The fall of the AI fortress: how DeepSeek is cracking the Stargate

(edited 01/28 09:45 CET to reflect new information that surfaced since the time of writing)

TL;DR -- DeepSeek is not only exposing weaknesses in the Stargate project but also catalyzes a fundamental shift in the global power dynamics of AI. They are showing that innovation can emerge from constraints, and that open-source models and optimization are key paths for the future of the sector. This disruption is opening up new opportunities for Europe to exploit fresh gold mines in the AI ecosystem. The focus of the AI industry will soon shift to efficiency, accessibility, and the democratization of AI. 

The Stargate project was announced this week, a massive $500 billion AI infrastructure private initiative involving OpenAI, Oracle, and SoftBank. It is backed by the Trump administration, which sees in it an opportunity to solidify U.S. dominance in AI. By announcing such a massive investment, the backers of Stargate hope to strengthen their “moat”, which OpenAI has been tirelessly working to establish. But DeepSeek's irruption into public awareness just a few days later is significantly threatening this ambition.

The Chinese company has developed AI models that rival the performance of U.S. tech giants,  but at a fraction of the training cost. Their DeepSeek-V3 model was trained in just two months for $5.6 million, whereas OpenAI reportedly spent $100 million to train its GPT-4 model. (Note: the total cost is estimated, depending on sources, to be 2x to 3x that -- still pennies compared to the budget required by other players)

“Brute force” approaches can be outsmarted 

DeepSeek just demonstrated that AI innovation does not necessarily depend on massive investments in cutting-edge hardware, and have put efficiency back as a key factor for success in AI, in a clear demonstration of the potential inefficiency of high-investment models like those pursued by Stargate. DeepSeek utilized less advanced chips and optimization techniques at multiple levels to achieve comparable, or even superior, results to those of US models. This approach challenges the notion that AI advancement requires access to the most advanced and expensive technologies.

Challenging global AI power structures

The success of DeepSeek quickly caused a 16.5% drop in Nvidia's stock, a key technology partner in the Stargate project, triggering widespread concern about the sustainability of massive AI investments in the U.S.. Meanwhile, Chinese tech stocks connected to DeepSeek's business model have surged.

The Trump administration's rushed announcement of the Stargate project underscores a sense of urgency to maintain U.S. leadership in the face of China's rapid AI advancements. The project’s goal is to create state-of-the-art AI infrastructure in the U.S. with over 100,000 American jobs. One can only wonder how busy top executives at OpenAI, Microsoft or NVidia are these days, rewriting a narrative to save faces and to secure the momentum they had built behind Stargate’s 500 billion commitments.

Many say that DeepSeek's emergence challenges the technological dominance of the United States in AI, proving that China is a significant player in the field; I'd rather repeat what Yann LeCun wrote yesterday: it challenges the notion that proprietary approaches to AI development are the only valid way and demonstrates that open source development can lead to equally performant models (despite current U.S. export restrictions on NVidia’s most advanced chips). This could significantly further global collaboration on AI development, and could lead to the integration of Chinese technology into the global tech ecosystem, a taboo topic so far. 

DeepSeek's success also suggests that large models may become commoditized, with decreasing market value as open-source models are emerging as likely game-changers by offering customizable and cost-effective solutions at the same level of performance. Their innovation democratizes access to AI by reducing development costs, opening up the field to new entrants and challenging the investment models of Western tech giants. And by encouraging global collaborations, it could trigger the emergence of a "second-hand market" for older hardware, further pushing infrastructure players to adjust to this new reality. The race for computing power and investment may become less relevant as optimization becomes a competitive advantage.

The Jevons Paradox: a cautionary note on efficiency and the environmental footprint of AI

Unfortunately, efficiency gains such as those implemented by DeepSeek won’t necessarily translate into a reduction of the environmental impact of AI. The Jevons paradox highlights that increased efficiency in resource use can lead to higher consumption of that resource, not less. This means that even if DeepSeek - and those who follow on their tracks of innovation through optimization and open source -  has achieved more efficient AI models with fewer computing resources, this efficiency could paradoxically lead to increased demand and usage of AI, potentially increasing environmental impact.

Constraint as catalyst: a defining moment in the AI revolution

Some like to say that necessity drives innovation; the DeepSeek phenomenon also shows that creativity thrives under constraint. A lesson to be pondered in other areas of the AI revolution. 

It is of course too early to tell if Stargate will crumble or adapt and survive; regardless, the apparition of DeepSeek will be remembered as a strongly disruptive event in the AI saga.

-- Benoit

#DeepSeek #OpenIA #Stargate #OpenSource analysis #foodforthought #perspective #ArtificialIntelligence #Innovation #TechnologyStrategy #FutureOfTech #Leadership #economy #ecosystem #ai