17. RMB10B Burned to Create the First Foundation-Model Stock: Zhipu vs. MiniMax — A Conversation with 庄明浩

Summary

• Zhipu and MiniMax passed the HKEX listing hearing on the same Thursday, with CICC involved as sponsor on both; they could list as early as January 2026 in a race to become the first foundation-model stock. One filed its prospectus on Friday and the other on Sunday, making clear that both knew the market would compare them head-to-head; 庄明浩 opened with his verdict: “This business is just too hard at this stage. Too hard. That’s the biggest takeaway.”
• “1%” is the common yardstick for comparing Chinese and US foundation models. OpenAI has been discussed at a $500B-$700B valuation, versus roughly $3.5B-$4B for China’s leading model companies; the 2 speakers also used “1%” to describe R&D spending and revenue. Even newly listed A-share GPU companies are roughly 1% of Nvidia by comparison. In the private-market era, the consensus was that the China-US gap should be 1/10 to 1/7: “We don’t know which side is wrong or which side is right. All we can say is that if it exists, it is rational.”
• The financial positions are worlds apart: Zhipu lost RMB1.8B on a non-GAAP basis and more than RMB2B on a GAAP basis in the first half, generated less than RMB200M in revenue, spent RMB1.1B on compute, and had RMB2.5B in cash, making its IPO need acute; more than RMB6B of the “over RMB8B” Xiaohongshu publicity figure was credit capacity. MiniMax uses a dollar structure, gets roughly 70% of its revenue from overseas, could generate $60M in 2025 while losing $200M, and still has $1B in cash. 庄明浩 thinks MiniMax may not be the one under greater pressure if it cannot get listed in January.
• The training-cost myth does not survive scrutiny: the $1.28M cited in DeepSeek’s paper covered only R1’s final single training run, while Zhipu’s first-half and MiniMax’s first-9-month training compute costs disclosed in their prospectuses were both above RMB1B, excluding inference. The Information’s estimate of OpenAI’s 2025 R&D costs was $13.4B, or roughly RMB100B. Raymond’s education analogy: “Fifth grade costs RMB100B; fourth grade costs RMB1B.” Average R&D compensation was RMB810K at Zhipu and RMB940K at MiniMax—“consistent with China’s reality, but clearly inconsistent with this round of AI competition,” 庄明浩 says.
• Traditional valuation anchors have limited use; investors are left with comparables or a seat-at-the-table framework. SenseTime had RMB3B-RMB4B of revenue at listing, far more than either company today. PS is a reference point but not a high-weight input; Raymond expects analysts may train investors to look at 2027 forward PS. The alternative is to treat foundation models as “atomic-bomb projects” and value the endgame seats—Mistral, which 庄明浩 considers 2 rounds behind and absent from SOTA rankings, still commands a $10B valuation on Europe’s geopolitical need for an indigenous model.
• The biggest strategic question concerns 杨植麟 and Kimi’s investors: Kimi may continue on the purest AGI path, get acquired at some point, or already have filed confidentially and be meeting cornerstone investors. 庄明浩 expects 张一鸣, Alibaba, and Tencent to emerge from the prospectuses more committed to their existing strategies. Raymond thinks they may not read them at all: “Nothing to worry about. Wait for your stock to fall, then buy you.” 梁文锋 continues along DeepSeek’s original path. Asked whether he is personally funding DeepSeek, 庄明浩 replied: “That is what the outside world sees right now.” Raymond estimates annual spending at RMB500M-RMB1B, but says the model’s sustainability is “hard to assess.”
• The listing window is full of crosscurrents: Zhipu has only CICC as sponsor, while MiniMax has CICC and UBS; neither has a US bank, and US investment restrictions could bar some dollar funds from participating. 庄明浩 said 4 IPOs broke issue price on the day of recording, even though the market’s core sentiment this year had been “if it lists, it goes up.” His execution advice to later entrants: an IPO gives you “only one breath”; lawyer 朱莉’s experience is that it is “rare to see a company succeed on its second listing attempt.”

Deep dive

1. Same-day hearings, shared sponsor: both chose a head-to-head comparison

• Zhipu and MiniMax passed the HKEX listing hearing on the same Thursday and, according to the timetable, could ring the bell as early as January 2026. CICC is involved as sponsor on both; one filed its prospectus on Friday and the other on Sunday. The 2 speakers believe both companies understood they would be compared “one-on-one” and effectively forced investors to decide which was stronger.
• Raymond’s analogy was a martial-arts tournament: whether the company trained in the Wudang or Emei school, used expert models or another attention mechanism, does not matter much. Public-market investors ultimately look at the financial statements. The contest is “flesh-on-flesh and extremely bloody.”
• Their reactions were highly sensitive to reading order. 庄明浩 had not finished both prospectuses; after reading Zhipu first, Raymond felt “the sky was falling.” By Sunday, MiniMax had given him some relief: “At least there is a path. It’s OK.”

2. 1%: the common yardstick for China-US comparisons

• Raymond’s earlier presentation posed the question: “If OpenAI is worth $500B-$700B today, what should China’s leading model company be worth?” His current read is roughly 1%. 庄明浩 extended that ratio to R&D spending, valuation, and revenue. On the US side, estimates had ranged from $350B-$500B or higher; on the Chinese side, roughly $3.5B-$4B. Chinese companies spent only a few hundred million dollars to get to where they are today.
• The same comparison extends to hardware. Raymond thinks A-share GPU companies including Moore Threads and Biren Technology are also roughly 1% of Nvidia. In the private-market era, the assumed China-US gap was 1/10 or 1/7. “Today it has become 1%.” But “we don’t know which side is wrong or which side is right. All we can say is that if it exists, it is rational.”

3. The reaction from 张一鸣 and peers: “Nothing to worry about—wait for the stock to fall, then buy you”

• 庄明浩’s read on Sam Altman: he would be surprised that China could build these capabilities with so little money, while also seeing the reality of Chinese To C and To B revenue. The reaction could be “some concern, and some relief.”
• For ByteDance, Alibaba, and Tencent, 庄明浩 believes the prospectuses reinforce their existing conclusions. These companies may look strong on rankings and under different definitions, but their non-revenue operating metrics are not enough to generate a massive short-term impact. Volcano Engine has just held its conference, DingTalk is about to hold one, Alibaba is pushing the Qwen App and 蚂蚁阿福, and Tencent is releasing related products. The prospectuses will not change their established strategy for next year; they may make them even more determined.
• Raymond’s response: the giants may not read the prospectuses at all. Their conclusion would simply be: “Nothing to worry about. Wait for your stock to fall, then buy you.”
• 梁文锋 is the least changed variable. From before DeepSeek’s breakout to after it and through today, he has largely stuck with the original strategy and continued on his own path.

4. 杨植麟 faces the hardest choice: pure AGI, acquisition, or another filing

• The bearish case is that only 2 of the previous generation’s “4 AI dragons” broke through, with poor outcomes. If 2 companies emerge again this time, the third will be in a difficult position. The bullish case is that Kimi may currently be the purest AGI contender; if it continues to receive investor and external support, it could become a much larger opportunity. But the path is “too dangerous and too difficult.”
• Raymond offered another scenario: if Kimi continues pursuing AGI while temporarily neglecting commercialization, it could be acquired at some point because of cash-flow or shareholder pressure. He explicitly said this was “all speculation, with no basis whatsoever.” 庄明浩 added that the 2 listed companies could also be acquired, potentially at fairer prices.
• 庄明浩 thinks this weekend was particularly agonizing for 杨植麟, 杨雨桐, and Kimi’s major outside investors. The investor bases of Kimi, MiniMax, and Zhipu also overlap substantially, making the mood even more complicated.

5. Zhipu’s cash position: RMB1.8B loss, RMB2.5B cash, and credit capacity behind the RMB8B publicity figure

• Zhipu lost RMB1.8B on a non-GAAP basis and more than RMB2B on a GAAP basis in the first half, generated less than RMB200M in revenue, spent RMB1.1B on compute, and had only RMB2.5B in cash. Raymond therefore sees a “very strong” need for the IPO.
• After the prospectus was posted, a series of publicity pieces on Xiaohongshu said Zhipu had more than RMB8B on its balance sheet. 庄明浩 explained that more than RMB6B of the figure was credit capacity. He believes “some force is pushing” KOLs to emphasize ample funding without clearly stating the actual cash position: “Only we care. Ordinary people don’t know.”
• Raymond added that listing first would provide a clear market price. That creates a valuation anchor if the company is later acquired and gives it a stock-issuance tool when acquiring others. He compared it with a previous episode in which 庄明浩 discussed Tencent Music paying a huge price for Ximalaya.

6. Prospectus craft: everyone finds a “No. 1” category, but the rankings have broken down

• Drawing on his investment-banking experience, Raymond said a prospectus usually needs a category in which the company can claim to be “No. 1”: JD.com was China’s No. 1 B2C platform, iQiyi was No. 1 in long-form video, and Bilibili was No. 1 in PUGC. The key is to define the category narrowly enough.
• The foundation-model market is difficult to slice. Zhipu therefore leans mainly on token usage from OpenRouter, while MiniMax relies mainly on Artificial Analysis, along with SOTA rankings from other benchmarks.
• 庄明浩 thinks the industry’s real problem is that “all the metrics and scoring systems have stopped working.” OpenRouter is relatively better, but covers only one third-party market; direct official access and large-account usage may not be captured, and large customers can generate substantial volume. The crudest metric is token volume; the most outcome-oriented metric is revenue.
• At a deeper level, each company’s strengths are diverging by use case. Anthropic may skew toward To B, enterprise services, and specific workflows; xAI may benefit from data linkage with Twitter; Google may emphasize the experience of combining text and multimodality. Baseline scores still cannot be too low, but use-case differentiation matters more, weakening general-purpose rankings.
• MiniMax’s chosen angle is that its multimodal capabilities are “far ahead” of other Chinese vendors, with the company claiming to sit in the first tier across text, image, video, and voice.

7. Zhipu’s industry report: a RMB100B Chinese enterprise market by 2030, 80% locally deployed—“too much like a guess”

• Raymond explained that prospectus industry reports are usually commissioned by the company from industry experts or consulting firms. They may cost RMB500K-RMB800K and become the foundation for the industry section, which is why many investors treat them as a window into management’s market view.
• Zhipu’s report forecasts a Chinese enterprise AI market of roughly RMB100B by 2030, still below OpenAI’s revenue at the time. It estimates that 80% will be locally deployed and 20% cloud-based.
• 庄明浩 thinks the figure was “too much like something pulled from thin air.” In a parallel universe without DeepSeek or the open-source release of R1, the same company might have used a much larger forecast when listing. After DeepSeek, if local deployment is more important in China, it could theoretically capture most of the market. He cannot tell whether the number is too high or too low; it simply looks casually written.
• He agrees that local deployment could account for a high share given the Chinese government, large enterprises, and the country’s overall level of digitization, but he is not sure it reaches 80%. The forecast is also too closely tied to Zhipu’s state-capital, To G, and To large-B narrative to be judged on the number alone.

8. From A-shares to Hong Kong: foundation models came after GPUs and humanoid robots

• Raymond added that Zhipu filed with the CSRC in April 2025, with CICC as its counseling institution, targeting an A-share listing. He cannot determine exactly why it later switched to Hong Kong.
• 庄明浩 used an analogy: listing in the A-share market was like the Xiongnu crossing the Great Wall, discovering by chance a hole they could pass through, only to find it sealed when they tried again. Raymond thinks that is probably too extreme, but the market’s direction did change in Q4.
• In terms of timing, GPUs were first tier, humanoid robots second, and foundation models further back. The A-share capital pool and liquidity are limited, retail investors dominate, and regulators have also signaled that there are “too many,” that only the best should be selected, and that the pace should be controlled. Raymond thinks this environment may have pushed Zhipu toward Hong Kong at some point, though only as one possible explanation based on “mutual adaptation.”

9. Even MiniMax’s $300B global market is too small; the reverse triangle is “wishful thinking”

• MiniMax’s industry report forecasts a global market of $300B by 2030, with 75% from apps and 25% from MaaS—APIs, cloud services, and related offerings. In practical terms, 75% is more consumer or prosumer-oriented and 25% more enterprise-oriented.
• 庄明浩 thinks the figure is far larger than Zhipu’s forecast but still too small. OpenAI is expected to generate $200B in revenue in 2030; based on compute services, he estimates another segment at at least $100B. Those 2 figures already total $300B, before Google, xAI, and other model and application companies are included. If the market is only that large in 5 years, today’s bubble should already be over, because several trillion dollars of data-center investment cannot be earned back.
• The 75/25 split points to the question of who captures value: APIs or applications. Raymond said industry maturity is typically depicted as a reverse triangle, with infrastructure taking the least, models in the middle, and applications capturing the most revenue and profit. AI today looks more like an upright triangle: infrastructure and chips take the most, models come second, and applications may even run at negative gross margin.
• 庄明浩 thinks the move from an upright triangle to a reverse triangle is somewhat wishful, supported mainly by “the most correct platitude”—that application companies are closest to users. At that point, it was still unsettled whether the model itself would remain the product.
• Raymond added that mobile internet is only one example. The value of major infrastructure such as railways, fiber optics, and China’s high-speed rail may be captured through positive externalities by other parts of the economy; infrastructure companies do not necessarily earn the most. The mobile-internet template cannot simply be applied.

10. The opening sentence sets the frame: Day One Global vs. a national strategic asset

• MiniMax’s prospectus opens with “a global AI foundation-model company,” consistent with roughly 70% overseas revenue and a dollar reporting structure. Zhipu opens with “China’s leading artificial-intelligence company,” then emphasizes its founding mission of “pursuing general-AI innovation in China.”
• Raymond believes Zhipu is backed by state capital from Beijing, Hangzhou, Chengdu, Zhuhai, and Shanghai. The overall logic is: “I am building a national strategic asset. China must have a seat, and I will play that role.”
• 庄明浩 estimates that overseas revenue accounts for roughly 10% of Zhipu’s total and believes it probably came from an order by a sovereign-model alliance in a Southeast Asian country. MiniMax is Day One Global: roughly 70% overseas revenue and 30% domestic.
• The prospectuses also show that the companies’ business mixes differ from prior perceptions. Zhipu has substantial To G, To large-B, and local-deployment exposure, while its receivables also reveal a large-enterprise customer base. MiniMax is clearly To C, with revenue concentrated in chat and gacha-style products.
• Their business models are mirror images. Zhipu’s local-deployment business carries high gross margin but requires heavy training costs. MiniMax has more than 20M MAUs but must absorb the inference cost of users running its app, leaving it with a relatively low overall gross margin.

11. Fifty lights: no wrong order in the US, but China has to choose

• Raymond compared the business matrix with 50 lights: domestic and overseas markets, To G/B/C, APIs, local deployment, private cloud, Talkie, and more all need to be occupied over time. The question is how a startup decides which lights to switch on first.
• 庄明浩 called that conclusion “particularly useless.” A prospectus is only a snapshot of the present, reflecting a short stretch of road each company reached through its capabilities, resources, strengths, unexpected factors, and convenience. Each company may have switched on only 3 or 4 lights. Within the full 50-light process, sequence has limited explanatory power and no absolute right or wrong.
• The key difference is constraint. The leading US vendors may want everything—all the lights switched on. Chinese companies have to choose because of the 1% resource and cost problem.
• Looking back from 2030, Coding will certainly remain. Multimodality and Agents may each be composite lights made up of many bulbs. In Raymond’s framework, advances in memory and context for pure language models are more technical-layer developments, not separate product or application lights. AI for Science remains an open question.

12. Zhipu is catching up: GLM Coder priced against Anthropic, AutoGLM open-sourced

• Raymond recalled that after Anthropic restricted China-related access, Zhipu quickly launched GLM Coder with a direct pricing challenge: “You charge $20; I’ll charge RMB20.” When Cursor later launched Composer, some people also believed it was based on a Zhipu model.
• Raymond said Zhipu had nearly disappeared from his radar, which made the speed of its response particularly impressive. Zhipu later “claimed” close to RMB100M in revenue and 150K users, leaving him with some hope for the company.
• After the Doubao phone became popular, Zhipu open-sourced AutoGLM, its earlier product for controlling a phone. 庄明浩 thinks Zhipu had previously carried a somewhat state-enterprise-like image and lacked internet-native instincts in To C products, but these 2 moves show it is catching up—and appears to be doing so reasonably well.

13. The Agent battlefield: Manus ARR roughly equals the combined revenue of the 2 listed companies

• Being able to do something and getting users to buy and pay for it are 2 different things. Raymond considers the revenue of Manus and Genspark relatively clear and says Manus’s ARR is roughly equal to the combined current-year revenue of the 2 listed companies. Both companies will realize that the Agent light must be switched on and pursued.
• That brings the debate back to whether the model or the product captures value. Some originally thought models would absorb the products built on top of them, but as Claude Code and GPT Codex advanced rapidly in 2025, Cursor also built its own moat. 庄明浩 thinks the Agent question remains unsettled, with model vendors and user-facing Agent companies both competing for the economics.
• The contest could become more intense in 2026. Raymond observed that most of MiniMax’s App revenue in the first 9 months came from newly launched Agent products.

14. RMB1B-plus training costs versus $1.28M: reconciling the 2 sets of numbers

• Raymond’s confusion was straightforward: self-media accounts often say Chinese companies can train a large model for “a few million dollars,” yet Zhipu’s first-half and MiniMax’s first-9-month training compute costs were both above RMB1B, excluding inference, salaries, rent, and other expenses. Even after DeepSeek, the costs had not fallen in the financial statements as expected.
• 庄明浩 broke it down: the $1.28M in DeepSeek’s paper covered only R1’s final single training run. From the base state to final delivery, a model may need many training runs, with the number driven by evaluation, performance, competition, and leaderboard requirements. “Practicing problems also counts as training.” The $1.28M cannot be compared directly with OpenAI’s total investment.
• A more structural explanation is that a prospectus is a snapshot while the business and R&D program keep changing. To chase SOTA, a company may run multiple language, multimodal, and next-generation foundation-model projects in parallel, like a Gantt chart, while trying to stay 1 to 1.5 steps ahead. The costs stack up.
• The visible businesses disclosed today are only 3 of the 50 lights, but the models are being built for many more. The companies therefore have to carry more costs than the current product set would suggest.
• The Information estimated OpenAI’s 2025 R&D costs at $13.4B, or roughly RMB100B. 庄明浩’s education analogy is that OpenAI is a fifth-grader and MiniMax a fourth-grader: fifth-grade tuition is RMB100B, fourth-grade tuition RMB1B, but moving from fourth to fifth grade may still cost RMB1B. The cost curve is not necessarily linear.
• Domestic chips do affect costs, but 庄明浩 does not think they can move China from 1% of US costs to 1/10. Domestic chips only began to be used gradually in 2025. DeepSeek’s and Huawei’s testing “should have been around May or June this year, in the summer,” so the effect exists but is not enough to change the order of magnitude.

15. Is 梁文锋 personally funding a national strategic asset?

• Raymond estimates that even the most frugal model company would need 100 to 200 people and substantial compute, putting annual spending “not below RMB500M; we don’t know whether it is above RMB1B.” He also believes DeepSeek has not raised outside funding, while its API revenue does not appear particularly large.
• Raymond therefore asked whether 梁文锋 was personally funding such a “national strategic asset.” 庄明浩 replied: “That is what the outside world sees right now.”
• 庄明浩 said that after DeepSeek’s breakout, state capital, local state capital, market-based funds, dollar investors, and large companies “should all have approached them.” DeepSeek rejected all of them without excluding any category, continuing to bear annual spending of several hundred million RMB on its own terms.
• Asked how long that model could last, 庄明浩 said only that its patience and sustainability were “hard to assess.”

16. Per-capita pay of RMB810K/RMB940K: “consistent with China, inconsistent with this AI cycle”

• Raymond annualized the salary component of R&D expense and divided it by the number of R&D employees. He calculated average R&D compensation of RMB810K at Zhipu and RMB940K at MiniMax, or less than $150K on average.
• 庄明浩 thinks that is “consistent with China’s reality, but clearly inconsistent with the conditions of this AI competition.” MiniMax’s CEO said on 老罗’s podcast that many team members were not returnees from the US but domestic graduates in computer science, mathematics, and related fields who moved into the industry after several years of R&D. The company also emphasizes its youth: the average employee is born in the 1990s and around 32 years old.
• The figure could change sharply in 2026. 庄明浩 said Tencent had “possibly” been hiring ByteDance employees at 2x their salaries. Top graduate students and PhDs are also commanding high pay in the relevant recruitment market, and more competitors will push up talent costs.
• Cheap Chinese APIs have a flip side: To B gross margins. Zhipu cut prices to compete and its To B gross margin was negative. MiniMax has a broader mix of text, voice, and video models, each with different pricing and gross margins, so the blended figure looks better than a negative margin.

17. Listing timing is a matter of luck: RMB fund exits, old-share SPVs, and a dried-up private market

• 庄明浩 thinks Zhipu faces pressure from RMB-fund exit cycles in addition to cash flow. If some funding came from 5-year funds, the large investments raised in 2023 and 2024 could make the listing timetable a consideration. It is less direct than cash flow, but still a factor.
• He also mentioned that substantial information about Zhipu old-share SPVs had circulated in the market. At some point, capital may have been available to take shares at an approximately RMB16B valuation, while Zhipu’s last round was roughly RMB23B-RMB24B. Those investors may have strong exit incentives.
• MiniMax should be in better shape than Zhipu based on cash flow and the patience of its dollar investors, 庄明浩 thinks, and may be in better condition in 2026 even under linear extrapolation. But Hong Kong is its main listing route, and the board may decide to try while sentiment remains favorable and the company is still in an uptrend.
• Raymond pointed to the more practical backdrop: China’s private market “has no money left; no one is providing liquidity anymore.” Financial investors, state-backed investors, and internet giants have already deployed their capital, making another financing round difficult. “Going to the Middle East” cannot be the only bet.
• Raymond’s investment-banking experience shows how much an IPO depends on sentiment. At one roadshow update, he opened by discussing the Syrian war, a US equity selloff, and rising oil prices; people in the room thought he was an idiot. But an IPO is an even higher-risk asset than other risk assets. Bilibili’s April 2018 listing coincided with the first trade war, while the weather also changed abruptly. The issuance window is fundamentally unpredictable.

18. The race between competitors—and the possibility that Kimi is already on the road

• Among direct competitors, the first listing sets the valuation anchor for the second. If the first company prices at 10x, the next may be asked why it is not at 8x. Daily Fresh Food and Dingdong Maicai, the early JD.com versus Alibaba distinction between B2C and C2C, and Lyft listing before Uber all illustrate the issue.
• Raymond thinks a company needs to know where it stands in the industry. Ideally, the market consensus should be that it is No. 1—not that the company thinks it is No. 1 while the market ranks it eighth.
• Raymond raised one explicitly speculative possibility: if Kimi is heading toward a Hong Kong IPO, it may already have filed confidentially or be meeting investors. Hong Kong cornerstone investors can review prospectuses and financial data under confidentiality agreements, so some may already have seen Kimi’s materials and chosen to wait. He stressed that he does not know whether Kimi is good or bad.
• 庄明浩 added that if the market is particularly weak in January and neither company can complete its listing, MiniMax may not be the one under greater pressure.

19. No US banks, 4 IPOs below issue price on the recording day: undercurrents in the window

• Zhipu’s sponsor is CICC; MiniMax has CICC and UBS. Raymond noted the absence of a US bank. The 2 speakers discussed US restrictions on American investment in Chinese chip, AI, and robotics projects, and speculated that some institutional investors funded primarily in dollars may therefore be unable to participate.
• Raymond said: “If I were CICC, I wouldn’t know how to sell this.” He contrasted that with SenseTime’s listing, where almost every investment bank participated, like “filling an entire banquet table.” CICC working on 2 companies in the same sector and pushing them toward market at nearly the same time also struck 庄明浩 as remarkable.
• 庄明浩 said all 4 companies that listed on the recording day broke below issue price. Raymond had not noticed the specific situation, but said that if the claim was accurate, market sentiment may have shifted. The core Hong Kong-market mood this year had been: “If it lists, it lists—and then it has to go up.”
• Raymond added that if the pricing of the 2 GPU companies in Hong Kong was indeed relatively disciplined when Biren filed, and even those prices could not hold, the marginal impact on institutions, retail investors, and the market’s positive sentiment could be meaningful. Neither speaker could confirm that conclusion.
• There were rumors of 300 to 400 companies waiting to list in Hong Kong, while HKEX has emphasized selecting the best and controlling the pace. A large move in US equities could also transmit to Hong Kong. “Whether you get listed really is a matter of fate.”

20. Valuation methodology: PS has limited use; the fallback is comparables and 2027 forward

• While preparing, 庄明浩 looked at CloudWalk and said it had a market capitalization of HK$70B at the time, underscoring how difficult traditional valuation methods are to apply. SenseTime had RMB3B-RMB4B of revenue at listing, far more than either company today, and its losses were much smaller relative to revenue. The 2 companies today have only a few hundred million RMB of revenue against losses of RMB1B-RMB2B.
• PS can provide a reference, he said, but not a meaningful high-weight pricing input. Investors are left with comparables: 1% of OpenAI or Anthropic, or existing Hong Kong names such as SenseTime and CloudWalk, combined with the mindset of the counterparty. “That’s the only way to compare them.”
• Raymond speculated that if he were running the deal, analysts might educate investors to look at 2027 forward PS: start with 2025 revenue, assume 2026 growth and product rollout, estimate 2027 revenue, then apply a multiple. That resembles the 2-year forward PS framework used for early Youku and Tudou, when advertising revenue was small but expenses and CDN costs were high. It remains only his speculation.
• 庄明浩 noted that on a PS basis, neither Zhipu nor MiniMax has a clear competitive advantage over OpenAI. Raymond therefore described OpenAI as “the relatively less extreme case among the extremes.”

21. Atomic-bomb projects and seats at the table: another valuation framework, plus Mistral’s sovereign premium

• Raymond proposed valuing foundation models as projects. If building a model just slightly worse than OpenAI would cost a Chinese company $1B, while Americans might think it costs $100B, and only 5 to 10 companies ultimately possess that capability, the value of a seat at the table can be backed out along with each company’s probability of getting there.
• 庄明浩 sees this as a survival race: companies must keep up and cannot repeat a grade. Raymond says investment requires “extraordinarily long patience”—not evaluating the present with today’s framework, but believing that a team can eventually reach the table. 庄明浩 summarized the situation as: “At least 3 out of 6 have made it through.”
• Mistral is another reference point. 庄明浩 believes it has never appeared on an SOTA leaderboard and is already 2 rounds behind, yet it still commands a $10B valuation, with its largest investor ASML holding roughly 12%-13%. The premium reflects Europe’s geopolitical desire to own its own model and avoid dependence on the US and China.
• 庄明浩 said Mistral only needs support from ASML. A public listing, by contrast, requires hundreds or thousands of investors to buy in simultaneously, with those investors influencing one another.

22. Is “Star Wars” a VC investment? The complicated feeling of “buying bullets for the state”

• Raymond thinks the China-US race for frontier capability gives these projects the character of “Star Wars” and the “Manhattan Project.” For an angel investor, 戴雨森 might see Kimi as a question of whether the outcome is 50x or 500x. For mid- and late-stage investors, the choice is much harder.
• 庄明浩 asked what mid- and late-stage investors should fund if not these projects. Raymond suggested that VC firms may need to formally decide whether they must increase their foundation-model exposure. In the US, a fund absent from the list of OpenAI, Anthropic, Cohere, SpaceX, and similar companies may attract little attention; in China, a fund absent from domestic GPU, humanoid-robot, or foundation-model companies may struggle to raise its next fund.
• He took the question back to 2023: should a $500M fund invest $50M or even $100M into a company already valued at $1B, taking a 2%, 3%, or 4% stake even if it knows the investment may not generate much profit?
• Raymond’s feeling is two-sided. He is proud of Chinese companies for advancing world-class capabilities with 1% of the money; they have become extremely “thrifty.” But it can also feel “like the country is going to war and asking me to buy the bullets.” No one knows how many seats China will ultimately have or which companies will remain at the table.
• Both speakers think this generation of entrepreneurship is far harder than building mobile internet 10-plus years ago. 唐岩’s job in building Momo was mainly to build an app; today’s model founders also have to think about meeting royalty in Saudi Arabia, geopolitics, and the US presidential election.

23. For 杨植麟 and the next entrants: an IPO gives you “only one breath”

• 庄明浩’s advice to later entrants is that if they expect a large funding requirement, they should prepare for an IPO early and complete the audit as soon as possible. Raymond says those preparations are what “put time on your side.”
• 庄明浩 rejects the idea of simply waiting. He relayed an early judgment from the now-famous lawyer 朱莉: it is rare to see a company succeed on its second listing attempt. During an IPO, employees, investors, and partners are all tied to the same breath. Once difficulties emerge, “no matter how hard it is, you have to keep moving forward and hold the line.” Losing that momentum is not a matter of taking a short break.
• Raymond agreed that the company has to get through that one breath. The 2 also noted that companies Raymond had worked with had filed 3 Hong Kong prospectuses; when 庄明浩 asked whether all 3 had “expired,” Raymond did not elaborate and said that portion could be cut.
• 庄明浩’s final view is that both sides have compelling arguments. Founders have to decide under tight constraints and in a very short window, without knowing whether the decision is right or wrong. If a competitor has already answered the listing question, then even delaying is an answer that must be given at this point. The IPO is only one of many interlocking problems.