In a short time, Chinese startup DeepSeek has rewritten the rules of global AI development. While the United States has long dominated AI innovation across the board, a small company from Hangzhou, China, has caused a global shockwave over the past few days. In this article, I dive into this new model—and spent a few days testing it myself.

The launch of DeepSeek feels like a classic Sputnik moment—an unexpected breakthrough that jolts the world awake and signals the beginning of a new era of technological progress. Just as the Soviet Union launched the first satellite in 1957, sparking the space race, DeepSeek may well mark the beginning of a new phase in the AI race.

I tried an earlier version of DeepSeek back in November, but it didn’t leave much of an impression. This newest release, however, left me stunned—especially when comparing it to the major American players.

How DeepSeek Stands Out from U.S. AI Models

Look at the biggest, most powerful models—GPT, Gemini, LLaMA—and the cloud infrastructure and chips required to run them. Nearly all of it is in the hands of U.S. companies. Out of nowhere, this Chinese startup emerged with an AI model that, on several fronts, outperforms its American competitors:

  • Training costs: While U.S. models reportedly required hundreds of millions of dollars to train, DeepSeek claims to have done it for just $6 million.
  • Performance: In independent benchmark tests, DeepSeek outperformed Meta’s LLaMA 3.1, OpenAI’s GPT-4o, and Anthropic’s Claude Sonnet 3.5 on accuracy—across complex problem-solving, math, and coding.
  • Cost-efficiency: DeepSeek is 98% cheaper to use than GPT or Gemini.

“To see the DeepSeek new model, it’s super impressive in terms of both how they have really effectively done an open-source model that does this inference-time compute, and is super-compute efficient. We should take the developments out of China very, very seriously.” – Satya Nadella, CEO of Microsoft

It’s always worth looking at actual numbers in AI. We’ve been so focused on developments in the U.S. with familiar tools like GPT and Gemini, but behind the scenes, China has been building aggressively. Last year alone, China filed 38,000 AI patents (compared to 6,300 in the U.S.), has the largest active AI user base, and ranks second only to the U.S. in the number of launched AI models.

Antifragility in Action

But what struck me most was something Nassim Taleb—one of my favorite authors—describes as antifragility: systems or entities that grow stronger through stress, limitations, or adversity. China has been severely restricted by U.S. sanctions, especially around chip imports necessary for running AI models. But DeepSeek is a perfect example of antifragility—it was forced to become more creative and efficient, ultimately surpassing those who didn’t face such limitations.

“Necessity is the mother of invention. Because they had to figure out work-arounds, they actually ended up building something a lot more efficient.” – Aravind Srinivas, CEO of Perplexity

The results are already clear. DeepSeek became the most downloaded app in recent days, caused a $1.2 trillion drop in Western AI stock valuations, and made the American Stargate project look outdated by comparison. Meta has reportedly launched multiple “war rooms” to study how DeepSeek developed its model so efficiently—especially after DeepSeek announced it would invest another $60 billion into AI.

More Temu Trash or TikTok Brilliance?

Looking at the model overall, I see several clear advantages over GPT 4o:

  • Open-weight model: DeepSeek-R1 is open-weight—its training data isn’t public, but the algorithms can be studied and modified. That’s not possible with GPT-4o, which is fully closed-source.
  • Chain of Thought (CoT) reasoning: The model solves complex problems step-by-step, much like humans do. This makes it better at multi-step reasoning tasks. In coding tasks, DeepSeek not only provides the code but also explains how components work together—great for beginners.
  • Mixture-of-Experts (MoE) architecture: With 671 billion parameters, only 37 billion are activated per task, making it highly efficient in terms of computing power and energy usage.
  • Open source & low cost: DeepSeek is open-source and largely free to use—unlike GPT’s paid models. It can even run locally (on a MacBook, for example), reducing costs and privacy concerns, and offers cheaper API access.

“We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive – truly open, frontier research that empowers all.” – Jim Fan, Senior Research Manager at NVIDIA

Battle of the Bots

All that sounds great—but does it actually work better? After a mediocre test back in November, I gave DeepSeek a proper second chance—running side-by-side comparisons with GPT across a range of simple and complex daily tasks.

Where DeepSeek Shines

  • Creativity & adaptability: DeepSeek really stands out in creative tasks. Writing vivid character descriptions or catchy stories felt faster and more natural than with GPT. It easily adapts to the required tone and style—whether formal, playful, or anything in between—while GPT often needed multiple steps or prompts to do the same.
  • Coding help: I tested it with some buggy scripts. DeepSeek spotted issues quickly and offered not only accurate fixes but also clear, beginner-friendly explanations.
  • Speed & efficiency: Thanks to its MoE architecture, DeepSeek delivers fast, detailed responses—even for complex tasks. It was noticeably faster than GPT in most tests.

Where It Falls Short

  • Accuracy on niche topics: For very specific or historical topics, DeepSeek sometimes gave incomplete or incorrect answers. I noticed more hallucinations than with GPT.
  • Handling sensitive content: DeepSeek tends to avoid politically or historically sensitive issues—like the Tiananmen Square protests or the Nanking massacre—likely due to Chinese government influence.
  • Limited support & documentation: DeepSeek’s help resources are far less comprehensive than GPT’s, which can be frustrating for new users looking to get started. I struggled to find decent tutorials or explanations.

Other Cool Use Cases I’ve Seen

  • Easily build an app that scrapes YouTube channels and generates trend reports
  • Watch videos of how the model handles advanced reasoning tasks
  • Create a custom “ready-to-play” game in minutes

“We Recommend Ourselves” Syndrome?

Naturally, there’s some skepticism. The biggest critique? The claim that such a powerful model was trained with so little hardware. The only report with concrete figures comes—unsurprisingly—from DeepSeek itself.

And since DeepSeek is Chinese, some users worry about how their data might be processed or stored. Especially when handling sensitive information, that concern could become a barrier—even though there’s no concrete evidence of data misuse.

Still, no matter how you look at it, DeepSeek represents a massive leap in AI development. It offers real opportunities for Europe and other regions by lowering the barrier to advanced AI. The focus on efficient models requiring fewer resources makes high-quality AI accessible for small businesses, researchers, and emerging markets—especially important in Europe, where access, transparency, and support for underfunded startups are priorities.

Driving Global Competition

DeepSeek also fuels global AI competition. By pushing forward despite trade restrictions, it shows that innovation doesn’t require unlimited budgets or resources. This could inspire other regions—including Europe—to pursue smarter, more efficient paths to innovation.

In just a few weeks, DeepSeek has accelerated the pace of AI innovation and created a more level playing field. I’m very curious to see how American competitors—and policymakers—will respond.