What happened
The Institute for Progress (IFP) published a report projecting that China's total AI capacity will expand from approximately 4.6 million to 8.8 million H100-equivalents by 2027. Despite this growth, the report states that US-designed chips will account for about 65% of this capacity. A significant portion of this growth is attributed to remote access via overseas cloud networks, which is expected to increase from roughly 1 million to 2.8 million H100-equivalents. The report also notes that while the US maintains a 16-fold advantage in training compute, leading Chinese models lag American rivals by only 6.5 months, partly due to techniques. Concurrently, the Institute for AI Policy and Strategy released a report urging stricter controls on cloud access and chip smuggling, while Huawei announced its Ascend 960DT chip will launch in Q1 2027, three quarters ahead of schedule.
The Institute for Progress (IFP), a Washington-based research group, released a report projecting that China's total AI capacity will double from approximately 4.6 million to 8.8 million H100-equivalents by the end of 2027. The report defines H100-equivalents as a metric benchmarked against Nvidia's data-center graphics processing units. This projection includes domestic chips, legal imports, smuggled processors, and compute accessed remotely via overseas cloud networks.
A key finding is that US-designed chips will account for about 65% of China's total capacity in 2027, indicating continued heavy reliance on foreign hardware. Remote access to overseas cloud networks is expected to swell to 2.8 million H100-equivalents by the end of 2027, up from roughly 1 million this year. The report also notes that while the US maintains a roughly 16-fold advantage in AI training compute over China, leading Chinese models lag their American rivals by only 6.5 months.
IFP identified , where developers train systems on the outputs of superior AI models, as a key factor helping China close the performance gap. China has previously rejected US allegations of 'aggressive, malicious' AI distillation as groundless. In light of these findings, IFP urged US policymakers to impose stricter controls on overseas cloud access, chip smuggling, and legal exports.
A separate report published on Tuesday by the Institute for AI Policy and Strategy stated that US-designed power remained legally available for rent to foreign entities otherwise barred from buying the hardware. This group called on the US Bureau of Industry and Security to restrict Chinese access to cloud capacity large enough for training frontier AI models or running them at scale, while advocating for stronger checks on customer identity and ownership.
Chinese tech giants are simultaneously bolstering domestic alternatives. Huawei Technologies announced that its Ascend 960DT chip, designed for AI model training, would become available in the first quarter of 2027, three quarters ahead of its original schedule. Huawei also revealed that more than 1,000 of its Atlas 900 A3 SuperPoDs had already been deployed, and its Atlas 950 system had entered large-scale commercial use. However, Epoch AI identified high-bandwidth memory as the main constraint on Huawei's output, estimating that Huawei's chip output this year would deliver less than 4% of the computing power of Nvidia's global shipments.
Why it matters
This data provides a concrete, quantified assessment of the geopolitical and technical landscape of AI infrastructure. It reveals that despite aggressive domestic chip development and export controls, China's AI advancement remains structurally dependent on US hardware, particularly through cloud access and smuggling channels. The finding that model performance gaps are narrowing faster than gaps suggests that efficiency and are critical competitive levers. For policymakers and industry leaders, this underscores that current export controls may be insufficient to prevent capability convergence, as remote cloud access and secondary markets continue to supply critical compute resources to Chinese entities.
The report provides a quantified baseline for understanding the efficacy of current US export controls. The fact that US-designed chips still constitute 65% of China's projected 2027 capacity suggests that existing restrictions have not eliminated dependency, but rather shifted the supply chain toward cloud access and secondary markets.
The narrowing performance gap, despite a massive disparity, highlights the strategic importance of software efficiency and . This implies that future competition may hinge less on raw hardware acquisition and more on algorithmic optimization and data utilization strategies.
The reliance on remote cloud access presents a new vector for policy intervention. Traditional chip export controls do not fully address the ability of foreign entities to rent resources from US-based cloud providers, creating a loophole that both IFP and the Institute for AI Policy and Strategy have identified as a critical area for potential regulatory tightening.
Huawei's accelerated timeline for the Ascend 960DT chip signals a significant push for domestic self-sufficiency. However, the constraint on high-bandwidth memory production suggests that domestic hardware may not be able to fully replace US chips in the near term, maintaining a hybrid infrastructure model.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Monitor US policy responses regarding cloud service provider restrictions and enhanced identity verification for foreign entities. Track Huawei's actual production volumes of the Ascend 960DT and whether high-bandwidth memory constraints limit its impact. Observe if the 6.5-month model performance gap narrows further in subsequent benchmarks, and watch for any new enforcement actions against chip smuggling or unauthorized cloud access.
US policymakers may respond to these reports by introducing new regulations targeting cloud service providers, requiring stricter identity verification and usage monitoring for foreign entities renting large-scale resources.
Huawei's ability to scale production of the Ascend 960DT will be a key indicator of China's domestic hardware progress. Watch for announcements regarding high-bandwidth memory supply chains and whether Huawei can overcome the constraints identified by Epoch AI.
The 6.5-month performance gap between leading Chinese and American models is a critical metric. Future benchmark releases from major AI labs will show whether this gap continues to narrow, potentially driven by techniques or other efficiency improvements.
Enforcement actions against chip smuggling and unauthorized cloud access may increase. Look for new legal cases or regulatory penalties targeting individuals or companies involved in circumventing export controls.