DeepSeek’s Disruption in the U.S. AI Industry

This week, DeepSeek, a Chinese AI startup founded in 2023, upended the blueprint of the AI Industry. Their model is industry leading and was built quicker and cheaper (by a lot) compared to the competition including Anthropic’s Claude, ChatGPT, Gemini, and just about everything else built to date. The results were immediate and dramatic: Silicon Valley is facing a heightened era of competition, Wall Street is questioning where it should focus its investments, and the US-led AI landscape is in scrambling to understand what their next step should be. Here’s why.

The Fallout: Wall Street’s Crisis of Confidence

DeepSeek’s release of its AI model, R1, has not just captured headlines - it has shaken investor confidence. The shared benchmarks of R1 show remarkable performance, rivaling that of OpenAI’s GPT-4 but at a fraction of the cost and hardware requirements. The stock market’s response was swift and brutal:

Google’s stock dropped 2.8%.

• Microsoft fell by 3.8%.
• NVIDIA, a critical supplier for AI hardware, saw its market cap plummet by $600 billion at one point, the largest one-day loss in US history.

This response underscores an uncomfortable reality: the bedrock assumptions underpinning Silicon Valley’s AI dominance are now in question. Wall Street, long aligned with Big Tech’s strategies, is beginning to doubt the necessity of costly data center and hardware investments that were once deemed essential. This can be seen not only in the tech stock price drops, but also the energy sector.

Why DeepSeek’s Model Is a Game-Changer

DeepSeek’s R1 began with R1-Zero, a model trained entirely through reinforcement learning. While this approach allowed it to develop strong reasoning capabilities, it came with major drawbacks. Outputs were often difficult to read, and the model sometimes mixed languages within responses, making it less practical for real-world applications. To address these issues, DeepSeek shifted its approach by combining reinforcement learning with supervised fine-tuning, incorporating curated datasets to improve readability and coherence. Problems like language mixing and fragmented reasoning were significantly reduced, making the model much more suitable for practical use. What this means is they started with a method where the AI learned all by itself, which made it incredibly intelligent without being able to express what it knew, then created a new way of teaching it by showing it examples of clear and correct responses so it could learn how to explain itself better.

Additionally, DeepSeek utilized distillation techniques to create smaller, more efficient models while maintaining high levels of performance. These distilled models, such as DeepSeek-R1-Distill-Qwen-32B, showcased impressive capabilities:

MATH-500 Benchmark: Scoring 94.3%, it demonstrated exceptional mathematical reasoning.

AIME 2024 Benchmark: Achieved 72.6%, excelling in advanced multi-step reasoning tasks.

GPQA Diamond: Scored 62.1%, reflecting strong factual reasoning capabilities.

Coding Benchmarks: While not optimized for programming, it achieved competitive scores, such as 57.2% on LiveCodeBench.

When compared to OpenAI’s o1 models, DeepSeek-R1 holds its ground. It slightly outperforms OpenAI o1 on mathematics benchmarks like MATH-500 (97.3% vs. 96.4%) and AIME 2024 (79.8% vs. 79.2%). While OpenAI retains a slight edge in coding and factual reasoning, DeepSeek’s cost-efficiency and performance make it a formidable competitor.

Janus-Pro-7B: Also a Threat to DALL-E and Other Image Models

DeepSeek’s Janus-Pro-7B also released on Monday (Jan 27th 2025) outperforms DALL-E 3 and Stable Diffusion in quality, efficiency, and accessibility. With a GenEval benchmark score of 0.80 (compared to DALL-E 3’s 0.67)). This is an open-source model that produces sharper, more realistic images while running on fewer resources with a development cost under $6 million and took just two month The accessibility and superior performance of Janus-Pro-7B pose a significant threat to proprietary models like DALL-E. By delivering state-of-the-art image generation with its open-source approach and reduced computational requirements, it erodes the competitive edge of established players that rely on exclusivity and high resource demands. Its practical applications across advertising, media, and education also further cement its potential to disrupt traditional market dominance.

So What Does This Mean?

DeepSeek’s achievement is impressive not only for its results but for how it only needs a fraction of the resources required by ChatGPT, Gemini, Llama, Claude, and its other US based competitors. This shift is as transformational as the transition from gas-powered cars to electric vehicles or from incandescent bulbs to LEDs. It’s a reminder that innovation often stagnates under the weight of entrenched practices and the sunken cost fallacy.

Consider Blockbuster’s stubbornness in the face of Netflix or Kodak’s slow pivot from film to digital. Silicon Valley now faces a similar moment of reckoning. The AI giants have long invested billions in chips, power, and expansive data centers. They didn’t need to. DeepSeek’s success proves these investments may not be as indispensable as once believed.

Borrowing and Building: The Apple Playbook

It’s important to note that DeepSeek didn’t start from scratch. Like Apple refining features that Android pioneered, DeepSeek built on existing AI innovations - polishing and optimizing them. The difference? DeepSeek delivered its model faster, cheaper, and arguably better (we’ll get back to “better” in a bit, as there are clear concerns to be raised.

Deepseek's Hidden Agenda and The Implications for Silicon Valley

DeepSeek’s dataset is censored, there's no quesito about that. As It's limiting information on politically sensitive topics, the U.S. Navy has now also issued a sweeping ban on DeepSeek, warning its members not to use the model for work or personal purposes due to "potential security and ethical concerns. " A spokesperson for the Navy's confirmed the ban, which was based on advisory input from their cybersecurity leadership. This action highlights the rising fears around data manipulation, espionage, and trust issues surrounding AI models originating from geopolitical rivals.

DeepSeek’s entrance in showing how AI can be done cheeper and better is shaking Silicon Valley's foundations, but it's also exposing a glaring double standard. OpenAI built its models using vast amounts of copyrighted material without permission and continues to dismiss the legal challenges as a hindrance to innovation. Yet, Open AI cries foul claiming that DeepSeek's success is unfair because its model may have borrowed from their own AI architectures and techniques.

The irony is hard to ignore. OpenAI and others justified their unrestricted use of copyrighted works under the banner of progress, arguing that access to large datasets was essential to AI development. Now, when faced with a faster, cheaper competitor like DeepSeek's - possibly inspired by their own work — they’re suddenly taking issue with what amounts to the same strategy they once defended. This inconsistency raises serious questions about who truly owns innovation and whether claims of copyright infringement are more about maintaining competitive dominance than protecting intellectual property.

There’s also a geopolitical dimension to this disruption. China has fewer restrictions on AI development, and DeepSeek’s rise signals the potential for the country to leapfrog the U.S. in AI innovation. While the U.S. government has recently loosened regulations (such as Trump killing Biden’s Ethical AI Executive Order), it may not be enough to counteract the momentum China has gained.

Yet, I Still Won’t Use DeepSeek

Despite the impressive benchmarks, it’s not going to be better to use and as a result, I won;t be using DeepSeek in its current (and likely future) state. Here’s why:

Censorship and Bias: DeepSeek’s dataset is censored. I tested this earlier this week on topics known to not agree with the CCP. The results? DeepSeek doesn’t provide information on topics like Winnie the Pooh, Taiwan, or Tiananmen Square. This level of control introduces the risk of outright lies rather than hallucinations. While these are sensitive topics to China, there’s no reason to believe it also won’t promote things that promote CCP’s views and are detrimental to the US, EU or other nations where China seeks a competitive edge or dominance. This makes the model not just inaccurate, but dangerous.

Privacy Concerns: The idea of installing an app from a Chinese AI company on my phone also raises serious red flags. With the potential for Apps to be able to access financial data, personal information, and other sensitive details connected to a user’s phone and active accounts, the risks far outweigh the benefits.

A Wakeup Call for U.S. AI and its Entrenched Models

DeepSeek’s R1 is not just an impressive innovation; it’s a challenge to the U.S. AI industry. The industry’s reliance on costly infrastructure and incremental innovation has left it vulnerable to faster, more agile competitors. The lesson here is clear: it’s time to rethink entrenched practices and assumptions. For instance, Meta is now reportedly scrambling to figure out how DeepSeek is doing it at a fraction of the cost. Otherwise, Silicon Valley may find itself in the same position as Blockbuster - watching the future unfold in someone else’s hands. Either way, DeepSeek has ensured that the generative AI landscape will never be the same.

Feel free to share your thoughts below!

This article was first published in my the Pandora's Bot Newsletter series. If news like this is of interest, don't forget to subscribe! And if you think your friends or colleagues would enjoy it, please share thepandorasbot.com!