In a world where “less is often greater than more,” the power of constraints to drive innovation is vividly illustrated by two seemingly unrelated stories: Mick Jagger’s iconic dance moves and DeepSeek R1’s groundbreaking AI platform. By achieving remarkable AI capabilities at a fraction of the cost of its rival, OpenAI’s ChatGPT, DeepSeek R1 has shaken up the tech industry—challenging the notion that only well-funded giants can lead in innovation.

Listen to the “How Mick Jagger and DeepSeek Turned Constraints into Breakthroughs” Deep Dive.

The concept that “less is greater than more” and how business limitations can be transformed into advantages is powerfully illustrated in the book, A Beautiful Constraint. The idea of finding beauty in constraints is introduced early on, with a compelling example from Mick Jagger’s career as the front man for the Rolling Stones. The authors reference guitarist Keith Richards’ autobiography, where Richards explains that, due to the small size of the stages and venues the Rolling Stones played early on, Jagger was confined to a space no larger than a small table. Rather than letting this limitation stifle his performance, Jagger turned it into an opportunity, crafting a distinctive and flamboyant dance style that would become synonymous with him and the band.

This principle of doing more with less resonated deeply with me, especially in light of recent developments in the tech industry. This week, with the launch of the AI Large Language Model, DeepSeek R1 shook the industry and stock market by demonstrating how constraints can drive innovation. Built for just 5% of the $1 billion+ cost required by OpenAI to develop its AI platforms, DeepSeek R1 is a testament to efficiency and ingenuity. According to reports, DeepSeek’s engineers needed only about 2,000 specialized computer chips from Nvidia to develop its platform compared to the 16,000 chips used by major American companies to do the same thing. To put this into perspective, Chinese engineers spent just $6 million on raw computing power to build the DeepSeek R1 system—a fraction of the $60 million that Meta reportedly invested in its latest AI technology. Last Monday, DeepSeek also became the most downloaded free app in the Apple App Store, overtaking ChatGPT.

The financial implications of this breakthrough were immediate and staggering: Nvidia, a key player in the AI hardware space, saw its stock plummet by nearly 17% in a single day, erasing approximately $600 billion in market capitalization. This dramatic shift underscores how DeepSeek’s innovative approach is not only redefining the AI landscape, but also challenging the status quo of resource-intensive development.

DeepSeek’s Competitive Advantage: Using Constraint to Spark Innovation

Over the past several days, using various writing prompts, I’ve been comparing DeepSeek R1 with ChatGPT-4 and Perplexity. DeepSeek’s responses consistently stood out for their depth and thoughtfulness, often providing more thorough and nuanced answers. By contrast, ChatGPT and Perplexity tended to deliver more basic responses, requiring me to follow up with additional prompts to achieve a level of detail and insight comparable to DeepSeek’s initial output. This distinction highlights DeepSeek’s ability to understand and address complex queries with greater precision and efficiency.

DeepSeek’s breakthrough in AI development showcases how constraints can spark innovation. Faced with limited resources compared to the tech giants, the company turned these limitations into advantages. It embraced efficient GPU use, clever data compression and selective model activation, proving that ingenuity can outweigh sheer financial power. By adapting to hardware restrictions and employing smart data strategies, DeepSeek created a powerful AI model at a fraction of the cost. This approach not only solved the company’s immediate challenges, but also opened doors for smaller players in the AI field.

When comparing prices and features, both DeepSeek R1 and ChatGPT delivered strong performances in addressing complex problems. Below are several benchmarks used to test the two platforms:

It’s important to note that while these benchmarks suggest a strong performance from DeepSeek R1, the overall comparison is nuanced. ChatGPT still maintains advantages in certain areas, such as multimodal tasks (i.e., supporting both text and image inputs). This is partly due to differences in token limits—the maximum amount of text a model can process in a single interaction. Specifically, ChatGPT boasts a larger context window of 200,000 tokens compared to DeepSeek R1’s 128,000 tokens.

Bigger Is No Longer Always Smarter

DeepSeek’s success story demonstrates that in the world of technology, creative problem-solving and resourcefulness can level the playing field. It does make you wonder: if OpenAI had faced similar constraints, would it have achieved comparable results with fewer resources? Does OpenAI’s virtually unlimited financial access make it less inclined to pursue cost-effective solutions? I’m not sure, but it certainly makes for an interesting conversation.

Morgan Stanley shared a different perspective: “Bigger is no longer always smarter. DeepSeek demonstrates an alternative path to efficient model training, diverging from the current arms race among hyperscalers by significantly increasing data quality and improving model architecture. DeepSeek now offers the lowest cost of LLM manufacturing, delivering frontier AI performance at a fraction of the cost, with output tokens priced 9x-13x lower than GPT-4 and Claude 3.5.”

When it comes to DeepSeek’s innovation, Bank of America highlighted this benefit: “Our view is that more important than the significantly reduced cost and lower-performance chips used to develop DeepSeek’s two newest models are the innovations introduced that enable more efficient—and less costly—training and inference in the first place. Lower AI compute costs should enable broader AI services, from autos to smartphones.”

Finally, in a recent email from The Neuron: “DeepSeek, the little AI that could … crash the markets, dominate the AI news cycle, inspire competition in China, inspire fear in Meta, inspire rage in OpenAI, and give Microsoft the ultimate revenge. Even as OpenAI and Microsoft actively investigate if DeepSeek trained its model off of OpenAI’s stolen data, Microsoft made DeepSeek’s top model (R1) available on the Azure AI Foundry (Microsoft’s AI cloud service).”

The notion that only companies with deep pockets can lead in AI innovation—or any innovation, for that matter—is being challenged. DeepSeek’s story demonstrates how constraints can not only motivate, but can also inspire teams to be agile, fearless and creative. When approached strategically, constraints can serve as catalysts for groundbreaking advancements—or iconic dance moves.

As always, thank you for reading (or listening). Questions? Feel free to email me here.

Image source: Photo by Solen Feyissa on Unsplash