A recent analysis by Bernstein has sparked a debate over the true cost of developing AI models, specifically targeting the claim made by DeepSeek, an AI firm known for creating cutting-edge language models. DeepSeek, which has positioned itself as a serious contender to OpenAI, recently stated that it built its advanced models for just $5 million. Bernstein, however, argues that this claim significantly understates the real costs involved in research, experimentation, and infrastructure.
Table of Contents
DeepSeek’s Flagship AI Models: Breaking Down the Technology
DeepSeek has made headlines for its two major language models: DeepSeek-V3 and DeepSeek R1. Both models push the boundaries of artificial intelligence, but the focus here is on DeepSeek-V3, which incorporates a Mixture-of-Experts (MOE) architecture. This design splits tasks across several smaller models to maximize efficiency, allowing for smarter, faster processing.
DeepSeek-V3 is no slouch when it comes to scale. With 671 billion parameters, it stands as one of the largest language models in the industry. However, only 37 billion of those parameters are active at any given moment, a feature that optimizes performance without taxing resources excessively. Another important technological aspect is its use of Multi-Head Latent Attention (MHLA), which reduces memory consumption and helps boost overall efficiency.
DeepSeek also employs cutting-edge techniques like mixed-precision training, specifically using FP8 computation, which aids in improving both the speed and efficiency of the model’s operations. The training of DeepSeek-V3 was a massive undertaking that involved 2,048 NVIDIA H800 GPUs, operating over a two-month period. In total, DeepSeek spent about 2.7 million GPU hours in pre-training and another 2.8 million GPU hours throughout the process.
Is $5 Million Enough to Build a High-End AI Model?
While DeepSeek’s figures might seem impressive at first glance, Bernstein’s report highlights some critical flaws in their cost calculations. DeepSeek’s claim that it trained its model for just $5 million assumes a GPU rental cost of around $2 per hour. However, this estimate omits several major expenses that are essential to building and deploying a cutting-edge AI model, such as research and development costs, infrastructure investments, and the constant experimentation required to refine these complex models.
According to Bernstein, even if the training cost is taken at face value, there are far more significant expenditures that need to be considered to build a truly advanced system. These include the salaries of researchers and engineers, the development of proprietary technologies, and the substantial infrastructure that supports these models.
DeepSeek’s Impact on AI Industry Players
Despite the questionable claims about costs, DeepSeek has undoubtedly made waves in the AI landscape. Bernstein suggests that the real beneficiaries of DeepSeek’s advancements might not be the company itself, but rather the businesses that provide the infrastructure that powers such AI systems.
Key sectors poised to gain from DeepSeek’s technology include data infrastructure, identity verification, cloud observability, and communication services. Some of the major companies that could see growth due to DeepSeek’s influence include:
- Confluent (CFLT) – Providing data streaming solutions, crucial for processing large-scale AI operations.
- Okta (OKTA) – Identity verification technologies, essential for securing AI applications.
- Datadog (DDOG) – Monitoring and observability tools for AI systems.
- Twilio (TWLO) – Communications infrastructure that enables AI models to interact seamlessly with users.
- Cloudflare (NET) – Services to optimize web delivery and security for AI-powered apps.
Twilio, in particular, has been thriving, with stock prices soaring by 93% over the past year. Investors are optimistic, and Bernstein’s analysis only reinforces this trend, predicting that software infrastructure providers are likely to be the biggest winners in the evolving AI race.
The Challenges of Scaling AI Models for Enterprise Use
While DeepSeek’s models are impressive, the broader picture for AI adoption remains more complex. Although DeepSeek’s approach demonstrates cost-effective training, the actual deployment of AI models at scale is still a slow process, particularly within large enterprises. The challenges are manifold, including high infrastructure costs and the ongoing need to validate AI models across multiple domains.
However, there is hope for more rapid adoption in the future. As AI models become more scalable and affordable, software infrastructure companies are likely to see accelerated growth. The competition between these firms will ultimately drive down costs and push the boundaries of AI applications across industries.
The Bottom Line: Who Really Wins?
In conclusion, while DeepSeek’s innovative AI models are certainly impressive, the claim that they were developed for only $5 million is questionable. The real financial winners in this AI race may not be the developers of the models, but rather the companies behind the infrastructure that makes these models possible.
For businesses and investors looking to capitalize on the AI boom, the focus should be on the firms that are powering AI applications rather than just the AI model creators themselves. As AI technology continues to evolve, infrastructure providers will likely remain at the center of the industry’s growth, even if the AI models themselves are becoming more accessible.
Final Thoughts
DeepSeek has certainly made strides in the world of AI, but as with any rapidly developing technology, the true costs and benefits are often more complex than they first appear. The $5 million figure may serve as a great marketing tool, but when looking at the long-term impact of AI advancements, it’s clear that the companies building the infrastructure to support these technologies may be the ultimate winners in the AI revolution.
1 thought on “DeepSeek’s $5 Million Claim: Is It Misleading? Bernstein Challenges AI Firm’s Cost Estimates best”