Around 100,000 Nvidia AI accelerators barred from sale to China are now running Tencent workloads inside Oracle-operated data centers across Southeast Asia, according to a Startup Fortune report published this week.

The figure — not yet confirmed by either company — would rank among the largest publicly reported workarounds to Washington's AI chip curbs. It also signals how aggressively Chinese hyperscalers are pursuing indirect paths to Nvidia's top-end silicon, two years into a US campaign to keep the most powerful AI training hardware out of Beijing's reach.

What's actually banned?

The chips fall under US Commerce Department restrictions that, since October 2022, have blocked sales of Nvidia's A100 and H100 accelerators to mainland China. Successive rounds added the A800, H800, L40S and the China-tailored HGX H20.

Nvidia has acknowledged in SEC filings that the curbs have cost it billions in addressable revenue. The company has repeatedly engineered compliant downgrades that have themselves come under fresh rounds of restriction, a treadmill that has driven Chinese buyers toward offshore capacity rather than domestic alternatives.

How does Oracle fit in?

Oracle operates cloud regions across Asia — Singapore, Tokyo, Seoul, Kuala Lumpur — outside Chinese jurisdiction but within reach of mainland customers. Renting capacity there lets a buyer effectively import restricted silicon by the hour, with workloads flowing to neighboring jurisdictions rather than a chip crossing the China border in a box.

US cloud providers have spent two years courting Chinese AI labs with Southeast Asian capacity. Oracle in particular has pitched itself as a neutral infrastructure layer for cross-border training runs, a strategy that puts it on a collision course with the same regulators it is trying to serve.

What does 100,000 H100s buy?

A single H100 packs eight HBM3 stacks at 80GB. Multiply that by 100,000 and you're looking at roughly 8 petabytes of high-bandwidth memory — enough to train frontier-scale models, not just run inference. Western AI labs are widely reported to operate clusters in the 50,000–100,000 H100 range, a benchmark that puts the reported Tencent arrangement in the same league.

Tencent has not publicly disclosed cluster sizes at that magnitude, though its in-house Hunyuan model family has rolled out multiple generations since its September 2023 debut. Whether the 100,000-chip figure represents net-new capacity or a reclassification of existing infrastructure remains unclear from the Startup Fortune report.

What changes next?

The deal sits in a regulatory gray zone. US export rules restrict the sale of controlled chips to China, not the offshore consumption of cloud capacity that happens to contain them.

But Commerce's Bureau of Industry and Security has opened a string of investigations into exactly this pattern, including subpoenas to Nvidia and Microsoft last year over alleged chip-routing through third-country data centers.

For Oracle, the report lands at a delicate moment. The company has been one of the most aggressive Western cloud vendors courting Chinese AI workloads, betting that infrastructure neutrality trumps political geography. A confirmed 100,000-chip lease would put that bet under a much brighter spotlight.

For Nvidia, the figure — if accurate — adds another line to the long ledger of restricted-revenue leakage. The company has declined to comment on specific customer arrangements in past earnings cycles but has repeatedly said it designs to the rule set in force on the day a chip ships.

The cloud-routing question, in short, is moving from fringe workaround to enforcement priority. Whoever blinks first — Commerce, the cloud provider, or the chip designer — will write the next rule book.