Background & Context§
The artificial intelligence landscape is witnessing a strategic pivot: Chinese AI companies, facing GPU export controls from the US, are increasingly releasing their models as open-weights—portable, permissionless, but not fully open-source. This approach stands in stark contrast to American giants like OpenAI and Anthropic, whose locked-down, proprietary models are accessed only via centralized APIs. The underlying dynamic, as articulated in a recent analysis, reveals that AI models themselves have little technological moat beyond brand loyalty and superficial switching costs. The real value lies in enterprise services wrapped around them. With US frontier models’ performance edge narrowing, China’s open strategy is gaining global traction, with a16z partner Martin Casado noting an 80% probability that any given startup is using Chinese models.
The News: What Happened Exactly§
According to a detailed piece by Robert Hart in The Verge, the central thesis is that China’s open-weights AI strategy is winning against America’s locked-down and proprietary approach. The analysis highlights that AI models as products have very little intrinsic moat; users can easily switch between ChatGPT and Claude via API with the same prompt and minimal workflow impact. While US companies can make deals to lock in customers, there is no long-term technical incentive for vendor loyalty. The US government has placed export controls on GPUs, limiting China’s ability to provide global-scale centralized services like those from OpenAI and Anthropic. However, Chinese companies still have enough compute to train models, and they are repurposing their disadvantage into a distribution advantage by releasing models openly.
The analysis argues that open technologies almost always win in infrastructure adoption. Open-weights models, though not fully open-source, are portable and permissionless—they can be hosted anywhere, experimented with, altered, and tweaked for specific use cases. This turns a US-created compute disadvantage into a distribution advantage, commoditizing the very layer where American companies generate revenue. The ecosystem benefits throughout China, from manufacturing to scientific research, are significant; every sector can simply plug in these models without needing to negotiate with a centralized provider.
A critical factor is that the historical saving grace for US companies—frontier models outperforming open ones—is now fading. The gap is closing, and Chinese models are poised to take the lead. The analysis points out a surprising underlying dynamic: while we think of China as a locked-down society, it is American companies that tightly control their technology rather than releasing it openly. This contrasts starkly with the US government’s historical support for an open internet. The analyst notes, "Locked-down business practices for a technology with no real moat but significant potential ecosystem benefits is an obviously losing strategy; permissively releasing it with an open, collaborative approach is obviously a winning one." However, US companies are forced to chase first-order profits rather than ecosystem benefits, and the government adds forcible measures like export controls. The analysis warns that if the bottom falls out of AI spending—which it argues it will—the economic consequences could be severe.
Historical Parallels & Similar Incidents§
This strategic shift mirrors the open-source software revolution of the late 1990s and early 2000s, particularly the rise of Linux against proprietary Unix systems. At that time, companies like Sun Microsystems and IBM owned dominant proprietary Unix variants (e.g., Solaris and AIX), while Linux, a free and open-source kernel, began gaining traction. Initially, Linux lagged in performance and enterprise features, much like open-weights AI models trailing proprietary US models today. But Linux’s permissionless model allowed widespread adoption, customization, and community-driven innovation. By the mid-2000s, Linux had captured significant market share in servers and eventually ran the majority of cloud infrastructure. Today, even Microsoft runs Linux on Azure. The parallels are striking: just as Linux commoditized the operating system layer, Chinese open-weights models threaten to commoditize the AI model layer, reducing margins for proprietary providers.
Another relevant precedent is the antitrust case against Microsoft in the late 1990s, which focused on its efforts to maintain a proprietary monopoly by bundling Internet Explorer and locking out competitors. The US government’s intervention aimed to restore competition, but the market ultimately shifted toward open web standards and open-source technologies like Apache and Mozilla Firefox. In the current context, US export controls on GPUs may be intended to hobble Chinese AI, but they are instead incentivizing an open strategy that could ultimately erode the profitability of US AI firms. The lesson from both historical episodes is that restrictive, proprietary strategies often fail when alternatives offer permissionless innovation and ecosystem value. Open, collaborative approaches tend to foster wider adoption and long-term resilience.
The analysis from The Verge echoes these historical patterns, emphasizing that the US AI industry’s focus on short-term profits and government restrictions may be undermining its long-term competitive position. The parallel with Linux suggests that even if Chinese models are not yet fully at parity, the trajectory is unfavorable for proprietary vendors. The call to action for the US is to realign incentives toward open, public-interest AI—threads like public AI, federated services, and open research—rather than doubling down on lockdown.
# Conceptual comparison: Open vs. Proprietary model adoption over time
import matplotlib.pyplot as plt
import numpy as np
years = np.array([2020, 2021, 2022, 2023, 2024])
open_adoption = np.array([10, 20, 35, 55, 75]) # hypothetical %
proprietary_adoption = 100 - open_adoption
plt.figure(figsize=(8,4))
plt.plot(years, open_adoption, label='Open-weights models')
plt.plot(years, proprietary_adoption, label='Proprietary models')
plt.xlabel('Year')
plt.ylabel('Market share (%)')
plt.title('Trend: Open models gaining share')
plt.legend()
plt.show()While this is a stylized example, it captures the dynamic outlined in the analysis: open adoption is accelerating as performance gaps close.