DeepSeek targets AGI in $10bn funding round

🔑 Enhanced Key Takeaways

•DeepSeek is currently in late-stage discussions to raise 70 billion yuan (approximately $10 billion) in its first major external funding round, which could value the company at around $45 billion.
•The company was founded in 2023 by Liang Wenfeng, who previously established the quantitative hedge fund High-Flyer, which initially provided DeepSeek's funding.
•DeepSeek's models, such as DeepSeek-V3 and DeepSeek-R1, are noted for achieving performance comparable to leading proprietary models like GPT-4o and Gemini-3.0-Pro, while reportedly incurring significantly lower training costs, with DeepSeek-V3 costing around $5.6 million.
•DeepSeek's V4-Pro, a 1.6-trillion-parameter Mixture-of-Experts system, and V4-Flash models, released in April 2026, are optimized for compatibility with Huawei Ascend and Cambricon silicon, alongside Nvidia GPUs, indicating a strategic focus on the Chinese domestic market.
•While DeepSeek promotes an open-source model strategy, some of its models are released under custom licenses that include usage limitations, distinguishing them from the conventional definition of open-source software.

📊 Competitor Analysis▸ Show

Feature/Model	DeepSeek	OpenAI (ChatGPT)	Google (Gemini)	Anthropic (Claude)	Mistral AI
Model Type	Open-source (with custom licenses) & Proprietary API; MoE architecture	Proprietary; GPT-series (e.g., GPT-4, GPT-4o)	Proprietary; Multimodal (e.g., Gemini Ultra, Pro, Nano)	Proprietary; Focus on ethical AI (e.g., Claude 3)	Open-source; Diverse model offerings (e.g., Mistral Large 2)
Key Strengths	Cost-effective, high-performance reasoning, coding, math, efficient training, support for Chinese NLP, agentic capabilities	Versatile, general-purpose conversation, creative content, robust code assistance, multimodal tasks	Advanced multimodal capabilities (text, image, audio, video), integrated with Google ecosystem, strong for productivity and complex data analysis	Excels in reasoning, long-form content, ethical safeguards, long-context handling, PhD-level science reasoning, agentic tasks	Customizable, efficient, high-performance open-source models, privacy-focused
Benchmarks/Performance	DeepSeek-V3/R1 comparable to GPT-4o and Gemini-3.0-Pro; DeepSeek-V3-0324 outperforms GPT-4.5 in math/coding	GPT-4 and GPT-4o are leading general-purpose AI tools	Gemini Ultra surpasses human experts on MMLU benchmark; Gemini 1.5 Flash offers extended context	Claude 3 leads in coding, PhD-level science reasoning, and agentic task completion	Mistral Large 2 is a top-tier reasoning model
Pricing/Accessibility	Free open-source models for self-hosting; competitive API pricing ($0.27-$0.55/M input tokens, $1.10-$2.19/M output tokens)	Paid tiers (e.g., ChatGPT Plus ~$20/month); API access	Paid tiers (e.g., Gemini Advanced ~$20/month); API access	Paid tiers (e.g., Claude Pro ~$20/month); API access	Mostly free open-source models; hosting costs may apply
Context Window	DeepSeek-R1/V3: 128K tokens	Varies by model (e.g., GPT-4 Turbo has 128K)	Gemini 1.5 Flash offers extended context windows	Hundreds of pages of text	Mistral Large 2: 128K tokens

🛠️ Technical Deep Dive

Architecture Foundation: DeepSeek models are large-scale language models based on deep neural networks, primarily utilizing a decoder-only Transformer architecture.
Mixture-of-Experts (MoE): DeepSeek-V3 and DeepSeek-R1 employ a Mixture-of-Experts framework, allowing dynamic activation of relevant sub-networks ("experts") for a given task, enhancing efficiency. DeepSeek-V3 has 671 billion total parameters, with only 37 billion activated per token during inference.
Multi-head Latent Attention (MLA): This mechanism optimizes the traditional key-value cache (KV cache) by compressing attention matrices into smaller latent representations, significantly reducing memory usage and improving inference latency.
Multi-Token Prediction (MTP): DeepSeek-V3 incorporates an MTP objective, enabling the model to predict multiple tokens at once, which densifies training signals and improves performance on complex benchmarks.
Context Length: DeepSeek-R1 inherits a 128K context length from DeepSeek-V3-Base, achieved through a two-stage extension process utilizing the YaRN (Yet another RoPE extensioN method) technique.
Training Efficiency: DeepSeek-V3 utilizes FP8 mixed precision training and an auxiliary-loss-free strategy for load balancing, contributing to its economical training costs (reportedly around $5.6 million for full training).
DeepSeek-V3.2 Innovations: This version introduced DeepSeek Sparse Attention (DSA) for scalable long-context reasoning, large-scale reinforcement learning for specialist models, and a "specialize-first" approach using distilled data.
DeepSeek-V4: The V4-Pro model is a 1.6-trillion-parameter Mixture-of-Experts system, with the V4 family optimized for Huawei Ascend and Cambricon silicon in addition to Nvidia GPUs.

🔮 Future ImplicationsAI analysis grounded in cited sources

DeepSeek's state-backed funding and "open" model strategy could establish a new template for AI development outside the U.S. tech ecosystem.

The involvement of China's National Artificial Intelligence Industry Investment Fund, combined with DeepSeek's controlled "open-source" approach, suggests a model for national AI capability building that balances community engagement with strategic control.

The company's focus on cost-efficient, high-performing models will intensify competition in the global AI market, particularly for open-source alternatives.

DeepSeek's demonstrated ability to achieve top-tier performance at a fraction of the training cost of Western counterparts pressures other AI labs to find more efficient development methods.

DeepSeek's optimization for Huawei Ascend and Cambricon silicon could accelerate the development and adoption of non-Nvidia AI hardware in China.

By explicitly supporting domestic chip alternatives, DeepSeek is contributing to China's efforts to reduce reliance on U.S. export-controlled advanced chips for AI training and deployment.

⏳ Timeline

2015-06

Liang Wenfeng co-founds High-Flyer Quantitative Investment Management.

2023-05

Liang Wenfeng announces High-Flyer's pursuit of AGI and launches DeepSeek.

2023-07

DeepSeek is officially founded as an AI company.

2024-12

DeepSeek unveils its V3 model, a 671 billion parameter MoE architecture.

2025-01

DeepSeek releases the DeepSeek-R1 reasoning model.

2026-04

DeepSeek releases its V4-Pro (1.6T parameters) and V4-Flash (284B parameters) models.

DeepSeek targets AGI in $10bn funding round

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (26)

👉Related Updates

PACA: Open-source tool for ancient fossil coordinate mapping

The Hidden Costs of Solo Sales Handoffs

Hotel groups launch ChatGPT apps for travel booking

Nvidia may guarantee $250bn for OpenAI data center