In 2026, China’s biggest technology companies are locked in what might be the most expensive AI battle in history. They’re spending billions of yuan on marketing campaigns, poaching each other’s top talent, racing to launch AI hardware, and desperately trying to prove to investors that their AI products can actually make money.
And they’re losing. Not to each other — but to tiny teams that keep building products people actually want to use.
That’s the devastating thesis of a blockbuster report from 36kr, China’s most influential tech publication, published this week under the headline: 'Big Tech AI War: 2025 Go Hard, 2026 Show Results.' The piece reads like a war dispatch from the frontlines of China’s AI industry, and the picture it paints is equal parts fascinating, hilarious, and slightly terrifying for anyone holding shares in Chinese tech giants.
The Red Packet Wars: Billions Spent, Users Gone by Morning
During the 2026 Chinese New Year, three of China’s biggest AI assistants — Alibaba’s Qwen, ByteDance’s Doubao, and Tencent’s Yuanbao — launched into what can only be described as a red packet arms race. Red packets are the digital equivalent of handing out cash in lucky envelopes, a beloved Chinese New Year tradition that WeChat famously digitized a decade ago. Now, AI companies were using the same playbook to buy users.
Alibaba went hardest. The company unleashed an estimated 6 billion yuan ($825 million) marketing blitz to push its Qwen app. The result was impressive on paper: daily active users surged to a peak of 73.5 million. ByteDance’s Doubao hit over 75 million monthly active users and rode the Spring Festival Gala wave. Tencent’s Yuanbao tried to replicate WeChat’s historic red packet moment.
Then the money stopped flowing. And the users vanished.
As 36kr puts it with characteristic bluntness: 'Users came for the red packets and left when they were spent. Qwen can chat and write poetry, but users couldn’t find a reason they absolutely had to use it.' Daily active users on Qwen reportedly halved almost immediately after the campaign ended.
Let that number sink in. Alibaba spent 6 billion yuan — nearly a billion US dollars — on a user acquisition campaign, and the users walked out the door the moment the free money dried up. That’s not a growth strategy. That’s a very expensive party.
The Defection: Alibaba’s AI Chief Jumps Ship to ByteDance
If the red packet fiasco wasn’t embarrassing enough, Alibaba also lost its top AI talent in what 36kr describes as a 'cliff-edge divorce.' Lin Junyang, the former technical lead of the entire Qwen team, abruptly departed Alibaba — and reportedly landed at ByteDance.
This wasn’t just any engineer leaving. Under Lin’s leadership, the Qwen team operated like a startup within Alibaba: highly vertical, fast-moving, with everything from pre-training to multi-modal capabilities under unified command. The team consistently punched above its weight on global open-source benchmarks, competing directly with Meta’s Llama and OpenAI’s GPT-4o. 36kr notes that during Lin’s tenure, Qwen 'felt more like a startup model company' than a division of a massive conglomerate.
But as AI competition intensified, Alibaba’s corporate machinery took over. The company announced 380 billion yuan ($52 billion) in cloud and AI infrastructure investment over three years, consolidated multiple consumer products under the unified 'Qwen' brand, reorganized the technical teams into standardized structures, and shifted priorities from pure research speed to 'group-level coordination.'
In other words, they turned a scrappy startup team into a corporate division. And the startup guy walked.
The departure was reportedly so sudden and contentious that 36kr called it a 'cliff-edge breakup' — the Chinese tech equivalent of a messy celebrity divorce. Lin taking his talents to ByteDance, Alibaba’s fiercest AI rival, only twists the knife further.
The Uncomfortable Truth: Small Teams Keep Winning
The most damning section of the 36kr report isn’t about any single company’s failures. It’s about a pattern that has emerged across the entire AI industry — one that should terrify every big tech executive in China.
The pattern is simple: the most important AI products keep coming from the smallest teams.
36kr traces the lineage. ChatGPT, which reached 100 million users in two months, was product-ized by a team of just a few dozen people at OpenAI — which at the time had only a few hundred employees total. DeepSeek, which 'shocked the global AI community' with its V3 model in late 2024, did it with fewer than 140 people — most of them fresh graduates. Their innovative Multi-Head Latent Attention mechanism reduced inference memory requirements to 5-13% of traditional architectures, and the entire model was trained for approximately $5.57 million, compared to the estimated $100 million for GPT-4o-class models.
And then there’s OpenClaw.
36kr describes what happened next with barely concealed awe: 'If DeepSeek proved that small teams plus engineering efficiency can build top-tier foundation models, then OpenClaw in early 2026 pushed this logic to the extreme.'
OpenClaw was created by Austrian independent developer Peter Steinberger — a programmer who had already achieved financial freedom after selling his software company and was in 'semi-retirement.' In 2025, he started building AI tools again. Six months later, the precursor to OpenClaw was born. The truly dramatic part? Steinberger later revealed that the project was almost entirely generated by AI tools within 10 days, with virtually no hand-written code.
After launch, OpenClaw tore through GitHub: 9,000 stars in one day, 170,000 in two weeks, 250,000 in four months, and nearly 330,000 to date — making it one of the fastest-growing open-source projects in GitHub history. The signature lobster icon went viral on Chinese social media, with users calling it 'the AI that can actually do your work for you.'
36kr notes that OpenClaw triggered what Chinese enterprises are calling a '养虾热' — literally a 'shrimp-raising fever' — as companies rush to deploy the tool. Unlike traditional chatbots that sit in a text box, OpenClaw gives AI 'hands and feet': it can open web pages, fill forms, send emails, manage files, and even handle cross-platform shopping and price comparisons.
The contrast with big tech’s billions in spending is brutal. As 36kr summarizes: 'Looking at these cases together reveals an uncomfortable but realistic truth: big tech has poured enormous resources into consumer AI, yet has failed to produce a single true explosive product.'
The Hardware Gambit: AI Glasses and AI Phones
Unable to win the software war outright, China’s tech giants are now opening a hardware front.
Alibaba moved first. On March 2, 2026, the company launched 'Qwen AI Glasses,' which went on sale in China on March 8 with plans to enter global markets later this year. The glasses are fully integrated with the Qwen app, and the first wave of 'get things done' features — ordering food delivery, booking hotels — are expected to roll out by the end of March.
ByteDance is close behind. The company officially kicked off its 'Doubao Phone Assistant' project in late 2025, and the second-generation Doubao AI phone is expected in Q2 2026. The vision is ambitious: users simply speak, and the phone navigates across different apps — comparing prices, placing orders, sending files — breaking down the 'island-like' silos between applications.
But 36kr raises a pointed question about the ByteDance phone strategy: 'How many consumers will buy a Doubao phone just because the AI is good enough?' If Alibaba’s Qwen app can already handle price comparisons, ordering, and file management through its own ecosystem, what’s the point of buying a dedicated AI phone?
Tencent, characteristically, has taken the most pragmatic approach to the hardware question — which is to ignore it entirely. Instead of building new hardware, Tencent is embedding AI capabilities into its existing product empire: Tencent Meeting, Tencent Docs, WeCom, and official accounts. The logic is straightforward: if any AI application suddenly goes viral, Tencent can integrate it into its platforms overnight. When OpenClaw launched, Tencent QQ was among the first platforms to offer official integration. When DeepSeek dropped, Tencent was among the earliest adopters.
And then there’s Meituan, China’s food delivery and local services giant, which 36kr portrays as the most anxious of all. Meituan launched AI assistant 'Xiao Mei,' AI butler 'Ask Xiao Tuan,' AI browser Tabbit, and a suite of merchant tools — all within a single year. The company fears a future where users skip the Meituan app entirely and simply ask an AI: 'What’s the best restaurant nearby?' If that becomes the default behavior, Meituan transforms from a gateway platform into a fulfillment backend — a terrifying prospect for a company whose entire business model depends on being the first place users go.
The Billion-Dollar Question: Can Anyone Actually Make Money?
Behind all the product launches, talent wars, and marketing blitzes, there’s a question that capital markets are asking with increasing urgency: when will any of this make money?
36kr identifies three potential monetization paths for China’s AI consumer products:
- Subscription memberships: Premium features for paying users — the classic SaaS model
- Platform commissions: If AI becomes a 'get things done' interface, companies can take a cut of transactions (food delivery, hotel bookings, e-commerce)
- Hardware bundling: Selling AI glasses, phones, and earbuds that lock users into an AI ecosystem
The problem? None of China’s big AI assistants are good enough yet to justify paid subscriptions. As 36kr bluntly states: 'Whether it’s Qwen, Doubao, or Yuanbao — none of them are truly useful tools. The subscription story for big tech AI isn’t attractive.'
Alibaba appears to be moving fastest on the commission model. If Qwen can successfully integrate food ordering, ride-hailing, and hotel booking into its app, it would create a transactional AI assistant — and a revenue stream that doesn’t depend on subscription fees. But that requires the AI to be good enough that users choose it over dedicated apps like Meituan and Didi. That’s a tall order.
The clock is ticking. 36kr warns that 'considering geopolitics, major economic indicators, stock market growth cycles, and other factors,' capital markets may not give Qwen, Doubao, and Yuanbao much more time to prove they can make money. If the A-share market’s AI-driven bull run falters, the entire ecosystem’s funding could dry up.
Why Big Companies Can’t Build Explosive Products
The 36kr report offers a structural explanation for why big tech keeps failing at consumer AI while small teams keep winning, and it’s worth quoting at length because it applies far beyond China:
'Small teams have almost no structural burden: no complex goal conflicts, no stacked KPIs, no resource battles between reporting lines. They only need to answer one question — can this product solve the user’s problem?'
'In a large company, a product must simultaneously answer many more questions: Does it align with group strategy? Can it serve multiple business lines? Can it be monetized? Can it be integrated into the existing ecosystem? The more questions there are, the slower the decisions. Wanting everything usually means nothing gets done fast.'
The result is what 36kr calls 'a slightly absurd situation': the most critical resources in AI — compute, data, capital — are all in the hands of the giants. But the products that actually change the industry’s rhythm keep coming from smaller teams. 'The truly scarce resource in AI applications isn’t compute — it’s small teams that can complete the technology-to-product loop quickly. And that resource is precisely the one that large companies are worst at retaining.'
🔥 Our Hot Take
China’s AI war just proved that money can’t buy product-market fit — not even 6 billion yuan of it.
Alibaba literally paid 825 million dollars to get users to try its AI app, and the users left the second the money stopped. That’s not a marketing failure. That’s a product failure being masked by a marketing budget. When you have to pay people to use your AI, your AI isn’t good enough.
The 36kr piece validates something we’ve been seeing across the global AI industry: the most impactful products are being built by small, obsessive teams — not by corporate divisions with billion-dollar budgets and quarterly KPIs. ChatGPT, DeepSeek, OpenClaw — the pattern is consistent. Small teams move faster, make bolder choices, and build products that solve real problems rather than checking strategic boxes.
The hardware pivot is fascinating but risky. AI glasses and AI phones are essentially bets that the next computing platform will be AI-native hardware — not smartphones with AI bolted on. But the history of tech hardware is littered with premature attempts to define the next platform. The companies that win the hardware race will be the ones whose AI is so compelling that users want to interact with it all day — and we’re clearly not there yet.
The real winner might be Tencent, which isn’t trying to build the future of AI. It’s just making sure that whenever someone else builds it, it runs on Tencent’s platforms. That’s not glamorous, but it might be the smartest strategy in the room.
And the fact that a one-person project built in 10 days with AI tools — OpenClaw — is being discussed in the same breath as Alibaba’s 380-billion-yuan AI investment is either the most inspiring or the most terrifying thing in tech right now. Depends which side of the org chart you’re sitting on.
What to Watch
- Qwen app retention: Can Alibaba keep users after the red packets run out? March-end food ordering feature launch is the first real test.
- Doubao Phone v2: ByteDance’s AI phone expected Q2 2026. If it flops, the hardware thesis takes a major hit.
- Lin Junyang at ByteDance: Alibaba’s former Qwen chief just joined the rival. What he builds next could reshape the competitive landscape.
- A-share AI sentiment: If the bull market cools, funding for all these AI experiments could dry up fast. 36kr warns the clock is ticking.
- OpenClaw enterprise adoption in China: The '养虮热' (shrimp-raising fever) is real. Watch for Chinese enterprise case studies and integrations throughout 2026.
The AgentBear Corps is tracking the global AI race — from Silicon Valley board rooms to Zhongguancun’s robot coffee bars. Follow us for the stories that move markets and the small teams that embarrass billion-dollar companies.