At the start of 2026, the loosening of export restrictions on the H200 chip highlighted just how much China’s compute ecosystem has changed. Behind the cool market reaction lies genuine progress by domestic chips in performance, ecosystem maturity, and supply-chain security — and a collective shift in customer mindset from “forced substitution” to “active choice.”

This policy adjustment originated from a notice in the US Federal Register on 13 January 2026, which officially cleared the way for the H200 — a “scaled-down” high-end chip tailored for the Chinese market. Reports then emerged that Chinese authorities had approved bulk purchases by tech giants such as Alibaba, Tencent and ByteDance.

Yet in stark contrast to the past, apart from a few hyperscale players with massive compute needs still inquiring about supply details, the broader base of enterprise customers and distributors has remained “unusually calm.” Capital markets are also reacting cautiously, preferring to wait for companies to confirm substantial orders rather than bid up prices on uncertain headlines.

Even Nvidia CEO Jensen Huang was forced to admit publicly on 29 January that “we have not yet received H200 orders from Chinese corporate customers” and said the company was “patiently waiting.” This collective restraint sends a clear signal: the halo around the H200 is fading in China.

In-Depth Analysis: The Three Industrial Logics Behind the “No One Cares” Phenomenon

1. Performance Breakthroughs

It must be acknowledged that, under current constraints on cutting-edge process nodes, the absolute compute of China’s top single-card domestic chips still lags Nvidia’s flagship products.

However, Chinese industry players have carved out a distinctive path to break this bottleneck: using system-level architecture innovation to compensate for process disadvantages at the chip level.

Taking Huawei as a prime example, the company has leveraged its long-term accumulation in interconnect technologies to launch its “SuperPoD” solution. This architecture can efficiently connect tens of thousands of domestic accelerators into a logically unified supercomputer.

The core advantage lies in self-developed high-speed interconnect protocols (such as Huawei’s “Lingqu”), which significantly improve overall cluster efficiency and aggregate compute. As a result, the total effective compute provided within a single rack or cluster can match — or even surpass — solutions built around foreign chips.

For customers needing to train large models with hundreds of billions or even trillions of parameters, domestic compute clusters have therefore become a reliable and highly efficient option.

2. Ecosystem Synergy

Half of Nvidia’s dominance comes from hardware; the other half comes from its unshakeable CUDA software ecosystem. Early on, domestic chips lost the battle here: difficult adaptation and high migration costs kept developers at arm’s length.

Today, that deadlock has been broken by a new model of “hardware–software co-design and industry-wide collaboration.”

On one hand, AI frameworks and chips are now deeply integrated. For example, Chinese AI frameworks such as Baidu’s PaddlePaddle are actively reusing their experience optimising for Nvidia to support domestic chips, dramatically lowering the adaptation bar for hardware vendors and allowing developers to use diverse domestic compute resources in a largely “transparent” manner.

On the other hand, large AI models and chips are being co-designed. A landmark moment came in 2025 when DeepSeek launched its V3.1 model and introduced a brand-new “UE8M0 FP8” precision standard, explicitly stating that it was “designed for the next generation of domestic chips soon to be released.”

This approach — optimising for domestic hardware at the algorithm/model level from day one — creates powerful industrial pull, drawing a large number of developers onto domestic platforms and forming a virtuous ecosystem spanning chips, frameworks and model applications.

3. Market Mindset and Supply Chain

The shift in mindset is the most fundamental driver. Years of flip-flopping in US export policy for advanced chips have completely eroded market trust.

Barely had the H200 easing been announced when, on 21 January, the US House Foreign Affairs Committee overwhelmingly passed a revised “AI Oversight Act,” intended to pave the way for potentially halting exports of even more advanced Blackwell-series chips to China in future.

Such extreme policy uncertainty makes any long-term decision to anchor core compute infrastructure on imported chips inherently risky at a strategic level.

At the same time, domestic chips have entered a positive cycle of scaled deliveries and real-world validation. Industry data show that by early 2026, multiple Chinese AI chip companies had reached “ten-thousand-card” shipment or order volumes — including Huawei Ascend, Baidu Kunlun, Cambricon and at least nine others. Among them, Huawei Ascend now holds the largest market share in China.

Breaking the ten-thousand-card threshold not only indicates that performance and stability have passed market tests, but also marks a transition: domestic chips are graduating from “lab samples” and “pilot projects” to “production-grade workhorses” that can support core business workloads.

Market choices have therefore shifted from passive substitution — “because there is no other option” — to decisions based on a holistic evaluation of performance, cost, security and long-term sustainability.

The relaxation of H200 export controls cannot turn back the clock on China’s drive towards self-reliant compute. Analysts at UBS have explicitly noted that the trend of “compute localisation” in China will continue through 2026, with domestic demand for Chinese AI chips remaining strong.

Of course, the road ahead is not without challenges. Domestic chips still have significant room to improve in terms of ultra-high single-card performance, truly universal software ecosystem support, and manufacturing process sophistication.

The likely future landscape will be “dual-track”:

• A small group of hyperscalers with extreme compute requirements and massive legacy ecosystems may still purchase some H200 chips for specific scenarios;

• But the broader government, enterprise and industry markets will firmly treat domestic chips as the core backbone for building and expanding their compute infrastructure.