DeepSeek-R1: A Cost-Efficient Disruptor Shaking Global AI Markets

Memo: 27 January 2025

Background

A two‐year‐old Chinese AI startup, DeepSeek, is rewriting the rules of AI development. The firm’s latest reasoning model, DeepSeek‑R1, built at a reported cost of just $6 million,¹ has rattled investors and sent shockwaves through Silicon Valley. DeepSeek’s AI assistant became the most downloaded free app on the US iOS App Store, while US tech stocks — most notably Nvidia — suffered record losses, with Nvidia’s market value reportedly plummeting by up to 18% in a single day.

Founded by hedge fund entrepreneur Liang Wenfeng and funded by High-Flyer, a Chinese hedge fund, DeepSeek has focused on developing large language models (LLMs) through a research-first approach. Its latest model, DeepSeek-R1, is touted as capable of advanced reasoning — comparable to leading US models such as OpenAI’s o1 — yet was trained at a fraction of the cost (approximately $6 million versus over $100 million for competitors) using fewer and lower-tier Nvidia H800 chips due to US export restrictions.

Analysis

Key Findings and Their Implications

• Smarter AI Without Constant Supervision: DeepSeek found a way to make AI reason better without needing lots of labelled examples (where humans tell it what’s right and wrong). Instead, they used reinforcement learning (RL) — a process where the AI improves by trial and error, similar to how people learn from feedback.

  • What happened? The AI (DeepSeek-R1-Zero) initially got a 15.6% score on a key reasoning test but jumped to 71% after training with RL. With a technique called majority voting, it reached 86.7%.

  • Why is this important? Instead of relying on massive amounts of expensive, hand-labelled data, AI can now teach itself. This could make AI training cheaper and more scalable, potentially leading to self-improving systems that don’t require constant human oversight.

• Making AI Understandable and Useful: DeepSeek didn’t just make their AI smarter — it also made it more readable and coherent by feeding it better initial training data (cold-start data) and fine-tuning it with a multi-stage approach.

  • What changed? The AI’s responses became clearer, more structured, and better suited for real-world applications.

  • Why does it matter? Many AI models struggle to generate understandable and reliable answers, especially in fields like healthcare, finance, and law, where precision and clarity are critical. This improvement makes AI more practical and trustworthy for real-world users.

• Powerful AI in a Smaller, Cheaper Package: DeepSeek developed a method to compress its most advanced AI model into a smaller, more efficient version without losing much performance.

  • What does this mean? Instead of needing huge, expensive computer systems, smaller, cheaper versions of DeepSeek’s AI can run on less powerful hardware while still performing well.

  • Why is this a big deal? AI is often too expensive for many companies because running large models requires massive computing power. If smaller AI models can think just as well as bigger ones, more businesses and individuals will be able to access cutting-edge AI without breaking the bank.

• A Smarter AI Engine: DeepSeek-R1 is built on top of DeepSeek-V3’s innovations which includes design choices like Mixture of Experts (MoE) and load balancing to use computing power more efficiently.

  • How does it help? These design choices make the AI model faster, more reliable, and cheaper to train.

  • Why is this important? Prior AI models burn through millions of dollars in electricity and upfront hardware costs. DeepSeek’s improvements reduce these costs² while keeping performance high, making advanced AI more sustainable and scalable.

Disruptive Impact on Market Sentiment and Underlying Motivations

The launch of DeepSeek-R1 has spooked investors and tech giants alike, triggering a dramatic plunge in Nvidia’s stock that erased nearly US$600 billion in market capitalisation in a single day. Policymakers and business leaders are now re-evaluating long-term investments in costly AI infrastructure, recognising that smaller, lean startups can achieve powerful results with far fewer resources. Political figures, including President Trump, have called the development a “wake-up call,” warning that cost-effective innovations may undermine the competitiveness of firms reliant on huge capital expenditures. Meanwhile, Microsoft’s Satya Nadella described DeepSeek’s work as “super impressive,” highlighting that even established tech giants may need to adjust their investment strategies.

Geopolitical and Policy Implications

DeepSeek’s success has sparked debate in Washington over the efficacy of US export controls on advanced semiconductor technology. That DeepSeek built a state-of-the-art model with older, cheaper chips illustrates how innovation can outpace material constraints. This has prompted some policymakers to consider stricter controls or re-examine their broader approach to maintaining technological supremacy. DeepSeek’s decision to open-source its models under an MIT License further challenges the proprietary norms of US tech, prompting calls to support R&D efforts in this rapidly shifting global AI landscape.

Challenges and Limitations

• Sustainability and Full Cost Transparency: While DeepSeek’s reported training cost is remarkably low, questions remain as to whether this figure fully accounts for pre-research investments, infrastructure, and potential repeated training runs. Some industry experts suggest that the true costs may be higher, though still significantly lower than those of US counterparts.

• Intellectual Property and Ethical Concerns: DeepSeek’s aggressive use of open-source techniques and possible reliance on knowledge distillation from competitor models has raised questions about intellectual property rights and ethical boundaries. Moreover, the model’s embedded censorship to comply with Chinese regulations poses challenges for its global adoption, especially in democracies with strict data privacy norms.

• Long-Term Competitive Sustainability: The initial market shock may prompt US companies to innovate further and lower their own costs. However, DeepSeek’s agile, research-focused approach — if sustained — could compel a broader industry shift towards cost-effective AI solutions. The strategic balance between low-cost innovation and long-term performance remains an open question.

Future Significance

DeepSeek-R1 demonstrates that frontier-level AI can be achieved without astronomical budgets. This has several potential outcomes:

• Reassessment of Investment Strategies: US tech giants and investors may be forced to re-examine their capital allocation for AI, potentially redirecting investments towards optimisation and efficiency rather than brute-force scaling. In the longterm, DeepSeek-R1 ought to encourage more investment; if DeepSeek can produce a frontier model for less than a 1/10th the cost of OpenAI’s training runs, then this does not suggest that OpenAI is overspending, rather that OpenAI could achieve 10x more with its current expenditure. Investments in Nvidia may continue adjusting to a potential short-term decline in demand.

• Policy and Regulatory Adjustments: As US policymakers debate export controls and national competitiveness, DeepSeek’s success may prompt a review of current semiconductor policies and market access, potentially influencing both domestic and international regulatory frameworks.

• Increased Global Competition: By open-sourcing its technology through an MIT License, DeepSeek invites global collaboration and innovation, which may accelerate the pace of AI development and force incumbent players to adapt to a more competitive, cost-conscious market environment.

Conclusion

DeepSeek’s innovations challenge the idea that powerful AI must be expensive and suggests:

1. AI can teach itself reasoning skills instead of relying on costly human-labelled data.

2. Smaller, efficient AI models can perform well, making advanced AI more accessible.

3. Smarter AI designs reduce training costs, pushing the field towards cheaper and more sustainable AI.

If these breakthroughs continue, they could disrupt the AI industry, making high-performance AI available to more companies, researchers, and even individuals — not just tech giants with billion-dollar budgets. The ensuing market reaction underscores the disruptive potential of such breakthroughs, pushing policymakers to reassess export controls, back efficient AI infrastructure, and craft new strategies for maintaining global competitiveness in emerging technologies.

1. USD unless otherwise stated.

2. Ostensibly.