Русский 
English 
中文 
भारतीय 
Eesti 
Español 
Deutsch 
Български 
Türkçe 
Portuguesa Chinese artificial intelligence startup DeepSeek is making a strategic shift from pure software development to full vertical integration. According to Reuters sources, about a year ago the company initiated the development of its own specialized processor (ASIC), focused exclusively on inference (data output), not on training new models.
This maneuver signifies the company's maturity: from an algorithmic disruptor that broke the market with efficient architectures, DeepSeek is transforming into a full-stack technology giant, striving for full computational sovereignty.
⚙️ Strategic focus: why inference, not training?
The decision to focus on inference is technically and economically justified: 🔹 Economy of scale: Training a model is a one-time (albeit huge) capital expenditure. Inference, on the other hand, generates constant operating expenses (OPEX), which grow proportionally to the number of users. 🔹 Architecture optimization: Inference chips do not require the same colossal memory bandwidth and interconnects as training chips. This allows for the creation of more energy-efficient and cheaper specialized solutions, perfectly tailored to specific DeepSeek model architectures (for example, MoE - Mixture of Experts).
Currently, the project is in its early stages: DeepSeek is conducting closed negotiations with external partners on issues of architecture design, memory supplies, and contract manufacturing. At the same time, the company is actively hiring chip architect engineers, creating an internal hardware development department.
🌍 Geopolitical context: evolution of hardware dependence
Developing its own hardware is a direct response to the escalation of US export control, which DeepSeek founder Lian Wenfen called the main systemic risk for the company back in 2024.
The history of DeepSeek's hardware evolution clearly demonstrates this path:
- Nvidia era: The basic R1 model was trained on Nvidia H800 chips, a specially cut-down version for the Chinese market.
- Huawei era: In April 2026, DeepSeek introduced the V4 model, adapted for the Ascend architecture. Huawei confirmed that their processors were used for part of the V4-Flash training.
- Era of sovereign silicon: The proprietary chip is designed to finally break the dependence on both limited Nvidia supplies and potential bottlenecks in Huawei's supply chains.
🛠 Ecosystem intent: Chip + Code Harness
DeepSeek's hardware initiative fits perfectly into the recent news about the formation of a team to develop Code Harness - a tool for autonomous programming, designed to compete with Claude Code (Anthropic) and Codex (OpenAI).
Autonomous AI agents writing and checking code require ultra-low latency and high throughput when processing thousands of small requests per second. A proprietary inference chip, optimized specifically for such load patterns, will give DeepSeek a critical competitive advantage in the speed and cost of providing Code Harness services, making the product economically viable on a mass scale.
⚠️ Barriers and risks of implementation
Despite the ambitiousness of the plan, Reuters rightly notes that the success of the project is far from guaranteed. DeepSeek will face unprecedented challenges:
🔸 Manufacturing deadlock: Even with a brilliant design, chip manufacturing on advanced technological processes (less than 5 nm) requires access to ASML lithographic equipment and TSMC capacities, which are under strict control of US export restrictions. DeepSeek will have to rely on domestic Chinese factories (for example, SMIC), which are still lagging in the yield rate of good chips.
🔸 Tough internal competition: The Chinese custom AI chip market already has heavyweights with huge budgets, such as Alibaba (T-Head/Yitian series) and Baidu (Kunlun).
🔸 Global race: DeepSeek is entering a race not only with Chinese giants, but also with global players developing their own hardware (OpenAI, Broadcom, Anthropic), which have deeper pockets and access to the best global foundries.
🔸 Tough internal competition: The Chinese custom AI chip market already has heavyweights with huge budgets, such as Alibaba (T-Head/Yitian series) and Baidu (Kunlun).
🔸 Global race: DeepSeek is entering a race not only with Chinese giants, but also with global players developing their own hardware (OpenAI, Broadcom, Anthropic), which have deeper pockets and access to the best global foundries.
📌 Main points:
• DeepSeek is developing its own AI chip, focused exclusively on inference, not on training
• The project started about a year ago, negotiations are now underway with partners on design, memory, and manufacturing
• Main goal: reducing dependence on Nvidia (H800) and Huawei (Ascend) against the backdrop of US export restrictions
• Hardware development is synergistic with the launch of Code Harness: the proprietary chip will provide low latency and low cost for AI programming agents
• Key risks: limited access to advanced foreign factories (TSMC), high competition with Alibaba/Baidu and global players
