Inside China's AI Monetization Engine: Notes From Conversations With China AI Insiders
China Open Source No More? New Competitions in LLM, Z.ai & MiniMax Explosive ARR
I wanted to do a quick update on China’s AI ecosystem. For this, I spoke to three public investors in Hong Kong, one banker in Singapore, and some people from the labs. Zixuan Li from Z.ai was the only one who graciously offered to go on record for a tidbit of this, so I WANT TO EMPHASIZE, this piece cannot be wholly attributed to him at all. I would explicitly tell you what he said if he is mentioned.
Many of you may know Zixuan from this earlier episode recorded for Differentiated Understanding. So if you’re curious about their business model/ the first LLM company to go public, check out the episode below.
Announcement
Btw, if any of you are in Singapore in June, please join us at SuperAI June 10-11/ or catch me in town June 9-12. Representatives from Chinese labs and big tech will share the stage to discuss their businesses and development. (And of course, I’ll be moderating;p)
Find out more about SuperAI and the speakers’ information here:
SuperAI is the largest AI conference in Asia, bringing together over 10,000 attendees and 1,500 AI companies from more than 150 countries at Marina Bay Sands in Singapore. The conference is where East meets West in AI – spanning robotics & embodied AI, frontier models, AI infrastructure, biotech & healthtech, finance, and AI’s global impact. SuperAI takes place 10-11 June 2026 as the anchor event of Singapore AI Week (8-14 June).
Find out more about SuperAI and the speakers’ information here.
ANDANDAND DON’T MISS THIS: 10% discount through my link!
A few quick thoughts on recent announcements and rumors
Tencent
QClaw, one of Tencent’s clawbot products, will launch its international version tonight, and I recorded an exclusive interview with their Senior Product Manager, Shuyu Zhang, which I’ll release tomorrow around 6:30 am EST. She dives into why QClaw, the product design behind it, the safety guardrails, the advantages Tencent and thus WeChat’s ecosystem have, and demystifies in a cultural and social context why OpenClaw took off in China.
Alibaba
Ok, and on to the next. It was confirmed that Alibaba’s former Qwen leader, Lin Junyang, had a philosophical clash with Eddie Wu, Alibaba's CEO, and decided to leave. Eddie wants to push on AI commercialization. It was said that Lin Junyang firmly believes in open sourcing for the ecosystem, so he left. However, Alibaba released an impressive video model last week - code name happy horse.
For many of the labs now, they’re under pressure to make money, and especially for Alibaba, a publicly listed company, it has a fiduciary duty to its shareholders. However, on Lin Junyang, I guess if he truly doesn’t vibe/ agree with this direction, I understand his decision to leave. I’m sure the discussion was ongoing for a while and not just a sudden heat-of-the-moment decision.
DeepSeek
Based on multiple conversations, it is widely known that DeepSeek will release a new model in the coming weeks. The expectation is that it will take the number-one slot on the usual leaderboards, but the margin over number two will be narrower than last cycle, when DeepSeek sat roughly 15 points ahead of the next open-source model. Worth noting alongside this is a separate news that DeepSeek is fundraising at $10 Billion-Plus Valuation per The Information’s reporting.
To read: Nvidia’s Jensen Huang warns Huawei chips for DeepSeek AI models would be ‘horrible’ for US and China’s AI firms scaled up on open-source models. The next phase may be different, both by the SCMP.
Breaking down the monetization framework
So for the labs, the revenue model is really simple, as confirmed by the labs and investors; the thinking is simply Price x Quantity. The battle in the end is to find the optimal balance of who has pricing power and who can sell the most. Both variables have increased simultaneously, thereby explaining public disclosures suggesting triple-digit revenue growth at some Chinese labs, including MiniMax and Zhipu, though public ARR disclosure remains limited.
When I asked whether they are not scared that if the price goes up, then demand goes down? The analogy was awesome. “We’re no longer selling 二锅头s, we’re pivoting to 茅台s, so if we lose some customers, so be it. (the margins will be much higher).” How do I translate this? Basically, the former is a white liquor that burns in your insides and tastes like diesel, and the latter is what is used at state banquets and fancy people's weddings.
Is pricing power everything?
On the price side, China’s leading AI labs have been quietly nudging prices upward, especially Z.ai, which has explicitly announced. But it reflects a general trend in many of their coding subscription plans and the per-token rates they charge developers through their APIs. They are doing this from a position of scarcity rather than strength of marketing: demand for their models is outrunning the supply of GPUs they have to serve it. That changes who gets prioritized. Customers willing to pay the full sticker price, and especially those willing to sign multi-year commitments, are being prioritized or preferred over those hunting for day-to-day discounts. In plain terms, according to an investor, the labs are choosing better-quality revenue over bigger-looking revenue, which is the kind of move a business makes when it is trying to protect its margins rather than chase growth at any cost.
The subtler point, and the one most outside observers miss, is that the headline price of tokens is not really what generates the revenue in coding use cases. Most of the billing in a coding workload comes from something called “cache hits.” Here is what that means in plain English, which was explained to me: when a developer works within a large codebase, the AI has to reread much of the same code repeatedly as it answers follow-up questions. Rather than charging full price each time, the labs charge a much lower rate for that repeated content, and that is the cache-hit price. Because so much of a coding session involves re-reading the same files, those discounted cache-hit tokens end up being the bulk of what customers are actually paying for.
This matters for how you compare Chinese models to Western ones. Depending on whether you compare input, output, or cached-input pricing, top Chinese models can range from modestly cheaper to dramatically cheaper than Sonnet. On output pricing, GLM-5.1 is roughly 70% below Sonnet 4.6; on standard input pricing, the gap is closer to 50%.
If you judge Chinese labs by the sticker price, you will badly underestimate how much money they are really making per customer and how healthy their margins actually are.
The open-source game theory
Despite a growing mix of proprietary flagship releases from Qwen, Alibaba has not abandoned open source. It is now running a hybrid strategy across closed and open-weight models. The reason is multi-party game theory.
In a two-player game, both labs rationally close. In a three-player game, one player can always choose to open-source specifically to deny profit to the other two — the prisoner’s dilemma outcome, where at least one actor defects. As long as one frontier lab holds the line open, the others cannot permanently close. So, even though under pressure to profit, it is believed that the Chinese AI ecosystem will not be closed forever. The read is that at least one frontier-grade open-source model will persist, which caps how high closed-source pricing can go.
It’s been widely reported, shared, and recognized that Chinese labs, in particular, lean toward openness for non-ideological reasons: they need an ecosystem, brand recognition, and community traction that a closed Chinese model cannot easily earn. The interesting take is that despite it sounds like a philosophical choice, whether to open or close, that debate is driven by where your economic interest lies: if you own compute, you want everything open; if you own cash, you want everything behind an API.
Competition intensifies, but no one is backing down
No consolidation is expected. This is where I realized I might be wrong, as I’ve been saying there must be consolidation happening later in this year at this burn rate. One of the lab people told me that researchers at each lab are strong, the frontier techniques are widely understood, and new entrants like Xiaomi are increasing competitive intensity rather than reducing it. But no one has an interest in dropping out of the race willingly.
The part where it did align with my recent analysis is that Moonshot (Kimi) is probably the best company and the best AI. The thing is, I was told that “good AI = good company, not marketing.” ;p
As we’ve written before, Moonshot’s founder, Yang Zhilin, was a student of Z.ai's founder, Tang Jie, so much of their vibe or aura can be characterized as technology-first and deliberately uncommercial. Maybe Yang a bit more cool with his rock band meeting rooms and what not. MiniMax is described very differently: SenseTime-lineage DNA, more commercially aggressive, more mature on capital markets mechanics, more willing to hire professional operators and play the fundraising game. That explanation, while not always investor-flattering, arguably explains the valuation trajectory better than benchmarks do. If the thesis is that the API layer is the rent layer, a lab unusually good at commercialization attracts exactly the kind of capital chasing that rent, independent of model leadership.
The endgame analogy
So I asked, " What is ‘to win’? The endgame, in the insider’s framing, will look like the auto industry rather than like a single winner-takes-most platform. Some will be Rolls-Royce with a low number of sales but extremely high margin, and some will be Mercedes, BMW, and Audi sit at the premium-meets-volume sweet spot. Mass-market players generate volume at thinner margins. The strongest model does not automatically win the commercial market — the winner is whoever maximizes price times quantity at their chosen tier. Developer, consumer, and enterprise positioning are each a bet on where that product peaks.
And for most Chinese labs, their simple goal right now is to win on speed over the big techs. They’re going to lean into lean organizational charts so they can iterate and ship faster.
