Jordan Schneider tackles the most critical blind spot in the global artificial intelligence race: we simply do not know how many chips China actually possesses. While Washington debates export controls in the abstract, Schneider brings a rare, data-driven rigor to the question, converging on a startling estimate of 2.8 million H100-equivalents through two entirely independent methodologies. This is not a political rant; it is a forensic accounting of the hardware powering a superpower's future, revealing that our national security framework is currently built on hunches rather than hard numbers.
The Convergence of Uncertainty
Schneider's most compelling move is the decision to run two distinct estimation models—a supply-side "brute force" count and a demand-side deduction based on model training needs—only to find they nearly perfectly align. "Surprisingly, despite large uncertainties on both sides, we arrived at nearly the same number! Both estimates were roughly 2.8 million H100e, and the convergence of estimates suggests we may be on the right track." This statistical convergence is the piece's strongest evidence, lending credibility to a figure that would otherwise be dismissed as guesswork.
The author is quick to temper expectations, however, acknowledging that "the answer is unknowable for any tight ranges" and that the true number likely sits within an order of magnitude. This humility is refreshing in a field often dominated by alarmist speculation. The core of the argument rests on the idea that without a credible baseline, "the success of our export control regime and national security framework... can only be based on hunches." This framing effectively shifts the conversation from ideological posturing to the practical mechanics of enforcement.
A credible number is needed to understand how well export controls are working, to what extent we are ahead of China, and to track China's behavior.
Critics might argue that relying on "H100-equivalents" oversimplifies the complex architecture of modern AI, but Schneider anticipates this. He notes that while "simply calculating FLOPS does not give the whole story," it remains the only standardized metric available for comparison. This is a pragmatic concession to the limitations of current data, much like early economists using GDP despite its inability to capture informal labor or environmental costs.
The Anatomy of the Supply Chain
Breaking down the 2.8 million figure, Schneider reveals a supply chain that is far more porous than official narratives suggest. The calculation splits into three buckets: legal foreign purchases, smuggled chips, and domestic production. Legal purchases, while the most transparent, account for only a fraction of the total, with China securing roughly 460,000 H100e through official channels before regulations tightened.
The real shock lies in the smuggling estimates. Using a Monte Carlo method—a statistical technique famously used in physics to model complex systems with random variables—Schneider simulates thousands of scenarios to estimate illicit flows. "I ran a Monte Carlo simulation... similar to the one conducted by CNAS in its 2024 estimate, and the results suggest a median of 312,000 H100e were smuggled into China in 2025." This methodology transforms vague rumors of black markets into quantifiable risk, suggesting that China was likely able to illegally import as much compute as they were able to legally import.
The article details how the administration's export controls, initially designed in October 2022, were quickly circumvented by chipmakers producing downgraded versions like the A800 and H800. "To circumvent these restrictions, Nvidia produced the A800 and H800 for the Chinese market... BIS revised its export controls to restrict the A800 and H800 in October 2023." This cat-and-mouse game highlights a systemic failure: the rules were reactive rather than proactive, allowing a one-year window where high-performance chips flowed freely.
It is unclear how usable much of this compute is for large-scale clusters, as Nvidia would not service such clusters if the chips were to run into issues. However, the jury is still out on how easily non-Nvidia engineers can fix potential issues with Nvidia hardware.
This uncertainty regarding the usability of smuggled hardware is a crucial nuance. Just because a chip is in the country doesn't mean it can be clustered effectively. However, the sheer volume of hardware—estimated at 452,000 H100e from smuggling alone—suggests that even with maintenance hurdles, the aggregate compute power is significant.
The Rise of the Domestic Alternative
Perhaps the most strategic insight in the piece is the rapid scaling of China's homegrown industry. While often dismissed as inferior, domestic chips are now contributing nearly a third of the total compute supply. "For homegrown compute, China's champion is Huawei, with its Ascend 910B and Ascend 910C products... China has acquired roughly 600,000 Ascend 910Bs and 650,000 Ascend 910Cs."
When combined with other domestic players like Alibaba's T-Head and Baidu's Kunlunxin, the total domestic contribution reaches approximately 904,000 H100e. This challenges the assumption that US export controls will simply starve China of AI capabilities. Instead, they are accelerating a parallel ecosystem. The author notes that while "none individually comes close to Huawei's scale," the aggregate effect is "meaningful."
This shift mirrors the historical trajectory of other sanctioned industries, where isolation often breeds self-reliance. The data suggests that the US administration has inadvertently created a market incentive for Chinese firms to invest heavily in their own silicon, reducing their long-term dependency on American technology.
Ultimately, this analysis should not have required this much guesswork. Without a concrete answer, the success of our export control regime and national security framework, as well as our perceptions of our advantages against China, can only be based on hunches.
Bottom Line
Schneider's analysis is a vital correction to the current discourse, proving that the US government is flying blind regarding its primary strategic competitor's AI capacity. The strongest part of the argument is the rigorous convergence of two independent estimation methods, which lends weight to the 2.8 million H100e figure despite the inherent opacity of the data. The biggest vulnerability remains the "usability gap"—we know how many chips exist, but not how effectively they are being clustered or utilized. The immediate takeaway for policymakers is clear: without new mechanisms to monitor cloud providers and enforce "Know Your Customer" protocols, export controls will remain a blunt instrument in a world that demands surgical precision.