HomeArticle

Raised 50 billion yuan, Liang Wenfeng has gone all out on infrastructure building

凤凰网科技2026-06-09 16:00
The demand for tokens is immeasurable

Abstract: In the past, tasks such as building computer rooms, installing cooling systems, and adjusting PUE were more the exclusive domain of cloud providers and telecom operators. Now, model companies are getting directly involved, indicating an extreme pursuit of control over computing power and cost boundaries.

Recently, the recruitment activities of DeepSeek have been even more lively than the release of its models. IDC design and planning engineers, senior data center operation and maintenance engineers, senior delivery managers... A series of new positions have appeared on major recruitment platforms, making insiders exclaim, "They're really going all out."

Putting these scattered pieces of information together, a clear logical line emerges. The latest news says that DeepSeek's valuation has soared to 350 billion yuan. Currently, it is unhesitatingly investing the funds from its financing into two major directions: computing power infrastructure and upper - layer applications. Buying graphics cards, building computer rooms, developing infrastructure, and creating products all show Liang Wenfeng's ambition.

Focus on Applications with One Hand and Infrastructure with the Other

In this round of recruitment, the most significant and unexpected position is the "IDC Design and Planning Engineer" newly launched on the DeepSeek official website. The full - name is Internet Data Center Design and Planning Engineer, which belongs to the core technical position in computing power infrastructure. It is responsible for the full - process planning and design from the early site selection, scheme, layout to construction drawings and supporting facilities for implementation. In short, it is the core technical person - in - charge in the early stage of computer room construction. An AI company specifically recruiting for planning and design positions for self - built large - scale data centers is sending a very clear signal. Renting computer rooms is no longer enough; they are going to build their own buildings, and it is very likely to be a super - large - scale intelligent computing center ranging from MW (megawatt) to GW (gigawatt) levels.

Meanwhile, earlier, DeepSeek was also aggressively recruiting for its intelligent computing center in Ulanqab, Inner Mongolia. Senior operation and maintenance engineers and senior delivery managers are offered a monthly salary of 15K to 30K, along with a 14 - month salary. The move to build a computer room on the grassland and expand the infrastructure layout from software to hardware was quite unexpected.

Many people's first reaction is to go to the grassland to see the servers. It sounds like a joke, but those familiar with the industry know that this means the large - scale computing power cluster has entered the delivery and operation and maintenance stage, and heavy assets are being invested.

If the infrastructure positions are about going deep, then the other group of new positions is about reaching higher. On May 20th, we also found on the DeepSeek official website that two new positions, "Agent Harness Product Manager" and "Agent Harness R & D Engineer", had quietly appeared. On the same day, a more direct signal came from Deli Chen, a senior researcher at DeepSeek. He publicly recruited on social media with a very straightforward title: "Come to DeepSeek to Build Code Harness from Scratch", and the text even directly stated, "Aim to compete with Claude Code and build DeepSeek Code Harness."

Forming an in - house code agent team and targeting Anthropic's leading programming tool, Claude Code, clearly shows an intention to make in - depth breakthroughs from the underlying model capabilities to the developer toolchain and application layer.

On one hand, they are piling up servers on the grassland and recruiting for planning and design. On the other hand, they are forming a code agent team codenamed Harness. In this round of recruitment, DeepSeek has inadvertently revealed its two major strategies.

50 Billion Yuan in Ammunition for Infrastructure and Productization

The confidence for recruitment expansion is closely related to the recent highly - publicized financing. Previously, DeepSeek was mainly supported by the self - owned funds of Magic Square Quant, with almost no external capital involved. However, recently, according to multiple media reports, DeepSeek has opened its doors to external capital for the first time, with a valuation approaching 350 billion yuan (about $48 billion). Potential investors include giants like Tencent and Alibaba. Once the funds arrive, how to spend them becomes the core concern. Now, it seems that the answer is written in the latest headcount announcements: invest boldly in basic computing power and productization.

In terms of basic computing power, the appearance of IDC design and planning engineers, along with the recruitment for operation and maintenance and delivery in Ulanqab, indicates that DeepSeek is taking a heavy - asset route of "self - building + self - maintaining". According to the recruitment details, IDC design and planning engineers will have the opportunity to participate in the planning and construction of infrastructure from MW to GW levels. "Today's data centers have evolved from traditional computer rooms into large - scale industrial systems supporting AI training and inference." The purpose is clear: to prepare a sufficient computing power base for the continuous training of trillion - parameter models and future inference services.

In the past, tasks such as building computer rooms, installing cooling systems, and adjusting PUE were more the exclusive domain of cloud providers and telecom operators. Now, model companies are getting directly involved, indicating an extreme pursuit of control over computing power and cost boundaries.

In terms of upper - layer productization, the formation of the Harness team indicates the direction for commercial breakthroughs. In the current situation where model capabilities are becoming increasingly similar, development tools and agent products are the key to attracting developers and building an ecosystem. DeepSeek has transferred senior researcher Deli Chen to lead the effort of "building Code Harness from scratch", aiming to transform model capabilities into the most user - friendly programming tools for developers. It is obvious that they are competing with Claude Code and GitHub Copilot for the market.

Meanwhile, with the arrival of financing and the expansion of the team, the launch of positions such as delivery managers also implies that DeepSeek may be more active in serving enterprise - level customers in the future, gradually transforming its past laboratory - based research capabilities into engineering - based and deliverable product solutions.

Buying graphics cards, building computer rooms, deploying applications, recruiting engineers to work on the grassland, and having researchers focus on code agents. Liang Wenfeng, the most low - key leader in the AI field, is making a high - stakes bet in the most hardcore way, investing in both computing power infrastructure and productization.

This article is from the WeChat official account "Phoenix Tech", author: Jiang Fan. Republished by 36Kr with permission.