If you believed that artificial superintelligence was just around the corner, would you tell your laboratories to pass on advanced American chips? That was Alvin Graylin’s test for what a government really believes in an August 18 episode of Moonshots with Peter Diamandis. Chinese officials, he said, “are not behaving like they believe that ASI is around the corner.” He pointed to instructions against buying H-200 chips and to rules that constrain public AI services.
The hosts pressed the opposite possibility: what if China is underestimating how quickly AI could become powerful enough to transform geopolitical competition? Graylin’s answer was that most Chinese institutions treat AI as a general-purpose technology that will take years, potentially decades, to spread through the economy—not a finish line that one laboratory crosses to win permanently.
That distinction runs through his May essay, “Misdiagnosing the U.S.–China AI Race”. Artificial general intelligence, or AGI, means broadly capable AI; superintelligence would exceed human capabilities. Graylin questions the assumption that reaching either first guarantees lasting dominance. His alternative measure of advantage is more practical: whose models get adapted, whose industries use them, and whose developer communities keep growing?
Longer timelines, concrete adoption targets
Graylin contrasted the imminent-AGI expectations he sees in parts of Silicon Valley with Chinese regulators’ willingness to retain restrictions. He acknowledged exceptions among Chinese laboratories pursuing AGI, but said most institutions behave as though society needs time to absorb the technology.
China’s 2023 interim generative-AI measures do combine development with controls. They cover politically prescribed content restrictions, lawful training-data sources, privacy and protection against minors’ excessive dependence. Their scope is services offered to the public within China, not every model made by a Chinese developer. Security assessment and algorithm filing are required for services with public-opinion or social-mobilization capabilities; research and applications not offered to that public fall outside these particular measures.
Graylin reads the continued restrictions as evidence of a longer institutional timeline. The hosts questioned whether rising capabilities might eventually force Beijing to abandon those constraints. He rejected the idea that Chinese officials simply do not understand superintelligence, pointing to safety researchers and discussions that inform regulators.
The more concrete expression of his deployment argument is China’s AI Plus policy. The State Council portal’s August 2025 summary names six application areas: science and technology, industry, consumption, well-being, governance and international cooperation. It also calls for stronger models, data supplies, computing capacity, open-source ecosystems and talent development.
Its numerical targets are narrower than Graylin’s description on the podcast. They call for penetration of next-generation intelligent terminals and AI agents—devices and systems that can carry out tasks—to exceed 70% by 2027 and 90% by 2030. They are not targets for the percentage of companies adopting AI, nor measurements of adoption already achieved.
Graylin characterized America’s approach as focused on supplying the best models and chips without enough attention to what happens afterward. But America’s AI Action Plan, released in July 2025, explicitly addresses adoption. It identifies organizational distrust, regulatory complexity and unclear governance as barriers, and proposes regulatory sandboxes—controlled settings for testing new applications—and sector-specific standards. It also recommends support for small businesses using open models, workforce retraining and federal deployment.
The contrast is therefore one of emphasis, not deployment versus no deployment. Both plans address the technology’s supply and its use; Graylin argues that spreading it through industry deserves more weight in how competition is judged.
Provincial champions, not a single command
To explain how a national priority turns into business activity, Graylin took apart the image of a leader simply ordering laboratories to produce AGI.
In his account, central plans set broad directions, such as clean energy, robotics and AI. Provincial and city governments then look for local companies that can advance those goals, offering stipends, recruiting help and other benefits. The result is competition among “provincial champions and city champions,” rather than one centrally prescribed technical approach.
A host supplied an entrepreneurial version of the same story. His son had returned from China after asking entrepreneurs what mattered most to business success. Rather than teamwork or a business plan, the answer he heard was knowing what the government would focus on next. Graylin likened working against that direction to swimming upstream.
National priorities, in this telling, shape which opportunities receive support. Companies and local governments still compete over who can deliver them.
Open weights let adoption travel
Graylin described China’s open-model movement as an “emergent strategy,” not an original government order. Drawing on conversations with friends at DeepSeek, he said its leadership chose to share its model without being instructed or funded by the government to do so. Officials initially questioned giving away such valuable technology; the company’s international recognition then helped turn the decision into something worth celebrating and copying.
He also acknowledged that successful closed models coexist with open ones in China. His account of how official attitudes changed is distinct from the documented endorsement: in his July 2026 World AI Conference keynote, Xi Jinping advocated open-source development alongside monitoring, emergency response and continued human control.
The practical mechanism starts with weights: the learned numerical settings that make a trained model work. Releasing them lets other developers download, adapt and run the model, subject to its license. Open weights do not necessarily mean every ingredient of training is available or that the release meets a full open-source definition.
For Graylin, that distribution model solves two problems. Outside developers can modify and improve a laboratory’s work. And overseas operators can supply the hardware needed for inference—the computation performed whenever a trained model answers a request.
A cloud provider abroad can host a Chinese-developed model on its own machines. The originating laboratory can gain users without financing every server that serves them. The hardware and operating costs do not disappear; more of them are borne by the organizations deploying the model. Users, meanwhile, gain an alternative to relying on a single provider’s hosted service.
Alibaba’s Qwen family offers an indicator of that reuse. Hugging Face’s summer 2026 report examines Hub activity during January–July 2026, with edits incorporating early-August releases. It reports roughly 180–210 new Qwen derivative repositories a day: developers publishing versions built from existing models, rather than merely watching new launches.
The report also notes continuing U.S. strength in smaller models and embedding models, which represent information numerically for uses such as search. Its download figures measure activity on Hugging Face, not unique users, model quality or total commercial market share. Private deployments, APIs and other distribution channels sit outside that measure.
Restrictions that also create incentives
Graylin’s sanctions argument has two sides. He explicitly said chip export controls have slowed China and made laboratories’ work harder. Labs he visited described inadequate computing resources as their biggest problem, and he said shortages constrain how many users Chinese services can support.
But he argues that scarcity also pushes algorithmic efficiency: finding ways to achieve useful results with less computation rather than relying mainly on larger computing budgets. Open releases extend that effort by letting outside researchers adapt the models and overseas operators host them.
He sees a parallel incentive in domestic hardware. Shortly after restrictions on high-end American chips, he said, semiconductor-industry contacts received calls offering government funding, resources and help finding customers. Chinese GPU executives told him that products previously passed over for being less efficient and generations behind suddenly had buyers.
His conclusion is not that restrictions imposed no cost. It is that their immediate constraints also strengthened incentives to improve domestic alternatives—a causal interpretation he draws from those conversations, rather than a measured net assessment of the controls.
Asked how the American AI ecosystem could make itself indispensable, Graylin proposed high-quality open-source releases. That would let it compete for the same developers and operators, not just for the highest model scores. In his account, advantage can accumulate when an overseas provider chooses a model to host or a developer chooses one to adapt—without either waiting for anyone to declare AGI.