AI Chatbots

By: Hasan Ahmad Rabbani
Grade 8
AI Chatbots are everywhere.
It seems like the only thing companies are talking about and news outlets are reporting on. Whether you are the average person or a major tech company, AI is the main conversation and the competition is more intense than ever. Gemini, Co-Pilot, Grok, Mixtral, Meta AI, Samsung Galaxy AI, Apple Intelligence, and the current king, Chat-GPT are all examples of chatbots that have been released in the past five years, and companies have not
been focusing on anything else since. Companies have been investing billions of dollars into AI, hiring thousands of employees, obtaining state-of-the-art hardware, and pouring time into perfecting their models. Until January 20th, 2025, the day DeepSeek AI officially launched and forced American tech companies to set record losses and rethink their structure.
AI is an extremely difficult and expensive project that companies are spending billions of dollars to fund.
Although these companies try their best to keep costs low, DeepSeek has proven that these companies are not doing enough. Microsoft, Google, and Open AI all spent upwards of a billion dollars developing and sharpening their AI models, while DeepSeek used only 6 million. They did three things that other companies failed to do. DeepSeek hired subordinately, about 400 compared to Open AI’s 4000. Having a smaller team means fewer salaries to pay and benefits to offer, along with needing less office space and being able to use every worker’s skills efficiently. DeepSeek also used the sluggish and aged Nvidia H800 chips while its competitors spent hundreds of millions of dollars on Nvidia’s Blackwell chips. Finally, DeepSeek did it all in China, while its competitors focused on building everything in America. Building anything in China is much cheaper than building it in Europe or America because of lower wages, high availability for STEM workers, government grants, cheaper cloud infrastructure, and regulations for data collection.
DeepSeek demonstrated that building AI isn’t as difficult as these tech giants are making it seem.
Earlier, we mentioned DeepSeek’s older hardware as a cost-saving strategy but it wasn’t their choice. The hardware used to code and run AI is distinct to the computer parts running on your PC. Enterprise-grade chips like Nvidia’s B200 are specifically designed to have more Tensor Cores and CUDA Cores for AI tasks compared to regular computers that use Shader Cores and SIMD Units. Although our computers might be faster at browsing the web and running games, these chips are better at multitasking and AI. Every year, the AI Chip leader Nvidia makes substantial improvements to their chips, however, due to sanctions by the American government, Nvidia can’t sell their fastest and most powerful chips to Chinese companies. So these older chips were the best DeepSeek could get. These older chips were less power-efficient, required more cooling, and performed 150% worse on average, yet miraculously, the DeepSeek team was still able to code an AI model as powerful as other tech companies. While DeepSeek used only 2000 AI computers, which may seem plenty, Open AI, Microsoft, and Google all used upwards of 50,000 computers to code their models. Bottlenecked by thin numbers and far outdated chips, the 400 DeepSeek employees still managed to pull off this tough project, proving their immense talent.