International AI company reveals a smarter challenger to ChatGPT and major AI corporations
By NAREN KRISHNA JEGAN — science@theaggie.org
Artificial Intelligence (AI) is defined as “technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.” From being a vision in the 1900s as a way to automate our human lives to now being integrated in everyday life as educational tools, such as Khan Academy and Google, AI has grown massively.
Part of the recent movement that spurred the interest in AI was a natural learning processing model called ChatGPT. From its inception by the company OpenAI, ChatGPT has quickly grown to popularize large-scale AI adoption, with sources currently estimating around 300 million users weekly across the globe. ChatGPT’s ability to process human inputs, compare it to an astronomical dataset that it was rigorously trained on and output a (mostly) accurate, relevant response caught on early with most users. As such, the field of AI has grown exponentially to include players such as Google’s Gemini, Microsoft’s Co-Pilot and even Facebook’s Meta.
Recently, a Chinese company called DeepSeek released a natural language processing model with responses rivaling those of GPT4, ChatGPT’s latest model. The kicker? They had a lower six-million-dollar budget in comparison to the $100 million dollar budget allocated for GPT4. Moreover, DeepSeek was able to produce a high-level product using lower-powered chips.
NVIDIA is one of the world’s leaders in producing high-level graphics processing units (GPU). These chips have various purposes such as gaming, data processing and AI. GPUs act as high-performance engines, accelerating AI computations by handling multiple tasks simultaneously. Unlike regular computer chips (CPUs) that do one task at a time, GPUs can handle many tasks at once, making them perfect for training AI models and making fast predictions. This speed boost is why AI, like chatbots and image recognition, runs much faster and more efficiently with GPUs. NVIDIA chips are used by a variety of AIs such as ChatGPT, Gemini and Meta’s Llama for these reasons.
NVIDIA has developed various high-level chips such as the A100 and H100 that are currently used in the United States models. The U.S. government placed restrictions on the types of chips and their processing power that NVIDIA could export to China. To comply with U.S. regulations, NVIDIA developed modified versions — H800 and A800 — specifically for the Chinese market. These chips have reduced interconnect bandwidth, which slightly limits their performance in large-scale AI training but still makes them powerful for AI workloads.
The rise of DeepSeek and its ability to produce competitive AI models using lower-powered chips has raised concerns about the need for expensive, high-end GPUs. Additionally, increasing global competition and U.S. government regulations limiting NVIDIA’s ability to export its most powerful chips to China have also impacted investor confidence. As a result, NVIDIA’s stock has seen a drop of a staggering $600 billion as markets react to the shifting AI landscape. Analysts predict that while NVIDIA remains a leader in AI hardware, emerging players like DeepSeek may influence future chip demands, forcing the company to adapt its strategies.
Memory efficiency plays a crucial role in AI development, as training and running large-scale models require vast amounts of memory storage. UC Davis Emeritus Professor of Computer Science Dr. Norman Matloff explained how DeepSeek was able to tactically optimize their performance using memory.
“Traditional AI models typically use a 32-bit floating-point (FP32) format, which allows for storing numbers between zero and four billion,” Matloff said. “DeepSeek has opted for an eight-bit floating-point (FP8) format, which only stores numbers between zero and 255. While this might seem like a limitation, it is actually a strategic advantage. FP8 reduces memory requirements by a factor of four, meaning that for the same AI application, only a quarter of the memory is needed. […] It is important to note that this does not impact processing power.”
Let’s use an analogy. A 32-bit system is like using a large book for each piece of information, even if it’s just a short note. It takes up a lot of space but can hold a lot. An eight-bit system is like using a small notebook for each note. It holds less, but you can fit more of them on the shelf. Switching from large books (32-bit) to small notebooks (eight-bit) saves four times the space while preserving the same information.
DeepSeek AI does something similar — it finds ways to store and process information more efficiently, using smaller “notebooks” (lower-bit numbers) to save space and run faster. By leveraging FP8 precision, DeepSeek has demonstrated that high-performance AI can be achieved with significantly lower hardware demands, making AI development more cost-effective and accessible.
Why is this a game-changer? FP8 cuts down the cost of training and running AI models, making it easier for smaller companies to compete with giants like OpenAI and Meta. This shift makes AI more accessible and pushes the industry toward more innovation. DeepSeek’s R1 model combines FP8 with multi-token predictions, setting a new standard for speed. Whether it’s chatbots, voice assistants or search engines, users will get quicker responses without losing accuracy. With R1’s reasoning abilities, AI can take on more complex problems, explain its answers clearly and even double-check its own work.
The significance of DeepSeek’s breakthrough cannot be overstated. By leveraging FP8 precision rather than the standard FP32, DeepSeek R1 has demonstrated that AI models can achieve high performance with significantly reduced computational costs.
WRITTEN BY: NAREN KRISHNA JEGAN — science@theaggie.org
