- The U.S. has put licensing conditions on AI computing rented to Chinese companies. An overseas data center providing more than roughly 760 H100-equivalent GPUs of processing to a Chinese-headquartered company must secure a license before receiving covered U.S. chips. Existing providers have 90 days to obtain one.
- Nvidia has already assumed zero direct China data-center compute revenue in its outlook. Even so, Q4 guidance is 12% above Q3 revenue. The new exposure is Chinese demand served through overseas facilities, which Kiwoom estimates at $16.3bn a year.
- With chips in short supply, other customers could fill any openings. Kiwoom puts unmet demand at 2.8–6.5 times the potential loss. On that assumption, the first casualty is not current revenue but the headroom to raise future guidance.
- Jefferies was reported to see a potential $44bn–$62bn hit to next year's revenue. Its methodology differs, so the two estimates cannot be compared directly. Whether replacement demand materializes is decisive.
- The burden may fall harder on data centers than on chips. Chip orders can be redirected. A data center built around Chinese tenants cannot be relocated so easily. For Nvidia's outlook, the pace of U.S. data-center completion may matter more than China itself.
Hover over dotted-underlined terms (tap on mobile) for definitions. Magenta underlines mark the key concepts in this report.
On Nov. 16, U.S. time, the Commerce Department's Bureau of Industry and Security (BIS) announced new export-control rules. The licensing requirement now reaches U.S. AI chips supplied to overseas data centers serving Chinese companies. Nvidia fell 4.1% that day and the Philadelphia Semiconductor Index lost 5.3%.1 But the selloff also followed CEO Jensen Huang's departure, CFO Colette Kress's appointment as interim CEO and a broader decline in AI stocks. The rule alone cannot explain the market move.2
Nvidia's Q3 results the following day beat expectations. We examined those results in Research No. 14. Anxiety about AI shares persisted: the Philadelphia Semiconductor Index fell 10.35% over the five sessions through Nov. 19,3 and South Korea's KOSPI plunged 5.21% to 4,869.96 on Nov. 20.4
Selling accelerated on Nov. 20, U.S. time. Nvidia was already down more than 6% when Reuters reported during the session that the Securities and Exchange Commission (SEC) was investigating its accounting. The stock ended 14.0% lower; the Philadelphia Semiconductor Index closed down 10.7%.5, 6 The scope of the investigation has not been disclosed, and an investigation does not establish wrongdoing.
That raises the central question: will restrictions on Chinese customers reduce Nvidia's sales? The answer also matters to Korean chipmakers supplying Nvidia with high-bandwidth memory (HBM), the fast memory mounted beside AI processors. Kiwoom's thesis is that the bigger hit is to headroom, not revenue. That headroom is demand waiting for additional chip supply—potential sales that could support future guidance upgrades. The scenario assumes licenses are denied and chips remain supply-constrained.
Not just chip sales:
access to computing power
Unable to buy many U.S. chips directly, Chinese companies have been renting computing capacity in overseas data centers through the internet—known as remote access. The U.S. think tank Institute for Progress (IFP) estimates that 23% of AI chips used by China could be accessed this way by year-end.7 The new rule targets the route by which chips enter those data centers, rather than blocking remote computing services directly.8
Two separate license triggers:
either one is sufficient
Where U.S. chips enter data centers
Sources: Federal Register; Morgan Lewis8, 9
The Rule at a Glance
| Provision | Requirement |
|---|---|
| A · Overseas data centers serving Chinese-headquartered firms above the threshold | |
| Controlled transaction | A license to send covered U.S. advanced chips to those data centersThe restriction acts on chip transfers, not the remote service itself |
| Customer covered | Entities headquartered, or ultimately parented, in countries of concern such as China and MacauForeign subsidiaries are covered when the ultimate parent is Chinese-headquartered |
| Compute threshold | More than 12 million TPP of aggregate processing capacityRoughly 760 H100s · Connected facilities are aggregated · Applies anywhere outside the U.S. |
| Existing providers | Must obtain a license within 90 daysFiling an application is not enough; absent a license, subsequent chip exports face restrictions |
| License review | Case-by-case reviewApproval rates remain unknown |
| Exception | Computing used solely to recommend products or contentGenerative AI systems that create new content do not qualify |
| B · Chip shipments to five countries | |
| Countries | Indonesia, Malaysia, Singapore, Thailand and BrazilLicense required regardless of the compute provided |
| Exception | The chip recipient or end user (or its ultimate parent) is U.S.-headquartered and complies with monthly reportingHaving a U.S. company among the data center's customers does not itself qualify |
| Provisions common to both | |
| Timing | Effective Nov. 19, 2026 · Comments due Jan. 19, 2027An interim final rule takes effect before comments close · Eligible shipments already in transit have until Dec. 21, 2026 |
Sources: Federal Register; Morgan Lewis; Greenberg Traurig8, 9, 10 · For detailed conditions, expand “More on the Rule” below
How Large Is 12 Million TPP?
- About 760 H100 GPUsA single Nvidia H100, from the earlier Hopper generation, carries roughly 15,800 TPP.7, 9
- Three racks of the latest chipsThree GB300 NVL72 racks, each containing 72 Blackwell GPUs, would cross the threshold (Morgan Lewis).
- Most large generative-AI training clustersThe threshold is equivalent to about 37% of the 2,048 GPUs DeepSeek used to train one model (Kiwoom estimate).1
Direct China sales: already zero.
New exposure: China renting abroad
There are two distinct China exposures. Nvidia's forecast already assumes zero revenue from direct China data-center chip sales. What the new rule adds is Chinese demand for AI compute rented at overseas facilities. Mixing the two would double-count the regulatory impact.
Nvidia sold $2.4bn of licensed, older-generation Hopper chips in fiscal Q3 (Aug.–Oct.). Shipments are now suspended, and those sales are excluded from Q4 guidance.11, 12
Annual demand for Nvidia equipment that could be lost if licenses are denied. The calculation follows.
Nvidia Raised Guidance Despite Excluding China
| Metric | Value | What it means |
|---|---|---|
| Q4 revenue guidance (Nov.–Jan.) | $130bn ±3% | Assumes zero data-center compute revenue from China13Nvidia did not specify whether the China-related share of overseas data centers was also excluded |
| Q3 revenue (Aug.–Oct.) | $115.8bn | Q4 guidance is 12% above Q3 revenueQ4 has 14 weeks, one more than Q3; the per-week increase is 4% |
| Next fiscal year's (FY28) revenue growth outlook | ~70% → ~75% | Raised in November from the initial estimate issued in August12 |
How Kiwoom Gets to $16.3bn a Year
- 0.5GW × $32.5bn per GWThe model assumes overseas data centers serving Chinese customers add 0.5GW of capacity each year, with $32.5bn of Nvidia equipment per gigawatt.1
- Basis for the 0.5GW estimateDayOne, the largest data-center operator in Johor, Malaysia, expanded operational capacity at an annualized rate of roughly 0.7GW. Kiwoom multiplied that by the 69.2% share of revenue from its largest customer.14
- A proxy, not a measured figureThe model extrapolates from one operator to the broader overseas market and applies a revenue concentration ratio directly to added capacity. It also assumes the new power capacity would all be filled with Nvidia equipment. DayOne sells data-center space, power and cooling—not GPUs. The result is a sizing exercise, not an observed measure of demand subject to the rule.
Why $16.3bn Could Be Too High or Too Low
| Direction | Basis | Implication |
|---|---|---|
| Could be higher | Broader scenario: annualizing fiscal Q3 revenue from customers directly billed by Nvidia and headquartered in China or Hong Kong ($11.05bn), after subtracting mainland China sales (Kiwoom)1Geographic revenue follows the headquarters of direct customers, which may differ from the actual exposure under this rule15 | $34.6bn/yr |
| Could be higher | Country-list exposure (B): The $16.3bn estimate covers Chinese customers renting compute (A). In the five named countries, chip imports by data centers unrelated to China can also require licenses. | Delays in licensing could widen the impactAmount unknown |
| Could be lower | DayOne's largest customer: The Information cited sources saying the customer's work did not involve AI training or inference, and that its share of 2027 contracted capacity would fall below 50%.16Not confirmed by the company | The share actually covered by trigger A could be smaller |
Potential loss: $16.3bn.
Unmet demand: 2.8–6.5× as large
When chips are plentiful, losing one customer means losing its revenue. With supply constrained, another customer may take the capacity that opens up, according to Kiwoom's model.1
Evidence That Chips Remain Scarce
Two pieces of evidence underpin the argument. First, Nvidia's FY28 outlook appears calibrated to how much it can produce. It has committed to buy $203bn of components and production capacity for FY28 through purchase commitments. If cost of revenue is 27.5% of sales, that spending equates to around $738.2bn of potential revenue—almost identical to the $741.4bn implied by a 75% growth forecast.1, 11
| Measure | Calculation | Result |
|---|---|---|
| Company outlook, converted to dollars | Estimated FY27 sales of $423.7bn × 1.75 (75% growth) | $741.4bn |
| Implied by purchase commitments | $203bn ÷ 27.5% cost-of-revenue ratio | $738.2bn |
| Difference | Gap between the two estimates | 0.4% |
Sources: Nvidia; Kiwoom Securities calculations11 · 27.5% cost ratio = 100% minus the midpoint of next year's projected gross margin (72%–73%). Treating commitments as costs is a simplifying assumption, not proof of a hard revenue ceiling.
Second, signs of constrained supply appear across the industry.
| Indicator | Evidence | Source |
|---|---|---|
| Company commentary | Demand substantially exceeds supply; Nvidia still expects strong results despite new export controls (Nov. 17 earnings call) | Nvidia12 |
| Rental prices | The cost of renting Blackwell B300 chips has risen more than 60% since launch | Kiwoom Securities1 |
| Contracted backlog | At major cloud providers (hyperscalers), contracted backlog rose 142% year over year in Q3 | IBK Securities17 |
| Manufacturing bottleneck | TSMC's advanced packaging process (CoWoS) remains constrained despite capacity expansion | Yuanta Securities18 |
Demand waiting for chips
exceeds demand at risk
Annualized · Dashed segments show the range · All figures are Kiwoom estimates
Source: Kiwoom Securities calculations1 · See “Methodology” below for the formulas
- Unmet demand: $46.2bn–$105.9bnIn August, Nvidia said its customers' combined plans pointed to FY28 revenue roughly double FY27 levels. Kiwoom subtracts the company's guidance from that implied demand. These are not firm orders: construction schedules and financing could prevent them from materializing.
- Broader $34.6bn case: 1.3–3.1×The cushion narrows sharply, though even the low end of unmet demand still exceeds the modeled loss.
What Shrinks Is the Upside
If another buyer takes the freed slot, Nvidia's revenue holds—but the waiting list gets shorter. That unmet demand is the potential for further guidance upgrades when supply expands. Restrictions eat into that reserve.1
Put in growth terms, current guidance calls for 75% sales growth next year. If all unmet demand converted into revenue, the rate could theoretically reach 86%–100%. After deducting the modeled exposure, that range falls to 82%–96%. The chart measures the percentage-point upside above 75%.
Additional growth headroom:
+11–25pp → +7–21pp
Unmet demand ÷ estimated FY27 sales of $423.7bn · Bar endpoints show low and high cases
Source: Kiwoom Securities calculations1
- Potential demand at risk equals 77% of November's guidance upgradeNvidia raised its FY28 growth outlook from 70% to 75%, an implied increase of about $21.2bn. If freed capacity cannot be reallocated, the modeled loss could erase most of that upgrade (Kiwoom estimate).
For the stock, that distinction matters. Nvidia's valuation has benefited from earnings beats followed by higher guidance. Even if near-term sales hold steady, a smaller reserve of potential upside could unsettle the expectations priced into the shares.
Jefferies: Revenue Could Fall
Jefferies was reported to estimate that the remote-access rule could reduce Nvidia's data-center revenue by $44bn–$62bn next year.19, 20 We could not obtain the original research note and therefore could not verify its assumptions. Some reporting suggests the figure includes both direct restrictions and knock-on effects.21
| Question | Kiwoom Securities | Jefferies |
|---|---|---|
| What is measured? | Potential demand | Potential revenue loss |
| Scale | $16.3bn/yr ($34.6bn broader case) | $44bn–$62bn next year |
| Coverage | Chinese customers exposed to remote-access controls | Effects of the new rule (as reported)Components unverified |
| Replacement customers | Assumes others take the freed capacity | Not established by available reporting |
The pivotal question separating the estimates is whether other buyers can actually absorb the displaced orders. Yet the scope of each estimate—including possible licensing delays in the five countries—and the way revenue is counted may also differ. Replacement demand is easier to find while chips are scarce; once supply catches up, an empty slot can translate directly into lost sales.
Chips can find a new buyer.
Buildings stay put.
Chip allocations can be redirected to customers with sufficient power and cooling. A data center developed for a particular tenant cannot readily change location or contract terms. That is why overseas facilities reliant on Chinese customers could feel the pressure first.1
On the day of the rule,
data-center exposure took a heavier hit
U.S. close, Nov. 16 · The day before Nvidia's earnings release
Sources: Bloomberg; Kiwoom Securities1 · AI stocks sold off broadly that day, so the performance gap cannot be attributed entirely to the rule
- What the divergence suggestsKiwoom reads Nvidia's decline as pricing in demand that could be replaced, while GDS's larger loss reflected the prospect of finding new tenants for hard-to-repurpose data centers.
- GDS holds an equity stake in DayOneDayOne, the largest data-center operator in Johor, Malaysia, was carved out of Chinese data-center firm GDS's overseas business. Kiwoom identifies GDS as a U.S.-listed company whose disclosures offer a way to gauge exposure to the new rule.
- One customer generates 69.2% of DayOne's revenueIts F-1 filing describes the customer only as a technology company operating a globally popular short-form video platform.14 Media reports have identified it as ByteDance.22 Across all DayOne markets—Southeast Asia, Japan and Europe—99.3% of its 1,328MW under construction was already contracted.14 Reporting that this customer's workload was not AI-related is discussed in the Section 02 table.
- Its IPO is on hold, tooBloomberg reported on Nov. 20 that DayOne had paused plans for a U.S. listing that could raise as much as $5bn. The report cited investor concerns over digital-infrastructure valuations and said the company might revisit the IPO in 2027.23 All four Southeast Asian countries where DayOne operates are included in trigger B.
Contracts Are Also at Risk in Southeast Asia
Malaysia is one of the five countries subject to trigger B. Data centers there need a license to receive covered chips regardless of compute size, unless they qualify for the U.S.-headquartered exception described in Section 01.8 There is already evidence of contract disruption: reports in early November said Oracle had canceled a roughly $7bn, five-year agreement to rent up to 100,000 AI chips in Southeast Asian data centers to China's Tencent after U.S. government concerns emerged.24 That preceded the new rule, so the cancellation cannot be ascribed to it.
The outlook turns when
data centers cannot take delivery
Kiwoom argues the regulation alone is unlikely to force a cut in Nvidia's revenue forecast.1 Instead, it identifies three conditions to watch. Nvidia has also said its next-year outlook takes into account data-center readiness, end-customer demand and customer financing.12
| Condition | Warning threshold | Current reading |
|---|---|---|
| ① FacilitiesCompletion delays | More than another 0.9GW of delayed data-center completions among waiting customers within one year (base case)Without operating facilities, customers cannot take delivery even when they want the chips | Texas grid-connection queue averaged across one year: roughly 1.3GWAn illustration of potential exposure—not measured delays · See chart below |
| ② DemandNext earnings | At the late-February results, FY28 growth guidance drops below 75%, or the company does not reiterate that customer plans imply demand for twice this year's revenue | Guidance was raised in November (Section 02)The “twice this year” statement was not repeated |
| ③ FundingCustomer finances | If days sales outstanding (DSO) and customer-related guarantees both increase again next quarter | DSO: 60 → 64 daysMaximum guarantees: $108.5bn → $118bn11 |
Sources: Kiwoom Securities; Nvidia CFO commentary1 · Kiwoom also flags the risk that Chinese export curbs on critical materials could reduce chip production itself, a supply shock that waiting buyers cannot offset
If all Texas projects in the connection queue slip,
can waiting demand absorb the delay?
Gigawatts · Worst-case stress test, not an observed delay
Source: Kiwoom Securities, based on ERCOT data1, 25 · Assumes an entire year's queue is delayed, all capacity represents Nvidia-equipment demand and none is already reflected in guidance · $13bn equals 1.8% of the FY28 revenue outlook
- The threshold falls sharply in the broader caseSubstituting Kiwoom's broader $34.6bn risk case leaves just $11.6bn of unmet demand at the low end, or the equivalent of 0.36GW of project delays. That is less than one-third of the roughly 1.3GW annual Texas queue, meaning a much smaller delay could put guidance at risk.
- 6.6GW is queued capacity, not an actual delayTexas grid operator ERCOT has finished its reviews of 17 large-load projects with up to 6.6GW of demand, but grid-connection approval is still pending. The capacity is expected to be phased in over roughly five years; ERCOT is targeting Dec. 10 for its audit report.1, 25 Texas accounts for 32% of data-center capacity under construction in the U.S., according to Kiwoom (see NO.27).
- There are two separate permitting processesThe process that restarted around Nov. 17 involved the Texas Commission on Environmental Quality (TCEQ), not ERCOT grid-connection approval. The incoming state administration, due in January, has pledged to pause data-center approvals until related legislation is enacted.26
Look at the Structure of Customer Financing (Condition ③)
- Why receivables are taking longer to collectNvidia attributes the increase to uneven shipments of its new Vera Rubin chips during the quarter and longer payment terms offered to some highly rated customers.11
- $118bn in guarantees is the maximum exposureIt is not debt due today, but a potential obligation if specified conditions are met. Most of it—$105bn—relates to one Ohio data-center campus and takes effect only after the buildings are completed. The first guarantee is expected to become effective in FY29 (Feb. 2028–Jan. 2029). No payments have been made so far.11 For the broader financing structure, see Research No. 12.
Implications for Korea's Chip Industry
If Kiwoom's supply-shortage assumption holds, the Korean semiconductor exposure needs to be split in two.
As long as chips are scarce, Nvidia can shift capacity from Chinese customers to other buyers. For HBM suppliers, the more immediate variable is whether U.S. data centers are completed on schedule—not the China rule alone.
The rule also tightens due diligence for foundries and packaging firms making advanced chips for Chinese customers.9 Taiwanese industry publications say TSMC, Samsung Electronics and others could defer orders from Chinese fabless chip designers until U.S. licenses are granted.27
We covered when new memory fabs will begin production in NO.34, and Korea's AI-export exposure in NO.36.
Dates to Watch
- DEC. 10 · EARLY KST
Korea timeFOMC decisionMeeting Dec. 8–9, U.S. time - DEC. 10 · U.S.ERCOT audit report (target)Next step for 17 large-load projects awaiting grid connection
- JAN. 19, 2027 · U.S.Rule comment deadline · New Texas administration takes officeThe incoming administration has pledged a pause in data-center approvals
- AROUND FEB. 17, 2027 · U.S.90-day license deadlineCounting from the Nov. 19 effective date
- LATE FEB. 2027Nvidia fiscal Q4 earnings (expected)Next opportunity to update FY28 guidance
Sources: Federal Register; Utility Dive; FOX 7 Austin8, 25, 26 · Dates may change · Daol Investment & Securities suggests waiting for the early-December Texas audit update and FOMC decision before becoming more constructive28