The manufacturing boss stood on a stage in Beijing and did something unusual: he described both the gains and the doubt in the same breath. AI had produced ¥2.5 billion (≈ US$347 million / HK$2.72 billion, approximate) of comprehensive benefit inside TCL, he said. Then he warned that thousands of AI companies might not survive, because spending on technology does not automatically turn into revenue.
Li Dongsheng (李东生), founder and chairman of TCL, was speaking at the 17th Caixin Summit (财新峰会) on 12 November 2026, where he was a co-chair. His theme was that Chinese manufacturers have two routes out of pressure: applying AI and globalising. The remark that drew attention was the qualification, not the boost.
The three opportunities he named
Li described three ways AI is changing industry.
- AI as a competitive variable. He pointed to very large global investment in AI and argued that even companies outside technology must engage with it because the spending creates spillover across supply chains.
- Product and business-model change. He cited intelligent, personalised terminals — AI phones, AI glasses and embodied-intelligence (具身智能) devices — as categories opening new markets rather than merely adding features.
- Efficiency and organisational change. He said AI can improve market insight, user analysis, content production and customer service, and can cut demand-forecast error by 20 to 50 percent.
The third point is the one he stressed most for a manufacturer. AI is not a separate product line; it changes how a company plans, builds and ships.
What TCL says it got
Li offered figures from inside his own company. TCL has generated ¥2.5 billion (≈ US$347 million / HK$2.72 billion) in cumulative comprehensive benefit from AI over the past several years, through smart scheduling, predictive maintenance, machine-vision inspection and supply-chain coordination. He said the next step is copying mature applications across more factories rather than leaving them as pilots.
In research and development, TCL Huaxing built a vertical large model (大模型) for the display industry called Xingzhi (星智). Li said it raised software R&D efficiency by 30 percent, device R&D efficiency by 75 percent and mechanical simulation efficiency by 90 percent. The model is domain-specific, which is a deliberate choice: TCL did not try to build a general foundation model.
He also described an internal contest between "silicon-based employees" (硅基员工) and "carbon-based employees" (碳基员工). One group of 30 engineers worked alone; another group of six worked with a large model. Accounts of the first round differ — one report says the model-assisted group lost, another says it was faster but the human group produced better quality — but they agree that the second round ended in a tie. Li said he expects human-plus-model design to surpass human-only design as models improve. TCL says it has trained 500 "silicon-based employees" internally.
The warning, precisely stated
Li's caution had two parts. First, AI investment does not necessarily become commercial value. He pointed to the thousands of AI-related companies founded in China in recent years and asked how many would still be operating in two years unless they converted spending into a sustainable business. Second, cross-border AI use raises the compliance bar. For a company with global operations, data security, privacy, intellectual property and local rules all become part of the deployment cost.
His conclusion was a phrase that translates roughly as "push AI toward the real economy" (AI向实): embed the technology in actual business processes, turn pilots into repeatable methods, and measure the result.
That framing is more demanding than it sounds. A pilot can be run by an enthusiastic team with a budget. A repeatable method has to survive a different factory, a different manager and a different quarter.
Why a manufacturer's view matters
AI coverage is dominated by model labs, chip makers and cloud providers — companies whose revenue is directly tied to AI spending. A television and panel manufacturer is a customer, which makes its accounting more interesting.
When Li says ¥2.5 billion of benefit, he is describing an internal estimate of efficiency gains, not a new revenue line. That is the right way for a manufacturer to think about AI, but it is also a figure that is difficult for an outsider to verify. Efficiency is measured against a counterfactual: what the factory would have cost without the system. Counterfactuals are easy to draw favourably.
His scepticism about AI startups and his defence of AI in manufacturing are not contradictory. They are the same argument from two sides: value comes from application, not from the existence of the technology.
Honest limitations
All figures — ¥2.5 billion in benefits, 30/75/90 percent R&D efficiency gains, 20 to 50 percent forecast-error reduction and 500 trained "silicon-based employees" — are Li Dongsheng's statements at the summit, reported by Chinese financial media. They are not audited metrics, and no methodology was published.
The design contest is an internal anecdote, not a controlled study; reports disagree about the first round, and this article says so rather than choosing one version. The phrase "comprehensive benefit" is the company's wording and should not be read as net profit or revenue.
The speech was delivered at a business conference, not a scientific or regulatory setting. Nothing here is investment advice, and the article does not evaluate TCL's securities. Currency conversions use an approximate rate of US$1 to ¥7.2 and HK$1 to ¥0.92.
What readers can do now
- If you manage a company, ask for the denominator. An efficiency gain is only meaningful alongside what was spent to achieve it and how the baseline was set.
- If a vendor promises transformation, ask whether the pilot has been copied to a second site. Li's own argument is that replication, not the first success, creates value.
- If you are evaluating an AI project, write down the compliance cost at the start. For cross-border operations it is part of the business case, not an afterthought.
