DeepSeek’s R1: The New AI in Town

Prospect v Evans [2024] EWHC 1533 (KB)

February 8, 2025

Trans-Pacific Turmoil: How Trump’s Trade Policies could Impact Global Trade

February 8, 2025

Published by Imran Chaudhri at February 8, 2025

Introduction: The $600B Question

Here’s a one-sentence investor horror story to kickstart your read: On the 27th January 2025, Nvidia suffered a staggering 17% single-day drop in shares, wiping out nearly $600 billion in market value. Although the chip pioneer has shown tentative signs of recouping its previous valuation, it still has a long way to go before fully recovering from its record-breaking loss. Talk about a portfolio nightmare!

But what triggered this panic sell-off? Enter DeepSeek’s R1. The Chinese AI startup released its very own large language model (LLM) just a week before the biggest market upset in U.S. history. Boasting advanced capabilities that rival OpenAI’s ChatGPT-o1, Anthropic’s Claude 3.5, and Meta’s Llama 3.3 (yet developed at a humiliating fraction of the cost and time), it’s no surprise that its debut sent shockwaves through the tech industry, from Silicon Valley and beyond.

As a result, investors, industry leaders, and tech enthusiasts have been scrambling to assess DeepSeek-R1’s long-term impact globally. Is this the dawn of a new AI superpower, or just an overhyped breakthrough? What does this mean for Nvidia, which was once seen as untouchable in the AI arms race? With the stakes standing so high, it’s inevitably not just the future of generative AI (genAI) that will be affected. This article will recount the rise of DeepSeek-R1, its impact on Nvidia and the broader tech industry, and what it may mean for the future of genAI.

The AI Race and the Rise of DeepSeek R1

Ever since 2016, when the AI race began in earnest, the field has typically been dominated by US tech giants, who would roll out state-of-the-art models that would set the benchmark for LLMs across the world. OpenAI has long been a leader in this endeavour, with Anthropic, Google, and Meta trailing closely behind.

However, maintaining this competitive edge comes with a price, and an eye-watering one at that. For instance, the initial training for ChatGPT-4 cost more than $100 million as it relied on supervised learning, which requires large volumes of labelled data, computational resources, and multiple cycles of feedback and adjustments. While AI companies shy away from revealing the full expenses that go into training their models, experts estimate that these costs are growing exponentially behind the scenes, reaching into the billions. To facilitate these immense demands, tech companies depend on high-end chips capable of delivering the necessary processing power, memory bandwidth, and parallel processing capabilities, not only to sustain training phases but also to efficiently handle user requests.

Thus gave rise to the symbiotic relationship between LLMs and graphical processing units (GPUs). For years, it was widely assumed that achieving top AI performance required the most advanced chips available, subsequently fuelling Nvidia’s dominance in the AI hardware market and justifying the exorbitant costs associated with developing cutting-edge LLMs.

That is until DeepSeek’s R1 came to town.

The Chinese startup’s breakthrough model not only rivalled the performance of OpenAI’s ChatGPT-4 but did so at a fraction of the cost. Compare OpenAI’s $100 million setback to DeepSeek’s mere $5.6 million! This staggering efficiency was allegedly achieved through a combination of innovative techniques, such as knowledge distillation and the strategic use of lower-grade hardware. By using larger, more advanced AI models to train the smaller, more efficient R1, and pairing banned Nvidia H800 and A800 chips with cheaper, less sophisticated alternatives, DeepSeek was able to significantly reel in its training expenses without compromising output quality or capability.

But cost efficiency wasn’t the only factor behind DeepSeek’s meteoric rise. The company’s open-source approach set it apart from its US counterparts, where unlike proprietary models from OpenAI, Anthropic, and Google, DeepSeek released R1 with open weights, allowing developers worldwide to download, modify, and build upon its architecture. This move not only democratised access to cutting-edge AI but also sparked a wave of innovation, particularly among smaller players and researchers. The transparency and accessibility of DeepSeek’s model resonated deeply with the global AI community, further boosting its popularity.

As a result, its impact was immediate. DeepSeek’s AI assistant stormed the iPhone download charts, unprecedentedly lifting Asian tech stocks. Investors, once confident in the dominance of U.S. tech giants, began questioning the sustainability of massive AI spending. Meta’s announcement of a $60 billion capital expenditure plan for 2025, for instance, now faces scrutiny in light of DeepSeek’s cost-effective approach. Meanwhile, we already know what went down with Nvidia. All-in-all, DeepSeek’s rise underscored the importance of efficiency, ingenuity, and collaboration in the AI race, an important lesson that US tech giants are now scrambling to internalise.

Why Throw Stones at Nvidia?

It’s been established just how integrated Nvidia and its products are in supporting frontier LLMs, where for years, the company’s high-end GPUs, such as the H100 and A100, have been the gold standard and thus indispensable to AI developers. However, DeepSeek’s success challenged this narrative by demonstrating to the world that it is possible to develop and deploy an advanced AI model without relying on the most expensive hardware on the market. Its incredible performance cast doubts about the future demand for Nvidia’s flagship products, whose investors feared that if AI computing became more efficient, the need for the most expensive high-end chips and cloud infrastructure would drop, threatening Nvidia’s dominance.

Yet, this panic may be short-sighted. While DeepSeek’s breakthrough underscores the potential for cost-efficient AI development, it doesn’t spell the end of Nvidia’s relevance. In fact, the broader implications of cheaper AI could ultimately benefit the tech giant. As AI becomes more accessible, its adoption is likely to explode across industries, driving demand for cloud infrastructure and computing resources. Lower costs will help businesses, developers, and startups to integrate AI into their operations, creating a surge in demand for GPUs, even if they aren’t the most expensive models. Moreover, the sheer scale of AI adoption means that data centres, energy infrastructure, and chips will remain critical, ensuring Nvidia’s products stay in high demand.

Microsoft CEO Satya Nadella aptly summarised this dynamic: AI isn’t just getting cheaper; it’s getting better. Efficiency gains could unlock entirely new use cases, much like the falling cost of cloud computing enabled the mobile app revolution. For Nvidia, the challenge will be to adapt to this new reality, whether by diversifying its product line, lowering prices, or investing in innovations that align with the industry’s shifting priorities. While DeepSeek’s rise has undoubtedly destabilised Nvidia’s position, it also highlights the growing ubiquity of AI, a trend that could ultimately reinforce the chipmaker’s long-term relevance.

DeepSeek’s R1: Here to Stay?

Overall, it is undeniable that DeepSeek-R1 has thoroughly shaken up the AI industry, but the question remains: is it a sustainable competitor to established players in the West? On paper, DeepSeek-R1’s achievements are short of remarkable, seeing as it rivals the best LLMs in reasoning and coding benchmarks despite being developed at a fraction of the costs and resources.

However, DeepSeek’s success is not without serious caveats.

As a Chinese startup, it faces unique challenges, including severely restricted access to diverse and clean datasets and high-quality compute resources. Liang Wenfeng, DeepSeek’s founder, acknowledged that Chinese LLMs face a “training and data efficiency gap” as they often require twice the computing power and data to achieve results comparable to their Western counterparts. Additionally, DeepSeek’s open-source approach, while honourably democratising AI, exposes it to risks of intellectual property disputes and potential misuse of its technology. The ongoing investigations by Microsoft and OpenAI into whether DeepSeek secretly accessed and reverse-engineered their proprietary data further complicates its position and have reignited tensions over tech security.

Moreover, while DeepSeek-R1 has performed impressively, it remains to be seen whether it can sustain its momentum. OpenAI’s o3 model, though not yet public, has reportedly achieved groundbreaking results in PhD-level benchmarks, suggesting that the US giant is far from ceding its leadership. Furthermore, with President Trump and his administration launching investigations into DeepSeek’s supply chain on suspicions of the company exploiting US chip restriction loopholes, AI innovation is no longer just a matter of corporate competition but also national strategy. Even now, a newly proposed bill seeks to criminalise the use of DeepSeek in the US, with individuals potentially facing up to 20 years in prison and businesses fined as much as $100 million for violations. If enacted, this would not only stifle DeepSeek’s international reach, but further entrench AI development within geopolitical battle lines, making its ability to compete an even greater test of its longevity.

Conclusion

DeepSeek’s R1 is a testament to the power of innovation and efficiency in an industry often dominated by scale and spending. While its long-term viability remains uncertain, its impact is unmistakable: it has forced a reevaluation of what it takes to compete in the US-dominated AI race and emphasised the growing influence of Chinese tech innovation.

However, the road ahead is still fraught with uncertainty. The commoditization of LLMs appears inevitable, but the winners of this new era will likely be those who can balance innovation with scalability, transparency, and ethical considerations. For OpenAI and its peers, the challenge will be to adapt to a landscape where efficiency and openness are as important as raw performance, while for GPU giants like Nvidia, the focus will be on maintaining its relevance in a market that may no longer demand its most expensive chips.

With more to come, we can be sure that the story of DeepSeek-R1 is far from over and its legacy will undoubtedly influence the next chapter of AI history.