Is the AI ecosystem fragmenting into two isolated stacks, or is a more complex interdependence emerging where compute, models, and chips cross borders unevenly?

My position is that the commercially visible facts point to hardening decoupling into two AI stacks, with short-term leakage across the model layer but increasing closure at the strategic layers: frontier weights, compute, and chips. The strongest contrary evidence is real: U.S. developers and startups are using cheaper Chinese open-weight models. Rest of World reported U.S. developers using Minimax, Kimi, Xiaomi MiMo, Qwen, and DeepSeek to cut costs; one founder reportedly spent about $500 per month on Claude and ChatGPT for hard tasks and $200 per month on Chinese models for 90% of routine coding and voice-recognition work, while Airbnb and Cursor faced scrutiny after disclosures involving Chinese open models. That is meaningful cross-border adoption, but it is a thin, opportunistic integration: model weights are downloaded or routed through intermediaries because they are cheap, not because U.S. firms are committing their long-run infrastructure, data governance, or security posture to Chinese supply chains.

The first decisive signal is policy symmetry. BIS’s chip controls since October 2022 explicitly target China’s ability to buy advanced computing chips and manufacture advanced chips, and later rules expanded due diligence and model-weight-related controls. Now Reuters reports Beijing has held meetings on restricting overseas access to its most advanced AI models, including closed and open-weight systems, and discussed treating leaks of proprietary AI technology as national-security offenses under China’s legal framework. China’s 2020 Export Control Law already defines export control broadly to include transfers of controlled items from China to overseas parties and provision by Chinese persons or organizations to foreign entities. If Beijing applies that logic to frontier AI weights, the model layer will start resembling the chip layer: governed by licensing, national-security exceptions, and retaliation risk rather than open commercial exchange.

The second signal is redundant sovereign infrastructure. Reuters reported ByteDance is developing custom CPUs for AI infrastructure because chip prices and shortages constrain expansion, while Reuters also reported DeepSeek is developing its own AI chip as U.S. restrictions limit access to Nvidia’s best accelerators and help domestic Chinese alternatives such as Huawei gain share. These efforts may be early, and exact foundry or partner details are less transparent than the direction of travel, but the strategic logic is clear: Chinese frontier labs are not planning around durable Nvidia access.

The third signal is U.S. compute consolidation. Reuters reported Meta and Anthropic are discussing a compute lease worth up to $10 billion over two years; Reuters also reported Bank of America extended OpenAI a $520 million credit line as OpenAI prepares for an IPO. This is not a borderless AI market. It is a capital-intensive Western stack consolidating around U.S. hyperscalers, Nvidia-class accelerators, domestic cloud contracts, and financing markets. The stakes are whether AI becomes a shared software ecosystem or a strategic-industrial contest. The evidence favors the latter: cheap Chinese weights are a temporary arbitrage layer sitting above increasingly nationalized compute and semiconductor foundations.

B’s strongest point is that present-day China still has real bottlenecks, especially at the training-hardware layer. Reports that DeepSeek’s R2 training attempt on Huawei Ascend chips ran into technical failures and reverted to Nvidia for training are not trivial; they show Chinese substitution is incomplete, and they prevent a simplistic claim that China already has a fully sovereign frontier stack. Likewise, ASML’s continuing China revenue exposure is evidence that decoupling is not yet total autarky.

But B’s inference is too static. The question is not whether every dependency has already vanished; it is whether commercial and policy behavior is hardening into two strategic stacks. On that test, the failed Huawei episode supports my position more than B’s. If Chinese labs could rely on stable Nvidia access, they would not be under state pressure to train on Ascend, and DeepSeek would not now be reported to be developing its own AI chip to reduce reliance on Nvidia and Huawei. Reuters says U.S. restrictions have helped Huawei gain roughly half of China’s domestic AI-chip market and that DeepSeek is pursuing a custom chip despite its continued dependence on restricted foreign accelerators. That is a transition under constraint, not durable interdependence. (investing.com)

B also treats U.S. compute leasing as evidence against stack formation, but cross-firm contracting inside the same security bloc is different from cross-border dependence on a strategic competitor. A compute lease between Meta and Anthropic, or U.S. labs using U.S.-controlled capital markets and U.S.-approved data-center capacity, indicates scarcity inside a Western stack; it does not indicate trust in Chinese accelerators, Chinese clouds, or Chinese model-governance regimes. BIS rules now cover advanced computing chips and certain closed AI model weights, and the EAR includes restrictions on where controlled advanced model weights may be stored or transferred. That is the legal architecture of a controlled Western stack, even if capacity is rented rather than owned. (bis.gov)

Finally, the model-layer leakage that B emphasizes is real but fragile. An open-weight model is one whose parameters can be downloaded and modified; that makes diffusion fast but also makes governments nervous once the models approach frontier capability. Rest of World documents U.S. uptake of DeepSeek, Qwen, Kimi, Minimax, and MiMo because they handle routine tasks at much lower cost, including a $500-per-month U.S.-model spend paired with $200 per month on Chinese models for most tasks. Yet the same report notes regulatory scrutiny around Airbnb and Cursor and reluctance in regulated sectors due to data-security, censorship, and geopolitical risk. Reuters separately reports Beijing meetings on restricting overseas access to top Chinese models, including open-weight systems. The pattern is therefore not stable interdependence; it is leakage followed by policy containment. (restofworld.org)

B’s strongest argument is that the present ecosystem is not cleanly bifurcated: U.S. developers are actively using Chinese open-weight models, China still depends on Nvidia-class hardware for top training runs, ASML still has meaningful China revenue, and the Trump-era U.S. policy posture has shifted from the rescinded AI Diffusion Rule toward conditional licensing rather than a universal embargo. Put fairly, B is arguing that the commercial reality is “frictional interdependence”: actors route around scarcity wherever they can, and policy often follows rather than prevents leakage.

I agree that this defeats an absolutist claim that two fully sealed stacks already exist. But it does not defeat my actual claim: the evidence points to genuine, hardening decoupling at the strategic layers. The specific reason is that B treats any residual transaction as proof of integration, while I treat the terms of the transaction as decisive. A stable shared ecosystem requires reliable cross-border access, technical co-dependence, and low political risk. The facts instead show access becoming conditional, surveilled, revocable, and strategically distrusted.

The BIS rescission point is a good example. The May 2025 Commerce release did rescind the Biden AI Diffusion Rule, but it simultaneously announced stronger chip-related controls, warned industry about using PRC advanced computing ICs, warned about U.S. AI chips being used for Chinese model training and inference, and framed the policy as sharing American AI with “trusted” countries while keeping it from adversaries. That is not free interdependence; it is bloc-based permissioning. Conditional H200 sales to Chinese firms, with KYC and remote-access restrictions, likewise show that cross-border chip flows are becoming licensed exceptions rather than dependable market plumbing.

The Chinese side shows the same pattern. Reuters reports Beijing meetings on restricting overseas access to advanced models, including open-weight and unreleased systems, with possible security reviews and domestic-only treatment for the most sensitive frontier models. The fact that no final rule exists yet matters, but the policy direction is still toward treating model weights as controlled national assets. China’s Export Control Law already gives Beijing a broad legal template for controlling technologies and services transferred overseas or provided by Chinese persons or entities to foreign parties.

Finally, the hardware evidence is more consistent with decoupling-under-construction than durable interdependence. DeepSeek’s reported custom inference-chip effort is explicitly about reducing reliance on Nvidia and Huawei amid U.S. restrictions, and ByteDance’s custom CPU effort responds to chip shortages and the need to control AI infrastructure. Rest of World’s evidence of U.S. startups using Qwen, Kimi, DeepSeek, Minimax, and MiMo cheaply is real, but it is mostly model-layer arbitrage. The strategic layers—frontier weights, advanced accelerators, foundry access, and hyperscaler compute—are moving toward sovereign redundancy and permissioned exchange, not a durable shared stack.