中文
IF–THEN · Impact analysis of a clue

Analysis date 2026-09-17

If Nvidia’s plan to export H200 chips to China remains deadlocked due to continued opposition from Chinese authorities and channel restrictions, with actual deliveries still nowhere in sight, how will this affect Nvidia’s related financial performance and China’s progress toward self-sufficiency in AI chips?

The original premise that "H200 deliveries have been delayed indefinitely" has been partially falsified: ByteDance and Tencent each received approximately 10,000 units in initial physical deliveries in August 2026 ([21][26]). The actual situation is now "symbolic small-volume drip-feed + administrative freezing of large quotas," rather than a zero-delivery stalemate.

Even under the revised "persistent small-volume" scenario, the marginal impact on NVIDIA's current-period financials is estimated at no more than 1% of revenue: H200 revenue to China in FY27Q2 was less than 1% of data-center revenue (<$890 million, approximately 0.9% of total revenue) [23][24][31][35]. The company's Q3 guidance explicitly assumes no China data-center compute revenue, and it has already recognized $404 million in charges for excess H200 inventory over the past six months.

The true decision-relevant issue is second- and third-order transmission: NVIDIA's annual report acknowledges that exclusion from China's data-center market "helps competitors build larger developer and customer ecosystems," while Jensen Huang said, "We have effectively given that market to Huawei"—a long-term erosion of the ecosystem moat, not a current-period revenue shock.

The direction of China's AI self-sufficiency was locked in by the April 2025 H20 cutoff. Whether H200 is drip-fed or reduced to zero creates only a marginal difference in acceleration. The real constraints are Huawei Ascend's high-end memory (HBM) and advanced-process capacity, not H200 availability—DeepSeek training still relies on NVIDIA, while Ascend 950DT is used only for inference ([27][40]).

Market pricing: the central scenario (drip-feed/zero delivery) is highly priced into NVIDIA's share price, with China data-center revenue already internalized at zero. Pricing in the domestic computing chain is medium-to-high, with expectations running ahead of orders.

HOW IT FLOWS

Core causal chains

Administrative approval friction → delivery shortfall (primary current-period channel)High confidence
In 2026-01, the US BIS changed H200 treatment from "presumption of denial" to case-by-case review with quantity and security conditions; an executive order simultaneously imposed a 25% ad valorem tariff ([12][13]).In 2026-01, Reuters reported that Chinese customs verbally notified declaration agents that H200 was "not allowed to enter China," while officials instructed technology companies that purchases were "not necessary unless essential."In 2026-05, the US approved approximately 10 Chinese companies, including Alibaba, Tencent, ByteDance and JD.com, with a ceiling of 75,000 units each; Lenovo and Foxconn were distributors, but not a single unit had shipped ([18]).In 2026-08, China allowed limited quantities: approximately 10,000 units each for ByteDance and Tencent, implying an approval utilization rate of approximately 10%–13% ([26]).Beijing required most chips to remain outside mainland China and allowed use in Hong Kong; the remainder required item-by-item approval from the NDRC. Hong Kong was constrained by data-center and power capacity.Result: deliveries remain at symbolic small volumes, creating a drip-feed state of "very limited releases + large-scale quota freezes."
Delivery shortfall → NVIDIA current-period financials (limited impact)High confidence
H200 revenue to China was less than 1% of data-center revenue, or less than $890 million per quarter, approximately 0.9% of total revenue ([23][24][35]).FY27Q2 revenue was $96.22 billion, up 106% year on year; data-center revenue was $89.0 billion, up 117%; mainland China plus Hong Kong revenue was $7.88 billion, approximately 8.2% of total revenue.FY27 Q3 guidance was $108.0 billion and explicitly excluded any China data-center compute revenue; $404 million of excess H200 inventory charges had already been recognized over the prior six months.Hopper-series gross margins are below the company average. A complete loss of H200 revenue would effectively remove low-margin revenue, producing limited or even slightly positive effects on overall gross margin.Result: the NVIDIA current-period financial gap between S1 (drip-feed) and S2 (zero delivery) is smaller than market intuition suggests, involving mainly staged inventory write-downs rather than revenue elasticity.
Delivery shortfall → China's domestic AI-chip self-sufficiency (direction locked in, speed changes only marginally)Medium confidence
Lower H200 availability → higher domestic procurement budgets for cloud providers → greater order visibility for Huawei and Cambricon. Alibaba, Tencent and Baidu sharply increased 2Q26 capex, primarily toward domestic GPUs and memory ([33]).Software-ecosystem migration is becoming substantive: DeepSeek V4 has been deeply adapted to the Ascend 950PR instruction set; Alibaba Cloud Bailian and Tencent Cloud TokenHub went live on the day of release.The physical constraint remains HBM and advanced-process capacity: because of high-end memory shortages, Huawei's 950DT output this year is only in the "hundreds of thousands," and fulfilling DeepSeek's order for 160,000 units may take more than a year ([27]).Beijing's active release of limited H200 units in August 2026 is evidence that domestic supply cannot meet frontier training needs in the short term; DeepSeek training still relies on NVIDIA.Result: the self-sufficiency direction is unchanged—the April 2025 H20 cutoff already locked it in. H200 drip-feed or zero delivery creates only marginal acceleration; the true dividing line is CXMT HBM mass production and the commissioning date of SMIC's Huawei-dedicated new fabs.
Delivery obstruction → long-term erosion of the ecosystem moat (third order, most important but difficult to identify)Medium confidence
NVIDIA's FY2026 annual report acknowledged that it was "effectively excluded from competition in China's data-center compute market" and that this exclusion "helps competitors build larger developer and customer ecosystems, enabling them to challenge us globally."Jensen Huang said in May 2026 that Huawei was extremely strong and that "we have effectively given that market to them."If mainstream open-source models such as DeepSeek V4 treat CANN rather than CUDA as the first-class platform, switching costs for foreign chips will rise irreversibly.If Chinese models run globally with "non-US hardware as the optimal choice," NVIDIA's long-term pricing power will be eroded through the ecosystem.Result: the loss of long-term option value is difficult to price quarterly; it is a second- or third-order risk with insufficient current-period causal identification.
MACRO

Macro impact

MixedUS-China trade and technology policyMedium impact · 6–24 months

H200 is being used as a tradable political asset: the 25% ad valorem tariff plus a "factual tax wedge" from a 25% revenue share raises the landed cost for Chinese buyers, while China exchanges release quotas for US flexibility on chip equipment and EUV-related tools, and the US exchanges licenses for relaxed export restrictions on rare-earth magnetic materials and gallium. Delivery volume has therefore become a thermometer of US-China relations rather than a purely commercial variable.

Supporting evidence

  • The 2026-01-14 executive order imposed a 25% ad valorem tariff on specified advanced AI chips, including H200.
  • The US extracts 25% of sales revenue and requires the chips to physically transit the United States, creating the institutional basis for China's concern that their overseas retention could be altered.

Counter-evidence

  • If the US and China reach a package agreement—rare earths/gallium in exchange for chip equipment and EUV-related tools—H200 quotas could expand at once.
  • There is continuing political resistance in Washington to releasing H200; substantive S3 implementation would itself be unstable.
PositiveChina's technology sovereignty and supply-chain securityMedium impact · 12–36 months

New State Council supply-chain security rules require government systems to investigate dependence on foreign technology in critical infrastructure. H200 procurement has been brought under the "critical-infrastructure dependence" framework and is therefore subject to inherent political constraints, strengthening institutional support for domestic substitution.

Supporting evidence

  • Reuters 2026-05-14: State Council supply-chain security rules promote investigation of foreign-technology dependence in critical infrastructure.
  • Chinese customs verbally notified stakeholders that H200 could not enter China, and officials instructed companies not to purchase unless necessary.

Counter-evidence

  • Beijing actively released limited H200 units in August 2026, indicating that complete rejection is not rationally optimal for China.
  • The pace and enforcement intensity of technology-sovereignty policy remain uncertain.
MixedChina's AI-computing investment structureMedium impact · 6–18 months

H200 drip-feed does not change the total amount of China's AI investment—cloud providers' capex is an exogenously determined strategic investment—but changes its structure: overseas/Hong Kong computing versus domestic Chinese computing versus imported computing inside China. Offsetting measures include overseas capacity leasing, overseas-built clusters and accelerated domestic substitution.

Supporting evidence

  • Citi: Alibaba, Tencent and Baidu sharply increased 2Q26 capex; the largest spending may relate to domestic GPU procurement and global and domestic memory-chip procurement, making H200's < $890 million immaterial.

Counter-evidence

  • There is no official statistical breakdown of China's AI capex structure, and most H200 quantities come from anonymous sources.
  • The China revenue contribution from H200 versus H20 and other non-AI products has not been disclosed; a recovery in China revenue cannot be directly attributed to H200.
UnclearUSD/RMB settlement and cross-border fundsUncertain impact · Uncertain

The 25% revenue share and ad valorem tariff structure requires chips to physically transit the United States, potentially changing settlement routes and fund flows. However, there are no verifiable scale data, so no clear direction can be established.

Supporting evidence

  • The 2026-01-14 executive order stipulated a 25% ad valorem tariff; a 25% sales-revenue share was announced on 2025-12-08.

Counter-evidence

  • The source materials provide no verifiable data on settlement routes or cross-border fund volumes.
INDUSTRY

Supply-chain impact

NegativeAI accelerators (NVIDIA Hopper product line)Low impact · 6–18 months
Upstream chip design and supply

Deliveries to China remain at symbolic small volumes. Approximately 500,000 H200 units originally prepared for mainland customers must either be resold outside China or written down in stages; $404 million has already been recognized over the past six months, so further impairment is likely to occur gradually rather than as a one-time shock.

Supporting evidence

  • H200 revenue to China was <1% of FY27Q2 data-center revenue (<$890 million).
  • $404 million of excess H200 inventory charges was recognized over the past six months.
  • NVIDIA held approximately 500,000 H200 units (FT 2026-08-19).

Counter-evidence

  • NVIDIA's $108 billion Q3 guidance explicitly excludes China data-center compute revenue, reducing marginal information value.
  • The main bottlenecks identified by the company are supply-chain constraints—wafers, memory, land, power and racks—not demand.
PositiveChina's domestic AI chips (Ascend/Cambricon and others)Medium impact · 12–36 months
Domestic-substitution beneficiaries

Lower H200 availability raises domestic procurement budgets at cloud providers, improving order visibility for Huawei AI chips. Huawei AI-chip revenue is expected to rise from $7.5 billion in 2025 to $12.0 billion in 2026 (FT).

Supporting evidence

  • DeepSeek V4 has been deeply adapted to the Ascend 950PR instruction set; Alibaba Cloud Bailian and Tencent Cloud TokenHub went live on release day.
  • Alibaba, Tencent and Baidu sharply increased 2Q26 capex, primarily toward domestic GPUs and memory.
  • ByteDance's 2026 Ascend procurement reportedly exceeded RMB40 billion, based on industry compilation with low reliability.

Counter-evidence

  • The physical constraint is HBM and advanced-process capacity: Huawei's 950DT output this year is only in the "hundreds of thousands," and delivery of DeepSeek's 160,000-unit order may take more than a year.
  • DeepSeek training still relies on NVIDIA; Ascend is used only for inference.
  • Beijing's August release of limited H200 units shows that domestic supply cannot meet frontier training needs in the short term.
PositiveAI-chip foundries (SMIC)Medium impact · 12–36 months
一阶 (foundry)

SMIC is the main foundry for Ascend. Two new fabs dedicated to Huawei are scheduled to begin operations later in 2026; 7nm DUV multiple patterning raises wafer costs, while capacity expansion directly drives capex.

Supporting evidence

  • FT 2026-05-01: new fabs dedicated to Huawei are expected to begin operations later in 2026.
  • Most Ascend chips are manufactured by SMIC.

Counter-evidence

  • Advanced-process expansion is constrained by equipment and yields, and the commissioning schedule is uncertain.
  • SMIC's latest capacity and capex guidance was not verified in this analysis.
MixedHBM/high-end memoryHigh impact · 12–36 months
Critical bottleneck

High-end memory shortages limit Huawei 950DT output, and delivery of DeepSeek's 160,000-unit order may take more than a year. HBM shortages are the hardest physical constraint on domestic self-sufficiency, and H200 delivery cannot alleviate it.

Supporting evidence

  • Bloomberg 2026-09-04: because of shortages of high-end memory and other components, Huawei's 950DT output this year is at the low end of the "hundreds of thousands."
  • Discussion of the $404 million H200 inventory write-down and memory-price increases compressing gross margins.

Counter-evidence

  • CXMT's HBM progress is the single most important variable; mass production would ease the constraint.
  • HBM supply data lack official disclosure.
MixedServer ODMs and systems (Inspur, Industrial富联 and others)Medium impact · 6–18 months
Midstream systems integration

During H200 restrictions, domestic Ascend servers and supernode solutions can absorb demand, including the Atlas 950 all-optical interconnect design. However, H200 server contract manufacturing is constrained; Lenovo notified customers in August 2026 that orders could resume.

Supporting evidence

  • Inventec publicly confirmed in January 2026 that H200 approval in China was obstructed.
  • Lenovo partners notified customers that orders for H200-equipped products could resume.
  • Inspur, Dawning, Industrial富联 and FiberHome support Ascend server manufacturing and supernode solutions.

Counter-evidence

  • Domestic server volumes are not comparable with NVIDIA-based systems.
  • The supernode supply-chain list is largely based on broker notes and has not been verified.
PositiveHigh-speed connectors/liquid cooling/PCBs (supernode components)Medium impact · 12–36 months
二阶 (supernode components)

Because domestic chips have lower per-card compute, 16-card systems and supernodes with all-optical interconnects are used to compensate, driving demand for 224G connectors, liquid cooling, orthogonal backplanes and optical interconnects. Liquid cooling becomes mandatory.

Supporting evidence

  • Industry compilations mention Atlas 950 all-optical interconnects and 8 FP8 EFLOPS per supernode.
  • Liquid-cooling results lag orders, with the earnings inflection point likely in 2026Q3.

Counter-evidence

  • The list is largely sourced from broker notes or industry media and has not been verified through company announcements or exchange disclosures.
  • The liquid-cooling segment is subject to expectations running ahead of orders; valuation retracement risk is substantial if orders do not materialize.
PositiveHong Kong data centers and powerLow impact · 12–24 months
Demand-side channel

H200 may be used in Hong Kong, allowing chips to enter Hong Kong while facing data-center and power bottlenecks. First-order beneficiaries are therefore Hong Kong data centers and power operators rather than mainland China IDC providers.

Supporting evidence

  • FT 2026-08-19: Beijing wants most chips to remain outside mainland China and be usable in Hong Kong; Hong Kong lacks sufficient data-center and power capacity.

Counter-evidence

  • The scale is extremely small and represents thematic rather than fundamental exposure.
  • The source materials provide no specific data on new Hong Kong capacity.
MixedCloud-provider capex (Alibaba/Tencent/Baidu/ByteDance)Medium impact · 6–18 months
Demand side

Although US procurement approval was obtained, with a ceiling of 75,000–100,000 units per company, Beijing's instructions have slowed purchases. Capex is simultaneously shifting toward domestic GPUs and memory, demonstrating a decoupling between "China AI demand" and "NVIDIA revenue."

Supporting evidence

  • The US approved approximately 10 Chinese companies to purchase H200 ([14][18]).
  • Citi: 2Q26 capex rose sharply but was primarily directed toward domestic GPUs and memory.

Counter-evidence

  • H200 quantities and payment terms have not been officially confirmed.
  • The US approval ceilings are inconsistent: 75,000 units in May versus 100,000 units in August.
ASSETS

Most affected securities

Selected by business exposure and the causal chain, not a list of theme stocks to buy.

US · NVDA

NVIDIA

Mixed · Direct

H200 deliveries to China remain at symbolic small volumes. Current-period revenue impact is no more than 1% of revenue, but ecosystem-moat erosion caused by exclusion from China creates a long-term loss of option value.

Horizon
Current period to 36 months
Magnitude
Low
Confidence
High

Financial channels

  • Data-center revenue (China)
  • Inventory impairment
  • Product gross margin (Hopper below company average)
  • Long-term pricing power/ecosystem share

Valuation variables

  • Data-center revenue growth
  • Gross margin (affected by memory-price increases and product mix)
  • Inventory impairment
  • Long-term ecosystem share (difficult to quantify)
Priced in?

Fully priced in: guidance has excluded China data-center compute revenue for multiple consecutive quarters, and NVIDIA rose only approximately 2.8%–4.4% on 2026-05-14 after approval for 10 Chinese companies.

Supporting evidence

  • FY27Q2 revenue was $96.22 billion, up 106% year on year; data-center revenue was $89.0 billion, up 117%.
  • H200 revenue to China was <1% of data-center revenue (<$8.9 billion); China plus Hong Kong represented approximately 8.2% of total revenue.
  • FY27 Q3 guidance of $108.0 billion explicitly excludes China data-center compute revenue.
  • $404 million of excess H200 inventory charges was recognized over the past six months.

Counter-evidence

  • Suppliers such as PCB makers temporarily halted production in January 2026 to avoid inventory impairment, showing that the delivery shortfall has already reached the supply chain.
  • NVIDIA's annual report acknowledges that exclusion "helps competitors build larger developer/customer ecosystems"; the actual scale of the second-order risk may be underestimated.
HK · 0992.HK

Lenovo Group

Mixed · Direct

As a US-approved distributor, H200 delivery volumes directly determine its distribution and integration revenue. In August 2026 it notified customers that orders for H200-equipped products could resume.

Horizon
6–18 months
Magnitude
Low
Confidence
Medium

Financial channels

  • Distribution/integration revenue
  • Server shipments
  • Accounts receivable and inventory turnover

Valuation variables

  • ISG revenue growth
  • Server gross margin
  • Inventory turnover
Priced in?

Partially priced in: distribution is a limited share of Lenovo's total business, and H200-related revenue has not been quantified in company disclosures.

Supporting evidence

  • Reuters 2026-05-14: Lenovo and Foxconn were distributors approved by the US.
  • FT 2026-08-19: Lenovo partners notified customers that orders for H200-equipped products could resume.

Counter-evidence

  • Each delivery batch is only approximately 10,000 units, so the incremental distribution contribution to Lenovo's total revenue is limited.
  • The source materials provide no disclosure of Lenovo's H200-related orders.
HK · 0981.HK

SMIC

Positive · First-order

As the main Ascend foundry, H200 restrictions extend the domestic-substitution window. Two Huawei-dedicated fabs scheduled to begin operations later in 2026 should drive capacity and capex.

Horizon
12–36 months
Magnitude
Medium
Confidence
Medium

Financial channels

  • Wafer-foundry revenue
  • Capital expenditure
  • Capacity utilization
  • ASP (7nm DUV multiple patterning raises cost)

Valuation variables

  • Mature/advanced-process capacity utilization
  • Capex intensity
  • ASP and gross margin
  • Commissioning date of new capacity
Priced in?

Partially priced in: the domestic-substitution theme is widely recognized, while the capacity ramp schedule is the key variable.

Supporting evidence

  • FT 2026-05-01: two new fabs dedicated to Huawei are expected to begin operations later in 2026.
  • Most Ascend chips are manufactured by SMIC.

Counter-evidence

  • Advanced-process expansion is constrained by equipment and yields, and the commissioning schedule is uncertain.
  • SMIC's latest financial statements and capex guidance were not verified.
A · 688256

Cambricon

Positive · First-order

Lower H200 availability improves order visibility for domestic AI chips, while users with actual deployments in search, recommendation and advertising scenarios can absorb domestic-substitution demand.

Horizon
12–36 months
Magnitude
Medium
Confidence
Low

Financial channels

  • AI-chip sales revenue
  • Order visibility
  • Capacity and delivery cycle

Valuation variables

  • Revenue growth
  • Gross margin
  • Order visibility (constrained by SMIC capacity and HBM supply)
Priced in?

Medium-to-high pricing, with expectations running ahead of orders.

Supporting evidence

  • Industry notes mention Cambricon deployments in search, recommendation and advertising scenarios.

Counter-evidence

  • The latest financial statements and Siyuan 690/590 shipment data were not verified.
  • The core advantage of domestic chips over NVIDIA's low-end cards is interconnect, not compute.
Positive · First-order

As one of the packaging and testing providers for Ascend, growth in domestic AI-chip shipments should raise packaging-and-testing capacity utilization.

Horizon
12–36 months
Magnitude
Medium
Confidence
Low

Financial channels

  • Packaging-and-testing revenue
  • Capacity utilization
  • Capital expenditure

Valuation variables

  • Packaging-and-testing capacity utilization
  • ASP and gross margin
  • Advanced-packaging capex
Priced in?

Difficult to assess; exposure of packaging and testing to the domestic computing chain has not been quantified in company disclosures.

Supporting evidence

  • Industry notes identify Tongfu Microelectronics as one of the Ascend packaging and testing providers.

Counter-evidence

  • Source reliability is low-to-medium and has not been verified through company announcements.
  • The competitive landscape and order share in packaging and testing remain uncertain.
A · 688629

HuaFeng Technology

Positive · Second-order

The domestic supernode route drives demand for 224G high-speed connectors, while per-card value in the 950 series increases.

Horizon
12–36 months
Magnitude
Medium
Confidence
Low

Financial channels

  • Connector revenue
  • Per-system value
  • Capacity utilization

Valuation variables

  • Revenue growth (constrained by supernode shipments)
  • Gross margin
  • Per-system value
Priced in?

Medium-to-highly priced in, with expectations running ahead of orders.

Supporting evidence

  • Industry compilation mentions higher per-card value in the 950 series and validation of 224G.

Counter-evidence

  • The source is an industry compilation/broker note, not verified by company announcements or exchange disclosures.
  • Liquid-cooling and other supporting segments lag orders; the inflection point is likely in 2026Q3, with meaningful execution risk.
Positive · Second-order

Full liquid cooling is becoming mandatory for domestic supernodes, driving demand for liquid-cooling equipment.

Horizon
12–36 months
Magnitude
Medium
Confidence
Low

Financial channels

  • Liquid-cooling equipment revenue
  • Order-delivery schedule
  • Gross margin

Valuation variables

  • Revenue growth (constrained by order delivery)
  • Gross margin
  • Capacity utilization
Priced in?

Medium-to-highly priced in; earnings lag orders and the inflection point is likely in 2026Q3.

Supporting evidence

  • Industry compilation mentions that full liquid cooling has become mandatory and that segment earnings lag orders.

Counter-evidence

  • Source reliability is relatively low and has not been verified.
  • If liquid-cooling orders fail to materialize as expected, valuation retracement risk is significant.
A · 002916

Shennan Circuits

Positive · Second-order

Supernode PCB/CCL demand is increasing; the claim of monthly orders exceeding RMB200 million is circulating but unverified.

Horizon
12–36 months
Magnitude
Low
Confidence
Low

Financial channels

  • PCB revenue
  • Capacity utilization
  • Gross margin

Valuation variables

  • Revenue growth
  • Gross margin (affected by high-end product mix)
  • Capital expenditure
Priced in?

Difficult to assess; the relevant order size lacks verification in company disclosures.

Supporting evidence

  • Industry compilation mentions Shennan Circuits supernode monthly orders exceeding RMB200 million.

Counter-evidence

  • The claim has low reliability and has not been verified by company announcements or exchange disclosures.
  • Competition and order share in PCBs remain uncertain.
Mixed · Second-order

Inspur can absorb Ascend server contract manufacturing and supernode solutions, but H200 system-contract-manufacturing orders are constrained.

Horizon
6–18 months
Magnitude
Medium
Confidence
Medium

Financial channels

  • Server revenue
  • Gross margin
  • Inventory turnover

Valuation variables

  • Server revenue growth
  • Gross margin
  • Inventory impairment
Priced in?

Partially priced in: the domestic-server theme is recognized, but the order mix has not been quantified.

Supporting evidence

  • Industry compilation identifies Inspur Information as an Ascend server and supernode solution provider.

Counter-evidence

  • Domestic server volumes are not comparable with NVIDIA systems.
  • The source materials provide no disclosure of Inspur's Ascend-related orders.
Mixed · Second-order

HBM and memory support are the hardest bottlenecks for domestic self-sufficiency. HBM shortages are both a constraint and a substitution opportunity for domestic memory vendors.

Horizon
12–36 months
Magnitude
Medium
Confidence
Low

Financial channels

  • Memory packaging/module revenue
  • Capacity utilization
  • Gross margin

Valuation variables

  • Memory-module revenue growth
  • Gross margin (affected by the memory-price cycle)
  • Capital expenditure
Priced in?

Difficult to assess; domestic HBM progress lacks official disclosure.

Supporting evidence

  • Industry compilation identifies Tsinghua Unigroup International as an HBM/memory-support exposure.

Counter-evidence

  • The key variable in domestic HBM progress is unlisted CXMT, so the direct benefit to supporting names is uncertain.
  • Source reliability is relatively low.
HK · 9988.HK

Alibaba Group

Mixed · Third-order

As a demand-side cloud provider, Alibaba obtained US procurement approval with a ceiling of 75,000–100,000 units per company, but purchases slowed following Beijing's instructions. H200 drip-feed does not change total AI capex; it changes the structure, mainly toward domestic GPUs and memory.

Horizon
12–36 months
Magnitude
Low
Confidence
Medium

Financial channels

  • Capital expenditure
  • Cloud-business revenue and gross margin
  • Computing costs

Valuation variables

  • Cloud-business revenue growth
  • Capex intensity
  • Computing costs/gross margin
Priced in?

Partially priced in: the relationship between H200 availability and cloud-AI investment has not been separately priced in valuation.

Supporting evidence

  • The US approved approximately 10 Chinese companies, including Alibaba, to purchase H200 ([14][18]).
  • Citi: 2Q26 capex rose sharply, mainly toward domestic GPUs and memory.

Counter-evidence

  • The source materials provide no exact figure for Alibaba's 2Q26 capex.
  • H200's share of China's AI spending is too small to determine the direction of Chinese AI investment.
HK · 0700.HK

Tencent Holdings

Mixed · Third-order

Tencent received an initial delivery of approximately 10,000 H200 units and is also a major buyer of domestic Ascend chips; H200 drip-feed has limited impact on its AI-investment structure.

Horizon
12–36 months
Magnitude
Low
Confidence
Medium

Financial channels

  • Capital expenditure
  • Cloud-business revenue and gross margin
  • Computing costs

Valuation variables

  • Cloud-business revenue growth
  • Capex intensity
  • Computing costs/gross margin
Priced in?

Partially priced in: H200-related revenue and costs are immaterial relative to Tencent's overall scale.

Supporting evidence

  • FT 2026-08-19: ByteDance and Tencent each received approximately 10,000 H200 units ([26]).
  • Tencent Cloud TokenHub went live on the day of DeepSeek V4's release.

Counter-evidence

  • H200 delivery volume is extremely small relative to Tencent's total capex.
  • The source materials provide no exact figure for Tencent's 2Q26 capex.
SCENARIOS

Scenarios and signals

S1 Baseline: symbolic drip-feed (already occurring)

Premise: H200 deliveries to China remain at symbolic small volumes, at an annualized level of tens of thousands of units. Most remain outside mainland China or in Hong Kong, require item-by-item NDRC approval, and face Hong Kong data-center and power-capacity constraints.

NVIDIA: H200 revenue to China remains <1% of data-center revenue, with incremental impact of ≤1% of total revenue; inventory impairment is absorbed in stages ($404 million already recognized). China's AI self-sufficiency: the domestic-substitution window remains open but is constrained by HBM and advanced processes; direction is unchanged and speed is marginally accelerated, while DeepSeek training still relies on NVIDIA.

Signals to watch

  • NVIDIA 10-Q disclosure of China revenue breakdown and the amount and pace of H200 inventory impairment.
  • Number of NDRC approvals and announcements of additional Hong Kong data-center capacity.
  • Language in Alibaba/Tencent/Baidu quarterly capex disclosures regarding the share of domestic-computing procurement.
  • Ascend 950PR/950DT shipments and delivery cycles (currently 4 weeks versus 6 months+ for NVIDIA).

S2 Hard zero: formal Beijing ban + US withdrawal/nonrenewal of licenses

Premise: Beijing issues a formal, public and comprehensive H200 import ban rather than the current verbal guidance, while the US revokes or does not renew export licenses, or BIS restores a "presumption of denial."

NVIDIA: relative to S1, the incremental loss is capped at the loss of < $890 million per quarter of China Hopper revenue, approximately 0.9% of total revenue, plus potentially several hundred million to more than $1 billion of additional impairment on the remaining approximately 500,000 H200 units. Because low-margin revenue is removed, the impact on overall gross margin is limited or slightly positive. China's AI self-sufficiency: direction unchanged and marginal acceleration, but the real constraints remain HBM and advanced-process capacity; demand stimulus does not equal capacity creation.

Signals to watch

  • Formal Chinese Ministry of Commerce, General Administration of Customs or NDRC documents converting guidance into a ban.
  • US BIS announcement or Federal Register amendment to the license regime.
  • A new large H200 inventory impairment charge in NVIDIA's 10-Q/10-K.
  • Earlier-than-expected CXMT HBM mass production or commissioning of SMIC's Huawei-dedicated new fabs.

S3 Substantive release: Beijing permits part of the authorized ceiling

Premise: The US and China reach a package agreement involving rare earths/gallium and chip equipment/EUV-related tools; H200 quotas expand at once to tens of thousands of units annually, passing Beijing approval, Hong Kong capacity and US case-by-case review.

NVIDIA: China data-center revenue could recover from approximately zero to several billion dollars per quarter, with a theoretical ceiling of approximately 1 million units × approximately $27,000 = $27 billion, although this is not practically attainable. More likely would be several batches per year of tens of thousands of units each. China's AI self-sufficiency: Huawei penetration in core internet applications, especially training and multimodal inference, would be marginally delayed, but the direction would remain unchanged because policy uncertainty would continue to encourage domestic procurement as supply-chain insurance.

Signals to watch

  • US-China joint statement or package-agreement announcement.
  • Changes to the BIS customer list and quantity ceilings.
  • A significant increase in the number of NDRC approvals.
  • New Hong Kong data-center and power capacity coming online.

S4 Expansion of out-of-rule channels: large-scale smuggling/overseas capacity leasing

Premise: Smuggling channels or overseas capacity leasing—including overseas-built clusters and rented computing—expand materially, weakening the practical constraints of H200 export controls and turning the nominal stalemate into a formal label only.

NVIDIA: the gap between nominal prohibition and actual demand narrows, and some China exposure may be recovered through indirect channels, but compliance and regulatory risks rise. China's AI self-sufficiency: the urgency of domestic substitution declines marginally, but the supply-chain-insurance rationale continues to support domestic procurement; the split between training and inference—domestic inference and dual-track training—continues.

Signals to watch

  • Scale and pricing data for the overseas computing-leasing market.
  • ByteDance and Tencent progress in building overseas clusters.
  • US BIS enforcement actions against third-country transshipment.
  • Changes in NVIDIA's data-center revenue mix outside China, including third countries.
Independent verdict · Medium confidence

Under the revised scenario in which H200 deliveries to China remain at symbolic small volumes over an extended period, the supporting case has stronger evidence for the causal chain of "current-period financial impact of no more than approximately 1% of revenue, with effects concentrated in second-order ecosystem and long-term competitive dynamics." The opposing case is correct that the premise has been partially falsified and that third-order causal identification is insufficient, but these points do not overturn the conclusion. The balance therefore leans toward the supporting case, limited to the revised scenario; the original premise of a zero-delivery stalemate is invalid.

Confidence is medium because the decisive evidence supports the revised scenario rather than the original proposition. On one hand, the Reuters January 2026 customs notice that H200 was "not allowed to enter," the May report of "not a single unit shipped," NVIDIA's CFO statement that H200 had generated "no revenue," the August delivery of only approximately 10,000 units each, and $404 million of inventory charges over the past six months establish a measurable delivery shortfall and limited financial consequences. On the other hand, two hard weaknesses identified by the opposing case remain unresolved: FY27Q2 H200 revenue to China was <$890 million, approximately 0.9% of total revenue, showing that the marginal financial impact was inherently very small; and NVIDIA's annual report and Huang's comments are post-exclusion facts that do not identify whether the current stalemate caused ecosystem erosion. Beijing's August release of limited H200 units and DeepSeek's continued reliance on NVIDIA also directly challenge the acceleration narrative. Current-period causality is verifiable but of limited importance, while long-term causality is important but difficult to identify; the evidence therefore does not support high confidence.} ,

The case for

  • Administrative approvals and channel restrictions have clearly translated into an observable delivery shortfall, supported by reporting from Reuters, FT and others: a January 2026 customs verbal notice that H200 was "not allowed to enter China," US approval in May for approximately 10 Chinese companies with a 75,000-unit ceiling each but "not a single unit shipped," and NVIDIA's CFO explicitly stating that H200 had "not generated any revenue," until symbolic deliveries of approximately 10,000 units each to ByteDance and Tencent appeared in August—far below the authorized ceilings.
  • The conclusion that further tightening would have limited current-period financial impact is supported by hard data: FY27Q2 China plus Hong Kong revenue was $7.88 billion, H200 revenue to China was less than 1% of data-center revenue (<$8.9 billion), Q3 guidance of $108 billion explicitly excluded all China data-center compute revenue, and $404 million was charged against excess H200 inventory over the past six months.
  • The second-order ecosystem impact is not purely narrative and is supported by company documents and customer behavior: NVIDIA's annual report acknowledges that exclusion from China "helps competitors build larger developer and customer ecosystems," while Jensen Huang said NVIDIA had "given that market to Huawei"; DeepSeek V4's deep adaptation to the Ascend 950PR instruction set and simultaneous launch on Alibaba Cloud and Tencent Cloud point to irreversibly higher switching costs.
  • Citi's observation that Alibaba, Tencent and Baidu sharply increased 2Q26 capex, mainly toward domestic GPUs and memory, while H200 was "immaterial," provides independent corroboration that Chinese AI investment and NVIDIA revenue are partially decoupling.
  • The supporting case candidly discloses the evidence boundary and does not overreach: it acknowledges that HBM and advanced-process capacity, rather than H200 availability, constrain the speed of domestic self-sufficiency—Huawei 950DT output is only in the hundreds of thousands because of high-end-memory shortages, and DeepSeek training still relies on NVIDIA. The conclusion is therefore limited to marginal acceleration and long-term ecosystem effects, rather than current-period earnings.

The skeptical case

  • The premise has been partially falsified: section 1.2 of the research draft explicitly says that the "persistent stalemate and inability to achieve actual delivery" was broken in August 2026, when ByteDance and Tencent each received approximately 10,000 physical units. Describing the small volume as "delivery stagnation" conflicts with the verified conclusion cited by the supporting case.
  • The scale itself weakens the practical weight of the argument: the supporting case's own figures—less than $890 million and approximately 0.9% of total revenue—show that the scenario's marginal information value for NVIDIA's current-period results is extremely small. The claim that it is "important enough" must therefore rely on second- and third-order narratives.
  • The third-order evidence consists of post-exclusion facts or risk language, not causal identification that the "delivery stalemate caused ecosystem erosion." NVIDIA's annual report and Huang's comments show that China has already been lost, but cannot distinguish whether this was caused by the current H200 stalemate or by the earlier H20 cutoff and broader controls.
  • The H20 precedent has limited comparability: when H20 was cut off, China data-center revenue had not yet been excluded from guidance, whereas NVIDIA now assumes no China data-center compute revenue for multiple consecutive quarters. The structure of the marginal impact has changed.
  • There is direct counterevidence weakening the "Chinese self-sufficiency acceleration" narrative: Beijing's active release of limited H200 units in August 2026 indicates that domestic supply cannot meet frontier training needs in the short term; DeepSeek training still relies on NVIDIA; and the true bottlenecks are HBM and process capacity, meaning demand stimulus does not equal capacity creation.

What would invalidate this

  • The US BIS publishes a Federal Register notice formally revoking or declining to renew H200 export licenses to China, or changes the review standard from case-by-case review back to presumption of denial.
  • China's Ministry of Commerce, General Administration of Customs or NDRC issues a formal, public and comprehensive H200 import ban, distinct from the January 2026 Reuters-reported customs verbal notice and instruction to companies not to purchase unless necessary.
  • NVIDIA's SEC-filed 10-Q or 10-K no longer excludes China data-center revenue from guidance, or H200 revenue to China rises above 1% of data-center revenue.
  • NVIDIA records a materially larger H200 excess-inventory impairment in a single quarter, clearly exceeding the cumulative $404 million recognized over the past six months, or inventory is substantially cleared in a one-time event.
  • CXMT announces a breakthrough in HBM mass production, or SMIC's Huawei-dedicated new fabs begin large-scale production ahead of schedule, allowing Huawei Ascend 950DT output to escape the hundreds-of-thousands range reported by Bloomberg.
  • The US approves Blackwell or Rubin exports to China, causing the H200 issue to lose focus in policy and market pricing.
  • The US and China reach a package agreement—such as relaxed rare-earth and gallium exports in exchange for flexibility on chip equipment or EUV-related tools—that expands H200 quotas to a significant share of the US-authorized ceiling.
  • Alibaba, Tencent or Baidu explicitly discloses H200 procurement volumes or a reversal in the share of domestic-computing procurement during quarterly earnings calls, contrary to Citi's qualitative 2Q26 description.

Limitations

  • All H200 delivery quantities—approximately 10,000 units each for ByteDance and Tencent, ceilings of 75,000 or 100,000 units per company, and approximately 500,000 units of inventory—come from anonymous sources cited by Reuters, FT and other media. There is no official confirmation from China or NVIDIA, and quantity definitions differ between the May and August reports.
  • NVIDIA's $788 million FY27Q2 China revenue includes non-AI products such as gaming hardware. H200's specific contribution was not disclosed, so the possibility that the recovery in China revenue was driven by non-H200 factors cannot be excluded.
  • The cost basis of the H200 inventory impairment was not disclosed, preventing precise estimation of the additional impairment under a hard-zero scenario.
  • Key data—including Huawei's 2026 Ascend shipment target of 1.2 million units, ByteDance's Ascend procurement above RMB40 billion and a domestic share approaching 90% in 2026—come from industry research notes or social-media retellings, have low reliability and lack verification through company announcements or exchange disclosures.
  • The latest Cambricon financial statements and Siyuan 690/590 shipment data were not verified, so the assessment of competition among domestic AI chips lacks first-hand financial validation.
  • The definitions and product breakdowns of NVIDIA's FY27Q1 and FY27Q2 China revenue are incomplete. The assessment of decoupling between Chinese demand and NVIDIA revenue therefore relies on Citi's qualitative description rather than a reproducible quantitative series.
  • All H200 quantity data are monthly or quarterly snapshots from January through September 2026. The observation period is less than one year and is insufficient to assess the probability of policy reversals over a longer cycle.
  • This analysis is a hypothetical scenario exercise based on a research draft. It does not include a cross-check against sell-side consensus expectations or market-implied probability data, so the assessment of how much is already priced in relies on limited price-reaction evidence.

Research sources

  1. 1https://global.hk01.com/%E4%B8%AD%E5%9B%BD%E8%A7%82%E5%AF%9F/60313047/nvidia%E8%8B%B1%E4%BC%9F%E8%BE%BEh200%E5%9B%B0%E5%B1%80-%E7%BE%8E%E5%9B%BD%E6%94%BE%E8%A1%8C-%E4%B8%AD%E5%9B%BD%E4%B8%BA%E4%BD%95%E6%8B%92%E6%94%B6
  2. 2https://big5.sputniknews.cn/20260117/1069348489.html
  3. 3https://ieknet.iek.org.tw/ieknews/news_open.aspx?nsl_id=d133bfbdd8ab4efcbd32510c2732b89b
  4. 4https://news.pconline.com.cn/2070/20709812.html
  5. 5https://ieknet.iek.org.tw/ieknews/news_open.aspx/news_more.aspx?actiontype=ieknews&indu_idno=1&nsl_id=d8eb1d5f385643d590b7976fbaa5a651
  6. 6https://www.zaobao.com/realtime/china/story20260117-8114981
  7. 7https://ieknet.iek.org.tw/ieknews/news_open.aspx?nsl_id=9b7c9ef6e1fb4a5fbad1ede0cd24ec2f
  8. 8https://global.hk01.com/%E5%8D%B3%E6%97%B6%E5%9B%BD%E9%99%85/60313635/%E8%8B%B1%E5%AA%92-%E4%B8%AD%E5%9B%BD%E6%8B%92%E6%94%B6h200%E8%8A%AF%E7%89%87-%E8%8B%B1%E4%BC%9F%E8%BE%BE%E4%BE%9B%E5%BA%94%E5%95%86%E5%B7%B2%E5%81%9C%E6%AD%A2%E7%9B%B8%E5%85%B3%E8%8A%AF%E7%89%87%E7%94%9F%E4%BA%A7
  9. 9https://www.zaobao.com/news/china/story20260121-8134474
  10. 10https://www.hk01.com/%E4%B8%AD%E5%9C%8B%E8%A7%80%E5%AF%9F/60313047/nvidia%E8%8B%B1%E5%81%89%E9%81%94h200%E5%9B%B0%E5%B1%80-%E7%BE%8E%E5%9C%8B%E6%94%BE%E8%A1%8C-%E4%B8%AD%E5%9C%8B%E7%82%BA%E4%BD%95%E6%8B%92%E6%94%B6
  11. 11http://hkcna.hk/docDetail.jsp?id=101272478&channel=2808
  12. 12US Approves Nvidia H200 Sales to China, But Shipments Remain Stalled - US Approves Nvidia H200 Sales to China, But Shipments Remain Stalled
  13. 13US clears sales of Nvidia H200 chips to some Chinese firms - report - US clears sales of Nvidia H200 chips to some Chinese firms - report
  14. 14US clears Nvidia H200 AI chip sales to Chinese firms, but deliveries remain stalled
  15. 15US clears H200 chip sales to 10 China firms as Nvidia CEO looks for breakthrough - Toggle navigation
  16. 16Nvidia rises after report U.S. cleared Chinese firms to buy H200 chips By Investing.com - Nvidia rises after report U.S. cleared Chinese firms to buy H200 chips
  17. 17https://statementdog.com/news/16482
  18. 18US clears H200 sales to 10 Chinese firms, but not a single chip has shipped
  19. 19Nvidia sube : Estados Unidos autoriza la venta de chips H200 a 10 empresas chinas - 8a519d435a
  20. 20Nvidia sube: Estados Unidos autoriza la venta de chips H200 a 10 empresas chinas
  21. 21http://dev.taisounds.news/news/content/84/283950
  22. 22https://wdatacn.aastocks.com/sc/(Z(@@78bfg2689%d0%a0%d0%86%d0%a2%d0%86%d0%a2%d0%8e%d0%a0%e2%80%9d%d0%b2%d2%96%d2%b7%d0%a0%e2%80%a2%d0%b2%d2%96%d2%bb%d0%a0%e2%80%9d%d0%a3%d2%b7%d0%a0%d0%86%d0%a2%d0%86%d0%a2%d0%8e%d0%a0%e2%80%9d%d0%a1%d2%b3s1%d0%a0%d0%86%d0%a2%d0%86%d0%a2%d0%8e%d0%a0%e2%80%9d%d0%b2%d2%96%d2%b7%d0%a0%e2%80%a2%d0%b2%d2%96%d2%bb%d0%a0%e2%80%9d%d0%a2%d3%98%d0%a0%d0%86%d0%a2%e2%80%93%d0%a2%d0%8e%d0%a0%e2%80%9c%d0%b2%d2%96%d2%bb%d0%a0%e2%80%9d%d0%a2%d1%96s2%d0%a0%e2%80%99%d0%92%c2%ac%d0%a0%d0%86%d0%a2%e2%80%93%d0%92%c2%a0%d0%a0%e2%80%a2%d0%b2%d2%96%d2%bb%d0%a0%e2%80%9d%d0%a2%c2%abs3%d0%a0%e2%80%99%d0%92%c2%ac%d0%a0%d0%86%d0%a2%e2%80%93%d0%92%c2%a0%d0%a0%e2%80%a2%d0%b2%d2%96%d2%bb%d0%a0%e2%80%9d%d0%a2%d3%98hjl2689=acx%7B%7B%7Bthis%7D%7D%7Dxca))/stocks/news/aafn-con/NOW.1539260/latest-news/AAFN#1
  23. 23https://china.hket.com/article/4183562/%E8%8B%B1%E5%81%89%E9%81%94%E6%8A%AB%E9%9C%B2%E9%81%8E%E5%8E%BB%E4%B8%80%E5%AD%A3%E5%90%91%E8%8F%AF%E5%94%AEH200%20%E4%BD%94%E6%95%B8%E6%93%9A%E4%B8%AD%E5%BF%83%E7%87%9F%E6%94%B6%E4%B8%8D%E8%B6%B31-?mtc=30003
  24. 24https://news.mingpao.com/pns/%e7%b6%93%e6%bf%9f/article/20260828/s00004/1787849618209/%e6%81%a2%e5%be%a9%e5%b0%8d%e8%8f%af%e5%94%aeai%e6%99%b6%e7%89%87-h200%e4%bd%94%e6%95%b8%e6%93%9a%e6%94%b6%e5%85%a5%e4%b8%8d%e8%b6%b31#top
  25. 25https://wwwhk.aastocks.com/en/usq/quote/stock-news-content.aspx?symbol=BABA&id=NOW.1539260&source=AAFN&totc=1#1
  26. 26https://china.hket.com/article/4178863/
  27. 27https://113.28.163.186/tc/cnhk/news/china-hot-topic-content.aspx?id=NOW.1539260&catg=4&source=AAFN
  28. 28http://wdatacn.aastocks.com/sc/(Z(@@bfg1889%ef%bc%9cs1%ef%b9%a5s2%ca%bas3%ca%b9hjl1889XMF=acx%7B%7B%7Bthis%7D%7D%7Dxca))/stocks/news/aafn-con/NOW.1539260/latest-news/AAFN#1
  29. 29https://wap.seccw.com/index.php/Index/detail/id/49480.html
  30. 30https://wealth.firstbank.com.tw/investment-tips/trend-insight/news/news-detail?id=%7b2E982A77-00A5-44EA-8396-CB1E3CD4A00E%7d
  31. 31https://invest.hket.com/article/4183748/?mtc=40001&srkw=%E8%8A%AF%E7%89%87
  32. 32Nvidia Reduces China Revenue Dependence Amid Export Controls | KuCoin - Nvidia Reduces China Revenue Dependence Amid Export Controls
  33. 33https://www.163.com/dy/article/L5ILHM7005568W0A.html#1
  34. 34Nvidia Slashes China Exposure as Export Curbs Reshape AI Chip Sales - Nvidia Slashes China Exposure as Export Curbs Reshape AI Chip Sales
  35. 35https://news.qq.com/rain/a/20260827A08BJE00
  36. 36https://www.kucoin.com/zh-hant/news/flash/nvidia-reduces-china-revenue-dependence-amid-export-controls
  37. 37https://m.zhitongcaijing.com/content/contentimmediatelydetail.html?immediately_id=7834064
  38. 38https://www.etnet.com.hk/www/tc/ashares/news_detail.php?newsid=20260827756&page=1&category=%e5%85%a8%e6%97%a5%e6%96%b0%e8%81%9e
  39. 39https://zhidx.com/p/588403.html#1
  40. 40https://www.163.com/dy/article/L0986EMR05568W0A.html?spss=adap_pc#1
  41. 41https://www.mscbsc.com/viewnews-2333023.html
  42. 42https://finance.mingpao.com/fin/instantf/20260904/1788523693528/deepseek%e6%93%9a%e5%a0%b1%e5%9c%a8%e5%85%a7%e8%92%99%e5%8f%a4%e6%95%b8%e6%93%9a%e4%b8%ad%e5%bf%83%e9%83%a8%e7%bd%b2%e6%9c%80%e5%b0%9116%e8%90%ac%e7%89%87%e8%8f%af%e7%82%ba%e6%99%b6%e7%89%87
  43. 43http://wdatahk.aastocks.com/tc/(Z('ontesting='zzzzz%604719940456%60'))/usq/news/comment.aspx?source=ANUE&id=AN6597946&catg=1
  44. 44https://www.jiuyangongshe.com/a/209lcdn2vg1#1
  45. 45https://www.edgen.tech/zh/news/post/huawei-ascend-950-chip-orders-near-750000-units-after-v4-launch
  46. 46https://www.toutiao.com/w/1875867651900428/?wid=1789112994417
  47. 47https://www.stheadline.com/realtime-finance/3567837/%E8%8F%AF%E7%82%BAAI%E6%99%B6%E7%89%87%E9%8A%B7%E9%87%8F%E6%96%99%E5%B9%B4%E5%A2%9E%E5%85%AD%E6%88%90-%E6%9C%89%E6%9C%9B%E5%A5%AA%E4%B8%AD%E5%9C%8B%E5%B8%82%E4%BD%94%E9%BE%8D%E9%A0%AD
  48. 48https://m.163.com/dy/article/L0986EMR05568W0A.html?spss=adap_pc&referFrom=#1#1
  49. 49https://news.mingpao.com/pns/%e7%b6%93%e6%bf%9f/article/20260905/s00004/1788538792286/%e5%bd%ad%e5%8d%9a-deepseek%e6%93%ac%e6%8e%a1%e8%8f%af%e7%82%ba16%e8%90%ac%e9%a1%86%e6%99%b6%e7%89%87

This page was generated by AI with web research from a user-submitted prompt and shared publicly by the submitter. It is not investment advice.

Browse more public analyses Analyze your own clue