What did Oracle cancel?
Oracle has scrapped a five-year contract to lease advanced AI chips to Tencent, according to the Financial Times. The report, published on 4 November 2026, said the US cloud provider acted after officials in the Trump administration raised concerns. Reuters, which could not independently verify the report, said Oracle, Tencent and the White House did not respond to requests for comment.
The deal was first reported by the FT on 1 October. It covered about 100,000 advanced AI chips hosted in multiple Oracle data centers in Southeast Asia. The five-year lease was valued at about $7 billion, with roughly 30% payable upfront, or about $2.1 billion. The FT described it as Tencent's largest overseas lease deal.
The October report cited two people familiar with the arrangement. It also said the prepayment weighed on Tencent's free cash flow. Tencent reported negative free cash flow of RMB13.8 billion for the June quarter.
The chips were not readily available to Tencent inside China because of restrictions from Washington and Beijing, according to the FT. Tencent planned to use the capacity first for training larger models, then for inference. Spare capacity could eventually be offered through Tencent Cloud, Network World reported.
Neither company ever publicly confirmed the contract. That silence matters. The new report says the deal is off. It does not say whether a contract was signed, whether Tencent paid any upfront sum, or whether money will be returned.
Why did Washington object to the Tencent deal?
US export controls restrict shipments of the most advanced AI chips to China. They generally do not cover remote access to chips installed in data centers in third countries. That gap is at the heart of the dispute.
Cassia King, a researcher at the Institute for AI Policy and Strategy, spoke to CNBC in August. She said the regime controls physical AI chips, not remote access to them. A Chinese company can therefore rent compute abroad without importing hardware.
Remote access is not the same as ownership. A Chinese company can rent capacity from a cloud provider that owns the hardware, often through a local subsidiary or partner. Aolani, the Singapore-based provider linked to ByteDance, told CNBC that its customers do not own the chips behind its service. They have no future claim on them and no physical access, it said. The company said any permitted access to its services is fully compliant with applicable regulations.
The gap has already caused friction. White House official Michael Kratsios accused Moonshot AI of using Nvidia GB300 chips through a facility in Thailand. CNBC also reported that ByteDance was working with Singapore-based cloud provider Aolani to access compute in Malaysia. The Wall Street Journal first reported that arrangement in March.
Chinese AI developers have released a wave of models that compete with US systems. Access to advanced compute abroad has been a factor in that progress, according to industry watchers quoted by CNBC. That helps explain why Washington treats a cloud lease as a national security question.
The pressure is widening. Nvidia has told some cloud customers to take on less business from Chinese companies, The Information reported. The guidance is not mandatory and applies to future contracts, according to The Decoder. Nvidia also asked customers to disclose the identity of end users, the report said.
For Washington, a lease like Oracle's would let a Chinese hyperscaler train frontier models on American chips without buying them. For Tencent, it was a way around a supply squeeze at home. Those two readings explain why a commercial contract became a political problem.
Why does Tencent need offshore AI chips?
Tencent is spending heavily on AI infrastructure. Capital expenditure reached RMB52.8 billion in the June 2026 quarter, up from RMB31.2 billion in the previous quarter, according to its results. Free cash flow was negative RMB13.8 billion, even as revenue rose 11% to RMB204.78 billion.
The company is building models and agents for WeChat, which has more than 1.4 billion users. It launched the Hy3 model and is developing a larger Hy4. It has also been testing an AI assistant called Xiaowei inside WeChat. Compute is the constraint on all of it.
Domestic suppliers cannot fill the gap quickly. Huawei's rotating chairman Eric Xu said in September that the company cannot produce enough AI computing equipment to meet demand in China. It is limiting overseas sales, he said. Huawei plans to have the Ascend 960DT ready in the first quarter of 2027 and the 960PR ready later that year. The 970 and 980 are scheduled for 2028 and 2029.
Xu said he believed Ascend's share of China's AI chip market was bigger than Nvidia's, though Huawei did not provide data. Even so, Huawei says clusters of about 100,000 chips have become standard for training some of the largest models. Tencent's reported lease was for a similar scale.
Huawei is trying to close the gap with systems rather than single chips. Its Ascend 960 supernode is designed to connect up to 4,096 processors. A new architecture called Peerium is meant to scale to as many as one million processors. More than 5,200 developers were active each month on Ascend software. More than 40 models had been trained on the platform, Huawei said.
Tencent is not alone in looking abroad. CNBC reported that ByteDance, Alibaba and Tencent had accessed Nvidia compute remotely through countries including Thailand, Malaysia and Japan. Tencent's cloud business grew at a low-twenties percentage rate on AI demand. The company also raised prices for cloud customers.
Beijing also has a say. The FT reported that senior officials in Guangdong, where Tencent is based, visited the company. They asked it to support domestic Chinese chips. That pressure runs in the opposite direction to Tencent's need for Nvidia-class hardware.
What does the cancellation mean for Oracle?
Oracle is one of the biggest beneficiaries of the AI data centre boom. In its fiscal first quarter, revenue rose 30% to $19.3 billion. Cloud infrastructure revenue rose 121% to $7.4 billion, and remaining performance obligations reached $664 billion, up $209 billion from a year earlier.
The company booked more than $30 billion of additional AI cloud contracts in the quarter. It delivered more than 300,000 GPUs to customers. It also reported negative free cash flow of $5 billion as it invested in capacity. It raised $20 billion of equity through an at-the-market program.
A $7 billion lease spread over five years is worth about $1.4 billion a year. That is modest beside Oracle's guidance of at least $90 billion in revenue for fiscal 2027. But the upfront payment would have helped fund new data centers before the capacity was fully occupied.
Oracle has told investors that demand for AI cloud training and inferencing exceeds supply. That claim underpins its spending plans. A $7 billion lease would have been a visible sign that overseas demand formed part of the backlog. Its annual value is small beside the total, however.
Losing the contract may also reduce political risk. A deal that appeared to route advanced chips to a Chinese company could complicate Oracle's relations with Washington. Oracle competes for cloud and AI work there. Oracle did not respond to Reuters' request for comment.
How would the US close the cloud loophole?
The main legislative vehicle is the Remote Access Security Act. The House passed it 369-22 on 12 January 2026. The bill would give the Commerce Department's Bureau of Industry and Security authority to regulate remote access to controlled items.
It has not passed the Senate. Senators Dave McCormick and Ron Wyden introduced a companion bill in December 2025. The Senate Banking Committee has been considering whether to hold a markup. As of CNBC's August report, the bill was still pending.
The bill would not ban remote access outright. It would extend the Export Administration Regulations to remote access, giving BIS authority to license or restrict transactions. Latham & Watkins notes that this could substantially expand compliance work for cloud providers and their customers.
Legislation alone would not close the gap. BIS would still need a rule defining what compute is covered, who is barred, and how providers verify customers. Michelle Nie of the Center for a New American Security told CNBC that cloud providers would bear the compliance cost. That cost would come from know-your-customer checks.
The Information reported in August that BIS was working on a remote-access rule. Cassia King said the agency could move in a matter of days with White House support. Writing an enforceable rule is the hard part, she said.
Enforcement is already informal. Nvidia's guidance to cloud customers, reported in November, suggests the administration is trying to shape behavior through suppliers. The formal rule is still being written. That approach is faster, but it depends on voluntary disclosure.
What is still unconfirmed about the Oracle-Tencent deal?
The central claim rests on a single outlet. The FT cited people familiar with the matter. Reuters republished the key points but said it could not independently verify them. No company or government official has confirmed the cancellation on the record.
The chip vendor is unknown. Oracle's cloud offers accelerators from Nvidia and AMD, and neither the FT nor Network World identified the processors. Reports refer only to advanced AI chips, a term that covers several generations of hardware.
It is also unclear whether the contract was signed. The FT said in October that Tencent had signed a lease. Oracle and Tencent never confirmed that. Whether Tencent made the 30% upfront payment, and whether it is recoverable, has not been reported.
The pressure campaign is described in general terms. Reuters' summary of the FT report said US officials approached Oracle. It did not name the agencies or the legal authority involved. It also said Guangdong officials asked Tencent to support domestic chips, without naming the officials.
The absence of on-record confirmation is common in export-control cases. Companies often avoid discussing customers that could attract regulatory scrutiny. Governments rarely confirm informal pressure. That makes the FT's sourcing central to the story. The report should be read as a well-sourced account, not a confirmed termination.
Finally, canceled may not mean permanently. The reports do not say whether the contract was terminated, suspended, or left open for renegotiation. The details should be treated as reported, not established.
What should AI teams and investors watch next?
Watch the Senate. If the Remote Access Security Act moves through the Banking Committee, cloud providers will face clearer obligations. A BIS rule could arrive sooner and matter more, because it would define how existing controls apply to remote access.
Watch Nvidia's cloud partners. The company's request that customers reduce Chinese business and disclose end users is a signal. If other chipmakers and cloud providers adopt similar terms, access to offshore compute will become harder to keep quiet.
Watch Tencent's next results. Capital expenditure and free cash flow will show whether the company has replaced the lost capacity or slowed its AI buildout. Its model roadmap, including Hy4, depends on access to large clusters.
Watch Oracle's bookings. Remaining performance obligations and data-centre delivery will show whether one lost lease matters. Oracle has said demand exceeds supply, so freed capacity may be resold to other customers.
Watch Southeast Asia. DC Byte counted 31 planned data centers of 100MW or more across Malaysia, Indonesia and Thailand, against two today. JLL estimates that global data-centre capacity could roughly double to 200GW by 2030. Those facilities are the infrastructure behind cross-border compute.
For investors, the key variable is not the headline value of one contract. It is whether US policy makes offshore compute materially more expensive or slower to obtain for Chinese AI companies. If it does, model releases may slow and cloud prices may rise in the region.
For AI teams, the practical lesson is that compute contracts are becoming political documents. Buyers should expect questions about end users, training locations, and whether a provider can lawfully serve them if rules change.
Practical checks matter. Buyers of GPU capacity should ask where the hardware sits, who owns it, and which entity signs the contract. They should also ask what end-user disclosures the provider requires. Contracts should say what happens if export rules change mid-term. Those questions are easier to answer before a cluster is booked than after a training run is under way.