The changing economics of artificial intelligence

The rapid rise of Chinese open-weight AI models is changing the economics of the global artificial intelligence industry. Models from companies such as DeepSeek, Alibaba, Z.ai and Moonshot are increasingly approaching the capabilities of leading Western systems, while offering substantially lower prices. This has prompted the question: if comparable AI capabilities can be delivered at a fraction of the cost, what happens to the business model built around expensive frontier systems?

The answer is more complicated than a simple contest between cheap Chinese models and expensive Western ones. And perhaps the more significant development is that the AI market itself is beginning to divide into different tiers, with different models suited to different kinds of work.

What is an Open-Weight Model? It is an AI model whose core components are publicly released, allowing anyone to download it. This lets users run the model on their own computers, study how it works, and modify it for their own specific needs.

Eric Benoist and Rita Boutros explore the rise of workload-specific AI strategies, and the implications arising from this shifting landscape.

Eric Benoist Tech & Data Research Specialist

Rita Boutros Tech & Data Research

Why Chinese models are becoming so competitive

Driven by intense domestic competition, Chinese developers are leading a shift toward highly efficient AI by finding clever ways to make large language models (LLMs) faster, smaller, and cheaper to run. Instead of relying solely on massive computing power, they are using breakthroughs like "efficient attention" and "mixture-of-experts" architectures, which act like smart shortcuts, allowing the AI to process long texts with less memory and activate only the specific "specialized" parts of its brain needed for a given task. And by adopting "multi-token prediction" to generate multiple words at once and "quantization" to compress the models into lighter formats, developers can run these powerful systems on standard, less expensive hardware without sacrificing performance. Ultimately, these innovations are turning AI from a resource-heavy luxury into a highly streamlined, cost-effective tool for everyday use. But these techniques are not exclusively Chinese. Many originated in, or were significantly advanced by, research and technology ecosystems outside of China. The importance lies partly in how effectively Chinese developers have combined and deployed them at scale. There is also a second factor at play: distillation – which is already a standard technique across the AI industry. This allows a smaller model to learn from the outputs of a more capable one, reducing the resources required to reproduce some of its capabilities. Government support adds another layer. Investment funds, computing subsidies and other forms of public financing reduce some of the costs of AI development and deployment. Together, these factors have helped Chinese developers compete aggressively on price. The gap between the release of leading Western models and competitive Chinese alternatives has narrowed from months to weeks, sometimes days.

The cheapest model is not necessarily the cheapest AI

Comparing AI costs by looking only at 'token' prices (price per word) is an important part of the story, but not the whole story. An open-weight model may be free to download, yet expensive to operate. Self-hosting a very large model requires GPUs, networking, electricity and specialist engineering resources. For smaller models and sufficiently high workloads, this may make economic sense. For near-frontier models, the infrastructure requirements can run into millions of dollars. Using an API avoids much of that upfront investment, but introduces usage-based costs and other considerations. Licensing terms can also affect the economics, particularly for large commercial deployments. But most importantly, the value of a model depends on what it is being asked to do. For straightforward tasks such as summarizing documents, classifying information or extracting data, a large difference in token price can translate relatively directly into savings. More complicated AI applications are different. An agent carrying out a multi-step task may need to make repeated decisions, call on external tools, process information and respond to errors. In these workflows, a small difference in reliability at each individual step can compound into a much larger difference in the likelihood of successfully completing the overall task. This means that the relevant measure is increasingly not the cost of a token, but the cost of successfully completing a task.

On some benchmark measures, the difference between leading Chinese open-weight models and Western frontier models is relatively small. But benchmark performance does not necessarily capture how models behave in complex, sequential workflows. When a task involves many interconnected steps, reliability becomes more important. A slightly less capable model can generate additional errors that then require correction further down the workflow. This creates an economic case for paying a premium for frontier models when the cost of failure is high. For advanced research, complex software development or other high-stakes applications, the cost of an error may be much greater than the additional cost of using a more capable model.

AI is becoming an ecosystem, not just a model

Developers increasingly choose an AI product rather than simply choosing a model. Coding tools, for example, combine a model with systems for managing context, editing files, running tests, using tools, handling errors and retrying tasks. Over time, these environments accumulate integrations, workflows and user knowledge. That creates switching costs. A cheaper model may be technically competitive but still struggle to displace an established tool if it does not offer the same surrounding ecosystem. Trust is another consideration, particularly for businesses handling sensitive information. Data residency, security, governance and model behavior can all influence adoption. Open-weight models can address some of these concerns by allowing organizations to run models within their own environments, but the broader questions of trust and governance are still relevant. These factors help explain why lower prices have not simply translated into the wholesale replacement of existing AI platforms.

Is intelligence becoming a commodity?

At one end of the spectrum, AI capabilities are becoming cheaper and more abundant. For many everyday tasks, organizations may no longer need the most advanced model available. If an open-weight system can perform the task adequately at a much lower cost, the economic argument for using a frontier model becomes weaker. At the other end, highly capable intelligence remains difficult to substitute. Complex, multi-step and high-stakes tasks can still benefit from the strongest available models, particularly where reliability and quality matter more than the cost of individual tokens.

The result of this could be a two-tier market. The first tier consists of increasingly commoditised AI for the broad mass of everyday workloads. The second consists of premium intelligence for applications where additional capability generates enough value to justify the cost. This does not necessarily mean that either tier will permanently belong to one geography or group of companies. The techniques driving China’s cost advantage are increasingly visible and replicable, while Western laboratories are investing heavily in efficiency themselves.

The emerging model may be less about choosing one AI provider and more about deciding which model is appropriate for each job. A single system could route routine, high-volume tasks to a lower-cost open-weight model while sending complex or sensitive workloads to a frontier system. The choice could depend on task complexity, required confidence, data sensitivity or other factors. This would shift some of the value away from the model itself and towards the technology that decides which model should do what. It could also change how AI companies charge for their services. Per-token pricing makes different models easy to compare on cost and encourages AI to be viewed as a commodity. Outcome-based pricing – or charging for a completed task or result rather than the underlying computation — offers another possible model, although measuring and attributing the value of individual AI tasks remains difficult.

The current competitive landscape is less a simple race with a single finish line than a rapidly changing market.

Chinese open-weight models have demonstrated that significant AI capability can be delivered at very low apparent cost. Their advantage is supported by technical efficiency, intense competition and substantial state backing, but some of those foundations may prove difficult to sustain.

Western frontier providers continue to have important advantages in capability, developer ecosystems and trust, but maintaining the frontier requires enormous and continuing investment.

The practical implication is that AI strategy is likely to become increasingly workload-specific. The relevant question is not simply which model is the most capable or which is the cheapest, but which combination of capability, cost, reliability, deployment and control delivers the best result for a particular task.

The AI market may be moving towards a world in which intelligence is partly a commodity and partly a premium service. As the cost of basic capability continues to fall, the value of AI may increasingly migrate towards the products, workflows and services built around the models themselves.