DeepSeek, the Chinese AI startup behind the popular open-source models DeepSeek-V3 and reasoning model R1, has quietly begun developing its own inference chips to lower costs and reduce dependence on NVIDIA GPUs. The early-stage project, which started about a year ago, signals a major strategic shift as the company seeks to control its infrastructure expenses amid rapid growth.

Why Is DeepSeek Building Custom AI Chips?

For AI companies operating at scale, compute costs have become the dominant expense. Industry analysts note that "compute-related costs may account for more than half of operating expenses for many AI companies". As DeepSeek's user base grows, the company faces mounting pressure to optimize how it serves millions of real-time requests efficiently. Unlike training new models, which requires intense bursts of computing power, inference (the process of running a trained model to answer user queries) demands sustained throughput and energy efficiency. Custom-designed chips can be optimized specifically for these patterns, potentially slashing the cost per query significantly.

DeepSeek's move mirrors a broader industry trend. OpenAI, Meta, and Microsoft have all explored in-house silicon to escape what many view as the NVIDIA premium. By designing chips tailored to their specific workloads, these companies hope to gain both cost savings and competitive advantages in deployment efficiency.

What Are the Technical and Strategic Challenges Ahead?

Chip development is not a quick or inexpensive undertaking. The journey from initial architecture design to tape-out (sending designs to manufacturing) and mass production typically takes over a year, meaning DeepSeek's custom chips won't reach the market soon. The company is also pursuing its first major external funding round, reportedly aiming to raise approximately $7 billion at a valuation between $52 billion and $59 billion. If successful, chip development and AI infrastructure investment would be key priorities for those funds.

DeepSeek faces unique hurdles that many Western competitors do not. U.S. export controls have restricted access to advanced chipmaking tools and technologies, limiting what Chinese companies can develop domestically. DeepSeek may need to rely on domestic foundries like SMIC, which currently lag behind TSMC in manufacturing process nodes. This gap could limit the performance of DeepSeek's chips and extend the development timeline considerably.

How Are AI Companies Pursuing Vertical Integration in Chip Development?

The shift toward custom silicon reflects a broader industry recognition that controlling the entire technology stack, from software to hardware, provides strategic advantages. Here are the key components of this vertical integration strategy:

• Inference Optimization: Custom chips designed specifically for serving user queries rather than training models, enabling better energy efficiency and lower latency for real-time applications.
• Cost Control: By manufacturing their own silicon, companies reduce reliance on third-party GPU suppliers and can better manage long-term infrastructure expenses as user demand scales.
• Competitive Differentiation: Proprietary chip designs allow companies to implement unique architectural features that give them advantages in speed, power consumption, or specialized AI workloads.
• Supply Chain Independence: Developing in-house chips reduces vulnerability to supply shortages, export restrictions, or price increases from dominant suppliers like NVIDIA.

DeepSeek's quiet push into chip development represents a strategic bet on long-term infrastructure independence. The inference chip market is projected to grow substantially as AI adoption accelerates across industries. If DeepSeek can successfully navigate the technical challenges and geopolitical constraints, custom silicon could provide a meaningful competitive edge in deployment efficiency and cost management.

However, the immediate landscape remains largely unchanged. NVIDIA GPUs will continue to dominate AI infrastructure for the foreseeable future. DeepSeek's success will depend on whether its engineering team can overcome the technical complexity of chip design, secure access to advanced manufacturing capabilities despite U.S. export controls, and execute a multi-year development roadmap while simultaneously scaling its AI services to millions of users.