When the disruptor gets disrupted: How Chinese open-source AI is eating its own industry

It’s ironic to watch an industry disrupting itself. Artificial intelligence was supposed to be the great disruptor – the technology that would hollow out call centres, law firms, accountants and coding jobs. That might well have started to happen. But meanwhile, something stranger is happening.

AI is disrupting AI. The very companies that promised to remake every other industry are now getting a taste of their own medicine, courtesy of a wave of free, open-source models pouring out of China.

The process is nothing new. In the 1990s, Harvard professor Clayton Christensen noticed something odd about how great companies fail. It wasn’t bad management. It was good management – companies doing everything right, listening to their best customers, chasing higher margins. But that left them blind to cheaper, “worse” alternatives creeping up from below.

The pattern has repeated itself regularly. In the late 1960s, mini steel mills such as Nucor made cheap, unglamorous rebar – fashioned from steel rods embedded in concrete to give it tensile strength – while giants like US Steel stuck to premium products like sheet steel, until the mini-mills climbed upmarket and swallowed the industry.

Honda entered the US with small, “unmanly” motorcycles that Harley-Davidson didn’t bother competing with. Within five years, Honda had overtaken every US manufacturer, including Harley-Davidson, in motorcycle sales. Japanese compact cars, digital cameras, Netflix’s mail-order DVDs and Amazon’s online bookstore were all dismissed as inferior before they ate the industry incumbents alive.

The mechanism is always the same: a cheaper, initially worse product wins the low end of the market, improves faster than anyone expects, and eventually becomes good enough for everyone. By the time the incumbent notices, it’s too late to respond without cannibalising its own profits.

Singaporeans have lived through their own version of this. When GrabTaxi launched in 2012, it wasn’t attacking ComfortDelGro’s taxi business directly — it was just an app solving a small problem: the hassle of hailing a cab in the rain. The private-hire cars that followed felt like an inferior, slightly dodgy option – an inconsistent, unlicensed feeling; someone’s personal car. Nobody at ComfortDelGro probably lost much sleep over it.

A decade later, flagging a cab on the street felt almost quaint, and ComfortDelGro spent years playing catch-up with its own app and fleet. Many of its drivers also became Grab drivers. The niche had quietly become the market leader, which then expanded its offerings and spread across South-east Asia.

That’s exactly what’s happening in AI right now – except the incumbents being disrupted are themselves AI companies.

China’s open-source wave

For the past three years, US AI labs built their businesses on a simple premise: The best models live behind a paywall, and customers will pay a premium for the best. That premise is now under serious strain.

Chinese labs – DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu’s GLM – have flooded the market with models that are free to download, cheap to run and increasingly hard to dismiss as second-rate. A year ago, Chinese open-source models barely registered on global usage trackers. Today, they account for close to a third of all AI model usage on major aggregator platforms, and Alibaba’s Qwen alone has overtaken Meta’s Llama to become the most downloaded open-source model family in history.

The price gap is what makes this a textbook disruption story rather than just a competitive skirmish. Developers say that they pay about US$10 (S$13) for an hour of coding work on Claude, versus less than 50 US cents for comparable work on DeepSeek. Multiply that gap across an entire industry of cost-conscious developers and startups, and you get exactly the low-end entry point Christensen described – a product incumbents considered beneath them, quietly building a beachhead.

Most AI users – including the broad middle of the corporate world – don’t need to develop a blockbuster drug or a new semiconductor architecture or do complex risk analysis. As the analyst Stephen Innes has pointed out, they want AI to summarise documents, translate reports, search internal knowledge bases, draft customer responses, classify data, assist with compliance, help with routine coding or automate a repetitive workflow. For them, “good enough” is good enough.

But just like the mini-mills, open-source models aren’t staying at the low end. They now meet top coding benchmarks that once belonged exclusively to Western frontier labs. Companies like Airbnb and the coding platform Cursor have disclosed using Chinese open-source models in their infrastructure. Even Microsoft has explored cheaper open-source alternatives to power parts of its own AI products. The disruption has moved from “good enough” to “pretty good”.

Singapore’s own verdict

The rising interest in open-source AI is not confined to cost-conscious firms. For years, Singapore built its home-grown language models for South-east Asia called SEA-LION on the back of Meta’s Llama. In late 2025, it switched to Alibaba’s Qwen instead. At less than half the size of its Llama-based predecessor in terms of parameters (a measure of a model’s size, which also determines how much memory it uses), SEA-LION now tops the regional leaderboard for understanding South-east Asian languages and Qwen offers a version that can run on a powerful laptop, which matters enormously for the region’s smaller developers and enterprises which don’t have access to industrial-scale computing.

Malaysia has made a similar bet, building its own sovereign AI ecosystem on Chinese open-source models.

Value for money aside, US AI labs have been sporadically subject to export control actions, including a since-reversed order barring Anthropic from serving one of its own models to foreign users. Unrestricted by such regulations, open-source models carry less of this particular geopolitical risk. Anybody anywhere can use them.

If it’s free, who’s paying for it?

But there is a nagging question. If these models are free, how can they make money? The world’s most downloaded model, Qwen, is free, so how can its parent Alibaba monetise it? And if its strategy is working, why has Alibaba’s stock price tanked about 35 per cent this year?

Maybe the answer is that Qwen isn’t really the product – it’s the hook. Alibaba gives the model away to pull developers and enterprises onto Alibaba Cloud, where it sells the computing, hosting and enterprise services that actually run the model. It’s the same logic as free razors selling blades and a free Android operating system licensed to phone manufacturers at no cost, selling Google’s ad and services layer instead.

There’s some evidence that Alibaba’s strategy is working. The AI-related portion of its cloud business has posted 10 straight quarters of triple-digit year-on-year growth, even as the broader Cloud Intelligence Group’s overall growth has been a steadier 36 per cent to 40 per cent. It’s also layering paid products on top of the free foundation – enterprise agent platforms and its own AI chips – in a deliberate shift from pure giveaway to monetisation.

But the stock has been held back by Alibaba’s enormous spending on infrastructure and chips. Geopolitics adds another layer of risk: In June, the Pentagon added Alibaba to a list of companies it says are linked to China’s military, which doesn’t bar US investors from holding the stock but does restrict Alibaba’s access to US government contracts and has made institutional investors more wary. Alibaba is suing to have the designation overturned.

But it is not uncommon for disruptors to be viewed with scepticism initially. Wall Street doesn’t reward a strategy of “giving away your product” until years later, once the cannibalised low end has quietly become the whole market. Mini-mill investors weren’t cheering when Nucor started making cheap rebar steel. Nor was GrabTaxi of great investor interest in 2012. It was a scrappy start-up fighting for survival in a taxi app war. Investor interest came only after its disruption had already started working – GrabTaxi’s Series A round in April 2014 raised US$10 million; by December that year, SoftBank alone put in US$250 million.

The pattern repeats

What makes this moment especially interesting isn’t just that cheap AI is beating expensive AI. It’s that the industry built to disrupt everyone else is discovering it isn’t exempt from the same forces. The paywalled frontier model, like the department store and the integrated steel mill before it, assumed its advantages were permanent. They rarely are.

Whether Chinese open-source models keep climbing towards the top of the market, the way mini-mills and Japanese carmakers eventually did, remains an open question. But the early chapters of this story – the overlooked entry point, the explosive usage growth, the incumbents’ slow and uneasy response, and a monetisation strategy that looks reckless until it doesn’t – read like they were lifted straight from Christensen’s original playbook. AI, it turns out, is not immune to the very disruption it was supposed to unleash on everyone else.