Tencent reportedly leases 100,000 AI chips in Oracle's Southeast Asia data centers

Tencent has signed a five-year deal to lease about 100,000 advanced AI chips across several Oracle data centers in Southeast Asia, the Financial Times reported, in a story TrendForce relayed on October 1, 2026. Neither company has commented on the report, per Tom's Hardware, so treat every number below as reported, not confirmed.

The reported price, per chip-hour

The deal is estimated at about $7 billion, with roughly 30% paid upfront, per TrendForce's summary of the reporting. Tom's Hardware works that out to about $1.60 per chip per hour over five years. The arithmetic is short: 100,000 chips x 24 hours x 365 days x 5 years is about 4.38 billion chip-hours, and $7 billion divided by that is $1.60. Roughly 30% upfront is about $2.1 billion paid before most of those hours run.

Tom's Hardware puts that rate around 43% below the roughly $2.80 per GPU-hour that research firm SemiAnalysis reported in August for a one-year Nvidia H100 contract. It is also about 48% below the $3.09 per GPU-hour for a three-year B200 contract shown in a 2025 filing by Japanese data center operator Datasection. Market pricing is moving the other way: the FT reports longer terms and larger upfront payments for leases like this one.

Nobody has said which chips

No chips are named. Tom's Hardware notes the FT says such leases could reach Nvidia's top processors, but that this does not mean the deal involves Nvidia hardware. Its own read is that the price fits Hopper-class chips (H100 and H200) at volume better than Blackwell. As a sanity check, it prices 100,000 B200s at $5.44 billion in raw server cost, about 78% of the reported deal value, before buildings, power, networking, and financing.

Where the compute physically runs

The chips stay in Oracle data centers outside China, and Tencent rents time on them. The countries, sites, and start date are undisclosed, per Tom's Hardware, which notes Oracle has two Singapore cloud regions and has announced one in Malaysia. The FT says the biggest clients of Southeast Asian data centers are ByteDance and Alibaba, and that Chinese firms use the capacity especially for AI training.

The spending is visible in Tencent's own numbers. Its second-quarter results, released August 12, say "Capital expenditure was RMB52.8 billion, up 176% YoY." Free cash flow was negative RMB13.8 billion; excluding prepayments for compute procurement, Tencent says it would have been RMB37.6 billion.

What the export rules cover, and what they do not

The license requirements for advanced AI chips were set out in the Bureau of Industry and Security's October 2023 rule, effective November 17, 2023. For the most powerful data-center chips (ECCN 3A090.a) it requires a license to ship to destinations in Country Groups D:1, D:4, and D:5 that are not also in A:5 or A:6. It also requires one for exports to an entity "headquartered in, or whose ultimate parent company is headquartered in, either Macau or a destination specified in Country Group D:5" (the group that includes China), even when that entity sits in another country. Shipping those chips to a Tencent facility in Singapore would trigger it.

Renting hours on chips that never move is the gap. The same rule only asked Infrastructure as a Service providers for comment on possible rules, "including know your customer requirements," and BIS wrote that it "continues to evaluate how it may approach this through a regulatory response." A January 2026 trade-law note puts the current position plainly: BIS "does not consider the provision of cloud computing services to be exports."

That could change. The Remote Access Security Act (H.R. 2683), which would authorize BIS to regulate remote access to controlled items, passed the House 369-22 on January 12, 2026, and still needs the Senate and a signature. The Information also reports Commerce is drafting a rule to block Chinese firms from renting compute in third countries such as Thailand and Singapore, though a Baker McKenzie attorney told it such a rule cannot be enforced under current law, per Tom's Hardware.

Why a build studio cares

This lease splits one workload across three jurisdictions: a Chinese customer, a US-headquartered provider, and data centers in a Southeast Asian country nobody has named. That split is what the data sovereignty and residency map in our Full Audit records for a client's own stack: for each data flow, which region processes it, which provider runs that region, and under which country's law. "Our vendor uses Oracle Cloud" does not fill a row on that map. "Inference runs in a named Singapore region, on a US provider" does.

The Tencent case adds a column we also fill in: which rule or policy each placement depends on today. That is a dependency we record, not a legal opinion. Here it is a line between shipped chips and rented hours, with a House-passed bill and a reported draft rule aimed at it. When a client's AI features run on rented GPUs in a third country, a rule change can force that compute to move, and that dependency goes into the vendor vanish-risk score alongside domain and hosting control.

Next step: read the Tom's Hardware analysis for the pricing math, then section D of the BIS October 2023 rule for the cloud question it left open. If your product runs AI workloads on rented compute and you cannot name the country it runs in, write to us at hello@gattyworks.com and we will map it.