Join our everyday and weekly newsletter for the latest updates and exclusive content for top -class AI in the field. More information
AI was Houpana last week, when Deepseek, Chinese startup AI, announced its latest language model Deepseek-R1, which seems to correspond to the capabilities of the American AI system in a fraction of the cost. The announcement launched a large market that deleted almost $ 200 billion from the market value of NVIDIA and caused heated debates on the future of artificial intelligence development.
The narrative that has rapidly appeared that Deepseek fundamentally disrupted the economy of building advanced AI systems and reportedly reached only $ 6 million that US societies spent billions. This interpretation feels shock waves through Silicon Valley, where companies such as OpenAi, Anthropic and Google justify massive investment in computational infrastructure to maintain a technological advantage.
But in the middle of the market turbance and brass subtitles, Dario Amodei, co -founder of anthropic and one of the pioneering scientists for today’s large language models (LLM), he published a detailed analysis that more prospect of Deepseek’s successes. His blog post goes through hysteria to provide some fundamental knowledge about what Deepseek has newly achieved and what it means for the future of AI development.
Here are four key knowledge from Amodei’s analysis that transforms our understanding of Deepseek’s announcements.
1. The narration of the “$ 6 million” is missing a key context
According to amodei, Deepseek’s development costs need to be perceived. Directly challenges popular interpretation:
“Deepseek doesn’t make $ 6 million for billions of AI.” I can only speak for anthropic, but the Claude 3.5 Sonet is a medium model that costs several $ 10 million for training (I won’t give an exact number). Also, 3.5 Sonet was not trained in any way that included a larger or exhausting model (unlike some legends).
This shocking revelation will fundamentally shift the story around the cost -effective Deepseek. When considering that Sonnet was trained 9-12 months ago and still overcomes Deepseek’s model about many tasks, this success occurs more in accordance with the natural development of AI development rather than a revolutionary breakthrough.
The timing and context also depends on the importance. After historical trends in reducing costs in the development of AI – which amodei estimates about 4 times a year – Deepseek’s cost structure is widely on trend than dramatically in front of the curve.
2. Deepseek-v3, not R1, was a real technical success
While markets and media focused intensively on the R1 Deepseek R1, Amodei points out that the significant innovation of society has come before.
“Deepseek-V3 was actually a real innovation and what people should make them notice a month ago (we certainly did it). It seems that the preliminary model is close to the performance of the most important tasks of the most modern models, with a much less training.
The difference between V3 and R1 is essential for understanding the true technological progress of Deepseek. Representatives of V3 Real Engineering Innovation, especially when controlling the cache ‘key value’ of the model and shifting the boundaries of the method of experts (MOE).
This insight helps to explain why the dramatic reaction of the R1 market was incorrectly placed. R1 Estaily added the V3 Foundation based on V3 – a step that is currently taking more companies with its models.
3. Total Company for Business Investments reveals another picture
Perhaps the most revealing aspect of amodei analysis concerns Deepseek’s overall investment in development.
“It was reported that it was definitely that it was true-that Deepseek’s news had 50,000 chips to generate hopper, which I would guess that it was within the factor of ~ 2-3 times what the main American companies have.” These 50,000 Hopper chips cost ~ $ 1B. Thus, the total expenses of Deepseek as a company (unlike expenditure on the training of a single model) is not extremely different from American AI laboratories.
This revelation dramatically tells the narration around Deepseek’s resources effective. While the company may have achieved impressive results with individual model training, its overall investment in the development of AI seems to be roughly comparable to American counterparts.
The difference between the cost of the model training and the overall business investment emphasizes the importance of nails of essential resources in development. This suggests that although engineering efficiency can be improved, the remaining remaining in AI still requires the required capital investment.
4. The current “point of crossover” is temporary
Amodei describes the current moment in the development of AI as unique, but fleeting.
“We are in an interesting” point of crossover “, where there is a temporary case that several companies can produce good models of thinking,” he wrote. “It quickly ceases to be true, because everyone is moving up on the scaling curve on these models.”
This observation provides a major context for understanding the current state of competition AI. The ability of more companies to achieve similar results in the ability to think is a time phenomenon rather than the new status quo.
The consequences are meaningful for the future of AI development. Since companies continue to expand their models, especially in the field of learning about strengthening resources, Differenaten is probably again based on who can invest most in training and infrastructure. This suggests that although Deepseek was an impressive milestone, it was fundamentally changed by the long -term economy of advanced AI development.
Actual cost of building AI: What Amodei’s analysis reveals
Detailed analysis of the success of Deepseeks Amodei will decrease during weeks of market speculation to show the real economy of building advanced AI systems. His blog post systematically dismantles the panic and enthusiasm followed by Deepseek’s announcement, and shows how the cost of training a company of $ 6 million for a constant march of artificial intelligence.
Markets and media attract to simple narrations, and the story of Chinese society dramatically undermines the cost of AI AI development has proven to be irresistible. Yet the collapse of Amodei reveals a more complicated reality: the overall investment of Deepseek, especially its reported $ 1 billion on computing hardware, reflects the expenses of its American counterparts.
This moment of cost parity between the US and the Chinese development of AI refers to what Amodei calls the “crossover point” – a window of temporary temporary, where more companies can achieve similar results. His analysis suggests that this window will close as the AI capacity proceeds and the demand for training will intensify. The field is likely to return to the preference of organizations with the deepest resources.
Building advanced AI remains expensive efforts and careful research Amodei shows why measurement of its actual costs requires investigating the entire scope of investment. Its methodological deconstruction of Deepseek’s successes can ultimately prove more importance that the initial announcement that such turbulence has caused in the markets.
