The generative AI industry in the U.S. is getting a wake-up call, as are jittery investors who have seen their stocks rise with the AI boom in the last few years. That’s because a Chinese startup, DeepSeek, upended conventional wisdom about how advanced AI models are built and at what cost.
The company reported in early 2025 that its models rival those of OpenAI’s Chat GPT, all for a reported $6 million in training costs. That triggered a record $600 billion single-day drop in Nvidia’s (NVDA) stock and forced investors to rethink their AI-based bets going forward.
For anyone investing in AI, understanding DeepSeek’s rise is critical for navigating a new era in this sector. We get you up to speed below.
Key Takeaways
- The recent unveiling of DeepSeek-R1 spooked AI investors, leading to a massive sell-off in chipmakers.
- Its models suggest that smart engineering can slash AI development costs, a problem for U.S. AI firms that have spent hundreds of billions on their own projects.
- Open-source, affordable models could expand AI adoption, creating new prospects for investors.
What Is DeepSeek?
DeepSeek is a Hangzhou, China-based AI research company founded in July 2023 by former hedge fund executive Liang Wenfeng and backed by quantitative investment giant High-Flyer Quant. It has prioritized algorithmic efficiency and open-source collaboration to challenge the AI dominance of U.S. tech giants.
Since its launch, DeepSeek has released a series of impressive models, including DeepSeek-V3 and DeepSeek-R1, which it says match OpenAI’s o1 reasoning capabilities at a fraction of the cost. In addition, DeepSeek’s models are open source, meaning they are freely available for anyone to use, modify, and distribute.
Most crucially—and potentially most devastating for competitors—DeepSeek achieved these advances despite U.S. export restrictions on advanced AI chips, such as Nvidia’s H100 and A100 models.
In fact, those limits might have been a godsend for the company’s breakthroughs. DeepSeek says it found workarounds for greater chip capacity and applied other crucial strategies that competitors are now likely poring over in detail:
- Mixture-of-experts (MoE) architecture: Activating only a subset of parameters per task (e.g., just 5% of all available tokens), slashing computational costs.
- Reinforcement learning: Training models through trial-and-error feedback, improving reasoning skills.
- Distillation techniques: Compresses large models like GPT-4 into smaller, cheaper versions.
This combined approach enabled the company to train its models using about 2,000 Nvidia GPUs over 55 days at a cost of around $5.6 million, a fraction of what U.S. tech giants are spending.
How Does This Affect AI Investing?
DeepSeek’s success challenges the prevailing idea fueling massive investments in AI in the U.S.—that AI development requires endless piles of cash for massive spending on Nvidia-type chips and other expensive technology.
Major tech stocks in the U.S. had significant declines on news of these developments. Nvidia’s stock plummeted nearly 17%, the largest single-day loss in U.S. stock market history. Other stocks in the AI space, including Microsoft Corporation (MSFT), Alphabet Inc. (GOOGL), and ASML Holding NV (ASML), also plummeted.
The DeepSeek moment also creates opportunities for investors in the AI space. Here’s what savvy investors are likely to do:
- Rethink the “bigger is better” mentality. Investors might want to seek out companies that are investing in more efficient training methods and energy-efficient technology, not those blindly expanding capital-intensive GPU clusters. Firms leveraging decentralized GPU clouds may have an edge.
- Prepare for a surge in AI access: The barriers to AI adoption seem to be falling. Investors should watch for emerging AI applications and companies developing real-world AI products, rather than simply betting on foundational model providers.
- Focus on software: While investors have driven AI-related chipmakers like Nvidia to record highs, the future of AI may rely more on software changes than on expensive hardware.
- Monitor geopolitical risks: DeepSeek’s success will likely intensify U.S.-China tech tensions. Investors should monitor policy shifts, such as tighter export controls.
DeepSeek’s Generative Error Problem
Investors should also stay updated as experts get a look under the hood at DeepSeek. An early study from NewsGuard, which rates the trustworthiness of news and information sites, included reasons for significant concerns about DeepSeek’s reliability. Despite topping App Store downloads, the Chinese AI chatbot failed accuracy tests 83% of the time, placing it near the bottom of evaluated AI chatbots—ranking 10th out of 11 competitors.
NewsGuard’s assessment uncovered several critical issues:
- 30% of responses contained outright false information
- 53% of responses provided non-answers to queries
- Only 17% of responses successfully debunked false claims
- The chatbot also tended to parrot Chinese government positions, even when answering questions unrelated to China, such as giving China’s diplomatic positions on irrelevant queries.
DeepSeek’s approach to accuracy thus seems to shift responsibility to users, with its terms of use notifying them to “proactively verify the authenticity and accuracy of the output content.”
The Bottom Line
DeepSeek may be a harbinger of a less costly future for AI. This could mean pivoting to a focus on software changes over the brute force of more and more expensive technology, open-source collaboration, and scalable infrastructure. But it also means looking past the hyped-up headlines and assessing whether DeepSeek offers something new and different or, given some early tests of its abilities, if it’s just another AI-produced hallucination.