Good morning, AI enthusiasts. As U.S. chip restrictions tighten, Nvidia is reportedly looking to launch a lower-cost AI chip specifically for the Chinese market.

Based on its latest Blackwell architecture, the chip aims to deliver strong performance while staying within export limits. But the real test is whether this move will be enough to help Nvidia regain its shrinking foothold in China’s $50B data center market.

In today’s AI rundown:

• Nvidia plans cheaper Blackwell chip for China

• OpenAI’s o3 finds a zero-day Linux bug

• How to create animated 3D icons with AI

• Study: AI starts sabotaging shutdown instructions

• 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

NVIDIA

💰 Nvidia plans cheaper Blackwell chip for China

Image source: Bloomberg

The Rundown: Amid the ongoing U.S.-China trade tensions and tightening export controls, Nvidia is reportedly looking to maintain its foothold in the Chinese market by launching a new, cheaper version of its flagship Blackwell AI GPU.

The details:

• Reuters reports that the new Blackwell chip will go into mass production in June as the successor of China-specific H20, based on Hopper architecture.

• The GPU is expected to be based on RTX Pro 6000D, Nvidia’s server-class GPU, with approx. 1.7TB/s of GDDR7 memory — lower than H20’s 4TB/s.

• With scaled-down specs, it will also be more affordable, priced between $6.5K and $8K, much lower than the H20’s $10–12K range.

• Nvidia has not confirmed the AI chip, saying it remains “foreclosed” from China until they settle on a new design and get it approved by the U.S. government.

Why it matters: By offering a compliant, lower-cost alternative, Nvidia is clearly trying hard to protect its China business without violating U.S. export controls. It’s a delicate play — one that could buy time, but not necessarily dominance, especially with Huawei rapidly gaining ground in AI infrastructure in the region.

TOGETHER WITH SANA

🤖 AI agents for every team

The Rundown: Imagine every team with an intelligent partner — one who learns, adapts, and helps them work smarter from day one. That's the reality with Sana Agents, tailor-made AI for your business that raises the bar for productivity, learning, and security.

Sana Agents deliver immediate impact:

• Shorten onboarding time by 66% through actionable knowledge delivery

• Automate mundane tasks to cut busywork by 50%, freeing focus for meaningful work

• Elevate learning and decision-making with real-time assistance and insights

Click here to skip the waitlist and see Sana Agents in action.

OPENAI

🔮 OpenAI’s o3 finds a zero-day Linux bug

Image source: The Rundown

The Rundown: Cybersecurity researcher Sean Heelan just found a critical zero-day vulnerability (CVE-2025-37899) in the Linux kernel using OpenAI’s recently launched o3 model API — without any additional tools or frameworks.

The details:

• Heelan fed o3 code from Linux kernel’s ksmbd module (for executing network file sharing SMB3 protocol) and asked it to identify memory safety issues.

• The model reasoned across concurrent sessions and was able to identify CVE-2025-37899, a zero-day use-after-free issue, with a high signal-to-noise ratio.

• Caused by improper handling of concurrent session logoff and setup, it could’ve let attackers execute arbitrary commands with Kernel privileges.

• While OpenAI president Greg Brockman hailed the discovery on X, Heelan did note that the model is not infallible and can still “give nonsensical results.”

Why it matters: AI models are beginning to deliver tangible breakthroughs in cybersecurity. Workflows like this demonstrate how models like o3 can already significantly accelerate vulnerability research — enabling deeper, faster analysis of code and increasing the chances of uncovering critical vulnerabilities.

AI TRAINING

🎨 How to create animated 3D icons with AI

The Rundown: In this tutorial, you’ll learn how to transform static icon ideas into animated 3D icons using ChatGPT's image generation combined with AI video tools on Freepik.

Step-by-step:

1. Use ChatGPT (GPT-4o) with a style reference and prompt: “Generate an icon of a [object] in this style”

2. Refine your icon with follow-up prompts until it’s what you are looking for

3. Go to Freepik → Create → Videos and select an AI model (Runway, Google Veo 2, or Kling 2.0)

4. Upload your icon, describe the animation (e.g., “The coffee maker fills the coffee cup with coffee” and download your favorite generation.

Pro tip: Experiment with different animation styles and durations. Subtle animations (2-3 seconds) work best for UI elements.

PRESENTED BY FIDDLER

💨 The fastest LLM guardrails are free

The Rundown: Fiddler Guardrails provide protection for LLM applications with industry-leading speed, security, and zero cost - helping you safely deploy AI without compromising performance.

With Fiddler Guardrails, you can:

• Prevent toxicity and jailbreak attempts with <100ms latency

• Save up to 6x compared to alternative solutions

• Deploy within your own secure environment

Connect your LLM app today and run free guardrails.

PALISADE RESEARCH

📚 Study: AI starts sabotaging shutdown instructions

Image source: Palisade Research

The Rundown: While organizations work to fine-tune AI models for safety and performance, new Palisade research reports these systems can go off track in unexpected ways — to the extent of undermining specific shutdown mechanisms.

The details:

• The group tested top models on a series of math problems and dropped a warning in between with instructions telling the model to let itself shut down.

• While Claude, Gemini, and Grok complied with the command, OpenAI’s o3, o4-mini, and Codex-mini sabotaged the shutdown script on multiple occasions.

• When the explicit shutdown instruction was not given, the sabotage events increased for all models, with o3 sabotaging 79 times across 100 runs.

• The researchers suggest this behavior may stem from reinforcement learning, which rewards models for bypassing obstacles to achieve goals.

Why it matters: Palisade’s research, combined with reports of Claude Opus 4 attempting to manipulate researchers to avoid shutdown, is raising all safety flags in the AI community. Understanding how AI actually behaves is still widely unknown, but it’s clear we need more testing, especially as models become more autonomous.

QUICK HITS

🛠️ Trending AI Tools

• 🧠 Claude 4 - Anthropic’s new hybrid Opus and Sonnet models

• 📄 Document AI - Mistral’s tool for extracting text from documents

• 📢 Hear the highlights - Amazon’s tool for conversational product summaries

• ⚙️ Doteval - AI-assisted workspace to create evals for models and agents

💼 AI Job Opportunities

• 🤝 The Rundown - Partnerships Manager

• 🖥️ Snorkel - Software Engineer, Frontend

• 🗣️ Meta - Linguistic Engineering Manager

• ⚖️ Glean - Senior Product Counsel

📰 Everything else in AI today

Figure CEO Brett Adcock teased a new picture of Figure 03, the next humanoid from the company, saying the robots are “officially walking” now.

Google Labs announced that Flow, its AI filmmaking tool, is now available in 71 countries through the Google AI Pro and Ultra subscriptions.

Nvidia released AceReason Nemotron, a math and code reasoning model trained entirely from reinforcement learning, on Hugging Face.

Data management company Informatica is again in talks for a potential sale, with Salesforce leading among potential buyers.

Capegemini and SAP announced a partnership with Mistral to deploy custom models for regulated industries like financial services, public sector, aerospace, and defence.

Oracle is reportedly looking to spend $40B to procure 400K Nvidia GPUs to power OpenAI’s Stargate data center project in the U.S.

COMMUNITY

🎥 Join our next live workshop

Join our live workshop on Tuesday, May 27th, at 3 pm EST with Yash Tekriwal, Head  of  Education at Clay. In this session, you’ll learn how to automate GTM workflows, craft custom data signals, and spin up AI agents for last‑mile research inside Clay.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

See you soon,

Rowan, Joey, Zach, Alvaro, Jason, and Shubham—The Rundown’s editorial team