The Export Controls Worked in One Market and Handed DeepSeek the Other
On July 7, Reuters reported that DeepSeek is developing a proprietary inference chip. That item reads like a technology story. It's not. It's the last move in a three-year sequence that splits the global AI market into two segments with different rules, different suppliers, and no mechanism for recombination under current policy.
The US Commerce Department's successive rounds of export controls: H100s, then H800s, then successor restrictions, were designed to deny Chinese labs the compute required to sustain frontier AI development. At the hardware layer, they worked. Nvidia's chip sales in China stalled, and local chipmakers like Huawei took share on the back end. Chinese labs couldn't buy the restricted chips directly. The compliance infrastructure that grew up around the controls: CFIUS reviews, executive orders and federal agency procurement bans, FedRAMP certification requirements, legally excludes Chinese AI models from regulated procurement: government contracts, defence applications, certified enterprise deployments. That's the segment the controls actually defend, and they defend it by mandate.
What the controls didn't do was prevent Chinese labs from finding a different path to capability. Denied access to premium compute, DeepSeek and Qwen trained on larger clusters of lower-spec hardware, then compressed inference cost until the price disadvantage of restricted hardware disappeared. DeepSeek's V4 Flash entered commercial markets at $0.14 per million input tokens, against GPT-5.4 at $2.50. Approximately 18x cheaper. US labs can't match that gap because their cost structures reflect the hardware and training investment required to maintain frontier capability claims in the regulated segment. The restriction that was meant to create a capability ceiling instead created an efficiency pressure that became a commercial weapon.
US AI labs have recognised the segmentation and positioned around it. Regulatory submissions from OpenAI and Anthropic support export control regimes while seeking procurement preferences in government and defence contexts. Both positions are consistent under bifurcation: if Chinese models are legally excluded from regulated markets by mandate, the case for procurement preferences is a case for protecting the only segment the controls actually defend. The commercial developer market is a different problem, and the submissions are largely silent on it. That's not a strategic gap. It's an acknowledgement that the commercial segment faces a structural pricing disadvantage the controls cannot address.
The one remaining lever by which controls could have reasserted relevance in the commercial market was hardware dependency over time. The implicit logic: pre-control Nvidia chips degrade, stockpiles deplete, and Chinese labs eventually face a capability ceiling the restrictions impose retroactively. That logic required Chinese labs to remain dependent on Nvidia architecture. It was plausible in 2022. Zhipu AI and DeepSeek's sustained investment in compute infrastructure throughout the control period should have flagged it as fragile. The spending was visible. Nobody recalibrated.
The proprietary inference chip severs the dependency entirely. It's not a product launch: the chip is still in development. But the development itself forecloses the retroactive leverage the controls framework was implicitly banking on. Once proprietary inference hardware is in production, the commercial price advantage isn't contingent on stockpile depletion. It's structural. The hardware lever that could have reversed commercial displacement no longer exists.
The bifurcation US semiconductor policy created is now the structural architecture of the global AI market under current policy. Regulated procurement belongs to US labs, secured by legal mandate and the compliance infrastructure the controls built. Commercial and developer markets belong to whoever prices cheapest, and that position is now locked. What US policy produced is exactly two markets, separate and distinct, with no path between them under the current control regime.