In July 2025, Japan’s Datasection told investors it planned to put 625 servers in Osaka. At eight Nvidia B200s per server, that came to 5,000 Blackwell processors. An intermediary had already lined up a customer for the computing time. The filing described one of the world’s largest cloud providers. The Financial Times identified it as Tencent.

Four months later, President Trump told 60 Minutes that the United States would keep the most advanced chips for itself. Put yourself in the shoes of the engineer at Tencent who needed to train a model. How much difference did it make to that engineer whether the machines belonged to Tencent or to a company in Japan that would send Tencent a bill?

For our Techstrong special report, China Couldn’t Buy the Chips, But It Could Rent Them, we followed the same customer through three separate decisions. In Osaka, the FT put the value of the service agreement at about $406 million over three years. Datasection’s filings give us an unusually clear view of the hardware. Tencent’s actual GPU hours remain undisclosed.

Download the Full Special Report (PDF)

Then came Washington’s decision to license direct sales of the H200, a generation behind Blackwell. Reuters identified Tencent among the Chinese companies cleared to buy. The inspection process brought the chips through the United States, where they incurred a 25% tariff on import value. Nvidia subsequently disclosed limited shipments under the licenses and said it had absorbed the tariff. Its filing leaves the buyers unnamed. The 15% figure readers may remember came from an earlier arrangement involving a different chip, the H20.

The third decision concerned an Oracle lease reportedly worth about $7 billion. It would have given Tencent access to as many as 100,000 advanced chips across Southeast Asia over five years. The FT reported that U.S. officials raised concerns with Oracle and that officials in Guangdong separately urged Tencent to use Chinese chips. The lease was canceled. Those 100,000 chips belong in an account of what governments can influence. Counting them toward Tencent’s available capacity would inflate the total by the size of a major AI cluster.

Tencent gives us a way to follow one customer through the system. The broader record runs from a Chinese university buying four months of time on an eight-GPU server to ByteDance’s reported role as the anchor customer for a Malaysian Blackwell build-out. Commerce had already described overseas model training in its own guidance. Washington understood the rental business, and the Oracle episode shows officials could act on a particular lease. The public record offers little explanation for where they drew the line.

There are defensible reasons to distinguish an older chip from a newer one, or a modest rental from a five-year commitment on that scale. What I find hard to defend is telling Americans that China was being denied a capability while measuring success by where the hardware landed. If an export policy is supposed to buy the United States time in the AI race, the measure ought to be how much computing power Chinese developers could actually bring to bear.

Our full report follows the buyers, operators and financiers across Asia and the United States. A companion investigation examines the Trump and Lutnick family business arrangements and what their public disclosures tell us about who retained an economic interest in decisions made in Washington.

Read the full Techstrong special report, China Couldn’t Buy the Chips, But It Could Rent Them.