Tencent has signed a five-year lease with Oracle for roughly 100,000 advanced AI chips, in a deal worth about $7 billion. The Financial Times broke the story on October 1, with Reuters following up. According to the report, this is Tencent's largest overseas compute lease to date, with the capacity housed across multiple Oracle data centers in Southeast Asia and roughly 30 percent of the payment due upfront.
The report says the chips are advanced Nvidia products that cannot be purchased inside China, though it does not disclose the specific model, how capacity is split across data centers, or the delivery timeline. Tencent and Oracle have not confirmed the contract terms through any public channel, so the figures below all come from the reporting.
Can't buy it, so rent it
Current U.S. export controls bar Chinese companies from directly purchasing the most advanced AI chips, but renting compute at overseas data centers falls outside that ban. The chips never leave the foreign facility — what Tencent gets is usage rights, not the hardware itself.
Other Chinese companies have already taken this route, typically at modest scale. A 100,000-chip order, locked in for five years with 30 percent paid upfront, pushes overseas compute rental from a supplementary tactic into a primary training resource. The report frames the goal as speeding up development of Tencent's AI models and agent tools so it can keep pace with other leading domestic model makers.
Southeast Asia is a common landing spot for this kind of arrangement: it's close to China's domestic networks, power and land are relatively abundant, the data centers are run by U.S. cloud vendors, and compliance with U.S. regulators falls to the lessor rather than the lessee.
A rough per-chip price
Dividing the reported total by the chip count gives a back-of-envelope number: $7 billion across 100,000 chips works out to about $70,000 per chip over five years, or roughly $14,000 a year. Assuming full utilization across 8,760 hours a year, that's about $1.60 per chip per hour.
That figure bundles in the data center, power, networking and operations, so it can't be compared directly to a chip's sale price. Long-term contracts are typically cheaper than on-demand rental, but the report doesn't say what discount, if any, this deal carries.
The upfront portion, at roughly 30 percent, comes to about $2.1 billion. For Tencent, that's a lump sum pushed out early in a five-year contract; for Oracle, it's construction capital in hand that can go straight into data centers and power infrastructure.
Oracle's list of big customers
Oracle's growth over the past couple of years has been driven mainly by cloud infrastructure orders, including a previously reported roughly $300 billion compute deal with OpenAI. Tencent's $7 billion doesn't move the needle much by comparison, but the customer is notable: a Chinese company bound by export controls, using the gap that current U.S. rules leave open.
How long that gap stays open isn't something the report addresses. U.S. officials have previously discussed bringing overseas cloud rentals under export-control rules; if that happens, what becomes of money already paid and compute not yet delivered would depend on contract clauses nobody outside the deal has seen. The report also doesn't include any comment from U.S. regulators on the transaction.
What's left for domestic rivals
Chinese AI labs looking for more compute broadly have three options: run their existing stockpile of Nvidia chips, switch to domestic accelerator cards, or rent capacity overseas. The first two both have ceilings — existing chips keep aging, and domestic cards are still catching up on software stacks and cluster scale.
Tencent has now pushed the third option to the 100,000-chip level. For peers like ByteDance and Alibaba, which are training large models of their own, the ceiling on overseas rental has effectively been reset, and more companies following suit wouldn't be surprising. The trade-off is just as clear: the compute sits in someone else's data center, under another country's laws, with five years of policy risk to manage.
For developers and everyday users, nothing changes immediately. Whether this accelerates the training pace of Tencent's Hunyuan model family is something that will only show up when Tencent next ships a model and the specs and benchmarks can be compared.
Sources: Financial Times (UK), Reuters, CocoLoop; chip count, contract value, lease term and prepayment share follow the reporting, and the per-chip and per-hour costs are rough estimates based on those figures.