Trung Quốc sở hữu các mô hình AI hàng đầu thế giới nhưng lại gặp khó khăn trong khâu vận hành do thiếu chip

 

  • Trung Quốc đã cho ra đời những mô hình AI hàng đầu như DeepSeek v3Kimi K2, vượt trội về khả năng mã hóa và kiến thức khoa học, thậm chí vượt qua cả ChatGPT 4.1Claude 4 Opus.

  • Tuy nhiên, điểm nghẽn lớn nhất không phải ở đào tạo mà là ở khả năng suy luận (inference) do thiếu chip bán dẫn cao cấp như Nvidia H20, dẫn đến tình trạng chậm, giới hạn sử dụng và mất kết nối thường xuyên.

  • Moonshot AI thừa nhận trên X: “Kimi K2 đang RẤT RẤT CHẬM”. DeepSeek hoãn ra mắt mô hình tiếp theo để tránh lỗi tương tự.

  • Tin vui đến vào giữa tháng 7 khi chính quyền Trump bất ngờ dỡ bỏ lệnh cấm xuất khẩu chip H20 sang Trung Quốc, giúp tháo gỡ phần nào nút thắt về năng lực tính toán.

  • Trung Quốc có lợi thế về nhân lực khoa học kỹ thuật, nguồn dữ liệu khổng lồ, năng lượng sẵn cóý chí chính trị, nhưng không có chuỗi cung ứng chip nội địa ổn định.

  • Trung Quốc đã lách luật bằng nhiều cách: nhập khẩu chip bị cấm trị giá 1 tỉ USD, phát triển chip nội địa như của Huawei, và tập trung phát hành mô hình nguồn mở qua nền tảng như Hugging Face để mở rộng ảnh hưởng dù thiếu hạ tầng.

  • Các mô hình như Qwen3 của Alibaba được tối ưu về hiệu suất, giúp chạy nhanh hơn và tiêu tốn ít tài nguyên hơn. Z.ai cũng ra mắt GLM-4.5 và 4.5 Air tập trung vào tốc độ và hiệu quả.

  • Trong khi đào tạo là chi phí một lần, thì inference là chi phí lặp lại – nếu thiếu chip inference, doanh nghiệp sẽ lỗ liên tục. Do đó, đây chính là nút thắt sống còn với AI Trung Quốc.

  • Mỹ đang có chiến lược hai mặt: vừa thắt chặt kiểm soát chip cao cấp, vừa mở lại xuất khẩu chip tầm trung như H20 để giữ Trung Quốc phụ thuộc vào công nghệ Mỹ thay vì phát triển nội lực.

  • Dù lệnh gỡ bỏ chip H20 có hiệu lực, Nvidia vẫn chưa đủ nguồn cung cho thị trường Trung Quốc cho đến cuối năm 2025. Do đó, ưu tiên hiện tại vẫn là mô hình nhẹ, chạy được trên máy tính cá nhân.

  • Nếu Mỹ duy trì xuất khẩu chip trong năm 2026, ngành AI Trung Quốc có thể bùng nổ trở lại, thoát khỏi nút thắt công suất hiện tại.

📌 Trung Quốc đang đứng đầu về mô hình AI nguồn mở nhưng bị kìm hãm bởi thiếu chip inference như H20. Quyết định bất ngờ từ Mỹ cho phép Nvidia xuất khẩu lại chip có thể tháo gỡ tạm thời, nhưng tương lai ngành AI Trung Quốc vẫn phụ thuộc vào nguồn cung chip và khả năng tự chủ công nghệ tính toán.

https://www.economist.com/science-and-technology/2025/07/30/china-has-top-flight-ai-models-but-it-is-struggling-to-run-them

China has top-flight AI models. But it is struggling to run them

Trump’s U-turn on chip-export controls could be a boon 

|5 min read
Six months ago DeepSeek, a Chinese artificial-intelligence (AI) firm, wowed the world with the v3 model and its successors. For the first time, a country other than America—and one that America had cut off from the supply of top-of-the-range semiconductor chips—was producing open-source models that rivalled those designed in Silicon Valley.
Despite the restrictions, Chinese firms kept training world-beating AI models—Kimi K2, unveiled in July by Moonshot AI, a Beijing-based lab founded by an alumnus of Google and Meta, rose straight to the top of the global leaderboards. With more parameters, as the connections between a model’s artificial neurons are called, than any open-source equivalent, Kimi K2 outperformed its Western rivals ChatGPT 4.1 on tests of coding ability and Claude 4 Opus on tests of science knowledge.
But for models to really impress, they need to be used. This is where chip restrictions have bitten the hardest. Shortages have affected the data centres AI labs need to run their systems once trained. Slowdowns, usage limits and dropped connections are becoming common. “We’ve heard your feedback—Kimi K2 is SLOOOOOOOOOOOOW,” Moonshot posted on X a few days after the launch. DeepSeek, meanwhile, has delayed the launch of its latest AI model to avoid similar performance issues, according to a report from the Information. And so both companies were given cause to celebrate two weeks ago, when the White House reversed its latest export controls, once again allowing Nvidia to sell its H20 chips in China. Making these available to tech companies there will remove the hurdles currently slowing their growth.
China is fertile ground for an AI boom: the country has millions of science and engineering graduates, spare grid capacity, the political will to build data centres as fast as concrete can be poured, and access to all the West’s public data sources and more of its own. It lacks a home-grown source of computing power, however, a fundamental constraint that has so far shaped the development of its industry.
In the past few months Chinese firms have found many ways to work around American restrictions. Banned chips worth $1bn have entered the country since April and domestic companies, such as Huawei, have developed chips to match Nvidia’s top-end offering in some respects (though at smaller volumes). A relentless focus on efficiency has also led to breakthroughs.
Limited access to chips also explains another feature of the Chinese AI sector that has baffled outsiders: the devotion to open-source releases. DeepSeek v3 and Kimi K2 are both available through third-party hosting services such as Hugging Face, based in New York, as well as to download and run on users’ own hardware. That helps ensure that, even if the company lacks the computing power to serve customers directly, support for its models is still available elsewhere. And the open-source releases serve as an end-run around hardware bans: if DeepSeek cannot easily acquire Nvidia chips, Hugging Face can.
Not all Chinese firms have been equally affected by the restrictions. On Friday Alibaba released the latest model in its Qwen 3 family, an open-source reasoning model called Qwen3-235B-A22B-Thinking-2507. The release brings Qwen, and Chinese AI in general, level with not just the best open-source AI models, but the best AI models full stop.
Alibaba’s system is around a quarter the size of K2, requiring commensurately less computing power to run, and, unlike DeepSeek and Moonshot, Alibaba has substantial cloud infrastructure behind it to keep the models working. Making models faster and more efficient to use is clearly the new game in the Chinese AI sector: on Monday another lab, Z.ai, released two models, called GLM-4.5 and 4.5 Air, explicitly touting their speed and efficiency.
But the canny workarounds and impressive models can stretch a resource constraint only so far. And since April, one limitation has bitten harder than any others: the loss of Nvidia’s H20 chips.
Successful AI companies must be able to do two things: train models and then run them, a process known as inference. The best-funded Chinese labs have continued to launch training runs of comparable scale to their Western peers. But inference has proved trickier. Whereas training data centres need monolithic clusters of top-end chips, inference is best performed by chips that balance power, energy efficiency and the ability to move data at speed. Until April, the H20 was the chip of choice.
Worse, while a training run is an upfront expense that can be recouped as revenue over the lifetime of the model, a company that loses money during inference has no opportunity to make it up. That means access to chips for inference, not training, is the bottleneck limiting the growth of China’s AI industry.
In response, the Trump administration has sent mixed signals. Its AI action plan, published in early July, doubled down on some chip controls, emphasising that denying adversaries access to “advanced AI compute” is a matter of both geostrategic competition and national security, and calling for novel approaches to enforcing export controls. At the same time, it has lifted the ban on H20 exports, arguing that it would be better for Chinese AI to rely on American companies for all their technology needs, including inference, than to develop an equivalent domestic capacity.
In the short term, such an easing will be cold comfort to China. Nvidia’s own supply constraints mean it will be unable to meet the country’s demand for chips until the last quarter of the year at the earliest. That means models which lean on efficient output and the ability to run on phones and laptops directly will continue to be prioritised for now. But if American exports pick up once more, then China’s AI sector could, at long last, start 2026 much less constrained. ■

Không có file đính kèm.

47

Thảo luận

© Sóng AI - Tóm tắt tin, bài trí tuệ nhân tạo