In less than four months, China's AI stock boom went into reverse. The STAR 50 index of Chinese technology shares fell about 40% from its peak, wiping out roughly US$480 billion in market value by late October 2026. Headlines quickly declared that "the AI bubble has burst".
For Australian business leaders deciding whether and how to invest in AI, it's tempting to read this as a reason to wait. That would be the wrong lesson. China's correction says a lot about where AI value really comes from, and very little about whether AI is worth adopting. This article explains what happened, why it happened, and the practical lessons for any business planning AI projects in 2026 and beyond.
The short answer
China's AI stock market bubble has deflated: investors are repricing companies whose valuations ran far ahead of their earnings. China's AI technology has not collapsed, and investment in it is still growing. The gap between the two is the real story: building and adopting AI turned out to be easier than making money from it. For businesses, the lesson is to invest in AI that solves specific, measurable problems, not in hype.
What actually happened
The STAR 50, which tracks 50 major companies on Shanghai's tech-focused STAR Market, fell around 40% from its peak in under four months. Its valuation dropped from nearly 250 times trailing earnings to about 100 times: a big reset, but still expensive. The broader CSI 300 index was about 20% below its June peak.
The sell-off hit companies across the AI supply chain:
| Company | Business | What happened |
|---|---|---|
| Unitree Robotics | Humanoid robots | Shares roughly halved after a heavily oversubscribed IPO |
| MetaX | AI chips (GPUs) | Fell sharply after a debut surge of around 700% |
| Moore Threads | AI chips (GPUs) | Fell sharply after a debut surge of around 400% |
| CXMT | Memory chips | Down more than 25% in a month |
| Zhongji Innolight, Eoptolink | Optical parts for data centres | Around 60% below their peaks, amid fears of US restrictions |
Two things are worth keeping in mind. The US$480 billion is lost market value, not cash leaving these companies' bank accounts. And the STAR 50 is a stock index, not a measure of the whole Chinese AI industry. As one fund manager put it, the correction will "put a stop for many of the companies trying to get themselves listed", which shows the impact falls hardest on fundraising, not on the technology itself.
Why the bubble deflated
1. Valuations ran ahead of earnings
Much of the boom was priced on future potential: Chinese firms becoming world leaders in models, chips, robots and data centres. When revenue and margins didn't arrive fast enough, those prices became hard to defend. A company can have impressive technology and still be overvalued.
2. Infrastructure spending is soaring
Chinese AI infrastructure spending is projected to reach about ¥932 billion (US$139 billion) in 2026, up 103% on the previous year, and more than ¥1.2 trillion in 2027. The strain shows in cash flow: Alibaba, Tencent and Baidu went from combined free cash flow of about ¥170 billion in 2025 to negative ¥16 billion in the first half of 2026.
3. Users are easier to win than revenue
China's six leading AI firms earn roughly 10% of the combined AI revenue of OpenAI and Anthropic. Consumer AI apps attract huge user numbers but few paying customers. The most telling figure: in the first nine months of 2025, MiniMax's consumer AI apps had a gross margin of just 4.7%, while its enterprise and developer platform earned 69.4%.
4. Open models and price wars squeeze margins
Many Chinese developers release open-weight models that anyone can host. That spreads adoption fast, but undercuts the developer's own pricing. After Z.ai released the weights of one of its models, its share of that model's traffic on the OpenRouter platform fell from 88% to 22% in about 20 days as other providers competed on price.
5. Chips and capacity add risk
US export controls limit China's access to advanced chips and equipment, raising costs and uncertainty for hardware makers. At the same time, there are long-running concerns that some government-backed data centres were built faster than demand for them.
A stock bubble, not a technology collapse
While share prices fell, investment in the technology kept rising. China's AI infrastructure spending is projected to keep growing sharply into 2027, and the state remains heavily involved: since 2023, government guidance funds and investment platforms have supplied 47% of equity investment in Chinese AI chip and server companies, with mostly state-owned banks adding another 14%.
In other words, investors are repricing expectations, while companies and governments keep building. Weaker start-ups may merge, shrink or close, and the survivors will be judged on revenue rather than announcements. That's a healthy correction, and one that matters well beyond China: China's AI infrastructure spending is still only around 15–20% of US levels, and the same question of whether returns can justify the spending is being asked of AI investment everywhere.
Five lessons for Australian businesses investing in AI
1. AI is getting cheaper, so use that to your advantage
Intense competition and open-weight models are pushing down the price of AI. Chinese developers now struggle to hold their pricing power, and that pressure ripples through the global market. For businesses buying AI rather than selling it, falling model and API costs make more projects viable. The cost that matters most is no longer the model; it's integrating AI properly into your systems and workflows.
2. Value lives in specific, practical use cases
MiniMax's numbers are a lesson for every buyer: a 4.7% margin on broad consumer apps against 69.4% on enterprise and developer services. AI pays off when it does a defined job for people who rely on it, such as triaging customer requests, moving data between your CRM and ERP, drafting documents for review, or flagging exceptions in financial data. Start from a workflow that costs you time or money today, not from the technology.
3. Expect many AI projects to fail, and design yours not to
The market correction mirrors what analysts expect inside businesses. Over 40% of agentic AI projects are forecast to be cancelled by the end of 2027 because of escalating costs, unclear business value or weak risk controls. The same forecast estimates that only about 130 of the thousands of vendors claiming "agentic AI" offer genuine agent capabilities. Define the business outcome, the cost and the success measure before you build.
4. Don't lock yourself into one model or provider
The AI market is volatile: models improve monthly, prices move, and providers rise and fall. Z.ai's experience shows how quickly traffic shifts to whoever offers the best price. Design your AI integrations so the underlying model can be swapped without rebuilding your application. A model-agnostic architecture protects you from price changes, outages and providers that don't survive the shakeout.
5. Treat data, privacy and governance as part of the build
Sending business or customer data to an AI model hosted overseas can create obligations. Under Australian privacy law, disclosing personal information to an overseas recipient means taking reasonable steps to ensure it's handled properly, and you can remain accountable for breaches. Know where each model is hosted, what data it receives and what the provider's terms allow. Government guidance on adopting AI safely, released in October 2025, sets out six practices and offers AI policy and AI register templates to help.
Hype signals vs value signals
Use this checklist when assessing any AI project or vendor:
| Hype signal | Value signal |
|---|---|
| "It uses the latest model" | "It saves X hours a week on this specific process" |
| User numbers and demos | Measured cost savings, revenue or error reduction |
| Locked to one AI provider | Model can be swapped as prices and quality change |
| Vague on where data goes | Clear data flows, hosting locations and privacy controls |
| Big-bang rollout | Small pilot with a success measure, then scale |
| "Agentic" branding on an old chatbot | Real automation of multi-step work, with human approval where it matters |
Start small, prove value, then scale
The correction in China rewards a simple discipline: prove value before you scale spending. For most businesses that means picking one workflow, building a focused AI agent or integration for it, measuring the result over a few weeks, and only then expanding.
This is exactly how we work at Black Vault. Our AI agent development and AI integration services start with a specific business problem, connect to the systems you already use, and are designed so the underlying model can change as the market does. And because the best way to judge any development partner is on real work, we'll deliver an initial portion of your project free of charge. You see a working result on your own data and processes before committing budget. Talk to us about a practical AI project.
If you're still weighing build versus buy for AI, our guide to custom software vs off-the-shelf walks through the decision.
This article is general commentary on market developments and is not investment advice.
Frequently asked questions
Click a question to see the answer.
Has China's AI bubble burst?
China's AI-linked stock market bubble has deflated sharply, but its AI industry has not collapsed. In October 2026, the STAR 50 technology index had fallen about 40% from its peak in under four months, erasing roughly US$480 billion in market value, and its valuation dropped from nearly 250 to about 100 times trailing earnings. At the same time, Chinese AI infrastructure spending is projected at about US$139 billion in 2026, up 103% on the previous year, with state-linked funding still central. The most accurate description is a correction in financial expectations alongside continued investment in the technology.
Why did Chinese AI stocks fall?
Chinese AI stocks fell because valuations had risen far faster than earnings, and investors began to doubt how quickly companies could turn technology and users into profit. Infrastructure costs are soaring: Alibaba, Tencent and Baidu's combined free cash flow swung from about ¥170 billion in 2025 to negative ¥16 billion in the first half of 2026. Monetisation has been difficult, with estimates suggesting that China's six leading AI firms earn about 10% of the combined AI revenue of OpenAI and Anthropic. Price competition from open-weight models, US chip export controls and concerns about excess data-centre capacity added further pressure.
Does China's AI correction mean the global AI bubble will burst?
Not necessarily. China's correction shows what happens when valuations run ahead of revenue, and the same question about whether returns can justify heavy AI spending is being asked in the United States and elsewhere. But China's situation has specific features: heavy state involvement in funding, US restrictions on advanced chips, and fierce price competition driven by open-weight models. A correction in one market does not prove another will follow. For businesses, the more useful conclusion is that AI investments are increasingly judged on measurable returns rather than potential.
Should my business delay investing in AI because of the AI bubble?
Generally, no. A stock market correction affects investors in AI companies, not the usefulness of AI for businesses that apply it well. In fact, competition is pushing down the cost of AI models and APIs, which makes more practical projects affordable. What the correction does suggest is investing carefully: choose specific workflows with measurable value, start with a small pilot, avoid locking into a single AI provider, and build in privacy and governance from the start. Projects designed this way are far less exposed to market swings than speculative, large-scale AI initiatives without a clear business case.
What kind of AI projects deliver real value for businesses?
The AI projects that deliver value solve specific, repeatable problems that already cost time or money. Common examples include triaging and responding to customer enquiries, moving data between systems such as CRMs and ERPs, extracting information from documents and invoices, drafting routine content for human review, and flagging anomalies in operational or financial data. China's experience supports this: MiniMax earned a 69.4% gross margin on enterprise and developer services but only 4.7% on consumer apps. Successful projects define a clear outcome, integrate with existing systems, keep a human in the loop where decisions matter, and measure results before scaling.
How can I reduce the risk of an AI project failing?
Start with a clearly defined business problem and a measurable success target, such as hours saved or errors reduced. Run a small pilot on real data before committing larger budgets. Design the system so the AI model can be replaced as prices and capabilities change, avoiding dependence on one provider. Map where your data goes, including whether personal information is sent to models hosted overseas, which can trigger obligations under Australian privacy law. Use the government's published guidance on AI adoption to set up governance. Finally, choose a development partner willing to prove value on a small piece of real work first.
Want AI that pays its way? Black Vault will deliver an initial portion of your AI project free, so you can see measurable results on your own workflows before you commit. Start a practical AI project.