U.S. President Donald Trump announced that Nvidia may export H200 AI chips to China under a revenue-sharing arrangement giving Washington 25 percent of sales. The decision is an attempt at strategic triangulation: Sell China something advanced enough that Beijing cannot reject it out of national pride, useful enough to disrupt its efforts at vertical integration in the semiconductor industry, but supposedly constrained enough to preserve U.S. advantage in AI deployment.
This logic deserves serious engagement. Denying China any and all chips, some believe, risks accelerating Chinese self-sufficiency, while unrestricted sales hand Beijing capabilities it cannot otherwise acquire. Threading this needle is a legitimate policy goal.
But the H200 is not the answer. The administration’s efforts at triangulation fail each of the logics underpinning its strategy, and risk undermining the allied coordination required to sustain U.S. export controls in protracted technology competition with China.
The Case for Calibrated Exports
The Trump administration has assessed that China needs chips good enough that its labs will actually buy them rather than default to Huawei. The chips need to be useful enough that Chinese developers integrate them deeply into their workflows, creating dependencies that slow indigenous alternatives from emerging; and they need to fall short of the United States’ cutting edge by enough margin to preserve the sources of U.S. advantage.
There is a self-contained logic here. Beijing banned imports of Nvidia’s less powerful H20 chip in April, partly to protect Huawei from competition. If the H200 proves compelling enough, then China might accept the dependency rather than force a complete break. Meanwhile, every dollar Chinese labs spend on U.S. chips is a dollar not invested in perfecting Huawei’s ecosystem. Revenue flows back to U.S. companies, funding the R&D that maintains technological leadership.
The logic of waiving chip restrictions on China rests on three pillars. First, exporting chips to China will slow their policy of domestic software and hardware integration. Second, exporting chips to China will increase a dependency on the U.S. in critical Chinese AI industry nodes. Third, the chips exported to China will be good enough that Chinese chips cannot compete.
The problem is that Beijing has seen this movie before, and knows what to expect. The administration imposed H20 restrictions in April, reversed them in July, then approved H200 exports in December. Chip design software got restricted in May, then rescinded after China retaliated with rare earth controls. Chinese authorities understand the fundamental volatility in U.S. chip exports.
More likely is that, even as it publicly welcomes the sale of advanced chipsets, Beijing will continue forcing Chinese labs to work with Huawei regardless of H200 availability. DeepSeek, despite owning thousands of advanced Nvidia chips and accessing Blackwell infrastructure through cloud providers, was compelled to collaborate with Huawei even though its Ascend platform reportedly failed certain training runs.
While the administration seeks to introduce a wedge via H200 exports, it will fail. While there will be major demand for H200s, the demand for Huawei chips in China will be sustained due to China’s enormous compute needs and policy framework that allows Chinese firms to absorb the costs of running comparatively lower quality, inefficient chips. The administration’s decision to greenlight H200 sales does not change these factors, which will continue to fuel China’s long term chip development.
The Technical Reality
Meanwhile, the claim that H200s are degraded enough to be safe does not survive scrutiny. A thorough Institute for Progress analysis pointed out that these chips deliver nearly six times the performance of H20s – which themselves required licenses and were approved only in limited quantities. Chinese labs could build H200-based training supercomputers matching U.S. capabilities at roughly 50 percent extra cost, a premium Beijing would subsidize through the same non-market mechanisms that support its electric vehicle and semiconductor industries.
For inference workloads, the performance gap nearly disappears. H200s achieve near parity with cutting-edge Blackwell chips on price-performance under realistic deployment scenarios. Nvidia advertises a 30-fold inference advantage for Blackwells, but that figure assumes best-case conditions for Blackwell clusters and worst-case configurations for H200s. Under comparable architectures, the gap shrinks dramatically. In addition, the inference use case’s emphasis on networked compute over resource-intensive compute means that selling H200s to China will not entrench enough significant CUDA dependencies to prevent Chinese alternatives from emerging.
The strategic cost compounds over time. The Institute for Progress estimated that even a conservative amount of H200 exports to China would shrink the United States’ 2026 AI compute advantage from 11-to-one to six-to-one, depending on how quickly developers adopt advanced Blackwell features. This matters because compute enables more sophisticated model training, supports more AI companies, and scales the inference infrastructure that determines commercial deployment success.
Huawei is not planning to produce an H200-equivalent chip until the fourth quarter of 2027 at the earliest. China faces severe manufacturing bottlenecks – reaching only 1-4 percent of U.S. AI chip production in 2025 and 1-2 percent in 2026. Selling H200s hands China capabilities it has no other way to acquire at scale for at least two years. These chips directly add to Chinese compute capacity rather than substituting for domestic production, because Huawei’s output will fall well short of demand regardless of U.S. exports.
The administration appears to discount this two-year window because artificial general intelligence is not imminent. But two years of additional compute accelerates Chinese progress in deploying AI systems and training more capable models. With China’s lead position in advanced robotics, drones, and embodied AI systems magnifying the manufacturing edge it has over the United States, waiving the sale of H200s would curtail the remaining compute advantage that the United States possesses. Every chip sold today makes China’s AI ecosystem stickier, while generating minimal U.S. leverage.
The Problem With Allies
The most serious damage of selling H200s is diplomatic. Japan and the Netherlands enacted semiconductor equipment export controls at Washington’s request, while South Korea has enacted similar measures for high-bandwidth memory. These countries sacrificed Chinese market share based on U.S. assurances that denial of strategic technologies serves collective security interests.
Now they will watch the United States sell advanced AI chips to China while collecting a revenue surcharge. The contradiction is stark. Why should Seoul restrict equipment sales when Washington sells the finished product? Why should ASML forgo Chinese customers when Nvidia profits from them?
The United States has asymmetric leverage in this sector because semiconductor supply chains concentrate in allied nations. But that leverage will dissipate once allies conclude that U.S. commitments are liable to shift with corporate lobbying.
Codifying Coherent Parameters
The administration’s stated goal of establishing clear parameters is defensible. Strategic coherence requires consistent thresholds rather than ad hoc negotiations. But moving the threshold from H20s to H200s contradicts the rationale that made the original restriction meaningful.
If the concern is maximizing disruption while minimizing indigenization pressure, the answer is holding the existing line. Adjusting restrictions upward as Chinese firms advance only works if China fails to recognize this as a temporary delay – and responds by easing its push for indigenous capabilities. There is no evidence this will happen.
The United States should not proceed with unrestricted H200 exports. If the administration insists on some sales, strict volume caps and extensive end-use verification should apply – not to create illusions of control but to acknowledge the inherent contradiction between selling to adversaries and constraining their capabilities.
A bipartisan group of senators recognized the problem with enabling continued sales of advanced chips to China. The SAFE Chips Act, introduced December 4 by Senators Pete Ricketts and Chris Coons, would codify current export control thresholds for 30 months, preventing the Commerce Department from licensing more advanced chips without congressional notification.
At a more basic level, the Trump administration must choose which objectives technology restrictions should aim to accomplish. Revenue extraction, strategic denial, and negotiating leverage each imply different policies – and pursuing all three simultaneously produces incoherence Beijing exploits with precision. The decision to sell H200s demonstrates how tactical considerations displace strategic thinking when that choice remains unmade.
Image credit: Shutterstock ID 2277418639. This article was originally published in The Diplomat.
Ryan Fedasiuk
Ryan Fedasiuk is an American Enterprise Institute (AEI) fellow and, formerly, the US government's main point of contact with the Chinese Embassy in Washington, D.C. His writing on China has appeared in Foreign Policy, War on the Rocks, and RealClearDefense.
Satvik Pendyala
Satvik Pendyala is a research associate at AEI.