Chinese AI researchers are becoming increasingly pessimistic about catching up to the U.S. in artificial intelligence. They cite a significant chip shortage stemming from U.S. restrictions, which prevents them from accessing advanced hardware like Nvidia's latest products. This gap may be widening rather than closing, despite some progress in specific areas.

Megaspeed International, a Singapore-based AI firm, has rapidly become the largest Southeast Asian purchaser of Nvidia chips. This surge has raised alarms in Washington over potential semiconductor smuggling into China, despite Nvidia's assurances that such diversion does not occur.

Jensen Huang, CEO of Nvidia, revealed concerns about China’s growing AI workforce during a private dinner in Taipei. He highlighted a vast talent gap between China and the US and criticized US export controls for inadvertently boosting China’s AI capabilities.

Nvidia has requested TSMC to ramp up production of its H200 AI chips to meet high demand from Chinese companies, which have ordered over 2 million chips for 2026. Despite regulatory hurdles, Nvidia anticipates significant revenue growth if it can fulfill these orders.

Top Chinese companies like Alibaba and ByteDance are training their AI models in Southeast Asia to access Nvidia chips, circumventing U.S. restrictions. This shift follows the U.S. ban on certain chip sales, prompting a rise in offshore training efforts. DeepSeek is an exception, training its model domestically while collaborating with Huawei on new AI chips.

Anthropic's CEO Dario Amodei argues that allowing Nvidia to sell GPUs to Chinese companies is akin to arming an adversary. He believes this decision could strengthen Chinese AI developers like DeepSeek and undermine U.S. technological leadership. Despite concerns, he admits that Chinese models have yet to compete effectively against American counterparts.

Nvidia is set to release a new AI chipset based on its Blackwell architecture for the Chinese market, priced between $6,500 and $8,000, significantly lower than its previous H20 model. The new chip will utilize conventional memory and simpler manufacturing processes, avoiding advanced packaging technologies from TSMC. This move comes as Nvidia adjusts to U.S. export restrictions while seeking to maintain its presence in China's data center market.

The Trump administration has halted its plans to restrict exports of Nvidia's H20 artificial intelligence chips to China following a dinner with CEO Jensen Huang at Mar-a-Lago. The decision comes after Nvidia pledged new U.S. investments in AI data centers, while Chinese companies have already placed significant orders for these advanced chips.

The U.S. government has announced new restrictions on the export of artificial intelligence chips from companies like Nvidia and AMD to China, aiming to hinder the country's advancements in AI technology. This move reflects a broader strategy by the Trump administration to combat China's growing capabilities in the tech sector.

Nvidia is working on a new AI chip built on its Blackwell architecture, aimed at outperforming its current H20 model available in China. Although U.S. President Trump has hinted at the possibility of allowing the sale of more advanced chips to China, regulatory approval remains uncertain due to security concerns. Samples of the new chip are expected to be delivered to Chinese clients as early as next month.

The U.S. government has imposed a fee on exports of Nvidia's H20 chip and AMD's MI308 to China, both significant for AI applications. Nvidia has indicated the export restrictions previously cost it $4.5 billion in a single quarter, while demand for the H20 chip in China remains high. AMD has not yet commented on the situation.

Two individuals have been arrested for attempting to smuggle AI chips from the U.S. to China, which raises concerns about national security and technology export regulations. Meanwhile, Nvidia has reiterated its stance against implementing kill switches for its products, emphasizing the importance of maintaining technological access.