Jerry Barnett - 28 Jan 2025

The artificial intelligence world has been rocked by recent claims from Chinese startup DeepSeek, which asserts it has developed an AI model rivalling those of tech giants like OpenAI and Google at a fraction of the cost. However, as the dust settles, scepticism is growing about the truth of these claims.

The DeepSeek Sensation

DeepSeek, a Hangzhou-based company backed by hedge fund High Flyer, made waves with the release of its DeepSeek-R1 model. The company claims that R1 matches or even surpasses the performance of OpenAI's ChatGPT o1 on multiple key benchmarks. What's more, DeepSeek asserts that it achieved this feat using only about 2,000 older generation Nvidia chips, at a cost of just $6 million in computing power. If true, this would mean that DeepSeek requires well under 10% of the processing power used by popular models from the likes of OpenAI, Google or Anthropic.

This announcement sent shockwaves through the tech industry. Nvidia, the leading manufacturer of AI chips, saw its stock plummet, losing nearly $600 billion in market value. The implications were clear: if DeepSeek's claims were true, it could upend the entire AI industry, challenging the massive investments made by US companies.

Scepticism Emerges

As the initial shock subsides, many experts are starting to cast doubt on DeepSeek's statements, the primary point of contention being the company's claim about its limited use of computing resources.

Alexandr Wang, CEO of Scale AI, publicly challenged DeepSeek's narrative. Speaking at Davos, Wang stated, "My understanding is DeepSeek has 50,000 H100s," referring to Nvidia's most advanced AI chips. This figure is dramatically higher than the 2,000 lower-power chips DeepSeek claims to have used.

Analysts have cast further doubt, pointing out the unlikelihood that a small startup could have discovered such enormous gains in efficiency that were meanwhile missed by very well-funded western research teams. Wall Street analysts Bernstein said “In short, we believe that 1) DeepSeek DID NOT “build OpenAI for $5M”; 2) the models look fantastic but we don’t think they are miracles”.

The Chip Question

Confusion is further fuelled by the context of US export controls on advanced semiconductors to China. These restrictions were specifically designed to limit China's access to cutting-edge AI technology. If DeepSeek indeed has access to 50,000 H100 chips, it would not only contradict their public statements but also raise questions about the effectiveness of these export controls.

Transparency Issues

Adding to the scepticism is DeepSeek's lack of transparency. While the company has open-sourced its model, allowing for some level of scrutiny, the details of its training process and resource usage remain murky. This opacity makes it challenging for independent researchers to verify DeepSeek's claims.

The Broader Implications

If DeepSeek's claims prove to be exaggerated or false, it could have significant repercussions. The company's assertions have already influenced market dynamics and policy discussions; the market shock alone might be seen as a positive outcome by Chinese state propagandists. US President Donald Trump referred to DeepSeek's rise as a "wake-up call" for the American tech industry.

Many experts argue that even if DeepSeek's claims are inflated, they still highlight important trends in AI development. The focus on efficiency and resource optimization, driven partly by export controls, could lead to genuine innovations in the field that would be beneficial globally: lower costs would reduce barriers to entry, allowing smaller companies to develop their own models. The environmental benefits of reduced power consumption would also be very significant.

Looking Ahead

As the debate continues, the world eagerly awaits more evidence to support or refute DeepSeek's claims. Regardless of the outcome, this episode underscores the intense competition in the AI sector and the geopolitical tensions surrounding technological advancement.

The true test will be whether DeepSeek can consistently deliver on its promises and compete with established players in the long run. As Sam Bresnick, a research fellow at Georgetown University's Center for Security and Emerging Technology, noted, "The real test of their effectiveness will be whether U.S. firms are able to continue to outpace China in coming years".

In the meantime, the DeepSeek story serves as a reminder of the need for transparency and verifiable claims in the fast-paced world of AI development. As the industry moves forward, the ability to balance innovation with credibility will be crucial for any company seeking to make its mark in this competitive landscape.