<div class="fs-article fs-responsive-text current-article article-body"><section id="china-rejected-nvidias-h200s-global"><h2>China Rejected Nvidia’s H200s, But The Global AI Demand Still Needs Every H200 Nvidia Can Build.</h2><p>President Trump confirmed on Friday what the markets had been pricing in since Wednesday. Speaking to reporters aboard Air Force One on the return flight from Beijing, he said China had “chose not to” approve purchases of Nvidia’s H200 because Xi Jinping “wants to develop their own.” The two-day summit produced agricultural commitments and not much else on chips. The 750,000-unit H200 license ceiling the Commerce Department cleared on May 14 still exists on paper. Not one chip has shipped.</p><p>The terms of the license were unusual. Under the framework Trump announced in December 2025 and Commerce formalized in January, Nvidia would have remitted 25% of revenue on each China sale to the US Treasury, every China-bound shipment would have passed through a US-headquartered third-party verification laboratory, and total volume could not exceed 50% of Nvidia’s domestic US sales of the same chip. Those terms describe more than an export permit. Whether or not the structure can be renegotiated in a future round, this version of it did not clear.</p><p><strong>The question Wall Street is asking is whether the failure is a problem for Nvidia. It is the wrong question.</strong> The real question is where 750,000 frontier accelerators that were nominally reserved for Chinese buyers end up in a world that has been compute-constrained for two years.</p></section><section id="demand-that-was-already-waiting"><h2>The Demand That Was Already Waiting</h2><p>Global frontier-chip allocation outside China has been supply-limited since 2024. Microsoft’s fiscal 2026 capex guide is $190 billion. Meta has committed $130 billion. Oracle’s RPO sits north of $500 billion, anchored by the Argonne supercomputer and the Stargate buildout. Hyperscaler buyers have been working through their Nvidia order books on multi-quarter waitlists, with no sign of a slack period.</p><p>Sovereign demand is the other half of the picture, and it has been growing faster than the press is tracking. Canada announced a trillion-dollar AI infrastructure initiative under Carney in April. MBZUAI and G42 in the UAE continue to expand allocations. Saudi Arabia, India, and several European sovereign-AI programs sit on order books that are not being filled fast enough. Every one of these buyers has been in the queue for the same chips the May 14 ceiling notionally diverted toward Chinese hyperscalers.</p><p>The H200 inventory that was theoretically reserved for the China list does not sit idle now. It moves to the next allocation in line.</p></section><section id="why-baseline-was-not-disturbed"><h2>Why The Baseline Was Not Disturbed</h2><p>The narrow read on the summit is that Nvidia lost an export ceiling. The structural read is that the ceiling was an upside scenario on top of a baseline that already excluded China.</p><p>Nvidia’s October 2024 export-control wave had already dropped formal China share from roughly 95% toward zero. The 13% of revenue China once represented has been out of the operating model for more than a year. The 750,000-unit H200 ceiling cleared on May 14 was the optimistic case, not the floor. Friday moves the optimistic case from “maybe” to “not this quarter.” <strong>The floor is undisturbed.</strong></p><p>The disconnect Wall Street is pricing in this week sits between the headline failure of a politically prominent deal and the underlying demand picture that did not change. Politically driven price weakness in the leading AI infrastructure provider is a function of summit theater, not of supplier behavior or hyperscale customer capex.</p></section><section id="china-cannot-fill-its-own"><h2>China Cannot Fill Its Own Demand</h2><p>The other half of the story is whether Huawei-SMIC can absorb the demand the May 14 deal would have served. The honest answer is some of it, but not most of it, and not soon.</p><p>The headline progress is real. SMIC’s yield on the Ascend 910C, manufactured on its 7nm N+2 process, has climbed from around 20% in late 2024 to roughly 40% by early 2025, with the line profitable for the first time. Huawei has publicly targeted 60% to reach industry norms. Cambricon is ramping in parallel. Huawei’s AI chip revenue is on track to roughly double to $12 billion in 2026.</p><p>The volume gap is the part the yield numbers do not close. Chinese hyperscaler AI capex demand for 2026 runs into hundreds of thousands of frontier-class accelerators. Huawei’s combined Ascend line is projected at roughly 1.6 million dies for the year, against a market that includes the 750,000 H200s the summit failed to clear and the demand growth Chinese hyperscalers were already absorbing through the Nvidia channel before the new framework existed. ByteDance’s 2024 purchase of 100,000 Huawei Ascend 910B chips delivered roughly 30% of the order as of last summer because of upstream supply constraints, particularly on high-bandwidth memory.</p><p>A domestic supplier improving fast but well short of frontier volume is not a Nvidia problem. It is a Chinese-hyperscaler problem.</p></section><section id="what-this-tells-investors"><h2>What This Tells Investors</h2><p>The signal worth holding from this week is not the summit theater or the rhetoric around it. It is that global frontier compute is supply-constrained, that hyperscaler and sovereign capex is committed for years against capacity that does not yet exist, and that the supply nominally diverted toward the May 14 ceiling absorbs into the line that has been waiting.</p><p>The companies positioned to benefit from a world where China cannot import frontier compute at scale and cannot yet produce frontier compute at scale are the same ones already running the global AI buildout. Nvidia, because every redirected H200 finds a sovereign or hyperscale buyer that was already waiting. Broadcom, because custom-ASIC programs at Google, Meta, and the cloud giants are not negotiable trade items and do not pass through Beijing’s customs at all. Cadence and Arm sit inside every advanced chip design that funds either path.</p><p><strong>The headline is the summit. The signal is the order book.</strong></p></section></div>

英偉達

半導體 3 倍做空 - Direxion

<p>中國拒絕英偉達的 H200 芯片可能看起來很重要，但全球對人工智能技術的需求依然強勁。儘管交易失敗，英偉達的庫存將被轉向其他買家，因為來自微軟和 Meta 等公司的需求持續增長。交易的失敗並未影響英偉達的基本收入，因為中國已經被排除在其運營模式之外。此外，儘管華為等中國公司正在改善其芯片生產，但它們無法滿足對前沿級加速器的高需求，這突顯了供應缺口</p>

Why Nvidia’s Failed China Deal Matters Less Than Wall Street Thinks

- 中國拒絕了英偉達的 H200 出口許可，市場需求依然旺盛。  
- 英偉達的 H200 需求集中於中國以外的超大規模及主權投資者。  
- 全球邊緣計算受到供應限制，未來潛在買家將接手 H200 庫存。

福布斯

2 倍做空 NVDA ETF - GraniteShares

iShares Future AI & Tech ETF

T-Rex 2X Long Microsoft Daily Target ETF

2 倍做空 NVDA ETF - T-Rex

1 倍做空英偉達 ETF - Direxion

2 倍做多英偉達 ETF - Direxion

Wisdomtree Artificial Intelligence And Innovation Fund - ETF

2 倍做多英偉達 ETF - GraniteShares

英偉達期權收益策略 ETF - YieldMax

動態半導體 ETF - Invesco

2 倍做多 NVDA ETF - T-Rex

1.5 倍做空英偉達 ETF - Tradr

半導體 3 倍做多 - Direxion

2 倍做多 META ETF - GraniteShares

標普半導體 ETF - SPDR

半導體 ETF - VanEck Vectors

費城交易所 半導體 ETF - iShares

微軟

甲骨文

中芯國際

寒武紀

NVIDIA Corporation

ORACLE CORP DEPOSITARY SH REP 1/2000TH PFD SER D

SMIC HK SDR 5to1

為什麼英偉達在中國的失敗交易對其影響比華爾街認為的要小