sees $1 trillion in orders for Blackwell and Vera Rubin through '27” - March 16, 2026
• CNBC - “Nvidia GTC 2026: Agentic AI Takes Center Stage” - March 20, 2026
• CNN Business - “Nvidia, the world's most valuable company, just sent another signal that AI agents are going to be everywhere” - March 16, 2026
• Tom's Hardware - “How Nvidia's $20 billion Groq 3 LPU deal reshapes the Nvidia Vera Rubin Platform” - March 16, 2026
• Tom's Hardware - “Nvidia GTC 2026 keynote live blog” - March 16, 2026
• IEEE Spectrum - “Nvidia Groq 3 LPU: Speeding AI Inference Tasks” - March 2026
• StorageReview - “NVIDIA GTC 2026: Ruby GPUs, Groq LPUs, Vera CPUs” - March 2026
• Newegg Insider — “NVIDIA GTC 2026: DLSS 5 Revealed and the $1 Trillion Bet” — March 2026
• CNBC — “Alibaba launches agentic AI tool for businesses with Slack, teams integration plans” — March 17, 2026
• Alibaba Group - “Alibaba Launches Wukong: An AI-Native Agentic Platform for Enterprises” - March 17, 2026
• Computer Weekly - “Alibaba Joins AI Agent Race with Wukong Launch” - March 17, 2026
• Bloomberg — “OpenAI Plan's Desktop App Combining Chat, Coding, and Web Browsing” — March 20, 2026
• CNBC - “OpenAI to create desktop super app, combining ChatGPT app, browser and Codex app” - March 19, 2026
• Bloomberg - “Google Testing Gemini AI App for Mac to Rival ChatGPT and Claude” - March 19, 2026
• 9to5Mac - “Bloomberg: Google begins testing dedicated Gemini app for Mac” - March 19, 2026
• VentureBeat - “Xiaomi Stuns with new MiMo-v2-Pro LLM nearing GPT-5.2, Opus 4.6 performance at a fraction of the cost” - March 2026
• BusinessToday - “Mystery behind 'Hunter Alpha' AI model revealed” - March 19, 2026
• labla.org - “AI Releases March 19-21: Xiaomi 1T Model, GPT-5.4 Free” - March 21, 2026
• Microsoft AI - “Introducing MAI-Image-2: for Limitless Creativity” - March 19, 2026
• Microsoft Security Blog - “Secure Agentic AI End-to-End” - March 20, 2026
• Microsoft Security Blog - “New Tools and Guidance: Announcing Zero Trust for AI” - March 19, 2026
• Anthropic - “Introducing The Anthropic Institute” - March 11, 2026
• eWeek - “Anthropic Launches Institute to Examine AI's Impact on Jobs, Security, and Society” - March 2026
• The White House - “President Donald J. Trump Unveils National AI Legislative Framework” - March 20, 2026
• CNBC — “Trump Administration Unveils National AI Policy Framework to Limit State Power” — March 20, 2026
• Fortune - “The White House has a plan for AI regulation, and it starts with keeping states out of it” - March 20, 2026
• NBC News - “White House Releases AI Legislation Framework” - March 20, 2026
• Transparency Coalition — “AI Legislative Update: March 20, 2026" — March 20, 2026
• CNBC - “Super Micro shares tank 33% after employees charged with smuggling Nvidia chips to China” - March 19, 2026
• CNN - “Co-Founder of tech company charged with diverting $2.5 billion in Nvidia AI chips to China” - March 19, 2026
• Fortune - “Supermicro's cofounder was just arrested for allegedly smuggling $2.5 billion in GPUs to China” - March 19, 2026
• CNBC — “OpenClaw's ChatGPT moment sparks concern that AI models are becoming commodities” — March 21, 2026
• TechCrunch - “Nvidia's version of OpenClaw could solve its biggest problem: security” - March 16, 2026
• S-GE - “Exa arrives in Zurich” - March 2026
• SWI swissinfo.ch - “Artificial intelligence in Switzerland: What's new in 2026" - 2026
• EU Council - “Council agrees position to streamline rules on artificial intelligence” - March 13, 2026