NextFin

Oracle Scraps $7 Billion Tencent Chip Lease as Washington Tightens the Nozzle on AI Compute

Summarized by NextFin AI
  • Oracle canceled a roughly $7 billion, five-year lease to provide Tencent with about 100,000 AI chips in Southeast Asia after US officials raised national-security concerns about the offshore compute arrangement.
  • The reversal follows a $20 billion stock-and-bond raise priced at an 8 percent equity discount and bond spreads above double-B averages, signaling investors are demanding a premium for Oracle's AI buildout risk.
  • US regulators have narrowed the offshore-access corridor, with BIS guidance, probes, and a draft rule targeting remote-access loopholes, while Nvidia told partners to sell at least 51 percent of services to non-Chinese clients.
  • Oracle's $523 billion backlog grew $68 billion in Q2, but losing the Tencent deal weakens the narrative that contracted demand is fully bankable, raising the cost of capital for future data-center deployment.

NextFin News - Oracle has canceled a roughly $7 billion, five-year agreement to lease about 100,000 artificial-intelligence chips to Tencent Holdings, abandoning what was reported as the Chinese technology giant's largest-ever overseas compute deal after US officials raised concerns about the arrangement. The reversal comes as Oracle is completing a $20 billion stock-and-bond sale to finance the same AI data-center buildout the Tencent contract was meant to fill — and the terms of that financing show investors are already demanding a premium for the risk.

The cancellation marks a sharp about-face from early October, when the lease was unveiled as a landmark structure: 100,000 advanced accelerators housed in Oracle data centers across Southeast Asia, with Tencent paying an estimated 30 percent of the roughly $7 billion value upfront. The hardware would never enter China; only the compute would flow. US officials told Oracle of their concerns about the Tencent arrangement, and the company walked away. Tencent did not respond to requests for comment.

On the day the deal was first reported, Oracle shares rose nearly 2 percent in premarket trading. Four weeks later, the same arrangement is dead, and the company is selling new shares at $116 — about 8 percent below where the stock closed before the financing announcement. The distance between those two prices is the market's verdict on how much geopolitical risk belongs in Oracle's AI story.

The Four-Week Reversal: From Template to Casualty

The timeline itself carries the news. On October 1, the lease was presented as a template for how American cloud providers could monetize AI capacity without directly exporting controlled chips to China. For a Chinese firm barred from buying the most advanced US-designed accelerators, renting the output from hardware parked in Singapore or Malaysia was a clean workaround. For Oracle, it was a guaranteed, upfront-heavy revenue stream from a customer with deep pockets and a pressing shortage of compute.

By November 4, the deal had been scrapped. In between, the US Commerce Department's Bureau of Industry and Security moved from rule-drafting to enforcement. BIS issued licensing guidance in June tightening know-your-customer screening for China-destined advanced computing items, opened probes into Chinese firms' offshore access to Nvidia chips in August, and drafted a rule to close what officials call the remote-access loophole. The Tencent cancellation is the first reported casualty of that campaign — and it is a deal that was not clearly illegal under the rules as written.

That distinction is the point. If Washington can unwind a signed, partially prepaid contract sitting in the gray zone of existing regulations, it can unwind any of them. The message to the rest of the industry is not that one tenant failed screening; it is that the corridor itself has narrowed, and that signed paper is no longer a reliable proxy for bankable revenue.

The pressure has not been limited to Oracle. On November 3, Nvidia told its data-center partners to scale back work with Chinese customers, requiring cloud clients to sell at least 51 percent of their services to non-Chinese companies. That rule applies only to future contracts and not existing agreements, but it draws the same line in the same sand: access to American-designed AI compute by Chinese firms, however structured, is now treated as a national-security exposure rather than a commercial transaction. Two of the sector's most important gatekeepers moved within 24 hours. That is coordination, not coincidence.

The $20 Billion Raise Was Priced at a Premium

Two days before the cancellation, Oracle priced concurrent offerings of $10 billion in common stock and $10 billion in senior notes to fund data-center capacity for customers including OpenAI, Meta, and Nvidia. The equity piece — 86,206,897 shares at $116.00 each, for net proceeds of approximately $9.9 billion — was priced at a discount to the market. The debt carried coupons ranging from 7.000 percent on the 2029 notes to 8.650 percent on the 2046 bonds, split across five tranches due between 2029 and 2046. The five-year tranche came with a spread of roughly 275 basis points over Treasurys, above the average spread available on double-B rated bonds of similar maturity.

Both offerings were expected to close on November 4, the same day the cancellation broke. The sequencing is unforgiving: investors priced the risk of Oracle's AI buildout before they knew the Tencent deal was gone, and the terms they demanded — an 8 percent equity discount, spreads above double-B averages — already embedded skepticism. The cancellation does not change the price Oracle got; it changes the price the market will demand next time.

Oracle's capital math has been the central tension of its AI buildout. In its second fiscal quarter, the company reported remaining performance obligations — its contracted backlog — of $523 billion, up 15 percent sequentially and $68 billion in a single quarter. Capital expenditure for that quarter was $12 billion, and full-year fiscal 2026 capex is now expected to reach about $50 billion, roughly $15 billion above the prior forecast. Management has said it expects spending to peak in fiscal 2027 and 2028 before the company can finance growth internally from fiscal 2029.

"Remaining Performance Obligations (RPO) increased by $68 billion in Q2—up 15% sequentially to $523 billion—highlighted by new commitments from Meta, NVIDIA, and others," said Doug Kehring, Oracle's principal financial officer.

Backlog is not cash. Converting $523 billion of contracted demand into revenue requires building data centers, buying chips, and securing power at a pace no cloud provider has ever sustained. The Tencent lease was one piece of that conversion engine. Losing it does not break the backlog, but it removes one of the cleaner stories management could tell investors: that demand is global, diversified, and prepaid. Instead, Oracle is funding the buildout by selling equity cheap and debt dear — precisely the financing mix that dilutes existing holders and raises the hurdle rate on every data center it builds.

The market reaction on the cancellation day reflected active repositioning rather than passive drift. Oracle shares opened at $122, down from a prior close of $126.19, and traded in a $121.35 to $128.14 range on volume of about 84 million shares — nearly triple the three-month daily average. The stock is down roughly 49 percent over the past year from its peak, and the average analyst price target of around $307 now sits far above the current level, a gap that measures how much of the AI optimism has already been repriced.

The financing also lands on a balance sheet that has been busy. Oracle has raised debt repeatedly to fund the buildout, with borrowings spanning multiple maturities and a mix of investment-grade senior notes. The company has emphasized preserving its investment-grade rating while prioritizing revenue-generating equipment over owned real estate and power. That strategy works only if the revenue on the other side of the equipment arrives as contracted. A canceled $7 billion contract does not threaten the rating by itself, but it is a data point in the direction investors were already leaning.

Why Offshore Leasing Was Always the Next Target

The structure of the Tencent deal exposed the gap in US export controls. Chips that cannot be sold to China can, in principle, be leased to a Chinese company if they sit in a data center in a third country. The hardware never crosses the Chinese border; only the compute does. For more than a year, that has been the industry's workaround, and it is precisely the channel the Commerce Department has been moving to shut.

This is a structural constraint, not a cyclical one. A cyclical problem — a customer delaying an order, a quarter of soft demand — reverses on its own when conditions improve. A structural problem changes the rules of what is sellable. Once Washington defines offshore access by Chinese tenants as a national-security exposure, the revenue from that channel does not come back when sentiment improves. It is permanently reclassified as off-limits.

The evidence for the structural read is in the sequence: licensing guidance in June, BIS probes in August, a draft rule in the fall, the Tencent cancellation on November 4, and Nvidia's 51 percent rule on November 3. Each step narrows the corridor. The administration is not waiting for the rule to be final before enforcing its intent. The Tencent deal was reported as Tencent's largest-ever overseas compute arrangement; if the largest deal in the category cannot survive, the smaller ones are not safe either.

The second-order effect runs through the cost of capital, not just the revenue line. Oracle is not funding this buildout from cash flow today. It is selling equity and debt into a market that is questioning the pace of AI spending across the sector. Every canceled contract weakens the story that backlog equals bankable revenue. If Washington can unwind a signed, upfront-paid lease, investors will discount the quality of the remaining backlog — particularly the portion tied to Chinese-adjacent demand. The risk is not idle capacity; it is a higher cost of capital at the exact moment Oracle needs $20 billion.

There is also a third-order channel through the peer group. If the offshore-lease model is closed for Chinese tenants, the addressable market for every US cloud provider with offshore capacity shrinks — not because demand disappeared, but because the customers who want it most can no longer be served. That compresses the revenue per watt that data centers can earn, which in turn lowers the return on the capital being raised today. The Tencent cancellation is one contract. The precedent it sets prices every future contract.

The Counter-Thesis: Demand Is Still Infinite

The strongest case against reading this as a crisis for Oracle is simple: demand for AI compute still far exceeds supply, and Tencent is one customer among many. Oracle's backlog grew $68 billion in a single quarter. If one Chinese tenant drops out, another hyperscaler or sovereign customer takes the capacity. The chips were never going to China; they were staying in Southeast Asia, where US rules still permit them to operate. Under that view, the cancellation is a speed bump, not a detour.

There is truth in that. Oracle's $523 billion RPO and cloud infrastructure revenue growth of 84 percent year over year in the third quarter show demand that is not dependent on any single counterparty. And the company has been explicit that it expects capital spending to peak in fiscal 2027 and 2028 before it can finance growth internally from fiscal 2029 — a roadmap that assumes the backlog converts, not that any one deal holds.

But the counter-thesis misses the financing channel. The equity offering was priced at $116, a discount of roughly 8 percent to the prior close. The bonds were priced at spreads above comparable double-B credits. That is not the pricing of a company whose backlog quality is unquestioned; it is the pricing of a company investors want to be paid more to fund. If the market truly believed the Tencent loss was idiosyncratic, Oracle would not have had to sweeten the terms. The discount and the spread are the market's own judgment, stated in dollars.

Nor does the demand argument address the marginal buyer. The backlog is real, but the next increment of capacity requires the next increment of capital — and that capital is being raised now, into a market that has just seen Washington void a signed contract. A backlog that was bankable at a 4 percent cost of debt may not be bankable at 8 percent. The bull case requires the backlog to convert at the margins Oracle modeled. The cancellation is evidence that at least some of those margins were exposed to a risk the model did not price.

The signal that would prove this read wrong is concrete: if Oracle closes the $20 billion offering on November 4 at the announced terms with strong oversubscription, and simultaneously replaces the Tencent capacity with a named non-Chinese customer at similar or better terms, then the market has decided the cancellation is a one-off. If the equity is upsized, the debt is delayed, or either tranche is re-priced richer, the structural read is confirmed.

What Comes Next: Three Scenarios

In the short term, the stock carries event risk into the closing of the offerings. Volume running at nearly triple the average signals active repositioning, and the intraday range — from $121.35 to $128.14 — shows both sellers and buyers engaged. The base case is continued volatility until the $20 billion raise settles and investors see whether the canceled capacity is re-let. Oracle's next earnings report is due December 14, and management will be pressed on whether the backlog remains intact and whether any other China-adjacent contracts are at risk.

Medium term, the terms of the financing will tell investors more about the market's appetite for Oracle's AI bet than any single contract. A clean close keeps the fiscal 2027–2028 capex peak on track. A difficult close — upsizing, delay, or richer pricing — forces either slower deployment or more dilutive financing later. The trading spreads on the new notes, once they begin seasoning, will be the clearest read on credit-market confidence; a spread that widens materially beyond the 275 basis points priced on the five-year tranche would signal that bond investors are repricing the buildout risk.

Long term, the question is whether the offshore-lease channel is dead for all Chinese tenants, or whether the Tencent deal was singled out. If other reported arrangements — similar leases to Chinese firms in third countries — start unwinding in the coming months, the structural constraint is confirmed across the sector, and Oracle, Nvidia, and every US cloud provider with offshore capacity will be pricing a smaller addressable market. If the Tencent deal stands alone, the damage is contained to one contract and one quarter of noise.

There is a fourth variable that sits across all three horizons: the policy rule itself. The Commerce Department's draft to close the remote-access loophole has not been finalized. If the final rule carves out existing contracts or third-country hardware already deployed, the damage is limited to deals not yet signed. If it reaches back into existing arrangements, the exposure is larger than one $7 billion contract. The text of that rule, when it lands, will matter more than any analyst note.

Oracle built its AI narrative on the idea that demand is infinite and borders are porous. The Tencent cancellation says otherwise: in the AI chip trade, borders are the whole point. The company can still fill its data centers. It just has to do it with customers Washington approves — and at a price the bond market is willing to fund.

Explore more exclusive insights at nextfin.ai.

Insights

Why did Oracle cancel Tencent chip lease?

How does US regulate offshore AI compute?

What is remote access loophole in chips?

How much was Oracle Tencent deal worth?

Where were leased Tencent chips located?

What rules did BIS issue recently?

How did Nvidia limit Chinese cloud work?

Why did Oracle stock drop after deal?

How much capital is Oracle raising now?

What are Oracle bond coupon rates?

How large is Oracle current backlog?

Is offshore leasing model now dead?

What risks do investors see in Oracle?

How does policy affect AI chip market?

Can Oracle replace Tencent capacity fast?

What if export rules tighten more?

Who are Oracle main cloud rivals?

Does backlog equal bankable revenue now?

What is Oracle fiscal capex forecast?

How will final Commerce rule impact deals?

Search, ask or analyze anything...