What is DeepSeek and How Impact on U.S Companies

What is DeepSeek AI?

The Rise of DeepSeek

DeepSeek, a private Chinese company, emerged in July 2023 under the leadership of Liang Wenfeng, a graduate of Zhejiang University, one of China's most renowned universities. With $8 billion in assets, Liang's innovative approach to AI began as part of a personal venture, funded through his hedge fund. According to MIT Technology Review, the company quickly made waves, largely due to its stockpile of Nvidia A100 chips, which, despite export restrictions to China, became the cornerstone of the DeepSeek AI model. This move was backed by research from Ben Reitzes, who praised DeepSeek's AI breakthroughs, particularly in learning and efficient memory usage. While some investors are still skeptical about its chips and their usage, DeepSeek is revolutionizing the AI landscape.

DeepSeek’s R1 Model and Its Power

DeepSeek's R1 model, with a staggering 670 billion parameters, is one of the largest open-source large language models to date, a feat made possible by its approach to cost-effective development. In comparison to OpenAI's O1 model, the R1 has shown remarkable power, excelling in areas like mathematics, coding, and reasoning. Its reasoning model simulates how humans process problems, producing responses incrementally while using significantly less memory. This has made DeepSeek a powerful player in AI, though it remains cautious about politically sensitive content, with its base model carefully trained to avoid controversial topics like Tiananmen Square. The company has focused on safeguarding information while delivering a helpful and harmless experience. The entire model was built cheaply compared to others in the field, costing an estimated $6m, far less than the $100m estimated for ChatGPT4. However, the exact details of how the company achieved this remain somewhat unclear.

How are US Companies Hit by DeepSeek's Impact?

Competition from DeepSeek

The rise of DeepSeek, a low-cost AI tool from China, has sent shockwaves through Wall Street. U.S.-based models such as ChatGPT and other generative AI apps now face a new competitor with cutting-edge technology, challenging the status quo. DeepSeek's ability to deliver high-performing AI models with limited compute resources is especially concerning for U.S. technology investors who are already grappling with the potential over-expansion of tech advancements in Silicon Valley. Many are now questioning whether this new wave of AI tools is advancing AI faster than the U.S. can manage and whether it has the potential to reshape AI models and the industry’s future.

Financial Impact on U.S. Tech

The financial markets have felt the weight of these concerns, with a significant stock selloff on 27 January, contributing to the Nasdaq dropping more than 3%. While some see this as an overreaction, the concerns about DeepSeek’s technology being cost-effective and disruptive are very real. Nvidia, for example, has seen its stock price plummet by 17%. The company, once the most valuable company in the world in terms of market capitalisation, has now fallen behind Apple and Microsoft. With DeepSeek’s efficient use of resources, it could continue to raise questions about the viability of Nvidia’s business model, especially as its hardware faces pushback from this new competitor.

Infrastructure and Future Demand

At the same time, U.S.-based companies are worried about how this technology might affect future demand for their infrastructure. While DeepSeek’s AI technology is promising, it’s clear that it has the potential to undermine established models in a way that would require significant adaptation from U.S. tech giants. The enormous demand for AI technology and the increasing need for efficient models places pressure on companies like Nvidia and others to rethink how they approach innovation and cost. Although DeepSeek’s model isn’t seen as revolutionary, it might just signal a tipping point in the AI landscape.

What's Nvidia tell about DeepSeek?

Praise for DeepSeek's AI Advancement

Nvidia has praised DeepSeek as a great AI advancement and a perfect example of Test Time Scaling. The company highlighted how DeepSeek's work demonstrates the power of using this technique, where models are created by leveraging available models and compute resources that are export-control compliant. This has been particularly useful in AI inference, which involves making decisions and predictions with the help of significant numbers of NVIDIA GPUs and high-performance networking. While DeepSeek has showcased impressive results, Nvidia points out that the easy part is using the training wheels of this technology, and scaling to real-world use is much more complex.

Nvidia on Scaling and Inference Costs

Nvidia further emphasizes that while training costs are lower, the inference costs—which involve having the model make ongoing decisions—are far higher and more labor-intensive. To make a comparison, Nvidia explains that it's like constructing buildings brick-by-brick, day after day, versus just drawing the blueprints. In addition, Nvidia suggests that DeepSeek's outage, if not caused by a cyberattack, might just be a result of surging demand and limited supply, which only strengthens their appeal to investors about the realities of scaling such AI models.