1. China LLM market of 5.6B RMB (<$1B) is tiny compared to Western countries (US and Europe ~$4-6B each), which highlights a fundamental problem with Chinese AI models - monetization.
2. Generally most Western LLM players suffer from the same problem - Anthropic also burned $5B+ on $1B revenue in 2024, but then in 2025 it should reach ~$9B revenue with $2-3B burn. Revenue should be increasing faster than costs in the future.
3. Totally agree with your analysis on Zhipu, it’s hard to efficiently monetize enterprise software clients, they will keep struggle with scale and market size.
4. On Minimax I slightly disagree. Because 70% of their revenue is not from China and it’s mostly B2C, the 5.6B RMB market size in China is not really relevant to them. They have much higher upside.
5. Do you think China’s government will let Zhipu and Minimax go bankrupt? With the crazy demand for these two companies among domestic (and even foreign) investors, they can keep raising rounds, although valuation and dilution might be an issue.
6. Have you looked at Moonshot? What is the unit economics situation there?
Thanks for the detailed thoughts, Denis. Let me push back on a few points:
On Anthropic’s economics: Sure, their revenue will eventually outpace costs. But look at the context. Anthropic burns $5B chasing a $40-60B enterprise market. Zhipu and MiniMax burn ¥27.6B annually chasing a ¥5.3B market. The competition is completely out of proportion to market reality.
And even if China’s LLM market hits the projected ¥101B by 2030, there’s another problem: ByteDance, Alibaba, Baidu, Tencent are all fighting for this space. They own the compute infrastructure. They control distribution. They can absorb ¥2B losses without blinking. What’s actually left for independent startups after platforms take their share? The addressable market keeps shrinking.
On MiniMax’s overseas revenue: 70% international sounds great until you realize it doesn’t change the fundamental problem. MiniMax spent 1% of what OpenAI spent and still needs emergency funding. Why? DeepSeek’s R1 launch forced every company to accelerate iteration cycles, regardless of where their revenue comes from. There’s no opting out of the compute arms race just because you sell overseas.
On the IPO question: I think we’re talking past each other. My piece isn’t about bankruptcy risk. You’re right the government won’t let them fail, and they can keep raising rounds. My question is different: does going public actually solve anything beyond buying another year or two? As I wrote, continuous dilution and no clear path to profitability make them questionable investments even if they survive as companies.
On Moonshot: The data I’ve seen is concerning. Kimi’s consumer base is dropping fast. They’re pivoting to API services and going international. Fine, but does this change the core challenge? Can they grow revenue faster than the compute investment required to stay competitive? Same question facing every independent model company in China.
Bottom line: when one competitor’s breakthrough forces everyone to match the investment, no moat works. Not government backing, not overseas revenue, not capital efficiency.
To be fair, almost all AI LLM companies, Western and Eastern, will look weak on traditional financial metrics for quite some time.
The valuation of these AI companies is driven mostly by:
#1. The ability to continuously build frontier‑level models.
#2. The ability to train and inference these models at competitive cost.
#3. The ability to keep raising large amounts of capital to fund compute and R&D.
#4. The perceived ability or potential to reach AGI, or at least AGI‑level capabilities.
In China, outside of the big internet platforms, there are only a handful of truly competitive, independent model companies right now: DeepSeek, Zhipu, MiniMax, and Moonshot.
So, coming back to your question, “Does going public actually solve anything?” Yes: accessing more capital via public markets (point #3) is precisely how they try to fund #1, #2, and eventually #4. Sure, most of them will probably still fail but that is inherently how startups and high‑risk tech cycles work.
Great points. I agree that western labs are in better shape. I’m a software engineer and I swear by Claude Code. The underlying models (Sonnet and Opus) could be twice as expensive and I’d still use them. The productivity boost is just too incredible. Anthropic has managed to differentiate themselves.
I just finished reading your article titled "Running Out of Runway," but I still have a few lingering questions. What are the chances that MiniMax will be able to further optimize after another capital injection through their IPO that will enable them to reach profitability? I mean they've already done it once, by reducing training costs from 1,365% in 2023, to 266% in 2025. How likely are they to be able to achieve a similar reduction (80%), if they are able to buy just a little more time? Also, what do you make of recent draft rules proposed by the Cyberspace Administration to make AI chatbots less addictive? How will this affect the future of MiniMax? Will these new proposed regulations, if implemented, end MiniMax as a company? What is the Chinese government trying to achieve with these regulations? Are they trying to shape the direction that AI chatbots take away from AI companions to chatbots that promote Chinese culture instead?
This reminds me of one of Buffet’s first investments: a department store that had 3 nearby competitors. He quickly understood the economics were poor because whenever one store spent some money, for instance to install an escalator, the others had to follow suit, even if that wouldn’t increase their sales. Btw I’m surprised the total market is so small (less than $1B). Thank you for the article.
You have to keep retraining models just to maintain current knowledge. This isn’t invention, it’s production.
The real metric: capital spent per month of model relevance. In this view, a model isn't an asset; it's a perishable good with a high rate of depreciation.
If competitors can force faster retraining cycles, they control everyone’s cost structure. DeepSeek didn’t just raise the quality bar—they shortened the lifespan of everyone’s manufactured goods.
This reframes the “profitability paradox” entirely: these aren’t high-margin SaaS companies burning cash on growth. They’re commodity producers with negative unit economics disguised by accounting categories.
Training costs classified as R&D should be COGS. The desperation isn’t about scaling too fast—it’s about being operationally unprofitable.
I think you’ve got it backwards. These companies are desperate for funding.
Zhipu had 9 months of cash left in mid-2025. MiniMax burns ¥2B monthly. December IPO filings came because they ran out of alternatives.
The ROI problem makes private investors walk away. Public markets become the only option when you’re burning this fast and can’t attract another private round at scale.
Awesome analysis!
A few thoughts:
1. China LLM market of 5.6B RMB (<$1B) is tiny compared to Western countries (US and Europe ~$4-6B each), which highlights a fundamental problem with Chinese AI models - monetization.
2. Generally most Western LLM players suffer from the same problem - Anthropic also burned $5B+ on $1B revenue in 2024, but then in 2025 it should reach ~$9B revenue with $2-3B burn. Revenue should be increasing faster than costs in the future.
3. Totally agree with your analysis on Zhipu, it’s hard to efficiently monetize enterprise software clients, they will keep struggle with scale and market size.
4. On Minimax I slightly disagree. Because 70% of their revenue is not from China and it’s mostly B2C, the 5.6B RMB market size in China is not really relevant to them. They have much higher upside.
5. Do you think China’s government will let Zhipu and Minimax go bankrupt? With the crazy demand for these two companies among domestic (and even foreign) investors, they can keep raising rounds, although valuation and dilution might be an issue.
6. Have you looked at Moonshot? What is the unit economics situation there?
Thanks for the detailed thoughts, Denis. Let me push back on a few points:
On Anthropic’s economics: Sure, their revenue will eventually outpace costs. But look at the context. Anthropic burns $5B chasing a $40-60B enterprise market. Zhipu and MiniMax burn ¥27.6B annually chasing a ¥5.3B market. The competition is completely out of proportion to market reality.
And even if China’s LLM market hits the projected ¥101B by 2030, there’s another problem: ByteDance, Alibaba, Baidu, Tencent are all fighting for this space. They own the compute infrastructure. They control distribution. They can absorb ¥2B losses without blinking. What’s actually left for independent startups after platforms take their share? The addressable market keeps shrinking.
On MiniMax’s overseas revenue: 70% international sounds great until you realize it doesn’t change the fundamental problem. MiniMax spent 1% of what OpenAI spent and still needs emergency funding. Why? DeepSeek’s R1 launch forced every company to accelerate iteration cycles, regardless of where their revenue comes from. There’s no opting out of the compute arms race just because you sell overseas.
On the IPO question: I think we’re talking past each other. My piece isn’t about bankruptcy risk. You’re right the government won’t let them fail, and they can keep raising rounds. My question is different: does going public actually solve anything beyond buying another year or two? As I wrote, continuous dilution and no clear path to profitability make them questionable investments even if they survive as companies.
On Moonshot: The data I’ve seen is concerning. Kimi’s consumer base is dropping fast. They’re pivoting to API services and going international. Fine, but does this change the core challenge? Can they grow revenue faster than the compute investment required to stay competitive? Same question facing every independent model company in China.
Bottom line: when one competitor’s breakthrough forces everyone to match the investment, no moat works. Not government backing, not overseas revenue, not capital efficiency.
To be fair, almost all AI LLM companies, Western and Eastern, will look weak on traditional financial metrics for quite some time.
The valuation of these AI companies is driven mostly by:
#1. The ability to continuously build frontier‑level models.
#2. The ability to train and inference these models at competitive cost.
#3. The ability to keep raising large amounts of capital to fund compute and R&D.
#4. The perceived ability or potential to reach AGI, or at least AGI‑level capabilities.
In China, outside of the big internet platforms, there are only a handful of truly competitive, independent model companies right now: DeepSeek, Zhipu, MiniMax, and Moonshot.
So, coming back to your question, “Does going public actually solve anything?” Yes: accessing more capital via public markets (point #3) is precisely how they try to fund #1, #2, and eventually #4. Sure, most of them will probably still fail but that is inherently how startups and high‑risk tech cycles work.
Great points. I agree that western labs are in better shape. I’m a software engineer and I swear by Claude Code. The underlying models (Sonnet and Opus) could be twice as expensive and I’d still use them. The productivity boost is just too incredible. Anthropic has managed to differentiate themselves.
Superb analysis - eye opening!
I just finished reading your article titled "Running Out of Runway," but I still have a few lingering questions. What are the chances that MiniMax will be able to further optimize after another capital injection through their IPO that will enable them to reach profitability? I mean they've already done it once, by reducing training costs from 1,365% in 2023, to 266% in 2025. How likely are they to be able to achieve a similar reduction (80%), if they are able to buy just a little more time? Also, what do you make of recent draft rules proposed by the Cyberspace Administration to make AI chatbots less addictive? How will this affect the future of MiniMax? Will these new proposed regulations, if implemented, end MiniMax as a company? What is the Chinese government trying to achieve with these regulations? Are they trying to shape the direction that AI chatbots take away from AI companions to chatbots that promote Chinese culture instead?
Thank you for the great work! Looking forward to see how the capital market reacts.
This reminds me of one of Buffet’s first investments: a department store that had 3 nearby competitors. He quickly understood the economics were poor because whenever one store spent some money, for instance to install an escalator, the others had to follow suit, even if that wouldn’t increase their sales. Btw I’m surprised the total market is so small (less than $1B). Thank you for the article.
You have to keep retraining models just to maintain current knowledge. This isn’t invention, it’s production.
The real metric: capital spent per month of model relevance. In this view, a model isn't an asset; it's a perishable good with a high rate of depreciation.
If competitors can force faster retraining cycles, they control everyone’s cost structure. DeepSeek didn’t just raise the quality bar—they shortened the lifespan of everyone’s manufactured goods.
This reframes the “profitability paradox” entirely: these aren’t high-margin SaaS companies burning cash on growth. They’re commodity producers with negative unit economics disguised by accounting categories.
Training costs classified as R&D should be COGS. The desperation isn’t about scaling too fast—it’s about being operationally unprofitable.
Increasing scale just makes the economics worse.
Thanks for the interesting article. How did you get to the CNY 2 billion monthly cash burn for MiniMax? Seems a lot higher than the numbers I’ve read.
In general the investment in AI does not show the ROI for any of the companies in the near future.
Are these companies over optimistic?
Is that the reason the companies reluctant to go IPO?
I think you’ve got it backwards. These companies are desperate for funding.
Zhipu had 9 months of cash left in mid-2025. MiniMax burns ¥2B monthly. December IPO filings came because they ran out of alternatives.
The ROI problem makes private investors walk away. Public markets become the only option when you’re burning this fast and can’t attract another private round at scale.