DeepSeek Secures Massive Capital to Scale Infrastructure and Model Efficiency
DeepSeek has expanded its second funding round to approximately 100 billion yuan, or 2.4 trillion yen, drawing major industrial investors including CATL, Geely Group, SMIC Juyuan, and Legend Capital. The company maintains a pre-funding valuation of about 500 billion yuan, or 12 trillion yen, and plans to apply for a listing on the Shanghai Stock Exchange’s STAR Market by the end of 2026.
Technical Innovations Driving Cost-Efficient Inference
The V4.1-Flash-1017 model uses a Mixture-of-Experts architecture with 552B backbone parameters and 196B Engram parameters, activating 8B parameters per token in prefill and 16B in decode. Technical efficiency gains include FP4 main KV cache and Compressed Sparse Attention 2 to reduce compute requirements. The model achieves 77.0 on the DeepSWE v1.1 benchmark, outperforming Claude Opus 5, which scored 74.0 on the same benchmark.
Strategic Expansion Through the STAR Market Listing
DeepSeek plans to apply for a listing on the Shanghai Stock Exchange’s STAR Market by the end of 2026, as part of its strategy to diversify fundraising routes and demonstrate compliance with government regulations. The influx of industrial capital signals a broader trend of Chinese manufacturers and technology firms investing directly in AI infrastructure to integrate model development with manufacturing, logistics, and energy management.
Balancing Rapid Growth With Service Stability Challenges
DeepSeek suffered a nearly five-hour global service outage on October 15, 2026, affecting users of its chatbot and API services for the V4 Pro and V4.1 Flash models. The company is offering one-week of consumption credits as refunds to users affected by the outage, covering any consumption between October 8 and October 14.
Future Outlook on Market Price Competition
DeepSeek recently launched Search and Fetch APIs to provide a first-party web layer for agent builders, with each endpoint priced at $0.75 per 1,000 successful requests. The company says its in-house engine serves hundreds of millions of requests daily, and agent harness calls via DeepSeek Harness, Claude Code, or Codex remain free until November 1, 2026. Peak pricing for the deepseek-flash model is $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens, with off-peak rates at half. As industrial capital continues to flow into AI infrastructure, the trend underneath is a growing pressure on proprietary model providers to justify premium pricing through performance or specialized features. Watch this space for how cost-efficient models like DeepSeek reshape enterprise AI adoption globally.
Industrial Investors Deepen AI Integration Across Sectors
CATL plans an additional investment of approximately 10 billion yuan (about 240 billion yen) in DeepSeek’s expanded funding round, exceeding its initial stake as a leader in EV and battery manufacturing.
New investors joining the round include Shixi Capital, SMIC Juyuan, CPE, Legend Capital, and the Geely Group, reflecting broad industrial participation beyond traditional tech financiers.
The investment signals AI’s penetration into industrial foundations, with funds directed toward integrating model development with manufacturing, logistics, and energy management systems.
Market Valuation and Investor Strategy
DeepSeek’s decision to expand its second round of fundraising from an initially planned 50 billion yuan to approximately 100 billion yuan reflects a significant shift in capital allocation. Despite this massive influx of cash, the company’s pre-funding valuation remains steady at approximately 500 billion yuan, or 12 trillion yen. According to reports, this strategy of securing substantial liquidity without increasing the valuation suggests a sense of urgency among investors to protect current technical capabilities and maintain market position, even while factoring in the potential risk of a future decline in value.
The investor lineup indicates that the company is effectively bridging the gap between software development and industrial infrastructure. Beyond the additional 10 billion yuan commitment from CATL, the round includes participation from Shixi Capital, CPE, Legend Capital, and the Geely Group. This broad industrial participation underscores a move toward integrating AI models directly into manufacturing, logistics, and energy management systems. By diversifying its fundraising routes through both private industrial capital and a planned listing on the Shanghai Stock Exchange’s STAR Market, DeepSeek aims to demonstrate compliance with government regulations while ensuring the resources necessary to sustain its competitive pricing models against global commercial alternatives.
Broader Implications for Global AI Development
The rapid expansion of DeepSeek’s capital and infrastructure is highlighting a fundamental divergence in international AI development strategies. According to industry analysis, China is currently accelerating its approach toward “infrastructure control through industrial capital,” where major manufacturers and technology firms integrate AI directly into the foundation of manufacturing, logistics, and energy management. In contrast, other regions, such as Japan, are prioritizing “practical data collaboration with privacy protection,” with organizations like the MS&AD Insurance Group participating in government projects that utilize proprietary encryption technology to process internal data safely.
For engineers and enterprise strategists, this shift presents significant implications for system architecture and tool selection. As DeepSeek continues to exert pressure on global markets through cost-efficient, high-precision models, the resulting price competition with existing commercial alternatives is expected to narrow functional gaps. This trend forces a critical evaluation of implementation environments, particularly as the reduction in inference costs makes high-performance AI models more accessible for large-scale industrial applications. The ongoing integration of AI into physical industrial sectors suggests that the future of enterprise AI adoption will be increasingly defined by these differing national priorities regarding data security and infrastructure control.
Technical Methodology and Reasoning Capabilities
The V4.1-Flash-1017 model incorporates reinforcement learning with self-feedback to facilitate end-to-end project execution, ranging from 3D modeling to game creation. The architecture utilizes a multimodal Mixture-of-Experts design consisting of 40 Transformer layers, split into a 20-layer causal encoder and a 20-layer decoder. To optimize performance, the system employs a Causal Encoder-Decoder structure where the decoder global KV is projected from final encoder hidden states, which nearly halves prefill compute requirements.
The model’s reasoning effort is managed through a 1 to 100 scalar, with API tiers mapping to specific values of 50, 75, and 100. While the 1017 checkpoint leads on benchmarks such as Automation-Bench and Agents’ Last Exam, it trails Claude Opus 5 on HLE and Terminal-Bench 4.0. The system is specifically tuned for integration with DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, with the company recommending the use of DeepSeek Harness v0.2.1-rc.1 or later for optimal functionality.
Sources
- DeepSeek Funding Exceeds 10 Trillion Yen and MS&AD's Participation in Domestic AI｜MCUBE USA
- DeepSeek suffers another major outage as company offers user refunds
- DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API
Photo: Matt Howard (CC BY-SA 2.0)
