DeepSeek was not built for $5 million: Bernstein challenges AI firms' claim

A recent report by Bernstein has examined the buzz around DeepSeek, an AI firm known for developing advanced language models. While the AI firm claims that it built an OpenAI competitor for just $5 million, Bernstein said that this figure is misleading and does not reflect the true costs involved in research, experimentation, and infrastructure.

DeepSeek’s key AI Models

DeepSeek has introduced two primary AI models: DeepSeek-V3 and DeepSeek R1. The V3 model is a large language model (LLM) that employs a Mixture-of-Experts (MOE) architecture, a method that distributes tasks among smaller models to optimise efficiency.

 

The model comprises 671 billion parameters, with 37 billion active at any given time. It also integrates Multi-Head Latent Attention (MHLA), which reduces memory usage, and mixed-precision training using FP8 computation, enhancing efficiency. To train the V3 model, DeepSeek used 2,048 NVIDIA H800 GPUs over a two-month period, consuming approximately 2.7 million GPU hours for pre-training and 2.8 million GPU hours in total.

 

While estimates suggest a training cost of $5 million based on a USD 2 per GPU hour rental rate, Bernstein’s report argues that this calculation excludes crucial factors such as research, experimentation, and infrastructure expenses.

 

Who benefits from DeepSeek’s innovations?

Bernstein predicts that AI infrastructure providers will gain the most from DeepSeek’s advancements. Companies in data infrastructure, identity verification, observability, and communications are expected to benefit. Key firms identified include:

 

·  Confluent (CFLT) – Data streaming solutions

 

·  Okta (OKTA) – Identity verification

 

·  Datadog (DDOG) – Cloud observability

 

·  Twilio (TWLO) – Communications tooling

 

·  Cloudflare (NET) – Browser delivery services

 

Twilio, in particular, has seen a 93 per cent stock return in the past year, with analysts raising earnings forecasts.

 

The future of AI adoption

While DeepSeek has demonstrated cost-efficient AI training, large-scale enterprise adoption remains slow due to high costs and validation challenges. However, as AI models become more scalable and affordable, software infrastructure firms could see accelerated adoption and growth.

 

While DeepSeek’s models are impressive, claims about their low development costs are exaggerated. The biggest winners in the AI race will likely be the companies powering AI infrastructure, rather than just the AI model creators themselves.