Welcome to the Halftime Show.

Lunar New Year. The one week a year when Shenzhen exhales. My first one in China in over a damn decade. The first time where I’m giving out red envelopes instead of receiving one. The tax of aging is felt more in no other country.

I ordered a jasmine milk tea (20% sweet, extra foam) through an AI chat app this week. It was free. Someone at Alibaba paid for it. It’s a gooood time to be a consumer. 

Let's lock in.

In Dec, Bloomberg reported that Meta's next flagship model (dubbed Avocado) is being trained on Alibaba's Qwen. The company that wrote the open source playbook needed the student's notes to build its next closed source product.

Big win on the infrastructure layer.

Consumer though? Whole diiiiiiifferent story.

Alibaba, Tencent, Baidu, and ByteDance collectively dropped $6.3 billion in subsidies this Lunar New Year - each fighting to become the default AI interface for daily life. 

The fight to become the OS of daily life - who will win and on what terms?

→ Meta's next flagship model is being trained on Qwen. Meta's TBD Lab is distilling Qwen, alongside Google's Gemma and OpenAI's open source variants, to train Avocado, a closed source model expected in spring 2026. (Bloomberg via SCMP)

→ Qwen showed the strongest post holiday retention of any platform. It maintained 22 million DAUs. Cash giveaways drive downloads. Utility drives retention. The platform that embedded itself deepest into daily transactions held the most users when the money stopped. (Fine Day Radio)

→ Qwen is the most downloaded open-source AI family on earth. 700 million downloads. 180,000 derivative models built on top. Overtook Meta's Llama by October 2025. By December, single month downloads exceeded the combined total of Meta, DeepSeek, OpenAI, Mistral, Nvidia, Zhipu, Moonshot, and MiniMax combined. When developers choose a foundation to build on, they're choosing Qwen. (Xinhua)

→ ByteDance launched Doubao 2.0 the same week. DeepSeek's next model is widely expected imminently. The domestic AI app war has no off season. Every major platform is racing to become the default AI gateway simultaneously. (InfoWorld)

→ Alibaba's cloud business grew 30% yOy driven almost entirely by AI. Every enterprise deploying Qwen needs Alibaba Cloud to run it. The open source giveaway was never charity, but a cloud acquisition funnel. (Motley Fool)

Most consumer founders think about user acquisition. Qwen thought about habit installation. There's a permanent gap between those two ambitions - and their viral Lunar New Year campaign is where it became visible.

Let's start with the architecture, because the campaign only makes sense once you understand what was built underneath it.

On Jan 15, Alibaba wired Qwen directly into the backend of an entire economy:

A user types "order me a coffee" and Qwen recommends the option, confirms the address, applies promotions, and completes payment through Alipay -without ever leaving the chat UI.

Most AI assistants are recommendation engines with a link at the end. Qwen is an execution layer with a payment system inside it.

Then they picked the right moment to detonate it.

Lunar New Year is the one week of the year when hundreds of millions of people pause their routines simultaneously. WeChat used that window in 2014 to make mobile payments ubiquitous. Alibaba ran the same play with AI.

In September 2025, OpenAI launched Instant Checkout- buy products directly inside ChatGPT, powered by Shopify's catalog. Toby Lutke called it a "new frontier." Over a million merchants were promised.

As of this month: roughly 30 are live.

Onboarding merchants was harder than anticipated:

So this is very much an architecture issue rather than strategic.

ChatGPT had to negotiate every integration it didn't own (painful) - retailer by retailer, data pipeline by data pipeline, payment rail by payment rail.

Each dependency = a point of failure. Qwen had none of those dependencies. Taobao's inventory is accurate because it's the same database. Alipay completes inside the conversation because it's the same company.

Owning the transaction layer and adding intelligence is a structurally different position than owning the intelligence and trying to acquire a transaction layer. That gap doesn't close on a product roadmap. It closes - if it closes - over a decade of building.

Now look at who they were competing against.

Doubao entered Lunar New Year as the market leader. 172 million monthly active users to Qwen's 100 million. ByteDance bought the CCTV Spring Festival Gala sponsorship for peak visibility.

Doubao got seen. Qwen got used.

After the subsidies ended, both platforms saw user drops. Qwen's was the smallest. The platform embedded deepest in daily transactions held the most users when the money stopped. Impressions don't retain users. Completed transactions do.

Order milk tea through an AI interface once - the next natural action is to order lunch the same way. Then book a flight. Then pay a bill. Each completed transaction is a proof point. Each proof point lowers the activation energy for the next one. Each habit is a switching cost. Each switching cost is a moat.

Visibility without utility is advertising. 

Utility without a moment is a feature. 

Qwen built the utility first, chose the moment carefully, and spent $432 million making sure 130 million people experienced both inside the same eleven days.

This is the most intense consumer AI competition on earth.

And this is the playbook so far: build the execution layer, wire it to real transactions, find the moment when hundreds of millions of people are available to form a new habit - and light the fuse. Sheeeeeeesh.