Why Are Leading AI Model Companies Developing Their Own Chips?

As generative AI competition heats up, leading model companies are collectively venturing into the hardware domain. Recent reports indicate that DeepSeek, a top Chinese AI firm, has launched an in-house chip development project codenamed 'inference chip.' Zhipu AI is also evaluating custom chips tailored for its GLM series models.

This is not an isolated case. U.S. companies OpenAI and Anthropic have previously revealed their own chip development initiatives. This trend signifies a shift in the AI industry’s competitive focus—from standalone model architectures to end-to-end co-optimization across 'model-chip-cloud.'

From 'One-Time Training Cost' to 'Ongoing Inference Cost'

According to industry observers like 'Gongfu Finance,' the primary motivation for model companies to develop their own chips is inference cost reduction. Unlike training, which is a one-time investment, inference costs grow linearly with user scale. Especially with the rise of AI agent applications, the number of tokens consumed per task can be hundreds of times higher than in traditional dialogue, leading to a dramatic surge in computational demand.

General-purpose GPUs, exemplified by NVIDIA products, offer flexibility but impose a heavy burden on AI firms due to hardware redundancy, resulting in excessive power consumption and margin premiums.

Custom chips (ASICs) eliminate such inefficiencies by optimizing hardware for specific operations. For example, if DeepSeek can hardwire its proprietary MLA (Multi-head Latent Attention) or fine-grained MoE architecture into silicon, it could achieve performance gains far beyond what is possible through software-level optimization alone.

Dual Challenges: Cost and Supply Security

Observers note that for Chinese firms like DeepSeek and Zhipu AI, chip self-development is not just about cost—it's also a strategic response to geopolitical supply risks. Due to U.S. export controls, access to NVIDIA’s most advanced chips is restricted. While domestic alternatives like Huawei’s Ascend chips are available, their production capacity remains limited. Self-developed chips are thus a necessary path to ensure independent compute supply and escape dependency on foreign technology.

Restructuring Business Models

The trend toward in-house chip development also signals a transformation in business models, especially for open-source model providers. A new model may emerge: 'free models, paid hardware.' For instance, companies could sell optimized 'appliance' systems tailored for specific models, converting open-source ecosystems into hardware revenue streams.

Despite significant challenges—such as high R&D costs (around $500 million for advanced chips), long iteration cycles, and software ecosystem barriers like CUDA—the industry consensus has solidified: the ultimate winner in AI competition will be the company capable of optimizing token cost down to the silicon level.