Deepseek to Build Custom AI Chips to Reduce Dependence on Major Chipmakers
Chinese AI startup Deepseek plans custom AI chips for pretrained model use, aiming to reduce reliance on major semiconductor firms and speed efficient deployment.
Chinese AI startup Deepseek is developing its own specialized processors designed specifically to run pretrained models, the company told Reuters through sources familiar with the plan. The move to build Deepseek custom AI chips seeks to reduce the startup’s dependence on established chipmakers and tailor hardware to its inference workloads. Company executives did not provide a public timetable, but three people close to the project described the effort as a strategic step toward tighter control over performance and costs.
Deepseek’s strategic rationale
Deepseek’s initiative reflects a broader trend among AI firms seeking hardware that matches their model profiles rather than relying solely on general-purpose accelerators. By creating Deepseek custom AI chips, the firm can optimize memory access patterns, precision formats, and dataflow for models it already trains or deploys. Sources told Reuters the chips are being designed for inference of pretrained models, where throughput and latency are primary constraints.
Design focus: pretrained-model inference
The planned processors will reportedly emphasize the specific computational patterns of pretrained models, such as large transformer architectures and dense matrix operations. Engineering choices may include quantization support, optimized matrix-multiply units, and high-bandwidth on-chip memory to reduce off-chip traffic. Those design priorities aim to boost efficiency for real-world serving tasks, cutting power consumption and improving response times for production deployments.
Supply-chain and geopolitical considerations
Deepseek’s push to produce its own silicon also addresses supply-chain exposure and commercial leverage. Dependence on a handful of major semiconductor vendors can create cost pressure and scheduling bottlenecks, particularly as demand for AI accelerators rises globally. Building Deepseek custom AI chips would give the startup more control over procurement, IP positioning, and potential export or partnership channels, according to people familiar with the plans.
Market implications for hardware providers
If Deepseek succeeds in fielding viable custom processors, the effort could signal mounting pressure on incumbent chipmakers to offer more specialized, customer-tailored silicon or service bundles. Hyperscale cloud providers and AI startups alike are experimenting with bespoke designs, and an additional player focused on pretrained-model inference could accelerate that shift. However, large suppliers still dominate fabrication, software ecosystems, and developer tooling, which will shape how aggressively customers move away from off-the-shelf solutions.
Technical and commercial hurdles
Designing a competitive processor is technically demanding and capital-intensive, requiring close collaboration with foundries, robust compiler support, and extensive validation against real workloads. Deepseek will need to secure manufacturing capacity, either through third-party fabs or by partnering with established semiconductor firms, and invest in software stacks to make its chips usable at scale. The transition from prototype to production volumes also brings testing, thermal management, and reliability challenges that can stretch timelines and budgets.
Funding, partnerships and next steps
To translate its design work into silicon, Deepseek will likely pursue commercial partnerships and additional funding, whether through venture rounds, strategic investors, or joint development agreements with chip manufacturers. The startup’s ability to align design goals with manufacturing economics will determine whether Deepseek custom AI chips remain a niche optimization or evolve into a broadly adopted product. Observers expect the company to prioritize integration with its existing model stack and to demonstrate clear efficiency gains before attempting wide deployment.
The development underscores an ongoing recalibration in the AI value chain, where software creators are increasingly eyeing hardware as a strategic lever. Deepseek’s effort to develop custom AI chips for pretrained models is part of a wider movement toward co-design of models and silicon that could reshape cost structures and performance expectations across the industry.
