12 February 2025
The release of the DeepSeek-R1 AI model on 20 January sent shockwaves across the world, wiping over $1 trillion off US tech stocks and soaring above ChatGPT in the Apple App Store.
Last week I looked at how DeepSeek-R1 works from a technical perspective, based on the research paper released by DeepSeek and analysis from data scientists.
This post will look at the broader implications of R1, focusing on:
Why R1 matters
Economic and market impact
Environmental impact
Geopolitical impact
Possible social impacts
1. Why R1 matters
Even if claims that DeepSeek has been produced for 5% of the cost of its competitors are misleading, DeepSeek-R1 represents a major progression from existing large language models (LLMs) for five main reasons:
It uses less energy
It uses fewer superprocessor microchips
It is open-source
It has made distilled versions available to run on systems as small as some PCs
It was produced in China, not the US
The extent to which the first two of these are true is unknown. If they are correct, the reduction in energy and reliance on rare earth minerals could have major economic and geopolitical implications.
Yet the final three points alone could revolutionise the AI industry, speed-up development of frontier systems, and lead to millions more people being included in AI design, potentially narrowing the “digital divide.”
Let's look at why that is...
2. Economic and market impact
The current global AI market is dominated by the US tech giants, who (with the exception of Meta) have chosen to keep their advanced models closed-source. The focus on generative AI (GenAI) and LLMs has pushed companies to develop increasingly large models, comprising billions of parameters.
"The global artificial intelligence (AI) market size was valued at USD 233.46 billion in 2024 […] The North America region dominated the AI industry with a market share of 32.93% in 2024."
Fortune Business Insights, Industry Reports, AI, 2024
According to Statista, China’s market share was 8% in 2024 (in contrast to the US’s 32.93%). However, DeepSeek overtook OpenAI’s ChatGPT as the most downloaded free app on Apple’s App Store on 27 January, marking a possible end to US AI dominance.
Following the release of R1, other Chinese firms released a number of AI models. Within days, Alibaba-backed start-up Moonshot released its Kimi k1.5 model, followed by Alibaba releasing an updated version of its Qwen 2.5 AI model, Qwen 2.5-Max, which it claims surpasses R1’s capabilities. ByteDance (parent company of TikTok) released its Doubao-1.5-pro model on 22 January, and Tencent (WhatsApp owner) continues to operate its Hunyuan model. Tencent claims that Hunyuan can perform as well as Meta’s Llama 3.1 but requires a tenth of the computing power.
Spending on AI has increased in China, with Goldman Sachs reporting a capital expenditure rise of 61% in the last 12 months albeit from a low base.
Following the release of R1, US tech stocks tumbled. Chip manufacturer Nvidia lost almost $600bn from its market value overnight, and Meta Platforms, Microsoft and Alphabet all saw major declines; analysts estimate the total losses across US tech stocks were over $1 trillion. Although some tech giants’ share prices have recovered some of their value, the effect of R1 on the long term market remains to be seen, with analysts predicting an end of AI hype investing in the US.
However, the most significant effects of R1 are likely to be on the direction of AI development, moving away from the gigafactories of recent US-based models, to a greater push for efficiency.
“Chinese players have been focused on driving the lowest cost, and also maybe trying to use minimal chips in doing the same tasks. I think over the last week, there's also been more focus on whether edge computing is becoming more popular, which could allow smaller AI models to run on your phone or computer without connecting to mega data centers.”
Ronald Keung, Head of Goldman Sachs Research Asia Internet Team
The low cost of R1 has also had an immediate impact. According to The Economic Times: “DeepSeek charges just $0.55 per million input tokens and $2.19 per million output tokens, which is far more affordable compared to OpenAI's API pricing of $15 and $60 for the same services.”
The pricing has led many to question the US AI market’s entire strategy, as argued by Jeff Sommer of the New York Times.
"One immediate question: Is the main approach to developing A.I. in the United States — pouring billions of dollars into chips and infrastructure — worth the expenditure for all companies if similar results can be achieved far more cheaply?"
Jeff Sommer, New York Times, 07 February 2025
3. Environmental impact
By demonstrating that advanced models may be developed with less energy and fewer superprocessor chips, DeepSeek had upended two key assumptions of global AI policy: that AI requires huge increases in energy production and extensive access to rare earth minerals.
National energy policy and AI
Some key national AI policies include:
US: The goal of President Trump’s Executive Order on Removing Barriers to American Leadership in Artificial Intelligence is “to sustain and enhance America’s global AI dominance” (section 1). Increased nuclear power plant capability is at the core of Executive Orders from Presidents Trump and Biden. The US also announced the OpenAI-led Stargate project on 28 January, building a series of huge AI data centres for approximately $500 billion, supported by increased nuclear energy infrastructure.
EU: The EU has pledged to create “AI gigafactories” according to its Competitiveness Compass, published on 30 January. Whilst the EU remains committed to decarbonisation, it is also preparing for a major increase in energy demand with an Affordable Energy Action Plan due to be published in the coming months.
UK: The AI Opportunities Action Plan also recognised the need for more energy to “power the increasing energy demands of AI” including greater reliance on nuclear power.
Energy efficiency
DeepSeek claims to require 50-75% less energy than Nvidia’s latest GPU units, due in part to its use of less-recent microchips, and 10-40 times less energy than competitors. These claims are pending third party review, but could represent a major re-thinking of national energy policies. For example, recent estimates suggest ChatGPT’s monthly carbon dioxide emissions may be over 260 tonnes (equivalent to driving a single car 1.3 million miles), and the Stargate project may require 15 GW of energy (over 3 times that of Greater London).
"In large economies like the United States, China and the European Union, data centres account for around 2-4% of total electricity consumption today. But because they tend to be spatially concentrated, their local impact can be pronounced. The sector has already surpassed 10% of electricity consumption in at least five US states. In Ireland, it now accounts for over 20% of all electricity consumption."
International Energy Agency, October 2024
If DeepSeek's claims that it requires far less energy to train and operate are accurate, countries could potentially rethink their energy strategies, spending less on increasing capacity or building new nuclear power plants. This could lead to the decarbonisation agenda of the Paris Climate Change Agreement returning to the forefront of geopolitics.
Water efficiency
Water usage is also a major factor in running large AI data centres. A 2023 study from the University of California found that ChatGPT-3 can use up to half a litre of water per query. Dr Venkatesh Uddameri, a Texas-based expert in water resources management, told the BBC that “a typical data centre can use between 11 million and 19 million litres of water per day, roughly the same as a town of 30,000 to 50,000 people.” This has prompted concerns over water shortages in the UK as a result of AI growth.
"DeepSeek's technology could mean predictions about AI's expanding resource use are exaggerated and some of the planned data centers might not be needed."
DW News, What does DeepSeek mean for AI’s environmental impact? January 2025
If producing DeepSeek-R1 genuinely required far less water and energy than competitors, the assumptions underpinning global environmental policy may need to be reconsidered. Even if these claims are untrue, the fact that its distilled models can be run on much smaller systems, including personal computers, means that individual users’ AI use may require far less energy.
Mining of rare earth minerals
Finally, the environmental impact of mining rare earth minerals to produce advanced chips is significant. If demand for such chips reduces as a result of efficiencies pioneered by DeepSeek (and this is a large unknown) then the devastating impact of mining could be reduced.
4. Geopolitical impact
Chips and minerals
Superprocessor microchips are a good place to start when considering the geopolitical ramifications of R1. The US and China have been engaged in a trade war over microchips and the rare earth minerals required to produce them since 2022, involving export controls and sanctions on both sides. These restrictions, according to DeepSeek, necessitated the innovations that led to R1’s efficiency.
However, MIT Technology Review reports that DeepSeek acquired a “substantial stockpile” (between 10-50,000) of Nvidia A100 chips before the US banned their export. Even so, DeepSeek claims that it used far fewer chips than its US competitors, being trained on 2,000 rather than the average 16,000 for Western models.
The need for fewer chips, and hence a smaller quantity of rare earth minerals, could be a benefit for Western AI companies, as the highest reserves are currently found in China, Vietnam and Russia.
Reserves of rare earths worldwide as of 2023, by country (1,000 metric tons REO)
In addition to the minerals required to make chips, their production is highly-specialised, with Taiwan still producing over 60% of the world's semiconductors and more than 90% of the most advanced semiconductors. In the first week of December 2024, China deployed the “largest fleet in decades” in the waters surrounding Taiwan. It is possible – but perhaps optimistic – to think that a reduced demand for advanced semiconductors could reduce pressure on those countries holding rare earth mineral reserves or with the highest rate of their production. These include Greenland and Canada.
US AI dominance
Most significantly perhaps, DeepSeek-R1 could be a “Sputnik moment” for the global AI industry, with the power balance shifting away from US tech giants. The US response to has included one Republican Senator proposing a bill to ban its use in the US, punishable with up to 20 years’ imprisonment, and a bipartisan bill to ban use of the DeepSeek app on government devices. South Korea, Taiwan and Australia have also banned the use of DeepSeek on government devices.
Privacy concerns have also resulted in action from EU privacy regulators. Italy’s Garante and Ireland’s Data Protection Commission have both placed temporary bans on the app, pending investigation into the use of EU citizens’ data by DeepSeek. Similar bans were imposed on ChatGPT when first released, though these were later lifted; the European Data Protection Board’s ChatGPT Task Force reported its interim findings in May 2024.
Although the online version of the model involves sending prompt information to China, and results are subject to Chinese censorship laws, it is not clear whether downloading an offline version of the distilled model results in data transfer to China when the device is next online. This question will likely form part of the data protection investigations in the EU.
AI nationalism
In spite of R1’s advances, AI nationalism appears to be increasing. As the AI Action Summit concluded in Paris yesterday, national sovereignty of AI systems remains a priority. President Macron unveiled a $100 billion pledge for private AI investment in France on Sunday, ahead of the summit.
“The future of AI is a political issue and an issue of sovereignty and strategic dependence”
President Emmanuel Macron, 09 February 2025
In contrast to expectations of a reduction in major AI infrastructure builds, Politico reports that many governments are looking to increase domestic AI spending, as they identify potential for innovation-led dominance in a more-open global AI industry.
"In fact, ahead of the AI Action Summit in Paris, governments have doubled down on a sovereignty-first AI strategy, with officials publicly rallying behind their AI 'national champions' and prioritizing strategies to race ahead on the development of the fast-moving technology as a way to hedge against tech dependency on other countries."
Politico, 10 February 2025
Yesterday, the EU announced at the AI Action Summit that it will invest EUR 200 billion for AI, with EUR 20 billion earmarked for gigafactories. According to Euronews:
“The Commission announced seven initial AI factories in December and will soon announce the next five. The gigafactories will have around 100,000 last-generation AI chips, around four times more than the AI factories currently being set up.”
Euronews, 11 February 2025
It seems that predictions that DeepSeek-R1 could reduce the appetite for gigafactories or AI mega-projects may have been premature. Instead, national participants in the Summit seem more committed than ever to developing domestic AI infrastructure.
5. Potential social impacts
Whatever the truth behind its efficiency claims, DeepSeek-R1 has made LLMs accessible via smaller, open-source models that can be run on some desktop computers. This has placed its underlying LLM technology in the hands of millions of academics and developers for the first time.
This potential “democratisation” of AI could have incredible implications for its development, as millions more people are able to advance the research.
The lower cost of DeepSeek-R1 potentially means that more companies could invest in using the technology in the workplace, speeding-up AI adoption more widely. The predicted shifts in working practices due to AI may therefore be accelerated, as could the potential flow of deepfakes and misinformation.
One thing that is unlikely to change is the need for a huge workforce of relatively low-paid data workers to annotate, label and transform the data needed for AI, and possibly to train the models. (You can watch my video for details on this serious workers’ rights issue, or read the work of Prof Antonio Casilli and the ILO’s AI Labor Disclosure Initiative for more information on the human labour cost of AI.)
The creation of more-efficient models could make the technology available in countries where advanced computing resources are scarce. The “digital divide” – in which around 40% of the world’s population has no access to computers – is of major concern to the UN, whose Global Digital Compact seeks to prevent parts of the world being "left behind." Whilst smaller, cheaper technology could potentially help bridge this gap, the plans of major economies to increase national AI spending may suggest these hopes are unlikely to be realised in the near-future.
Conclusion
DeepSeek-R1 has shaken the world of AI and challenged assumptions around the inevitability of US dominance and environmental impact of the technology. It could also democratise AI, leading to major advancements in model development, and increase its adoption more broadly.
Initial indications, however, suggest that countries are focusing more than ever on national strategies to develop major AI systems as a result of DeepSeek’s challenge to US dominance.
Seeing this opportunity, countries such as the UK, France and Canada have continued their push towards high-cost AI infrastructure projects. The EU’s own Competitiveness Compass places AI gigafactories at the centre of its economic growth strategy, with the announcement of over EUR 200 billion for AI development yesterday. Whilst President Trump acknowledged that the release of R1 was a “wake-up call” for the US tech industry, he has not wavered from his policy on establishing US dominance, including via the Stargate project.
Initially, the lesson of DeepSeek-R1 appeared to be that cheap, efficient, open-source AI models are the future. However, two weeks later, attitudes have already changed.
On its release, tech venture capitalist Marc Andreessen hailed DeepSeek as “one of the most amazing and impressive breakthroughs I’ve ever seen and as open source, a profound gift to the world”. Last week, Shyam Sankar, Chief Technology Officer of Palantir, accepted that DeepSeek has made basic AI cheaper. However, he added: “I think the real lesson, a more profound one, is that we are at war with China. We are in an A.I. arms race.”
DeepSeek has changed the world of AI in the past few weeks, making the technology cheaper, more efficient, and accessible to millions more developers. However, it does not appear to have changed governments’ fundamental attitudes towards AI: it is still a weapon in a global arms race.