The U.S. administration has made a sharp turn in its semiconductor policy: Donald Trump has confirmed that the latest NVIDIA GPUs under the Blackwell architecture will not be exported to China. The message, recently articulated unambiguously by the president — “we won’t give them to other people” — effectively kills expectations of easing restrictions and reinforces the strict bans, with exceptions for strategic allies in Asia, such as South Korea.
The shift comes after weeks of mixed signals. On one hand, it was leaked that senior officials within his own administration had blocked export initiatives toward Chinese companies; on the other, the president hinted that a window might have opened for “clipped” versions. The final outcome: a closed door for China on the Grace Blackwell family (manufactured at TSMC’s 4NP, interconnected via NVLink 5 and powered by HBM3e), and a green light for allied countries within specific bilateral agreements.
This turn consolidates Washington’s geopolitical view on AI technology: cutting-edge chips are strategic assets and their access is reserved for those within the shared interest perimeter. Technologically, the measure implies redrawing the AI supply chains, rethinking provider roadmaps, and most importantly, pushing China to rely even more on its domestic ecosystem.
What exactly did Trump say (and what does it mean in practice)
In a public statement, the president made it clear that the most advanced NVIDIA chips won’t be in anyone else’s hands. The phrase — “we won’t give them to other people” — has been interpreted as an explicit veto on China, breaking any hopes of reopening the gate to Blackwell. Close sources to the process also indicate that officials within his administration had blocked previous export moves.
In practical terms:
- Blackwell (including the Grace Blackwell that combines Arm “Grace” CPU with Blackwell GPU) will not reach China.
- The veto covers both training and inference with long context windows and token-heavy loads, where latency and bandwidth of HBM3e and NVLink 5 are critical.
- Asian allies do receive units: South Korea has signed for >260,000 Blackwell GPUs for its “AI factory” national initiative, with purchases involving major domestic firms (semiconductors, cloud, automotive).
- The rest of the world remains in an intermediate zone: preferred access for partners under specific agreements; restrictions on other markets where Washington fears technology transfer.
The move aligns with a national security strategy that prioritizes supremacy over advanced semiconductors versus rival powers; and with the goal of safeguarding leadership in AI as a matter of state.
The impact on NVIDIA: between global dominance and the Chinese gap
For NVIDIA, the decision is ambiguous. On one hand, it reaffirms its central role in the global AI architecture: Blackwell is, today, the de facto standard for frontier models in hyperscalers and labs; orders from allies like South Korea or Western consortiums flow. On the other, it closes yet again a massive market: China.
- “Lost” revenue: Jensen Huang — CEO of NVIDIA — has previously admitted that his market share in China could drop from 90% to 0% within a quarter, partly due to internal bans by Beijing on US technology use and the export controls from Washington. The Blackwell ban cements this gap.
- “Allied” horizon: the massive South Korean order and Western bids compensate part of China’s absence, but not all; China does not just purchase — it builds an ecosystem (data centers, models, services), pulling in complementary demands (storage, networking, software) that are now migrating to domestic providers.
- Product strategy: “Clipped” editions for export — explored previously with H20 during the Hopper era — are not likely on the table for Blackwell in China. This complicates “half-measures” and leaves NVIDIA with two separate worlds: full performance for allies, zero for Beijing.
From a stock market perspective, the political message supports the thesis that AI growth will remain centered in blocks aligned with the US, and that Western demand (and allies’) will continue to outstrip the supply of accelerators through at least 2026–2027. But the opportunity cost of “no China” will persist.
China: forced acceleration towards a domestic “Plan B”
For China, the door shutting on Blackwell is another impetus to double down on local silicon and own ecosystem development:
- Domestic accelerators: roadmaps like Ascend (Huawei) or designs by local startups will be prioritized, focusing on inference and serving within national platforms.
- Software stack: replacing CUDA with local runtimes, homegrown compilers, and compatible libraries, a transition that is slower than hardware代替品.
- Supply chain: without Blackwell and with limitations on H100/H200, Beijing will need to favor local NAND/DRAM, packaging, and backend capabilities within its own sphere, and decrease reliance on HBM (dominated by SK hynix, Samsung, Micron) and advanced packaging in Taiwan.
- Risks: increased use of substitutions could reduce efficiency in frontier training/inference and raise the TCO of AI for domestic providers, slowing projects requiring scale and optimal latency.
The likely scenario: two increasingly distant AI ecosystems with limited interoperability — one US/ally-based and another China with replacement technologies.
TSMC, HBM, and the computing economy: the technical subtext
The Blackwell ban is also a message to the supply chain:
- TSMC 4NP: the Grace Blackwell chips are manufactured on a TSMC node reserved for key clients. The veto reinforces that this capex is prioritized for the West and its partners.
- HBM memory: modern AI relies heavily on HBM (bandwidth, latency). SK hynix and Samsung set the pace; without Blackwell, HBM3e and future HBM4 allocations will go to US/alliances, tightening supply for Chinese suppliers.
- NVLink 5 and networks: low-latency interconnection between GPUs is crucial for models with enormous contexts. This architecture (NVLink 5, InfiniBand NDR, future lines) will be a strategic good… but scarce in sanctioned markets.
The resulting landscape indicates a decade where chip, memory, interconnect, and energy are negotiated in both technical rooms and diplomatic offices.
Implications for the global AI market (and why they matter)
- Chain tension: orders from allies will absorb most of the GB200/GB300 production; HBM and packaging bottlenecks will persist; timelines and pricing will remain .
- Technological bifurcation: AI will accelerate its split into two blocks with limited knowledge transfer and diverging standards.
- “Lock-in” innovation: the US reinforces the idea that the cutting edge in AI (chips, models, tooling) is not neutral, but a foreign policy instrument.
- Higher value for foundries and memory: TSMC, 3DIC^>and1 assembly houses, and HBM suppliers will continue to capture the theory and practice of marginal value in AI.
At a glance, the landscape looks increasingly political and expensive, as technology becomes even more intertwined with geopolitics and strategic interests.
Open questions (and plausible scenarios)
- Will there be “censored editions” for China?
In the short term, no. The president’s message closes that route for Blackwell. Will future GPUs with performance limits appear? It’s possible, but not currently on the table. - What about non-frontier inference?
Light inference may rely on local silicon or predecessor generation GPUs if authorized. However, the line for Hopper and earlier successors has already tightened in earlier control rounds. - Can China turn to the “grey market”?
In theory, yes; in practice, the volume needed to train and serve frontier models is hard to hide. Compliance programs have been strengthened, and suppliers have incentives not to take risks.
Conclusion: an increasingly political (and costly) map of AI
The Trump decision is not just a technical announcement: it’s another chapter in the geopolitics of computing. It prohibits Blackwell in China, rewards allies, and cements the idea that advanced AI is critical infrastructure. For NVIDIA, it means strengthening leadership in the West, ceding China, and filling that gap with domestic agreements (South Korea, Europe, Middle East). For China, it accelerates its domestic ecosystem development at the expense of higher costs in the short to medium term.
Ultimately, the veto refines a clear message: AI chips are now at the level of defense and energy, not just commodities but instruments of power leverage. Looking ahead to 2026–2027, it seems likely they will remain so.