The approach to AI in the US and China is in stark contrast.
But the long-term competition may not be between American and Chinese AI models.
It may be between two visions of AI itself: intelligence delivered as a cloud service versus intelligence embedded throughout the physical economy.
The US focus #
In the US, there is enormous focus on frontier models, pushing their capabilities ever higher and selling access to those abilities as a service.
The dominant business model for US AI is as a centralized cloud service. Customers generally do not own, modify or independently deploy the model. Instead they purchase access to LLM capability that is hosted and continuously improved by the provider.
As frontier models become more capable, the assumption is that greater capability creates greater value.
In this business model, competitive advantage for US AI companies comes from developing the most capable cloud models, building ecosystems around them, and integrating them deeply into customer workflows. This helps maintain a technological lead over competitors.
The strategic question becomes whether value remains concentrated in the model itself, or whether it shifts towards the deployment of intelligence throughout the wider economy.
The rise of open-source #
Open-weight models are large language models whose model weights are publicly released, allowing others to run, adapt and build upon them.
Using open-source, companies can deploy their own LLM on-premise, in a private cloud where they can show accountabilty for data regulation within a legal jursidiction. This answers the question some companies have that are nervous of placing sensitive data under the control of a third-party AI provider operating in a different legal jurisdiction. Nations may increasingly seek sovereign AI capabilities, preferring models that can be operated under domestic legal control rather than relying entirely on foreign providers.
As of September 2026, many of the open-weight models can comfortably run on a mobile phone: an edge device. Companies that ship products at scale, millions of phones, cars and robots, do not want a large ongoing cloud bill behind every transaction.
A language model that runs at the edge, using little power, without the need for an internet connection and no per-query cost, is the more favourable economic model.
Moving models to the edge relies on a technique called quantization. Internally, the values that encode the model are represented with less memory. This makes the model smaller, faster, and cheaper to run, using less power, while retaining approximately 95% of the model's capability.
Through quantization and the shift to the edge, if model capability becomes commoditized, competitive advantage may increasingly come from how effectively intelligence is embedded throughout the economy.
Open-weight models potentially shift value away from the model itself and towards deployment, integration, customer relationships and real-world data collection. With commoditized capability, competitive advantage may increasingly come from where intelligence is deployed rather than who created it.
Historically, we have seen this with micro-processors, Linux, and Android.
As Linux became commoditized, value shifted from ownership of the operating system to the services, support, integration and operational expertise surrounding it. The specific Linux distribution became less important than the reliability, support, certification and operational expertise surrounding it.
Commoditization helped create companies such as Red Hat, whose success came not from controlling Linux itself but from helping organisations deploy it effectively.
Open-weight AI may exert a similar pressure on the AI industry. If model capability becomes widely available, value may increasingly shift towards deployment, integration, data, products and customer relationships rather than ownership of the underlying model and selling access.
Open-weight LLMs weaken the economic assumptions behind AI-as-a-service.
China #
If we look to China, we see a completely different approach to AI. Partly because China frames AI as a collective national project, not a private corporate endeavour, owned by a few companies. Overall, AI is seen as a product, not a threat. China's official policy framing places much greater emphasis on AI as an means for economic development, industrial upgrading and national competitiveness. China treats AI safety and content governance as matters for state supervision.
They do not see AI as something that might "wake up" and "go rogue".
China emphasises safety through governance, through strict model supervision, content filtering, and alignment with Chinese values. For the Chinese, AI must remain subordinate to state objectives: it is a powerful tool, but one that is manageable.
Chinese official policy language tends to emphasize AI's economic, industrial and governance applications rather than framing AI primarily through the question of whether machines will become human-like or autonomous.
Productizing AI #
The US has focused on models and monetizing them as software. China has emphasiszed cost-optimization and the rapid deployment of AI into the rest of the economy.
By 2024, China's manufacturing value add (MVA), the value that China adds to its manufacturing sector based on the value of manufactured output minus the value of its inputs, was 32% of global MVA. The US was about 15%.
Since 2015, China has experienced strong production growth while for the US it has been flat.
China produces an enormous number of product categories (electric-cars, batteries, solar, electronics, ships, industrial equipment, robots, steel, and machinery) and is often globally dominant. China produces 75% of global electric-car production, compared with 6% in the US.
And electric-cars are only one example. China is a huge producer of batteries, smartphones, appliances, drones, industrial equipment, robots and other products that can become increasingly software-defined.
More and more manufactured products are becoming software-defined products. Vehicles, industrial robots, appliances and drones increasingly derive value from software updates and digital capabilities rather than from hardware alone.
Theses products are the type that can take advantage of AI, and China has the skills and experience to productize technology into manufactured goods.
Adding AI to manufactured goods gives China a huge incentive to optimise the LLM capability/cost ratio as it matters enormously for physical products.
If putting a frontier-model AI into a car costs $1,000, that is one thing. But if you can get a comparable capability for $100 that customers are happy with and are used to, you can put it into vastly more products. And what if the cost was $10 or $1? Frontier AI maximizes capability. Product AI maximizes capability per dollar. Embedded product AI maximizes capability per dollar at massive scale.
But the opportunity does not stop with adding AI to products. Once added, the deployed AI can collect real-world data, resulting in an AI more finely tuned to product use, further lowering cost to access more products.
In the long-term, the competitive question becomes who can create the fastest feedback loop based on AI to product to data to improved AI.
A manufacturing ecosystem equipped with AI can therefore become a self-reinforcing system: AI improves products, products generate data, data improves AI, and improved AI enables better products.
AI may increasingly favour those organisations capable of taking smart advantage of large volumes of real-world operational data rather than those possessing only the most capable models.
For both countries, it is not all upside #
China's demographics, an aging population, shrinking workforce, and rising labour costs, all create enormous incentives for automation. For China, AI-powered products and manufacturing at scale may be a demographic necessity, not only an economic opportunity.
The US leads in the intelligence layer: AI models, software, chip design, cloud infrastructure, research and venture-backed startups.
In addition, the US and its allies retain important advantages in AI-chip design, semiconductor equipment and leading-edge manufacturing. China is investing heavily to overcome these constraints, but it has not yet eliminated them.
Can China compensate for restricted access to the most advanced compute through cheaper models, more efficient inference and much greater deployment?
One unknown is whether access to cutting-edge compute remains a decisive advantage. If it does, the US may preserve a substantial lead. If efficiency improvements reduce the importance of frontier hardware, manufacturing scale and deployment could become more important.
Also, AI is an energy-intensive technology, meaning that access to abundant and affordable electricity may become a strategic advantage alongside compute and manufacturing capacity. This may increase the importance of efficient models, particularly when AI is deployed at scale across millions of products rather than concentrated in a small number of datacentres.
Two AI journeys #
This creates two very different routes.
The American route is to make intelligence extraordinarily capable and sell it as a service. The Chinese approach is to make intelligence sufficiently capable, extremely cheap, and put it into everything.
If the value of AI is in the intelligence itself, the US models may prove extraordinarily powerful. But if much of the value comes from applying enough intelligence, in the most appropriate way, to the physical world, the manufacturing ecosystem with its feedback loop becomes part of the AI advantage. It may not be about who builds the smartest AI. It may be about who can turn AI into the most useful things, at the lowest cost, at the greatest scale.
The long-term competition may not be between American and Chinese AI models. It may be between two visions of AI itself: intelligence delivered as a cloud service versus intelligence embedded throughout the physical economy.
And open-weight models may move value from creating intelligence to deploying intelligence.
Read next: What Tech Executives Need to Know About Working With LLMs Why leaders must treat LLMs as probabilistic systems requiring new forms of governance, oversight, and accountability.
Related Articles #
- LLMs fail when asked to interpret and safely update existing systems. Their value lies in text creation tasks (research, summary, explanation, deep dives), not system modification
- Agentic systems generate around 1.2 production defects per day, requiring extra fix capacity to stay stable
- LLM-centred agentic systems will always make mistakes, just like human engineers. Human-authored tests are all that can stop an agentic quality drop
- In a delivery system reshaped by AI, metrics are the only defence against blind risk
- LLM workflows need context, but unmanaged context growth becomes expensive and degrades results. How to keep usage controlled and effective
If this was useful, you can get more pieces like it in the Phroneses newsletter.