DeepSeek Raises 50 Billion Yuan: Unveiling Its AI Coding Strategy

On June 16, several media outlets reported that DeepSeek had completed its first external fundraising since its inception. The company raised over 50 billion yuan, bringing its post-investment valuation to over $50 billion. Founder Liang Wenfeng personally contributed 20 billion yuan, making him the largest single investor. Tencent invested 10 billion yuan, CATL (Contemporary Amperex Technology Co.) around 5 billion yuan, and JD.com, NetEase, and IDG Capital each invested 3 billion yuan. This marks the largest single funding round in the history of China’s AI industry.

Although DeepSeek has not officially disclosed the specifics of how the funds will be used, its product testing trends and job postings reveal the contours of its strategy. The company appears poised to focus its investments on AI coding and Agent development, aiming to rival Anthropic’s Claude Code.

Background of the Fundraising and Investor

Composition

DeepSeek had previously relied solely on internal funding from its founder for research and development. The acceptance of institutional investors in this round reflects a significant shift in the company’s growth phase. Investors include major Chinese tech players like Tencent, CATL, JD.com, NetEase, and venture capital firm IDG Capital. Tencent’s 10 billion yuan investment underscores its strong strategic commitment to the AI sector.

Founder Liang Wenfeng’s additional personal contribution of 20 billion yuan signals his intention to maintain control over the company’s management and vision. With a valuation of $50 billion, comparable to OpenAI and Anthropic, the high level of investor confidence in DeepSeek’s technological capabilities and market potential is evident.

Focusing on AI Coding

DeepSeek’s strategic focus lies not in acquiring end-consumers but in providing services to developers and businesses. According to leaked information from a forum on June 15, the company has launched a canary release (greatest release) of DeepSeek V4.1 for its web-based platform. Testers reported a “day-and-night difference” in the coding capabilities of V4.1 Flash, with significant advancements. The knowledge cut-off date has been updated from May 2025 to January 2026, with some users detecting updates extending to May 2026.

Job listings published in May further corroborate this strategy. For instance, the posting for an “Agent Harness R&D Engineer” highlighted the formula “Model + Harness = Agent.” The job description explained that the role involves converting state-of-the-art model capabilities into advanced Agent products, with Harness handling all tasks outside the model. This suggests that DeepSeek is intensifying its development of Agent products to compete with Anthropic’s Claude Code.

Several factors underpin the prioritization of AI coding. First, coding rigorously tests and enhances a model’s inference capabilities. Second, it represents the only commercially validated closed-loop scenario currently available. Third, coding is becoming foundational as an infrastructure for AI productivity. The U.S. export restrictions on Anthropic’s top coding model further underscore the strategic importance of this field.

Massive Investment in Proprietary AI

Computing Centers

A large portion of the funds is likely to be allocated to expanding computational capabilities. DeepSeek is transitioning from rented data centers to its own ultra-large-scale AI computing centers. In Inner Mongolia’s Ulanqab City, the company is recruiting staff for its AI computing center, including senior operations engineers and “IDC Design and Planning Engineers” tasked with designing ultra-large-scale AI computing centers ranging from megawatt to gigawatt levels.

According to insiders, as of 2025, several vacant AI computing centers in Zhejiang Province are being made available to DeepSeek. While the company supplements its proprietary computing capabilities with external supply, computational power remains a long-term challenge. Recent limitations on internet search functionality in “Expert Mode” and caps on generation counts highlight the ongoing strain on computational resources.

In comparison, Anthropic has adopted an aggressive strategy of “multi-cloud binding + dedicated clusters.” Its diverse supply chain includes GPUs, TPUs, and proprietary chips, with its computing capacity projected to grow from approximately 1.4 gigawatts by the end of 2025 to 7–8 gigawatts by the end of 2026. However, Anthropic’s CEO Dario Amodei recently acknowledged that the exponential surge in computational demand has outpaced its infrastructure.

DeepSeek’s decision to build its own computing infrastructure is driven by its need for custom optimization across multiple layers—algorithms, chips, networks, and frameworks. Standardized services from cloud vendors cannot meet these requirements. As a result of its multi-layer optimization, DeepSeek’s unit computation costs are significantly lower than Anthropic’s. For instance, the cost of its code models is approximately 1/138th of the input price of Anthropic’s Fable 5.

Enhancements in Model Layer and Harness Team

On the model layer, DeepSeek is heavily investing in post-training for V4.1, an incremental version of V4. Post-training costs are estimated to be only one-tenth to one-fifth of pre-training costs. The primary goal is to narrow the gap between V4’s coding capabilities and the top-tier level.

On the execution layer, the company is building its Harness team, led by trading systems expert Cui Tianyi. The team’s core mission is to realize the formula “Model + Harness = Agent.” They are tasked with complementing the model’s engineering capabilities and developing end-to-end engineering Agents.

The Harness system acts as the execution layer, managing tasks such as context handling, tool invocation, and error correction, essentially covering all tasks outside the model itself. The industry has long noted the challenge that while models can write code, they struggle to complete comprehensive engineering tasks independently. DeepSeek aims to bridge this gap with its Harness initiative.

Fundamental Logic Behind Betting on AI Coding

The current coding performance gap among top-tier models, as measured by the authoritative coding test SWE-bench Verified, is only 0.1 to 0.9 points. However, this small margin determines whether a model can independently complete engineering tasks. The U.S. export restrictions on Anthropic’s top coding models further confirm AI coding’s status as foundational infrastructure.

DeepSeek’s push for autonomous AI coding aims to democratize AI productivity. The current cost of its code model input is 0.2 yuan per million tokens—about 1/138th of Anthropic Fable 5’s cost. The company’s goal is to transform AI coding from a luxury to an everyday utility, making it akin to semi-public infrastructure accessible to all. This initiative represents a shift in the industry’s narrative from consumption to production.

Editorial Opinion

In the short term, DeepSeek’s development of proprietary AI computing centers could alter the supply structure of computational resources in China’s AI industry. Reduced reliance on cloud vendors may lead to a reorganization of the competitive landscape. If the coding capabilities of V4.1 are confirmed to have improved, competition with Anthropic in the developer market—both domestically and internationally—will intensify. Particularly, DeepSeek’s cost advantage could accelerate AI adoption among small and medium-sized enterprises and startups.

From a long-term perspective, AI coding has the potential to become a productivity infrastructure for software development. If DeepSeek’s low-cost strategy gains traction, the user base for advanced AI support will expand, helping to close the technological gap. However, U.S. export restrictions on chips could constrain Chinese AI companies’ hardware procurement, raising questions about the sustainability of DeepSeek’s infrastructure strategy. Whether the company can meet the explosive growth in computational demand will be a decisive factor in its future trajectory.

The editorial team will be closely watching how effectively DeepSeek’s Harness team bridges the gap between model capabilities and engineering implementation.

References

• Huxiu: “Where Will DeepSeek’s 50 Billion Yuan Be Spent?” — Published June 22, 2026