Featured image of post DeepSeek Pursues 500 Billion RMB Valuation with $1B Annual Revenue, Over 70% Compute Allocated to Model Training
AI News

DeepSeek Pursues 500 Billion RMB Valuation with $1B Annual Revenue, Over 70% Compute Allocated to Model Training

DeepSeek's annualized revenue reaches $1 billion; a new 50 billion RMB funding round targets a 500 billion RMB valuation.

Key Facts and Timeline

Key Facts and Timeline
Key Facts and Timeline|News screenshot

DeepSeek is finalizing a new 50 billion RMB ($7B) funding round targeting a 500 billion RMB ($70B) valuation, with completion expected by late October 2024. Shanghai STAR Market IPO preparation is running in parallel.

  • Annualized revenue run rate: $1 billion (≈6.7 billion RMB)
  • More than doubled from under $500M several months ago
  • New lightweight model V4.1-Flash launched September 10, priced ~60% lower than predecessor
  • V4-Pro peaked pricing increased starting August 17, 2024
  • Core revenue source: API model calls only; free consumer chat app, no ads

Revenue Surge With a Profitability Paradox

DeepSeek’s commercial momentum is evident in rapid revenue growth and capital inflow. Its API business achieves 82.9% gross margin—higher than Anthropic and OpenAI—demonstrating strong profitability potential.

A notable contradiction: despite robust demand and price hikes that raised customer costs 2.3-4.5x, revenue growth isn’t the priority. Founder Liang Wenfeng told investors that user numbers remain stable post-increase, yet the company continues reporting losses.

In 2024’s first seven months, DeepSeek earned 475 million RMB while burning 715 million RMB in net losses. The prior-year full year showed 935 million RMB net loss against 475 million RMB revenue in just seven months.

Resource Allocation Split: Training vs. Inference

The compute allocation reveals strategic priorities:

  • Over 70% of compute resources dedicated to model training
  • Less than 30% allocated to existing model inference
  • Training priority maintained despite API demand growth

To relieve inference pressure, DeepSeek is testing smaller models on consumer-grade GPUs. Internal tests show such chips can handle most daily user tasks, potentially freeing high-end GPUs for training workloads.

Training Compute Constraints and Domestic Chip Adoption

Training demands remain the harder bottleneck. Liang identified increasing domestic chip usage for training as a critical initiative.

Huawei is expected to begin delivering training chips to DeepSeek as early as Q4 2024, marking a milestone in China’s semiconductor supply chain for AI.

Implementation Recommendations

  • Developers/Enterprises: Test V4.1-Flash for cost-sensitive or general tasks; reserve Pro for complex workloads
  • Individual Users: Free consumer app remains fully functional for everyday use
  • Investors: Watch for sustainability of R&D-driven growth model amid high valuation

Bottom Line

DeepSeek pursues a unique strategy: subsidizing aggressive R&D with high-margin API revenue. Whether this path can transition from speculative valuation to sustainable dominance hinges on solving both training scalability and inference efficiency simultaneously.