Made in China AI: new tech, same old story?

When Beijing-based AI firm Zhipu released its open-source General Language Model (GLM) and it began competing with Anthropic’s Claude and OpenAI’s ChatGPT at a fraction of the price, a familiar question resurfaced. Are we watching a new kind of intelligence, or a very old story? Solar panels and electric vehicles are among the numerous Western inventions that China turned into cheap, plentiful goods, until their inventors could barely profit from them.

For American and European tech firms, like Microsoft or NVIDIA, the precedent is uncomfortable. Their stock market valuations rest on an assumption of scarcity: capable models being rare, costly and controlled by whoever trained them gives them an edge; a competitive advantage they can capitalise on financially.

Europe, which invented much of the solar industry and then watched its manufacturers disappear, knows what happens when that assumption fails. If a credible AI model can be downloaded for free, the premium on owning one evaporates.

The economics of deliberate loss: a case of déjà vu?

China’s strategy is not unfamiliar; we have watched it unfold twice. In both cases the science was Western, but the pricing was conquered by a method no private investor would choose: state-owned banks financed the factories regional party committees wanted, and those factories competed by selling below cost. In finance this is a negative net present value strategy, pouring in more money than you can expect to earn back.

Why would anyone do this? Because the decision-makers are not necessarily trying to make a profit. Andrei Shleifer and Robert Vishny showed that when politicians rather than shareholders steer a company, it serves political ends (employment, prestige, strategic dominance) rather than financial ones.

A private laboratory must earn a return or perish; a state champion need only serve its brief. When the two meet, the player who’s indifferent to losses sets the price. The result is abundance: plentiful supply in the form of alternative comparable models, low prices in the form of more flexible usage policies, and no lasting edge for the innovator, whose profit margins are suppressed. Considering that the two largest AI companies (Anthropic and OpenAI) are preparing for an IPO based on (currently negative) operating profits increasing significantly in the foreseeable future, this is a significant challenge.

Cheapness is also a factor in whether technology is adopted. American laboratories may keep their lead on the elite benchmarks (the standardised tests, such as the GLUE or the MMLU by which AI models are ranked) but those scores do not decide which system ends up embedded in the world’s software. Cost does. A model that is merely good enough and very cheap becomes the default others build upon, and defaults are hard to dislodge once products have formed around them.

But is the same story unfolding for AI?

The analogy deserves scrutiny. Innovation in solar panels and electric vehicles evolves in a discreet manner. New products equipped with better cell chemistry or a denser battery, compete based on other qualities, such as cost effectiveness.

In particular, the evolution of these technological inventions follows electronic engineer Gordon Moore’s 1965 observation (or Moore’s law) that the number of transistors on a chip roughly doubles every two years, so computing power climbs as unit costs fall.

Other technologies, like solar panels or EVs, follow similar curves, albeit at different rates, and see increasing innovation leading to throat-cutting competition and vanishing margins. Anything riding that kind of curve becomes a commodity: so plentiful and interchangeable that nobody can charge much for it.

Does AI ride the same curve?

The models behind artificial intelligence are loosely inspired by the brain: layers of “nodes” whose connections are adjusted until the network produces useful output. But one difference is decisive. An AI model learns once, during training, and thereafter applies what it has learned. Your brain rewires itself with every conversation; today’s AI does not. It might come close with algorithmic trading or with retrieval-augmented-generation, where a trained model adjusts its behaviour according to incoming data, but still, to make it cleverer, you must train it again with more data, more chips, more electricity. Ultimately, more resources.

A trained AI model is therefore a closed system: its abilities are determined the moment training ends. It may surprise us in how it recombines what it knows, but it is unable to pass its current level of performance without being rebuilt or retrained with more data. In a world where AI exhibits notable strategic advantages, like in defence, security, power distribution, etc., the race is on, and the demand for resources follows.

But is a techonology that tries to resemble the cognitive functions of a brain (AI), subject to Moore’s law, like a technology that exhibits no cognitive capabilities (batteries, for example)? Kurt Gödel’s incompleteness theorem offers a useful analogy here. Any closed system is a set of rules and, as such, it cannot prove itself from within. An AI model, in the current state of technology, is restricted in the same way, as it cannot widen its cognitive capabilities using its existing resources. To do so, it needs new resources (training, energy, computation) and, therefore, it is subject to a decay curve, similar to Moore’s law or the ones in solar panels and EVs, destined to meet the same fate.

But can AI deviate from that curve? Should AI models ever become evolutionary in nature (by learning as they process new information in a similar manner to the human brain), then the story changes: such a system compounds its advantage instead of competing for resources. This continuous evolution is the frontier that matters because the competitive advantage of the AI models would be based on which one learns better or faster and not on which one is cheaper.

Why it matters

Financially speaking, if the evolution of AI models remains discrete, its competitive edge will inevitably erode. A (state) champion willing to sell at a loss can push the price of “good-enough” intelligence toward its production cost and margins will fall as they did for solar cells. The implication is uncomfortable and liberating at once: durable returns will come not from owning another interchangeable model but from what is built on top – products, workflows and distribution that rivals cannot copy. As intelligence becomes cheaper, value migrates to whatever stays scarce. China leads the way here; automated drone deliveries (Meituan), driverless taxis (Baidu) or delivery robots (Alibaba) show that value emanates from efficient AI usage and not necessarily from how “smart” a model is.

This logic reframes the race. The industry’s instinct is to build ever-larger AI models burning ever more energy for marginal gains on a benchmark; yet sheer capacity is what abundance renders cheap. The more interesting prize lies with models that learn as they work, because no rival can download a moving target.

So, can AI become a commodity?

On today’s architecture, very probably – and the GLM moment is the warning shot.

The only way to stay ahead of abundance is to change the machine itself: to build intelligence that grows rather than a smarter, static model. Until then, the smart money might follow the example of solar panels: instead of financing the construction of better panels, money goes to how their use can benefit the real economy (energy independence, emissions reduction). Equivalently, instead of pushing for smarter models that might become abundant, profits will arise from smart applications using them, such as traffic control, medical screening, energy grid management, or anything that can create social benefits.

A weekly e-mail in English featuring expertise from scholars and researchers. It provides an introduction to the diversity of research coming out of the continent and considers some of the key issues facing European countries. Get the newsletter!