Industry & CompetitionGovernance & ComplianceMacro & Geopolitics

NVIDIA's First Post-Huang Earnings: Why It Halted China Shipments and When Its $118B Guarantees Actually Take Effect

After the US market closed on November 17, 2026—five days after founder Jensen Huang resigned—NVIDIA released its third-quarter fiscal 2027 results. By the headline numbers, the compute giant is still sprinting ahead: quarterly revenue reached $115.8 billion, doubling year over year, with next-quarter guidance set at $130 billion.

Yet on the very same day, Chief Financial Officer Colette Kress—newly installed as interim CEO—and NVIDIA’s board delivered two major new actions alongside one pivotal disclosure. First, even though Washington has not revoked its export licenses, NVIDIA voluntarily announced it will stop shipping all US-approved Hopper data center chips to China and wrote zero China data center compute revenue into its official Q4 guidance. Second, despite a petition with nearly 200,000 signatures demanding Huang’s return and Huang’s own public offer to come back, Chairman Harvey Jones opened the earnings call by flatly rejecting his reinstatement while announcing a standing board National Security and Export Control Oversight Committee chaired by a former Goldman Sachs vice chair, pulling compliance reporting lines away from business executives. Third, in the CFO Commentary filing submitted to the SEC and during the earnings call, Kress laid out for the first time the existing effectiveness boundaries governing NVIDIA’s $668 billion in commitments and guarantees, directly addressing investor fears that NVIDIA is backstopping construction-phase risk. In the same earnings report, shipment and governance decisions genuinely changed, while investors finally saw when NVIDIA’s hundred-billion-dollar guarantees actually take effect. Why did the new leadership shut its own China front door before Washington pulled the licenses—and under what conditions do those guarantees actually trigger?

Washington Hasn’t Revoked the Licenses—Why Did NVIDIA Shut Its Own $2.4B-a-Quarter China Front Door?

Last quarter, NVIDIA’s US-licensed Hopper data center chips for China—primarily the H200—still generated $2.4 billion in revenue. On the Q3 earnings call, Kress stated plainly: “Our guidance does not have any licensed Hopper data center products included, and we are no longer shipping licensed Hopper data center products to China.” The company did not disclose why it halted shipments or whether the move is permanent, but waiting for export-control relief is no longer part of NVIDIA’s operating assumptions.

A public company walks away from a legally approved revenue stream first because the money barely moves the needle anymore. That $2.4 billion accounted for just 2.2% of NVIDIA’s data center revenue last quarter. Meanwhile, NVIDIA’s next-generation flagship architecture, Vera Rubin, shipped $20 billion in its very first ramp quarter, backed by more than $2.5 trillion in non-cancelable cloud backlog from major hyperscalers. With TSMC wafer and advanced packaging capacity still tight, any capacity previously earmarked for China can be reallocated immediately to Western cloud giants waiting in line. Cutting off the China front door did not dent the outlook—Kress actually raised full-year fiscal 2028 revenue growth guidance to 75%.

More importantly, that tiny 2.2% channel had turned into a ticking liability on both physical delivery and US legal fronts. On the delivery side, a November Reuters investigation found that domestic import quotas and mandates to pair foreign GPUs with domestic chips left roughly 600,000 US-approved H200 chips stranded in Hong Kong warehouses. With official imports clogged, mainland grey-market prices surged, and Reuters subsequently confirmed that Northwestern Polytechnical University—a Chinese military-linked university long on the US Entity List—acquired H200 chips through a domestic middleman.

In Washington, the legality of the H200 channel also came under fire after Democrats recaptured both chambers of Congress in the midterm elections. The H200 sales had relied on a bespoke arrangement struck between Huang and the Trump administration: Taiwan-manufactured chips were first routed to the United States for security inspection, assessed a 25% tariff under Section 232 of the Trade Expansion Act of 1962, and then licensed for re-export to Chinese buyers. Democratic leaders on the Senate Finance Committee and House Ways and Means Committee recently sent a formal investigative letter to Commerce Secretary Howard Lutnick, labeling the route-through-the-US tariff mechanism an unconstitutional export tax, questioning why military-linked shell companies appeared on the approved buyer list, and probing whether the licenses were tied to more than $26 million in donations from NVIDIA toward White House priorities including the ballroom project.

On the same day NVIDIA reported earnings, the Wall Street Journal revealed that the SEC has opened a civil investigation into NVIDIA over whether its investor disclosures were misleading. That same day, 15 national security experts—including a former CIA Inspector General—wrote to the US Attorney General urging the Justice Department to investigate whether NVIDIA deliberately blinded itself to export-control violations. While no single regulatory action can be isolated as the sole cause of the halt, the calculus for Kress and NVIDIA’s independent directors was straightforward: continuing to ship into Hong Kong warehouses would only hand the SEC and DOJ fresh ammunition. Closing the front door before Washington revokes the licenses was the lowest-cost way to stop the bleeding.

If the Front Door Was Only $2.4B, Where Did the $104B Booked Through Singapore Go?

If NVIDIA’s China-related compute exposure were truly limited to $2.4 billion a quarter in H200 sales, Washington’s regulatory storm would never have escalated into a governance crisis that forced out the founder. What drew regulators to NVIDIA’s disclosures was a far larger offshore channel long routed through Singapore.

On earnings day, the Financial Times unearthed local corporate filings for NVIDIA’s Singapore subsidiary. The filings show that in calendar year 2025, external customer sales booked through the Singapore entity hit $104 billion—nearly half of NVIDIA’s global revenue during that period. Two accounting scopes must be kept strictly separate here: that near-50% share reflects the local subsidiary’s calendar-year statutory books and cannot be conflated with the parent company’s fiscal-year geographic segment disclosures. Even more tellingly, despite booking over $100 billion in annual revenue, the Singapore subsidiary recorded just $53 million in local pre-tax profit, and less than 2% of NVIDIA’s chips physically landed in Singapore. Singapore functioned as an offshore contracting and invoicing hub: Asian customers signed contracts and received invoices there while the hardware shipped to data centers in neighboring countries such as Malaysia.

What alarmed SEC investigators was an accounting disclosure switch NVIDIA made in the summer of 2025. Under its previous reporting method based on customer billing location, Singapore’s share of NVIDIA’s public segment revenue had climbed to 22%. Right during that period, US authorities uncovered how Singapore-based cloud firm Megaspeed was acquiring large volumes of NVIDIA chips and renting them to Chinese AI labs through Malaysian data centers; the Commerce Department queried NVIDIA and worked with Singapore police to raid Megaspeed. Immediately afterward, in July 2025, NVIDIA changed its SEC geographic revenue reporting basis from customer billing location to customer headquarters location—and Singapore, which had represented more than a fifth of reported revenue, vanished from the geographic breakdown.

Comparison of NVIDIA’s front-door H200 halt and the tightening of the $104B Singapore invoicing and Southeast Asia remote-compute side door

Reclassifying billing addresses did not stop Washington from closing the Southeast Asian remote-access side door. Recently, research group Unemployment Labs published a weight-forensics report on open-source AI models, showing that frontier Chinese models including DeepSeek V4 and Moonshot AI’s Kimi K3 carry floating-point hardware fingerprints of training on export-banned Blackwell chips accessed via Southeast Asian clusters. On November 16, the US Commerce Department issued a 117-page interim final rule on remote access to advanced chips. In an unusual disclosure, the official press release explicitly noted that the rule was written by and at the direction of the Treasury Department and the White House National Security Council—publicly sidelining Commerce Secretary Lutnick, who had drawn congressional fire over earlier approvals. The new rule requires US export licenses for shipping advanced AI chips to five countries including Singapore and Malaysia, and gives overseas data centers renting large-scale compute to Chinese firms a 90-day grace period to obtain a US license. Within days, both the H200 front door and the Singapore-invoiced, Southeast Asia remote-rental side door reached the end of the road.

Why Did a 200,000-Signature Petition and Huang’s Offer to Return Meet a Flat “No” From the Board?

Once the regulatory storm over the Singapore disclosure shift and Southeast Asian side door is clear, the board’s refusal to bring Jensen Huang back makes complete sense. After Huang resigned, an external petition gathered nearly 200,000 signatures, including more than 10,000 people identifying themselves as NVIDIA employees, alongside public backing from Elon Musk, Bill Ackman, and President Trump. Ahead of earnings, Huang himself stated that he would be honored to return if the board deemed it appropriate and would cooperate with any independent investigation. Yet Chairman Harvey Jones shut the door at the start of the call: “The NVIDIA board unanimously accepted his resignation and does not intend to reinstate Jensen as CEO.”

Beyond Huang’s personal workplace scandals and the internal-control sign-off crisis, a separate Financial Times and Reuters investigation published on November 17 exposed how deeply the Huang family had breached corporate governance boundaries during the Megaspeed probe. Huang’s daughter, Madison Huang—whose official role was Senior Director of Product Marketing for Physical AI and robotics—privately hired a crisis PR firm (later moved onto NVIDIA’s payroll) after Megaspeed came under US scrutiny, stepped outside her remit to direct NVIDIA’s public defense of Megaspeed, and excluded a senior corporate communications executive from the response group that included the General Counsel. Meanwhile, her husband, former professional skier Nico Caprez, was promoted twice in four months after joining the company, rising to director overseeing Asian international cloud partner deals and then Vice President of Global AI Infrastructure Growth, flying repeatedly to Asia to handle deals involving major client Megaspeed.

When the founder’s daughter and son-in-law were simultaneously handling Asian cloud deals for an investigated client and sidelining corporate communications leaders to steer the public defense, NVIDIA’s internal controls looked hollow in the eyes of the SEC and DOJ. Reinstating Huang under pressure would have signaled to federal investigators that the company refused to reform. On the same day it rejected his return, the board established its standing National Security and Export Control Oversight Committee, appointing former Goldman Sachs Vice Chair Suzanne Nora Johnson as chair and rerouting NVIDIA’s National Security and Global Trade Compliance teams to report directly to the committee and Interim CEO Kress rather than business executives. Only by stripping compliance authority away from business lines and keeping the Huang family out of decision-making could the board erect an institutional firewall ahead of federal probes.

As Commitments and Guarantees Hit $668B, When Do NVIDIA’s Guarantees Actually Take Effect?

While walling off geopolitical and family-governance liabilities, investors were equally focused on how CFO-turned-interim-CEO Kress would handle another financial flashpoint: hundreds of billions of dollars in commitments and credit guarantees supporting NVIDIA’s downstream compute ecosystem. In our earlier analysis of Huang’s resignation, we predicted that once this veteran CFO—who had internally opposed the $100B+ data center lease guarantees—took charge, she would tighten off-balance-sheet backstops and extended payment terms rather than endlessly subsidizing leveraged buyers with cash flow. The Q3 report and earnings call delivered a subtly different answer: Kress did not tear up or rewrite existing contracts, but instead put the pre-existing effectiveness conditions written into those agreements out in the open. Over the past two years, to help neoclouds (such as CoreWeave) and frontier AI labs secure land, power, and data center buildout financing, NVIDIA entered into large-scale compute commitments and credit guarantees.

Looking at the Q3 cash-flow and commitment figures, it is easy to see why Wall Street pressed for clarity. Although NVIDIA posted strong GAAP profits last quarter, free cash flow came in at $43.8 billion—more than $13 billion below Wall Street expectations—while accounts receivable jumped to $81 billion and average collection cycles stretched to 64 days, the slowest pace in four years. Meanwhile, NVIDIA’s combined commitments and guarantees swelled from $530.5 billion in Q2 to $668 billion: $518 billion in direct supply, lease, and investment commitments, $32 billion in neocloud agreements, and $118 billion in maximum gross guarantee exposure.

Breakdown of NVIDIA’s $668B commitments and guarantees and their separate existing effectiveness conditions

Reading the EX-99.2 filing closely, the $118 billion maximum guarantee exposure and the $32 billion neocloud compute commitments follow two distinct paths and should not be conflated. The $118 billion maximum gross guarantee exposure breaks down into three parts: $105 billion in credit support guarantees for SoftBank subsidiary SB Energy’s PORTS-Pike Technology Campus in Ohio (nine buildings totaling roughly 4.25 GW, exclusively hosting NVIDIA infrastructure under 20-year leases to OpenAI), $9.5 billion in power and systems credit support guarantees, and $3.5 billion in land, power, and shell guarantees for neoclouds. For the dominant PORTS-Pike project, the filing specifies that guarantees take effect phase by phase, with each becoming effective only when its related building is completed and ready for service (the first expected in fiscal 2029). Exposure therefore builds over time as buildings come online and then declines over each 20-year lease term as OpenAI makes lease payments; to date, NVIDIA has made no payments related to its guarantees.

The separate $32 billion in neocloud agreements is where the filing spells out financing, deployment, and acceptance contingencies: commitments take effect only if the neoclouds secure financing to purchase the data center infrastructure products and the related compute capacity is deployed, made operational, and accepted by NVIDIA. If those contingencies are not timely resolved, the amount and timing of commitments may be affected, and certain commitments could be reduced or eliminated. On page 12 of the Q3 earnings call transcript, an analyst asked point-blank whether this was a new way of doing business or simply something NVIDIA had been doing all along and was only now spelling out. Kress answered directly: “Is it new? No. Did we change anything? No.” Pointing to PORTS-Pike as an example, she explained that most of NVIDIA’s guarantees only begin when facilities are built and ready to go, adding on the call: “If all of that begins and unfortunately they are delayed or they have a challenge in terms of building that, our guarantee doesn’t exist anymore.” She also reiterated plans to utilize the remaining $227.3 billion share repurchase authorization through fiscal 2028 after funding supply-chain and strategic priorities.

The most failure-prone phase of AI infrastructure is precisely the construction window—securing land, grid interconnects, construction loans, and getting shells built on time. The real news in the Q3 filing is not that new management suddenly rewrote contracts, but that investors finally saw the boundaries that were already built into them: construction financing and buildout risk remained with partners and their lenders, and NVIDIA was never unconditionally backstopping unfinished data centers.

Putting every signal from November 17 together clarifies the shape of NVIDIA’s post-Huang transition. Jensen Huang’s era leaned on founder-led political diplomacy to navigate export controls and broad hundred-billion-dollar ecosystem narratives to fuel market expectations. With Vera Rubin shipping $20 billion in its debut quarter and cloud backlog topping $2.5 trillion, core AI compute remains a seller’s market where NVIDIA does not need to take regulatory risks to sell every chip it makes. So the same earnings report delivered two clear shifts: on one side, a Goldman-veteran committee chair and a career-CFO interim chief actively cut off legally contagious China Hopper shipments and pulled compliance under the board; on the other, detailed SEC footnotes spelled out the pre-existing boundaries of NVIDIA’s hundred-billion-dollar guarantees, easing fears of open-ended construction backstops. The compute empire’s cash machine is still running at full throttle—only the era of crossing geopolitical red lines on personal charisma has given way to institutionalized compliance and far clearer risk boundaries.