The $7 Billion Detour: Tencent Leases 100,000 Nvidia AI Chips From Oracle Without Ever Shipping Them to China

There is a hole in the wall that surrounds American AI chips, and on Wednesday the Financial Times put a price tag on it: roughly $7 billion. That is the estimated value of a five-year agreement under which Tencent, China’s largest gaming and social-media company, will lease computing capacity across multiple Oracle data centers in Southeast Asia — capacity that gives it access to approximately 100,000 advanced AI chips it is not permitted to buy.

According to the FT report, corroborated by Reuters and widely picked up on October 1, the deal is structured as a straightforward commercial lease. Tencent committed to roughly $7 billion over five years, paid about 30 percent upfront, and in exchange receives training capacity on hardware sitting in Oracle facilities across Southeast Asia. The chips themselves never enter China. Reported as Tencent’s largest-ever overseas cloud arrangement, the deal gives the company access to frontier-grade Nvidia compute precisely where US export law says it cannot have it: inside its own borders.

The loophole, precisely stated

US export controls on advanced semiconductors are physical. They regulate the movement of chips: which companies may buy them, in what quantities, and into which territories. Since the H20 restrictions and the broader China chip bans, Chinese AI developers have been unable to legally purchase Nvidia’s most advanced accelerators for domestic deployment.

What the controls do not directly regulate is access to compute. A Chinese company cannot import a rack of B200s into Shenzhen. But nothing in the current rules prevents it from renting time on those same chips if they sit in a data center in Malaysia, Singapore, or Indonesia — owned by a US company, operated by US-controlled cloud software, and paid for in dollars through a commercial contract.

That distinction — governing the chip but not the workload — is the gap Tencent has now industrialized. And it did so not through some gray-market intermediary or a cutout shell company, but through a contract with Oracle, a pillar of American enterprise software. The symbolism is hard to miss: the enforcement architecture of US technology policy now has one of its own cloud providers as a paying counterparty on the other side of the wall.

A pattern, not an isolated case

Tencent’s deal is the largest single instance, but it is not the first. Reporting throughout 2026 has documented Chinese AI firms testing the boundaries of overseas compute access. A CNBC investigation in August examined how Chinese companies were already accessing advanced Nvidia computing power outside China’s borders. Earlier reporting traced Tencent tapping Nvidia Blackwell B200 processors through a Japanese partner, Datasection, which owns the hardware and rents capacity back to a single major customer. ByteDance, meanwhile, has been reported to have secured tens of thousands of B200s in Malaysia.

The scale is what has changed. Those earlier arrangements were measured in the tens of thousands of chips at most, often through intermediaries of varying opacity. The Oracle deal is 100,000 chips, five years, ~30% upfront, with a US-listed counterparty and an estimated contract value that would rank among the largest AI infrastructure commitments of the year by any measure. When the workaround becomes the size of the original problem, it stops being a workaround.

Why Oracle said yes

Oracle’s motives are commercial and uncomplicated. The company is in the middle of the largest infrastructure buildout in its history, anchored by the reported $300 billion, five-year cloud contract with OpenAI that begins in 2027. Southeast Asia is its fastest-growing capacity frontier. Filling that capacity with a creditworthy anchor tenant willing to pay 30 percent upfront — $2 billion or so in immediate cash — is exactly the kind of deal that de-risks a capital cycle.

Whether it survives regulatory scrutiny is another question. The deal was structured to be lawful under current rules; whether it survives future rules is precisely what Tencent’s 30 percent upfront payment hedges against. If Washington moves to close the compute-rental gap — by extending export controls to cover US-person-provided compute services to Chinese entities, as some in Congress have already proposed — the upfront cash is already banked, and Oracle’s lawyers will have a signed contract to point to.

There is also a shareholders’ angle that US policymakers will find awkward. Oracle’s stock moved on the news precisely because the market reads the deal as low-risk, high-revenue: the same export-control regime that constrains Nvidia’s China sales is generating a windfall for another US company one step up the stack. Enforcement hawks get a weaker negotiating position; Oracle investors get contracted revenue. Both are reading the same rules correctly.

What it means for the AI race

For Beijing, the deal is a quiet vindication of patience. Chinese AI policy has spent two years pushing domestic alternatives — Huawei’s Ascend line, Cambricon, and the emerging supernode designs — while simultaneously letting its champions buy whatever frontier access remains legally available abroad. Tencent itself said in August 2025 that it had ample chips for training and was upgrading existing models rather than chasing new hardware; this deal suggests that stance has evolved as training demands for next-generation models have grown.

The message to Chinese AI labs is that the compute constraint is softer than the export-control headlines imply. Frontier training runs can be executed on leased overseas capacity; inference and domestic deployment ride on domestic silicon. DeepSeek’s open-sourcing of the Ascend programming stack just last week — TileLang, DeepGEMM, DeepEP, and the rest of the Huawei software suite — completes the picture: a two-track strategy in which foreign chips train what domestic chips will run.

For Washington, the deal lands at a singularly awkward moment. The same week saw the White House sign its voluntary “Super Intelligence” accord with six frontier labs, the Pentagon unveil Project Meridian, and the FTC open an industry-wide probe. The Tencent-Oracle lease is a reminder that the levers the administration prefers — voluntary commitments, moral suasion, procurement pressure — do nothing to stop a contract between a Chinese conglomerate and a US cloud provider that is, under current law, simply legal.

The closing question

Every export-control regime eventually meets its arbitrage. The 1980s COCOM restrictions on Soviet technology transfers were defeated by front companies; the Huawei entity list was blunted by redesigned chips and third-country sales. The compute-rental gap is the 2026 edition, with one important difference: this time the counterparty is not a shadowy intermediary in a third country. It is a company headquartered four miles from the Pentagon, with a market cap over half a trillion dollars, renting out the wall itself.

The bet embedded in the $7 billion is that Washington will not close the gap in time to matter. Five years is roughly two generations of AI models. If the pace of the last two years holds, the chips Tencent is renting will be obsolete — and the models they trained will already be running on something else — long before any legislative fix arrives. That is not a loophole being exploited. That is a loophole being scheduled around.