Opinions expressed by Entrepreneur contributors are their own. For the past few years, the AI industry has operated on a fairly simple assumption: Better models would create more valuable companies. The enormous investment going into chips, data centers and foundational models has largely been built around that idea. China is starting to test the assumption.
In October, Reuters reported that China’s STAR50 technology index had fallen 40% in less than four months, wiping out around $480 billion in market value. AI-related stocks had grown to represent 42% of Chinese market capitalization and 57% of trading turnover across mainland China and Hong Kong. Some of this is simply what happens when valuations run ahead of financial performance. But I think there is a more interesting change happening underneath the market correction.
AI is continuing to become more capable while the cost of accessing that capability is falling and competition between model providers is increasing. For businesses using AI, that is very good news. For every company investing billions to build the underlying technology, the economics may be more complicated. The AI industry could create enormous economic value without the companies building foundational models capturing most of it.
The pace of model development is still extraordinary, but the gap between the strongest models is getting smaller. The 2026 Stanford AI Index found that Anthropic, xAI, Google and OpenAI were all within 25 points of each other on the Arena leaderboard in March. Chinese models are also closing the gap, with Stanford describing the performance difference between leading U.S. and Chinese models as having “effectively closed.” That matters because when several providers can complete the same task well, the buying decision starts to change. Businesses begin looking beyond which model is technically the best.
Cost matters. Reliability matters. Speed matters. The ability to use a model inside an existing workflow matters.
I have seen this change in how we approach AI internally. When we started building more AI agents and automations, it was easy to focus heavily on the model. We would look at which model produced the best output and build around it. As the technology has improved, the question has become much more practical: What is the cheapest reliable way to complete this particular task?
The best model for analyzing a complex problem may be unnecessary for extracting information from a document. A cheaper model may work perfectly well for a repetitive classification task. Some workflows can even use several models, depending on what needs to happen at each step. Once models become components rather than the product itself, the economics of AI start to look different.
There is a tendency to assume that the companies responsible for the biggest technological advances will also capture most of the financial value those advances create. Technology markets do not always work that way. Consider what happens when the cost of completing an AI task falls significantly. A customer service team can process more conversations.
A developer can run more coding agents. A marketing team can analyse more campaigns and build more experiments. A small business can automate administrative work that previously required another employee or an external service. The economic value of all those activities can increase even while the amount paid to the model provider for each task decreases.
This is why aggressive competition between model providers is so important. As capabilities converge and prices fall, some of the value created by AI moves downstream to the companies applying it. For a business owner, cheaper models can mean better margins or the ability to automate a process that was previously too expensive. For a software company, it can make an AI feature viable for thousands of smaller customers rather than only large businesses.
The technology improves while the cost of using it falls. That can be great for the AI economy without necessarily being great for every company building AI infrastructure. This also changes where founders should look for defensibility. A model can have a significant technical lead and lose part of that lead within a few months.
Building a long-term company around access to one particular model therefore looks increasingly difficult. Customer relationships are harder to replicate. So are distribution, proprietary context and a deep understanding of how work is actually performed. This is especially relevant in business software.
Most customers are not trying to buy artificial intelligence. They are trying to get something done. They want to respond to a customer, prepare a proposal, reconcile an expense, schedule employees, create a campaign or understand why revenue changed. The model is only one part of delivering that outcome.
This is similar to a shift I have seen as we have become more AI-native as a company. The useful part is rarely the model in isolation. The difficult work is connecting AI to the right data, giving it enough context, building the workflow around it and deciding where a person still needs to intervene. That business context becomes increasingly valuable as the underlying intelligence becomes easier to access.
A company that already understands a customer’s workflow can potentially apply each improvement in AI much faster than a new entrant that first has to understand the customer. I think that is where some of the more durable advantages in the application layer will come from. There is another issue that founders building AI products need to understand early: Usage now has a cost. Traditional SaaS created an attractive economic model because the marginal cost of another customer using a feature was usually very small.
Whether somebody opened a dashboard five times or 50 times did not dramatically change the infrastructure economics. Generating documents, analyzing conversations, reasoning across data or running autonomous workflows consumes compute each time the task is performed. This means a customer using an AI product heavily can also become significantly more expensive to serve. It changes what good growth looks like.
A team can celebrate that customers generated one million AI actions this month, but that number has little meaning on its own. If every action costs more than the value it creates, greater adoption simply increases the problem. When I look at an AI workflow now, I am increasingly interested in cost per successful outcome. How much does it cost to complete the task?
How often does the AI succeed without intervention? How much human work remains? Does the task create revenue, save meaningful time or improve the customer experience? Those are much more useful questions than how many prompts or tokens were consumed.
This also puts pressure on the subscription model that has dominated software for the past two decades. SaaS largely trained customers to pay for access. Choose a plan, pay a monthly fee and use the software. AI introduces a meaningful variable cost into parts of the product.
That does not mean every software company needs to abandon subscriptions. But it does mean businesses need a clearer understanding of the relationship between usage, customer value and cost. For some products, a traditional subscription may still work perfectly well. Others may move towards credits, usage allowances or a combination of subscription and consumption pricing.
What matters is that pricing and infrastructure economics cannot be separated. If heavier usage makes a product less profitable, the pricing model eventually has to reflect that. If inference costs continue to fall, the opposite can happen: Previously expensive capabilities can be included more broadly, making the product more useful without damaging margins. This is one reason falling model prices could have a much bigger effect on software than many people expect.
The correction in China does not make me less optimistic about AI. If anything, it reinforces why I think the opportunity for companies applying AI is becoming larger. More competition between foundational models should produce better models, lower prices and more choices for businesses building on top of them. Stanford’s data already shows the performance gap narrowing, which means competition increasingly moves towards cost, reliability and usefulness in specific domains.
Founders should design around that direction of travel. Avoid making a single model the entire competitive advantage. Understand the workflows where the business has genuine expertise. Measure the cost of completing tasks rather than simply celebrating AI usage.
Build systems that can take advantage of cheaper or better models when they become available. Most importantly, focus on what happens after the model produces an answer. The biggest business opportunities may not sit inside the model at all. They may sit in the customer relationship, the proprietary context, the workflow and the distribution required to turn intelligence into something useful.
That is why China’s correction is worth watching even for founders who have no interest in Chinese technology stocks. The important signal is not that AI is losing value. It is that creating value with AI and capturing that value are two different things. As intelligence becomes cheaper and more widely available, that distinction will become increasingly important for anyone building a business around it.
For the past few years, the AI industry has operated on a fairly simple assumption: Better models would create more valuable companies. The enormous investment going into chips, data centers and foundational models has largely been built around that idea. China is starting to test the assumption. In October, Reuters reported that China’s STAR50 technology index had fallen 40% in less than four months, wiping out around $480 billion in market value.
AI-related stocks had grown to represent 42% of Chinese market capitalization and 57% of trading turnover across mainland China and Hong Kong.
