DeepSeek Apocalypse: AI giant burns billions of dollars, much of it wasted

In January this year, artificial intelligence startup DeepSeek released two breakthroughs through its new model R1, quietly redefining the economics of artificial intelligence. This model achieves top performance at 1/40 of the cost of the previous model. As of December 2024, DeepSeek's V3 large language model has reduced training costs by more than 90%.

Two of DeepSeek's breakthroughs attracted widespread attention: First, DeepSeek revealed that asking AI models to elaborate on their reasoning processes - a research approach known as chain-of-thought prompting - improved accuracy and efficiency. Second, DeepSeek uses artificial intelligence to generate its own data sets, completely independent of manual labeling of the data. While there are arguments that DeepSeek isn't as cheap as it claims, these breakthroughs have certainly ushered in a new era of AI economics.

The cost structure of artificial intelligence is changing dramatically. Every dollar of performance surge has had a profound impact on start-ups, enterprise applications, and infrastructure investments. This shift could upend market forces, ultimately helping nimble startups catch up with tech giants in the short term while boosting profit margins.

Tech giants have already invested more than $100 billion in AI infrastructure development, and it continues to increase. Now they must consider how to generate a return on these huge investments and maintain an edge over algorithms against more nimble, smaller market competitors. In the face of a rapidly changing environment, both tech giants and start-ups are facing a clear signal: seize the opportunity of technological progress quickly, or be eliminated.

The AI market landscape before and after DeepSeek

Before DeepSeek's rise, start-ups struggled to compete with infrastructure spending by tech giants, which poured billions of dollars into building huge data centers every quarter and gained huge advantages from advances in artificial intelligence technology. These giants not only have massive data resources, but also gather a large number of doctoral talents, and the progress of algorithms also depends on their strong technical strength. In addition, long-established distribution networks allow them to move products quickly to existing customers and accelerate technological progress through feedback loops.

Today, however, startups are big enough to compete with the tech giants. By 2025 alone, the cost of training models will fall by 95%, significantly reducing the infrastructure advantage of the tech giants. Reasoning costs have plummeted nearly a thousand-fold in the past three years and are expected to fall further in the future. The duration of the algorithmic advantage has been reduced to 45 to 100 days and may continue to decrease.

When training costs are no longer a key bottleneck, inference performance (that is, how well AI models perform in real-time applications) becomes a new focus. We are entering a new phase: smaller, cheaper models that offer comparable power to larger models and can run on lower performance Gpus, extending the life of older Gpus. If smarter AI products can be delivered at a very low cost, then startups finally have a chance to outperform the tech giants while increasing profits.

Efficient manpower allocation further strengthens the challenger's advantage. With no longer needing to hire large numbers of PhD-level talent to assemble a competitive AI team, startups can develop, optimize, and distribute models at a much lower cost than tech giants. And, because they are largely focused at the application level, challengers are able to enjoy higher profit margins in the same way that cloud startups gained an advantage by improving unit economics 15 years ago.

This trend is not just good for startups. It also puts companies like Nvidia at greater risk. After DeepSeek's announcement, Nvidia's stock price fell 12%, although it has since rebounded. The risks for chipmakers are heightened because demand is shifting from training-focused hardware to more efficient inference solutions. The rise of consumer-grade neural processing units (Npus) could accelerate this shift, allowing AI models to run natively on devices such as smartphones and laptops.

Artificial intelligence spending

What's good for challengers is bad for tech giants. Ai giants have almost instinctively tied DeepSeek's dominance to national security implications in an attempt to drum up support for its development of similar technology, ignoring the fact that US researchers, including at Stanford University, have been able to replicate and even surpass DeepSeek's technology. Looking ahead, companies investing huge sums in data infrastructure projects may ask: Was the huge spending on AI model research and development wasted? If cheap technology works just as well as expensive technology, why spend so much money?

Historical trends suggest that most AI advances have indeed relied on excessive capital investment in scale. The Transformer architecture was successful because of overtraining, exceeding what was considered algorithmically optimal at the time. New technological advances are proving that we can achieve the same performance at a lower cost. Although efficient solutions like DeepSeek have significantly improved efficiency, even so, the expansion of hyperscale cloud providers still requires larger data centers and must bear ballooning inference costs.

However, the tech giants are not sitting still. We're already seeing an arms race for DeepSeek's achievements, with the likes of Google's Gemini model, Microsoft's Azure AI Foundry and Meta's open source LLaMA all vying for dominance. Open source models can play a key role. Mark Zuckerberg, CEO of Meta, stressed the importance of personalized AI - that is, models tailored to the needs, culture and preferences of individual users. This vision aligns with a broader trend in AI development: smaller, more specialized models capable of delivering high performance without the need for a massive cloud infrastructure.

Startups win new chips

At the same time, open source and closed source giants have different goals, further enhancing the challenger's advantage. Open source models created by companies like Meta will continue to compete and reduce costs across the ecosystem, while closed source models try to charge higher fees through better technology. Startups can leverage the competition between the two camps to achieve the best price/performance ratio for each use, while increasing profit margins.

Regardless of the size of the business, the message is clear: Take advantage of the specific advantages available to them - market dynamics, computing power and talent - quickly or face failure. The cycle of technological progress is getting shorter, from the months or even years it used to take to set new performance standards, to DeepSeek's technological breakthrough suggesting it could now take as little as 41 days. Innovation is advancing at an unprecedented rate, and the fault-tolerance space is rapidly shrinking.