Success and failures aside, DeepSeek made the world realise how quickly and deeply a single AI model release can impact global events, and this raises three questions. First, how legitimate (and sustainable) are the massive AI investments in the West? Second, what risks and opportunities does open-source development pose? Finally, is it possible to balance growth and innovation with data privacy and security amidst a global AI race?

Western AI – the Emperor with no clothes?

The claim that DeepSeek’s model training cost a mere $6 million is questionable at best and blatantly false at worst: SemiAnalysis speculates about $1.3 billion in server capital expenditure, meaning that modest training costs were backed by way larger infrastructure expenditures. Even so, there is reason to believe that DeepSeek AI still costs a fraction of Western models’ development costs.

The next few weeks will surely reveal more details, as evidence emerges that the startup used datasets the big US players spent a fortune on, a situation that OpenAI is very unhappy about. Ironically, OpenAI has been accused numerous times of dubious data collection and copyright usage practices themselves. Perhaps the greatest paradox is that by putting a chokehold on technological exports to China, the US prompted Chinese engineers to innovate better and do more with less.

What matters more than the exact development costs, however, is the broader economic impact that cheaper LLMs will have on the AI industry. As venture capitalists and tech giants reassess their investment strategies, DeepSeek's scrappy approach suggests that the path to AI leadership might require fewer resources. This is already impacting how AI services are priced and delivered to end users and developers.

With API calls cheaper by an order of magnitude – DeepSeek reportedly charges just 1.4 cents per million tokens compared to Meta's $2.80 for the same output – the Chinese player is changing the nature of the game, lowering the barrier to entry for different market players not only in the West, but in the developing world as well. If DeepSeek becomes the platform of choice for budding AI developers, will it result in even larger losses for incumbents?

DeepSeek is neither the only nor the last “discounted” AI model to emerge from China. Other Chinese LLMs also have the advantage of being narrower in their use, which means that they need less computational resources to operate. This helps them keep their pricing competitive – Doubao 1.5-pro by ByteDance and Qwen plus by Alibaba both cost $0.30 per 1 million tokens.

Open-source dilemma

DeepSeek's success sheds light on another important aspect of the AI game – open vs closed-source systems. The Chinese startup partially based its breakthrough model on Meta's open-source Llama architecture, and in a bold move, released its own models in open-weights form. While Meta's Chief AI Scientist Yann LeCun celebrated this as proof that "open-source models are surpassing proprietary ones," the new state of affairs raises serious strategic concerns. With a tariff war seemingly imminent, will an AI war be waged in parallel, and where does it put all security concerns?

The situation at hand has forced Western AI leaders to reassess their positions on open-source technology. In the aftermath of DeepSeek’s debut, OpenAI's CEO Sam Altman admitted that his company might be "on the wrong side of history" regarding its open-source strategy, hinting at potential changes in the near future. This move would mean a return to open-source development, which was abandoned after GPT-3 due to safety concerns. Echoing this sentiment, Mark Zuckerberg emphasised that advancing “an American open source standard is crucial” to maintain the country’s global advantage.

However, market competition is only one side of the coin. Existing licensing schemes weren’t built for software capable of leveraging vast swathes of data from multiple sources, as Meta's VP for AI research Joëlle Pineau pointed out. This brings open-source AI into more or less direct confrontation with data protection efforts. Further, increased technological complexity also brings about liability issues – if an AI system based on open-source models produces harmful outputs, determining responsibility becomes nearly impossible when multiple contributors are involved.

Safety and security dilemma

For most Western observers, DeepSeek's rapid success raises critical questions about data privacy and security. These concerns mirror earlier debates about TikTok but with potentially greater implications. All data processed through DeepSeek’s models is stored on Chinese servers, raising serious concerns about the different ways this data can be put to use. Similarly, there exists the issue of censorship and bias.

Users and journalists have already noticed that DeepSeek's model refuses to respond to queries about sensitive topics within China, such as the Tiananmen Square massacre and Uyghur detention camps. It is worth noting that the thinking process of DeepSeek's model (which you can openly observe) is very advanced and probably not that biased; however, the chatbot sitting above the model is politically biased and does all the censoring. While self-censorship in AI models is nothing new, DeepSeek shows a yet unseen marriage of technology and political ideology.

Western AI firms, on the contrary, are tightly bound by various ethical and legal concerns, ranging from public outrages to emerging AI regulation. The EU has just kicked off the AI Act to protect fundamental human rights and ensure ethical AI development, but its implementation presents significant challenges, made acute by AI training data requirements.

By nature, LLMs need vast amounts of data to ensure adequate contextual comprehension and prevent bias or hallucinations. For data providers, the new regulation translates into new responsibilities, as they now must implement robust verification procedures to ensure their infrastructure is not supporting prohibited AI systems or collecting copyright-protected information.

Moreover, although strict in certain areas, the EU regulation is oddly loose in others. For example, it provides important exemptions for open-source AI systems, even though the industry hasn’t yet reached a consensus on what "open-source AI" is. This ambiguity creates opportunities for exploitation, with companies using loopholes to win legal exemptions.

Some experts, including Anthropic CEO Dario Amodei, warn of broader strategic implications this regulatory imbalance might bring. With China directing more technological focus, AI included, towards its military-industrial complex, DeepSeek's breakthrough could help China "take a commanding lead on the global stage, not just for AI but for everything". Especially considering China’s well-known cynical attitudes toward data privacy and broader societal implications.

The road forward

DeepSeek and the many more models that will inevitably follow it, signal an urgent need for global coordination in AI governance. The tech's dual-use potential, coupled with its rapid and uncontrolled proliferation, make it clear that no single nation's regulatory approach will be sufficient.

Today, talks of AI (and AGI in particular) being an "existential risk" are not as prominent as they were in 2023, when the big tech giants called for "a regulatory body overseeing AI to make sure that it does not present a danger to the public". However, while reduced alarmism should be embraced as a positive development, guardrails in the form of regulation are needed more than ever before.

As with nuclear energy, what's needed is an all-encompassing international framework that addresses AI development, deployment, and safety standards. This framework must balance innovation with security, establish clear guidelines for data protection across borders, and most importantly, create mechanisms for monitoring and enforcement.

The EU's AI Act can be seen as a strong enough starting point, but broader global consensus is needed for issues, such as data rights, the use of AI in military conflicts, and open-source AI development with possible proliferation of AI technology in rogue states. Otherwise, we might be facing a situation similar to that of a teenager being able to build a DIY nuclear reactor in their parents' garage.

For more startup news, check out the other articles on the website, and subscribe to the magazine for free. Listen to The Cereal Entrepreneur podcast for more interviews with entrepreneurs and big-hitters in the startup ecosystem.