China's AI Declares Chip Independence

The semiconductor war is officially on, and China’s AI giants are declaring independence. US export controls, intended to cripple China's technological advancement, instead ignited an unprecedented domestic arms race. Nvidia, once the undisputed king, saw its AI chip market share in China reportedly plummet from over 90% to a projected 8% by 2025, forcing local companies to innovate or perish.

This wasn't merely a business pivot; it became a national imperative. To achieve true technological sovereignty and insulate from geopolitical pressures, Chinese firms now recognize the strategic necessity of owning the full AI stack—from proprietary models to the foundational silicon. Matthew Berman aptly calls this a "no-brainer."

Leading the charge, DeepSeek, known for its open-source AI models, reportedly began developing its first in-house AI inference chip a year ago. Z.ai, behind the formidable GLM-5.2, also evaluates custom chip development, aiming to reduce reliance on foreign suppliers like Nvidia and Huawei. The soaring cost of AI inference further accelerates this push for domestic control.

This ambition transcends mere business strategy. Beijing now actively encourages and subsidizes the use of domestically produced AI chips, transforming what was once a corporate preference into a critical pillar of national security and technological self-reliance.

Chinese AI powerhouses now declare independence not just in rhetoric, but in silicon. DeepSeek, creator of globally recognized open-source models, initiated its own custom inference chip project roughly a year ago. This ambitious undertaking aims squarely at mitigating the prohibitive operational costs of running their advanced models at scale, a direct response to the monumental expense of AI inference.

Z.ai, developer of the formidable GLM-5.2—a leading open-source model acclaimed for long-horizon coding—mirrors this strategic pivot. Confronted by a staggering 27-fold surge in daily token usage, Z.ai actively explores custom hardware solutions. Their preliminary discussions with Chinese semiconductor design firms highlight the urgent need to alleviate immense pressure on existing computing resources, a development path that could span over two years.

This isn't merely a technical endeavor; it's a profound reorientation. Top-tier software and model companies are now descending the stack, directly engaging in complex hardware design. They seek not just to reduce dependence on foreign suppliers like Nvidia, but to gain absolute control over deployment costs and performance, forging a complete AI stack from algorithms to custom metal. This costly gamble, with an estimated $500 million per advanced chip and no guarantee of success, underscores the existential necessity of chip autonomy for China's AI future.

Make no mistake: DeepSeek's chip venture is a half-billion dollar gamble. Designing a single advanced AI chip can easily demand $500 million in investment, a staggering sum with absolutely no guarantee of success. This isn't just about capital; it’s a high-stakes bet on overcoming monumental technical hurdles, where failure means incinerated cash and lost time.

The technical gauntlet is brutal, pitting them not only against global leader Nvidia, but also formidable domestic rivals. Established champions like Huawei, with its potent Ascend chips, already command significant market share in China, having capitalized on Nvidia’s export control woes. DeepSeek must not merely build a chip, but engineer a superior, differentiated solution in an intensely competitive landscape.

Crucially, this isn't an immediate frontal assault on Nvidia's cutting-edge training dominance. Instead, DeepSeek aims for highly optimized, cost-effective silicon tailored specifically for their demanding inference workloads. This bespoke hardware promises to slash the soaring operational costs for their advanced open-source models, which you can explore further at DeepSeek. Their initial goal is strategic niche mastery, creating chips that serve their specific needs far better than general-purpose silicon.

The sanctions-fueled arms race now accelerates an inevitable bifurcation of the global AI landscape, creating two distinct technological worlds. US export controls decimated Nvidia’s market share in China, plummeting from over 90% to an estimated 8% by 2025. This vacuum is not merely being filled; it actively forges a parallel Chinese AI ecosystem, complete with bespoke hardware, advanced open-source models like DeepSeek-V3 and GLM-5.2, and its own emerging standards.

Chinese firms like DeepSeek and Z.ai are not just reacting; they are strategically building for self-reliance. DeepSeek’s initiative to develop an in-house inference chip exemplifies a "good enough" hardware strategy. This focus on running trained models, rather than the more demanding task of training them, could offer sufficient performance at a significantly lower operational cost, addressing the soaring expense of AI inference.

This strategic pivot secures a defensible, insulated AI market, independent of foreign supply chains. The immense financial risk — designing a single advanced AI chip costs an estimated $500 million — underscores the stakes. This shift fundamentally redefines the AI race, transcending the mere intelligence of an LLM. The new battleground is full AI stack ownership.

From the silicon wafer to the final application, controlling the entire supply chain dictates true technological sovereignty. DeepSeek’s and Z.ai’s half-billion-dollar gambles confirm this new reality: the future of AI belongs to those who control the metal beneath the models, not just the code on top.

Why are Chinese AI companies building their own chips?

To reduce dependency on US companies like Nvidia, bypass US export sanctions, and gain full control over their technology stack for better performance and lower AI inference costs.

Which Chinese companies are developing new AI chips?

Leading AI model developers DeepSeek, creator of DeepSeek-V3, and Zhipu AI, known for the GLM-5.2 model, are reportedly developing their own custom chips to power their platforms.

Are these new chips for training or running AI models?

The initial focus for companies like DeepSeek is on developing inference chips. These chips are specialized for efficiently running already-trained AI models, which is a major operational cost.

How have US sanctions affected Nvidia in China?

US export controls have severely impacted Nvidia's dominance. Its market share for AI accelerators in China has reportedly plummeted from over 90% and is projected to fall further.