> ## Content Index
> Fetch the complete content index at: https://nordisk-ai.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# DeepSeek's Costs Are Misleading
- URL: https://nordisk-ai.ghost.io/deepseek-s-costs-are-misleading/
- Published: 2025-02-13T08:55:11.000Z
- Updated: 2026-08-27T15:06:17.000Z
- Description: DeepSeek has sparked an AI frenzy, but is the low cost really realistic - or just a strategic bluff? The model may have been subsidized to quickly dominate the market. What does this mean for the AI race between China and the West?
- Author: Anders Hagen
- Tags: AI Models, AI, AI Costs, aisweden, article, DeepSeek, High-Flyer, China, Mixture of Experts, nordiskai, Nvidia, #Migrated-1787841634512, #wp, #wp-post, #Import 2026-08-27 14:41, #Import 2026-09-07 13:41

DeepSeek has quickly become one of the most talked-about AI models in the world. Its daily traffic has now surpassed established players like Claude, Perplexity, and even Google's Gemini. But despite the hype, few experts see the model as a revolution. It's more a matter of very skillfully executed technical optimization - and possibly strategic pricing.

#### DeepSeek Is Backed by High-Flyer, Which Invests Heavily

Behind DeepSeek is the hedge fund High-Flyer, an early AI adopter in the finance sector. Already in 2021, the company invested in 10,000 A100 GPUs, giving them a huge head start before US export restrictions took effect. DeepSeek was later founded as an independent company in May 2023 and has since grown rapidly, backed by High-Flyer's capital and resources.

DeepSeek and High-Flyer share both computing capacity and staff, and currently have access to around 50,000 Nvidia Hopper GPUs, including 10,000 H100s and an equal number of H800s, plus further orders for H20s. Their estimated server investment is $1.6 billion, with annual operating costs close to a billion dollars.

#### The Real Cost Behind DeepSeek

One of the biggest talking points around DeepSeek has been its claimed low training cost of $6 million for the V3 model. But according to experts, that figure is misleading. It only includes the GPU cost for the final training run and excludes all other costs, such as research and development, architectural innovations, and data collection.

Comparisons with other AI models show that the actual cost must be significantly higher. Anthropic, for example, spent tens of millions of dollars training Claude 3.5 Sonnet - and that's just part of their total development costs. DeepSeek's real budget is therefore estimated to be well over $500 million, showing the model is far from a low-cost product.

#### DeepSeek V3 - A Technical Breakthrough or Just Optimization?

DeepSeek V3 has been praised for its performance, but comparisons with OpenAI's GPT-4o show that the AI landscape is changing fast. AI labs have gotten significantly better at creating models with lower computing costs, and DeepSeek has been first to use certain optimization techniques, such as:

**Multi-Token Prediction (MTP):** Improves the model's learning by predicting multiple words at once.

**Mixture of Experts (MoE):** A model architecture where several specialized sub-models work together.

**Multi-head Latent Attention (MLA):** Reduces memory usage in the inference process by over 90%, drastically lowering costs.

These techniques have allowed DeepSeek to offer a competitive AI at low cost - but that doesn't mean their actual expenses are low. On the contrary, analyses suggest DeepSeek is subsidizing its inference to quickly gain market share.

#### The AI Race Between China and the West

DeepSeek's rapid progress has led to speculation about China's AI strategy. That China can develop a top-tier model for a fraction of the cost compared to OpenAI and Google is remarkable - but much suggests it's an illusion created by smart pricing and early hardware investments.

The model has, however, made it clear that Chinese AI labs are capable of competing at the highest level. It remains to be seen how Western players respond to this challenge, and whether DeepSeek can really keep its prices down in the long run.

**Sources:** [SemiAnalysis](https://semianalysis.com/2025/01/31/deepseek-debates/?ref=nordisk-ai.ghost.io), [Artificialanalysis.ai](https://www.artificialanalysis.ai/?ref=nordisk-ai.ghost.io), [a16z](https://www.a16z.com/?ref=nordisk-ai.ghost.io)