DeepSeek develops custom AI chips to cut Nvidia and Huawei reliance

DeepSeek is developing its own custom AI chips for inference to reduce dependence on Nvidia and Huawei, according to a Reuters report. The move follows OpenAI's Jalapeño chip and signals a broader trend among AI labs to control hardware costs and supply chains.

What happened

DeepSeek has been exploring in-house AI accelerators for about a year, with discussions with chip design, foundry, and memory partners, and is actively recruiting experienced chip designers. The focus is on inference chips, which become the recurring cost center once models are deployed. DeepSeek previously trained its R1 model on Nvidia H800 chips (later banned) and its V4 model on Huawei Ascend GPUs. The company recently raised $7.4 billion in funding. Rivals Alibaba and Baidu are also developing their own AI processors.

Why AI builders should care

If successful, DeepSeek's in-house silicon could serve as a case study for controlling inference costs, a major expense for deployed models. The trend toward vertical integration of hardware and software may affect pricing and supply dynamics for AI workloads. It also illustrates how AI firms adapt to export controls and hardware constraints.

Practical implications

AI teams should monitor DeepSeek's progress as it could inform compute procurement and build-vs-buy decisions for inference. The trend may accelerate vendor diversification strategies in constrained markets.

Caveats

Plans are early-stage and subject to change. No final design, timelines, or pricing have been disclosed. Comparisons to OpenAI's Jalapeño are contextual, not confirmation of DeepSeek's product specs.

FAQs

What is DeepSeek developing in terms of AI chips?

DeepSeek is developing in-house AI inference accelerators to reduce reliance on Nvidia and Huawei. The initiative is early-stage, involving discussions with chip design, foundry, and memory partners, and active recruitment of chip designers.

Why are companies building custom inference chips instead of buying off-the-shelf GPUs?

Inference workloads become recurring cost centers as models are deployed, prompting firms to seek cost control and supply stability. In-house chips can offer performance, efficiency, and stack control advantages for commercial deployment. OpenAI's Jalapeño chip co-designed with Broadcom is a parallel example.

How does DeepSeek's move affect Nvidia and Huawei dependencies?

DeepSeek aims to reduce reliance on both Nvidia and Huawei by pursuing its own silicon. Historically, DeepSeek trained R1 on Nvidia H800 chips and later used Huawei Ascend GPUs for V4. The strategic shift mirrors a wider industry trend to diversify hardware sources amid export controls.

What is Jalapeño and how does it relate to OpenAI's hardware strategy?

Jalapeño is OpenAI's custom inference chip co-designed with Broadcom. DeepSeek's reported chip plans are framed as following a similar playbook to gain hardware control and reduce Nvidia dependency.