Our Blog

Blog Index 

Oracle Scraps $7 Billion Tencent AI Chip Lease as Washington Closes the Overseas Compute Loophole

Posted on 7th Nov 2026 11:05:11 in Artificial Intelligence, Machine Learning

Tagged as: Oracle, Tencent, AI Chips, Export Controls, AI Infrastructure, Nvidia

Oracle has cancelled a five-year agreement to lease advanced AI computing capacity to Tencent, walking away from a contract valued at roughly $7 billion after US officials raised concerns about the arrangement. The Financial Times reported the cancellation on Wednesday, November 4, 2026, citing two people familiar with the matter. Oracle, Tencent and the White House did not respond to requests for comment, and Reuters said it could not immediately verify the report.

The deal would have given Tencent access to about 100,000 advanced AI chips housed in Oracle data centres across Southeast Asia — hardware the Chinese internet group cannot legally buy at home. It was Tencent's largest overseas compute lease with a US cloud provider, and it sat in a category that US export rules technically permitted: renting foreign capacity rather than importing restricted silicon. That permission, according to people briefed on the discussions, is now the part Washington wants to revisit.

The Deal That Vanished: 100,000 Chips, Five Years, $7 Billion

The structure of the agreement explains why it mattered well beyond the two companies. Oracle signed the lease earlier in 2026, committing capacity for five years. Tencent was to pay about 30 percent of the contract value upfront, a prepayment that helped Oracle finance the data centre build-out behind the deal. In exchange, Tencent secured a steady supply of cutting-edge accelerators in a region where US-designed hardware is legal to operate but not to ship into China.

For Tencent, the chips were not a luxury. The company's cash flow turned negative in the second quarter of 2026 at RMB 13.8 billion, or about $2.06 billion — the first negative figure in more than a decade. Chief financial officer John Lo attributed the swing to heavy AI infrastructure spending and prepayments for compute. Executives said the leased capacity would go first to training larger Hunyuan models and then to serving them, with the possibility of renting surplus capacity through Tencent Cloud later.

Tencent president Martin Lau has defended the spending publicly: "We're comfortable in making significant investments in AI because not only is there a substantial upside potential, there is also clear downside protection." The cancellation removes a large slice of that protection at precisely the moment Tencent is trying to keep pace with Alibaba and ByteDance on frontier model training.

Washington's Two-Part Squeeze: The Nvidia 51 Percent Rule

The Oracle cancellation did not happen in isolation. A day earlier, the Financial Times reported that Nvidia had told its cloud computing partners that at least 51 percent of their capacity must be sold to non-Chinese companies under new contracts. The requirement was communicated verbally through Nvidia's sales force and regional teams rather than in writing, according to five people cited by the newspaper. Existing agreements were left untouched.

Nvidia neither confirmed nor denied the policy when approached by AFP. The company said in a statement that "America cannot win the AI race unless America's AI infrastructure wins worldwide," arguing that US policy encourages cloud providers to build services on American technology.

The two moves fit together. US export controls have restricted sales of Nvidia's most advanced chips to China since 2022, but Chinese firms found workarounds: renting compute from cloud providers in Singapore, Malaysia and elsewhere, and in some cases smuggling hardware. Earlier this year, several senior employees at server maker Super Micro were charged with smuggling around $2.5 billion worth of Nvidia AI servers into China. The cloud-lease route was legal but achieved a similar outcome, and the new Nvidia rule narrows it without changing a single export regulation.

That is the important regulatory detail. No new statute was passed, no formal rule was published in the Federal Register. A supplier, under pressure from the administration, changed the commercial terms on which capacity is allocated. For companies that plan procurement years ahead, that is a harder risk to model than a published export control list.

Beijing Pushed From the Other Side

Tencent was not only under pressure from Washington. According to the Financial Times, senior officials from Guangdong province, where the company is headquartered, visited Tencent and urged it to support chips manufactured in China. Beijing has spent the past year pushing its largest technology companies toward domestic silicon, including a ban on foreign AI chips in state-funded data centres.

The result is a squeeze from both ends: Washington does not want Chinese models trained on American accelerators anywhere in the world, while Beijing does not want its champions dependent on American supply chains. Tencent has hedged by backing domestic chipmaker Enflame, which nearly tripled in its Shanghai debut in September 2026. But domestic alternatives remain capacity-constrained, and the gap between what Chinese fabs can supply and what frontier training demands is the central bottleneck the overseas lease was designed to bypass.

China's foreign ministry, asked about the report, said spokeswoman Mao Ning was not aware of it and added that Beijing hoped all parties would provide a "fair, transparent and non-discriminatory business environment" for companies from every country.

What Tencent Loses — and Why It Cannot Simply Buy Elsewhere

Tencent's options are narrower than they appear. Buying the chips directly is barred by US controls and, in parallel, by Chinese rules that keep foreign accelerators out of state-funded facilities. Leasing abroad was the practical middle path, and it is now visibly politicised.

Alternative suppliers exist, but each carries friction. Other US cloud providers face the same 51 percent pressure on new contracts. Non-US providers operating Chinese-made accelerators cannot match the performance of the newest American parts for frontier training. And building domestic capacity takes years, not quarters. Tencent can redirect spending toward Huawei and Enflame silicon, but that changes the economics of training runs rather than removing the constraint.

There is also a pricing signal worth noting. A rack of Nvidia's B300 chips reportedly sells for about $600,000 internationally, while the same class of hardware fetched around $2.8 million in China last month. That gap is the financial incentive behind every workaround Washington is now trying to close — and it explains why enforcement is moving from customs borders to contract terms.

Oracle's Problem Is Timing, Not Just $7 Billion

On its own, the lost contract is manageable. At roughly $1.4 billion a year, it represents about 1 percent of what analysts expect Oracle to earn in fiscal 2028. The company can, in principle, re-lease the freed capacity to other customers.

The timing is what stings. Oracle priced a $20 billion stock and bond sale two days before the cancellation report, with interest of up to 8.65 percent on part of the debt. Its shares have roughly halved over the past year as investors weigh AI growth against the cost of financing it. Free cash flow has been below zero for several quarters, and the company is spending on data centres more than it earns in a quarter. Its backlog stood at $664 billion in August, a figure that depends entirely on converting signed contracts into revenue.

Re-leasing the Southeast Asian capacity is also harder than it sounds. ByteDance and Alibaba are the biggest buyers of data centre capacity in that region. Under Nvidia's new rule, any new contract must sell most of its capacity to non-Chinese buyers — a requirement one Singapore-based cloud provider reportedly described as difficult to meet outside the United States. Oracle's fiscal second-quarter results in mid-December will show whether anything else moved alongside the Tencent deal.

The Real Shift: Compute Is Becoming Regulated Geography

For a decade, the mental model of AI regulation was a list: which chips, which countries, which performance thresholds. That model is being replaced by something more fluid. Location of the data centre, nationality of the end customer, and the identity of the cloud provider now determine what a company can train and where.

This has consequences beyond the two headline companies. Neoclouds — the GPU-rental specialists that grew up serving AI labs — must now police their own customer mix. Chinese firms that built training pipelines around overseas capacity must re-plan. And every cloud provider selling to multinational customers has to decide how it documents end use, because a verbal instruction from a chip supplier can become a contractual condition overnight.

None of this means the compute trade stops. ByteDance and Alibaba reportedly maintain substantial leasing arrangements with US cloud providers, and Oracle remains an important supplier to ByteDance in Asia through data centres in Malaysia. But the direction of travel is clear: the loophole is being closed administratively, one commercial relationship at a time, and the companies caught in the middle have to price political risk into contracts that used to be purely technical.

What This Means for Indian Enterprises and AI Buyers

India sits outside the direct line of fire, but the second-order effects are real. Indian enterprises renting GPU capacity from US or Southeast Asian providers should expect tighter allocation for any workload that could be re-exported or used by Chinese-linked entities, and more diligence requests at onboarding. Contracts that guarantee capacity in a specific region are worth more than contracts that promise capacity somewhere.

Three practical points for procurement teams. First, ask where the accelerators physically sit and which jurisdiction governs the provider's customer mix. Second, secure continuity clauses — what happens to reserved capacity if a supplier's upstream chip provider changes the rules mid-term. Third, treat single-supplier dependence on frontier hardware as a concentration risk, the same way you would treat a single-supplier dependency in any other critical input. Indian AI service companies serving US clients may also face new end-use questionnaires, because their customers now have to prove where compute went.

The broader lesson is that AI infrastructure is no longer a commodity market with a spot price. It is a licensed, politically mediated supply chain, and the terms can change without a law being passed.

What to Watch Next

  • Oracle's December earnings: whether the $664 billion backlog holds and whether management quantifies the Tencent impact.
  • Other cloud providers: whether the 51 percent rule spreads from Nvidia's verbal guidance into written contract terms across the industry.
  • Tencent's capex guidance: how much of the freed $7 billion shifts toward domestic silicon, and whether Hunyuan training timelines slip.
  • Beijing's response: whether Guangdong-style pressure on domestic chip adoption becomes formal policy for all major platform companies.
  • Enforcement: whether Washington moves from arm-twisting suppliers to formal rules covering overseas cloud leases, which would make the current arrangement permanent rather than ad hoc.

Sources

Recent Posts

Catgories

Archives

`
whatsapp me