Nvidia GB200 server architecture powered by 72 Blackwell GPUs.
What is Nvidia new server technology on Nvidia GB200 server?
Nvidia, a company that makes hardware for AI, launched a new server system called the GB200 NVL72. This system features 72 Blackwell GPUs. These GPUs are very powerful They process data at high speeds This server is designed for mixture-of-experts (MoE) models. MoE models break down tasks into smaller, expert parts, making them efficient Nvidia’s servers increase the speed of these models tenfold However, compared to previous generations, these servers have fewer chips, while newer ones have more Furthermore, they have faster connections, allowing data to flow faster.
How Chinese companies are taking advantage of this?
Moonshot AI is a Chinese company that created the Kimi K2 Thinking model Nvidia’s server makes this model ten times faster, but DeepSeek is also Chinese. Their models are affected by this as well. Other companies include 01.AI, Baichuan, and Alibaba’s Qwen. All of these use MoE models These companies then create open-source models DeepSeek released its model in 2025, which Moonshot AI released in July Companies like OpenAI and Mistral also adopt MoE. Performance is improved by having a larger number of chips in Nvidia servers and their simultaneous operation. Faster links transfer data faster, yet MoE models use fewer chips for training. However, speed is crucial for inference Inference means answering user queries Millions of users use them simultaneously This is where Nvidia’s technology shines. Companies like AMD are competing. They will launch similar servers next yea However, Nvidia still has the upper hand.
US-China Relations and Export Controls
The US restricts Nvidia’s advanced chips from being shipped to China for security reasons. However, Chinese companies access them through cloud services Or use older chips This news also shows that China is advancing in AI. Despite export controls, startups like Moonshot AI are growing rapidly. They compete globally. Now, we turn to another topic AI reading the human mind AI can now read human minds. It interprets thoughts from brain scans Scientists call this mind-reading This technology is based on brain-computer interfaces (BCIs). BCIs record brain activity, and then AI analyzes it.
How Mind-Reading Works of Nvidia GB200 server
BCI implants are placed in the brain For example, in the posterior parietal cortex, the region involved in thinking and planning The implant picks up signals The AI decodes them. Take the example of Nancy Smith She plays the piano. The implant first captures her thoughts and then takes action. This happens in milliseconds, and there are also non-invasive methods EEG measures electrical activity on the scalp AI extracts signals from noise This helps determine attention and decisions Another method is fMRI This uses brain scans. Stable Diffusion AI reconstructs images, then recreates what a person sees. Japanese researchers trained on 10,000 photos Using brain patterns and captions, the results are accurate.
Examples and Applications
AI creates images from brain scans, such as a clock tower or a teddy bear Matching the layout and content helps paralyzed people communicate through thoughts, read dreams, and study animal thinking. Companies like Neuralink and Synchron also develop implants, and Neuralink is testing them on volunteers. EEG headsets are becoming commonplace Companies like Apple are showing interest. Future prospects include AI foundation models trained on neural data, which will uncover common patterns This will work across different brains and could help with psychiatric treatment. Subconscious signals can be used for diagnosis. Feedback loops offer therapy, but there are risks Manipulation is possible Society must be prepared.
Discussing ethical concerns on Nvidia GB200 server
This technology threatens privacy. AI can read thoughts, including pre-conscious thoughts This impacts freedom Data security is crucial Companies can sell neural data and infer mental health or political views. Some countries have laws. Chile and four US states protect neural data. UNESCO and the OECD provide guidelines, but not enough. Regulation in consumer products is lacking Data remains vulnerable Experts call for stronger laws. The two topics are linked, as Nvidia servers make AI models faster. Mind-reading AI also requires powerful hardware to process large data sets MoE models can handle neural data, demonstrating that hardware and software work together as AI rapidly evolves.
About the Author
