Disruptive innovation: Strategic lessons from China’s Deepseek AI

The sudden emergence of the startup Deepseek to compete with players like OpenAI or Meta is a thunderclap in the AI world. Disruptors are being disrupted in their turn. It’s a perfect illustration of the special nature of disruptive innovation that makes it so difficult to grasp. Here are eight lessons we can learn from this case.

The launch of Deepseek sent shockwaves through the AI world. Coming out of nowhere, its model is on par with the very best, such as ChatGPT or Claude. America thought it had finally dominated AI, and now everything is being called into question. And it’s not just the AI players who are trembling, but also the suppliers of high-end chips (the famous GPUs), network equipment suppliers and data center managers, and even suppliers of electrical equipment.

Indeed, training the best AI models is currently extremely expensive. OpenAI, Anthropic, etc. spend over $100 million on computing alone. They need massive data centers with thousands of GPUs at $40,000 each. They also need huge amounts of energy to run these centers. This whole model is being challenged.

Deepseek has actually rethought everything. Traditional AI consists of coding numbers to 32 decimal. Deepseek has managed to use only 8, resulting in a 75% reduction in memory usage and a gain in speed and energy. They also developed a “multi-token” system. Normal AI reads like a child, one word at a time. Deepseek reads entire sentences at once. It’s twice as fast and 90% more accurate. When you’re dealing with billions of words, the impact on raw efficiency is significant. What’s more, instead of being a generalist AI, Deepseek uses specialized experts that only wake up when needed. As a result, while traditional models require 1.8 trillion parameters to be constantly active, Deepseek manages with 670 billion, of which only 37 billion need to be active at any given time.

The results are impressive: the cost of training drops from $100 million to $5 million. The number of GPUs required drops from 100,000 to 2,000, and they are simpler (game console GPUs instead of high-end data center hardware). The cost of the API (access to the system for third-party applications), which determines the price customers pay, is 95% lower. And it’s all open source. Anyone can check their work. Finally, Deepseek has done all this with a team of less than 200 people, while giants like OpenAi or Meta maintain armies of engineers at staggering costs.

What lessons can we learn?

Deepseek is a classic story of disruptive innovation. It has the usual eight ingredients.

1. Disruption isn’t about doing things better, it’s about doing things differently. Deepseek rethought every aspect of the AI model, not hesitating to challenge models that seemed axiomatic (e.g., the only way forward is to increase raw performance).

2. Disruption changes the value of resources. Until Deepseek, mastery of high-end GPUs was strategic for AI. The US controlled the industry with Nvidia, and the ban on selling to the Chinese ensured that control was maintained. Now, if we can create an equivalent of ChatGPT with regular GPUs, Nvidia’s value collapses and so does their control of the market. It’s a bit like Kodak, which controlled the photo market with its chemistry and distribution network, and saw both become irrelevant in a digital world.

3. Disruption lowers barriers to entry and reopens the game. Until now, you needed a lot of capital to play in the AI world. Now you need much less. This is particularly good news for all the losers in the first round, especially the smaller European players who didn’t have, and never could have had, the same amount of capital. They can now hope to get back into the game. Breaking away is thus the best weapon against rent-seeking. The moats that seem to protect the big companies (market position, barriers to entry, regulation in their favor, etc.) are more like streams that can be easily crossed by those who know how and understand that they must not attack the leaders on their own turf.

4. Disruptions can happen very fast. We can see how quickly AI is evolving, with the field being reconfigured only a few months after the places seem to have been assigned. What’s interesting here is the speed with which the initial disruptors (OpenAI, Meta, NVidia) are themselves being disrupted, in just a few months. Remember that SpaceX, which also has a low-cost disruptive strategy in space, took a good ten years to have a real impact on an industry dominated for decades by incumbents like Ariane and Boeing.

5. Disruptions defy predictions by challenging the linear evolution of the world. The AI model based on increasing performance implied significant energy consumption, which did not fail to arouse criticism based on alarmist predictions. These were based on the model that tomorrow would be a continuation of today, a classic example of breakthrough-blind forecasting. This inevitability has been shattered by the collapse of the model. It’s reminiscent of the apocalyptic predictions made about the production of horse manure and urine in New York at the end of the 19th century, given the huge number of horses used for transportation. There was no way out except to ban horses. Only a decade later, the automobile had replaced the horse and the problem had disappeared.

6. Disruptions are the key to market growth. The cost reduction initiated by Deepseek is likely to increase the use of AI. Two centuries ago, the reduction in the cost of producing cloth made possible by the introduction of the loom led to strong market growth. More people could afford it. It’s a classic phenomenon of democratization made possible by lower costs. Deepseek’s breakthrough opens the door to mass AI. This means that hardware requirements (and costs) will go down in relative terms (much fewer resources needed for the same performance), but probably not in absolute terms (much more power used because it’s cheaper). NVidia is therefore not on the verge of bankruptcy, even if its margins are likely to fall.

7. Responding to a disruption is difficult for existing players. It’s difficult for a player who has bet on a heavyweight model to go back on it because it’s to his advantage to do so. This is reminiscent of the difficulties traditional airlines had in responding to the low-cost disruption: low-cost had no secrets for them, but their business model, and in particular the assets mobilized, made it unattractive to implement.

8. Disruption makes economic warfare and sanctions difficult. Deepseek was born out of China’s reaction to the blocking of sales of high-end Nividia GPUs. This shows once again that all systems are adaptive in the medium term. As good entrepreneurs and excellent engineers, the Chinese have turned a constraint into an opportunity. In a way, the American blockade has created a formidable competitor. Politicians have lost, entrepreneurs have won.

The show must go on

Of course, Deepseek is not without its flaws. Being Chinese, it is forced to reflect the views of its government in the answers it gives (though American AIs also suffer from significant biases). But if it shows one thing, it’s that the game is far from over. Disruptive innovation is resistant to predictions, linear trajectories, taken-for-granted situations, and political manipulation. Get ready for more surprises…

🇫🇷 A version in French of this article is available here.

🔎 Sources for this article: X thread from Morgan Brown (@morganb) and The Economist article: DeepSeek sends a shockwave through markets.

Read my previous articles on similar topics: In praise of indirection, or how problems aren’t always solved by problem solving; The incumbent’s (difficult) response to a disruption: Google and ChatGPT.

📬 If you liked this article, why not subscribe?