Vol.81 Zhipu or MiniMax—Whoever Becomes the First Listed Stock, Both Are Damn Good — A Crossover with 苔藓之火
Summary
Zhipu and MiniMax are racing toward Hong Kong listings at the same time, turning China’s foundation-model contest from a leaderboard narrative into a bruising financial face-off. Both passed their hearings on Thursday, posted prospectuses just 2 days apart, and both brought CICC in as a sponsor, making side-by-side comparisons of revenue, losses, cash and commercialization inevitable. Raymond’s reaction was blunt: Zhipu first made him feel “the sky was falling,” while MiniMax then came as a relief.
China’s leading model companies are roughly 1% the size of their US peers in R&D spending, revenue and valuation. Working backward from OpenAI’s roughly $500B-$700B valuation, 庄明浩 put leading Chinese model companies at around 1% of that figure; Chinese GPU stocks show a similar ratio relative to Nvidia. The point is not that either side must be wrong, but that the market’s current answer is simply “1%”—far below the 1/10 often used in the past or the roughly 1/7 implied by exchange rates.
The prospectuses reveal 2 different businesses: Zhipu is betting on China, To G/large enterprises and local deployment, while MiniMax is betting on global markets, consumer applications and multimodality. Zhipu generated less than RMB2B of revenue in the first half, posted roughly RMB1.8B in non-GAAP losses and more than RMB2B on a GAAP basis, spent about RMB1.1B on compute and held around RMB2.5B in cash; roughly 70% of MiniMax’s revenue comes from overseas, with 2025 revenue potentially reaching about $60M and cash still around $1B. “The first listed model stock” is therefore both a race within one sector and a pricing exercise for 2 different strategic identities.
The claim that Chinese models can be trained for a few million dollars does not conflict with companies spending billions of yuan on training each year, because the former usually refers only to one final training run. 庄明浩 used the $1.28M disclosed for DeepSeek R1 as an example: that was the last single run, excluding repeated evaluation, leaderboard optimization and parallel work on multiple model lines; model companies must also stay 1 to 1.5 steps ahead of their current products. The real mismatch is that revenue has lit only 3 or 4 of 50 lights, while costs are already preparing all 50.
Zhipu looks more like a fund-raiser that needs to catch the IPO window, while MiniMax looks more like a company proactively adding a capital-market tool while conditions remain workable. Zhipu is burning cash quickly, and the maturity of its RMB funds and exit demands from legacy-share SPVs may add pressure; MiniMax has invested at least RMB8B cumulatively, but theoretically has enough cash for several years. After listing, both companies gain not only financing capacity but also a public price and stock they can issue, giving them another card in future acquisitions, mergers or takeovers.
Traditional valuation frameworks barely work here; the more workable narratives are forward PS, comparables and the value of securing a seat at the endgame table. SenseTime already had RMB3B-RMB4B of revenue when it listed, yet was still roughly 10 times larger than the 2 companies today; on 2027 forward PS, OpenAI, with its larger revenue base, might actually be “the least extreme of the extreme.” Another approach is to ask: if only 5 to 10 companies ultimately control an “atomic bomb”-level capability, what is a seat at the table worth today?
After Zhipu and MiniMax jumped the gun, the greatest pressure may fall on Kimi, which remains committed to an AGI path, but “list immediately, remain independent or get acquired” are all only conditional scenarios. 庄明浩 believes Kimi could go further if it does not prioritize commercialization, but it could also be acquired under cash-flow or shareholder pressure; Raymond stresses that once an IPO reaches a critical stage, it is hard to stop casually: “You only get one breath for a listing.” The most practical advice for the companies behind them is to complete audits and filings early, keeping time and optionality on their side.
Deep dive
1. Two Prospectuses Turn the Model War into a Financial Face-Off
Zhipu and MiniMax both passed their Hong Kong Stock Exchange hearings on Thursday. Zhipu posted its prospectus on Friday and MiniMax on Sunday, with CICC involved as sponsor for both. Raymond said this effectively “forces every investor to compare them,” taking the contest from technical schools of thought straight to revenue, losses, cash and operating data.
Raymond compared the earlier technical debate to rival martial-arts schools: mixture-of-experts, attention mechanisms, training data and GPU-saving tricks can all be discussed, but the secondary market ultimately asks only, “Which one is better?” This is a “very bruising, very bloody” fight.
The reading order itself changed the psychological anchor. Zhipu first made him feel, “This industry is genuinely too hard”; MiniMax then made him feel that at least there was a path he could recognize. Had the order been reversed, he guessed MiniMax’s numbers would have scared him first, followed by the feeling that “the sky was falling” after reading Zhipu.
2. The Market Prices Chinese and US Model Companies at 1%
庄明浩 came away with 2 strong impressions at once. On one hand, “This business is too hard at this stage”; on the other, Chinese companies have achieved today’s model capabilities while spending only a few hundred million dollars, perhaps just 1% of what leading US companies have invested.
His rough comparison was this: if OpenAI is worth $500B-$700B, leading Chinese model companies are around $3.5B-$4B; R&D spending, revenue and valuation all sit at roughly 1%. Chinese GPU companies listing around the same time show a similar ratio relative to Nvidia.
That is well below the China-US mapping the primary market had grown accustomed to. Investors might accept 1/10, or roughly 1/7 after adjusting for exchange rates, but rarely expected 1%. 庄明浩 did not claim either side was wrong. He simply said, “What exists is reasonable”—this is the market’s real feedback at this stage.
3. The Prospectuses Firm Up the Giants’ Convictions but Leave Kimi More Conflicted
Asked to imagine Sam Altman’s reaction, 庄明浩 thought it would be surprise that China could spend so little to achieve these capabilities, mixed with concern about Chinese efficiency and relief that Chinese companies still face obvious difficulties monetizing To C and To B.
张一鸣, Alibaba and Tencent may see not an escalating threat but renewed confirmation of the incumbents’ advantages. Volcano Engine, DingTalk, Qwen, 阿福 and Tencent’s own moves already show that the strategy is taking shape; the prospectuses will only reinforce one message: “Then let’s do it.”
梁文锋 may be the least perturbed. DeepSeek has not materially changed its pace of technical evolution or commercial expansion from before its breakout to after it. 庄明浩 sees its path as the most committed of the group, and 2 prospectuses are unlikely to push it off course.
The most delicate position belongs to 杨植麟 and Kimi’s investors. Pessimists may recall that among the previous generation’s “AI Four,” only 2 ultimately listed and the third was left in a difficult position. Optimists may argue that if Kimi remains the purest AGI-focused player, staying the course could correspond to a larger endgame opportunity.
4. What Being “First” Really Adds Is a Financing and M&A Tool
Raymond had spent months expecting Kimi to list first, but IPO execution can be delayed by countless details, audits, filings and management decisions. Over the long run, ringing the bell 2 weeks earlier may not matter, and the historical record of many “first stocks” is poor; sometimes a company races for first simply because it genuinely needs that first.
Listing first still has clear strategic value. The company gets a public-market price and a clear anchor in a takeover; if it wants to acquire someone else, it can issue stock to fund the transaction. Private shares in an unlisted company lack liquidity, and counterparties are often unwilling to accept them.
The alternative path 庄明浩 raised was explicitly labeled by both speakers as “pure speculation, with no basis whatsoever”: if Kimi stays committed to AGI and does not prioritize commercialization, it could eventually be acquired because of cash-flow or shareholder pressure. Zhipu and MiniMax could also be acquired, but a listing would give them a fairer price and more transaction tools.
5. Zhipu’s Cash Pressure Explains the Early Push
Raymond’s prospectus-level summary was stark: Zhipu generated less than RMB2B of revenue in the first half, lost roughly RMB1.8B on a non-GAAP basis and more than RMB2B on a GAAP basis, and spent about RMB1.1B on compute alone. With approximately RMB2.5B in cash, its need for an IPO is “extremely strong.”
A wave of weekend commentary emphasized that Zhipu had more than RMB8B of funding, but the 2 speakers’ breakdown showed that more than RMB6B of that was bank credit, not cash on hand. 庄明浩’s observation was that the PR decision to focus its clarification on “whether it has enough money” was itself worth investors’ attention.
Beyond cash burn, Zhipu’s RMB fund shareholders may face shorter time horizons. 庄明浩 mentioned that the market had previously seen a Zhipu legacy-share SPV with an entry valuation of around RMB16B, roughly 30% below the last round’s RMB23B-RMB24B. That capital may also have a more explicit exit requirement.
6. Model Leaderboards No Longer Prove Commercial Position
A prospectus has to identify a “No. 1.” JD.com can segment B2C, iQIYI can segment long-form video and Bilibili can segment PUGC; foundation models are harder to segment, so they have to use different data sources to demonstrate leadership. Zhipu mainly cites token usage on OpenRouter, while MiniMax leans more on Artificial Analysis and multiple SOTA rankings.
庄明浩’s judgment was blunt: “All the metrics and scores have stopped working.” OpenRouter can reflect third-party market calls but does not count official direct connections from vendors; large customers usually connect directly and generate the most usage, so topping the ranking cannot represent total demand.
If forced to choose an outcome metric, he would look at the most primitive “amount dumped into the pit”—the phrase used in the original—or revenue. He also acknowledged that model launches still need third-party rankings for distribution. The problem is that each company is betting on different capabilities: baseline scores cannot be too low, but enterprise services, common sense and memory, user habits, X-platform data, and the combination of text and multimodality may determine retention and monetization.
7. Zhipu’s Industry Report Is Strikingly Restrained on China’s Enterprise Market
The industry report cited by Zhipu forecasts China’s enterprise AI market at approximately RMB100B in 2030, with 80% from local deployment and 20% from the cloud. Raymond noted that this total market is even smaller than the current revenue scale attributed to OpenAI in the cited context, which surprised him at first glance.
庄明浩 was unwilling to say whether RMB100B is too high or too low because the number “looks too much like something pulled out of thin air.” Without DeepSeek and the open-source shock triggered by R1, he believed the same listing materials would very likely have described a much larger market.
The dominance of local deployment aligns closely with Zhipu’s positioning: government, large enterprises, China’s digitalization level, and its traditional To G and large-To B resources all support the narrative. But precisely because the report’s conclusion fits the company’s business so well, the independent credibility of the 80/20 split becomes difficult to verify outside the prospectus context.
8. MiniMax’s $300B Market May Still Understate the Endgame
MiniMax’s industry report targets the global market: approximately $300B in 2030, with 75% from app applications and 25% from MaaS, namely APIs, cloud model services and other more B-oriented revenue. This exactly matches the first line of its prospectus: “a global AI foundation-model company.”
庄明浩 instead thought $300B was “still too small.” Based only on the forecasts cited on the program, OpenAI could reach $200B in revenue by 2030, while compute services could generate at least $100B; those 2 items alone total $300B. Google, xAI, coding, image generation, legal and other application companies are not even included.
If the entire industry really produces only $300B in 5 years, the economics of spending trillions of dollars today on data centers do not work. This reductio ad absurdum means either industry reports are understating the market or today’s capex bubble is extreme; at least 1 of those assumptions needs to be revalued.
Raymond reminded listeners that industry reports are usually commissioned and paid for by the company through a consulting firm, at a cost of roughly RMB500K-RMB800K, and their framing tends to align closely with the issuer. They are better read as “how the company defines its market” than as a neutral forecast.
9. There Is Still No Answer on Whether AI Can Flip from an Upright Triangle to an Inverted One
庄明浩 used a triangle of value distribution to explain MiniMax’s 75% application assumption. After mobile internet and cloud infrastructure matured, infrastructure took less and applications took the most, forming an “inverted triangle.” AI today is an “upright triangle,” with chips and infrastructure taking the most, models next, and applications taking the least—or even operating at negative gross margins.
Whether applications can capture 75% in 2030 depends on whether that triangle can flip. Supporters say applications sit closest to the user, but as of the program date there was still no answer to “model or product?” Directly copying the historical structure into AI looks more like wishful thinking.
Raymond added counterexamples from high-speed rail, railways and fiber optics. Network infrastructure can create enormous positive externalities for society while the gains accrue to other industries and the broader economy; the entities that build the infrastructure may not capture the most profit through fares or service fees. AI may not replicate the mobile-internet sample.
10. The Companies Are First Selling 2 National and Market Identities
Zhipu defines itself as “China’s leading artificial-intelligence company” and emphasizes its work on China’s pursuit of AGI innovation since 2019. With state-owned capital from Beijing, Hangzhou, Chengdu, Zhuhai and Shanghai, its overall identity is close to a “national champion.”
MiniMax, by contrast, was “born global.” Roughly 70% of its revenue comes from overseas and 30% from China. The program said Zhipu’s overseas revenue is about 10%, possibly from a model-alliance order involving a Southeast Asian sovereign state, but offered no further evidence.
This determines the market-size, product and valuation narratives. Zhipu first has to prove, “China must have me; I carry this role.” MiniMax has to prove that global consumer applications, multimodality and MaaS can jointly form a larger revenue pool.
11. The Prospectuses Expose the Companies’ Real Revenue Mix
Zhipu’s To G and large-enterprise business is heavier than outsiders had previously understood, and its accounts receivable also reflect a large-customer profile. Local deployment carries high gross margins, but ongoing training costs are extremely heavy, leaving To B with negative gross margins.
MiniMax had often drawn attention for video, voice and API capabilities, but the prospectus shows that a large share of revenue still comes from chat and “gacha”-style payments. Raymond’s surprise was: “How is everyone making money from chatting? Everyone is making money from gacha.”
MiniMax bears the inference cost of a large consumer base, which has kept historical gross margins low. But it has multiple model categories—text, voice and video—with different prices and margins mixed into one number. The average may look better than a negative margin, while making a precise breakdown more difficult.
12. There Is No Consensus on Which of the 50 Lights Should Be Turned On First
Both prospectuses show companies progressively lighting up a matrix: domestic and overseas markets; To G, To B and To C; APIs, private deployment, consumer applications, multimodality and Agents. 庄明浩 sees only a snapshot of a dynamic business process, insufficient to infer a single correct path.
Founders’ backgrounds, resources, organizational character and accidental opportunities determine which 3 or 4 lights come on first; if the companies started over, the order might differ. Kimi may also have lit many lights that later went dark while continuing to explore reasoning, Agents and other new directions.
US giants have the capital to light everything. Chinese startups, constrained by 1% of the resources, have to make choices. Looking back from 2030, perhaps 40 of the 50 lights will have been collectively switched off, but no one has the answer sheet today.
13. Coding, Multimodality and Agents Look Different at Different Time Horizons
Asked which capabilities might still exist in 2030, 庄明浩 first said Coding would definitely remain, as would multimodality; Agents were also a major combination light. AI for Science might matter, though both hosts admitted they were not familiar enough with it.
But when pressed on which lights foundation-model companies must turn on by the end of 2025, 庄明浩 said, “Coding has no meaning anymore, and neither do Agents,” shifting to continued evolution of language models themselves, memory systems, context and AI for Science. The program offered different answers to different questions, so they cannot be reduced to a single list of “must-have” capabilities.
After Anthropic restricted some Chinese users, Zhipu quickly launched GLM code, using a strategy roughly equivalent to “you charge $20, I charge RMB20.” Raymond admired not only the price but also Zhipu’s response speed after its consumer presence had weakened.
The program said GLM code had already approached RMB100M in revenue and roughly 150,000 users. Cursor’s claim that it had launched its own Composer was also viewed by many as potentially based on Zhipu’s model. The 2 speakers did not treat that speculation as confirmed fact.
14. The Agent Battle Is Still a Boundary War between Model Vendors and Product Companies
After the Doubao phone launched, Zhipu also revived the AutoGLM it had previously built. Together with GLM code, Raymond believes the company is correcting its earlier image of being state-enterprise-like and weak on internet instincts—and that it is “keeping up reasonably well.”
The key Agent question is not whether it can be built, but who can get users to pay. Whatever their sales and marketing expenses, Manus and Jasper should have relatively clear revenue and ARR; in MiniMax’s prospectus, most of the revenue from its apps in the first 9 months of 2025 came from newly launched Agents, which contributed significantly.
“Model or product?” will be fought out more intensely in 2026. Claude Code and Codex are growing quickly, but Cursor is also building its own moat. Model vendors may move forward and absorb applications, while product companies may preserve value through workflows, distribution and user relationships.
Raymond’s judgment is that, across the market, having a model solve tasks is one of the direct ways a foundation-model company can generate revenue. Even with high marketing expenses and weak gross margins, that capability cannot be abandoned outside Coding.
15. A Million-Dollar Training Run Is Not a Billion-Yuan Annual Budget
The $1.28M in the DeepSeek R1 paper was, 庄明浩 explained, the cost of “the last single training run”: the final complete run that was trained and released publicly, not the full cost of the R&D project from start to finish.
Before delivery, a model may be trained repeatedly, with the number of runs affected by evaluation results, target performance, competition and rankings. “Gaming the tests” also creates training expense. A single successful figure cannot be directly compared with a company’s annual compute bill.
Companies also run language, multimodal and next-generation models in parallel, like a Gantt chart, staying 1 to 1.5 steps ahead of the current business. Revenue comes only from the 3 or 4 lights already switched on, while the technical foundation has to prepare for all 50. That is the root of the cost-revenue mismatch.
16. Inference Costs Mean Short-Term Costs Rise as Commercialization Succeeds
The billion-yuan figures discussed by the 2 speakers mainly concern training costs. Inference expense must be analyzed separately through cost of revenue and gross margin. MiniMax has more than 20 million monthly active users, and their app usage continuously consumes compute.
“A large user base” therefore does not automatically mean attractive profits. The more active the product, the larger the inference bill; if subscriptions, gacha or API pricing are insufficient, scale may initially magnify costs and losses. Zhipu and MiniMax expose the same pressure through training investment and inference burden respectively.
17. DeepSeek Is Personally Funding a National Champion
Raymond inferred from the 2 companies’ billion-yuan-scale training costs that even DeepSeek, China’s most frugal model company, could not have spent only a few million dollars. 庄明浩’s answer was “probably not”—its cost should not be much below RMB1B. He did not provide a definitive annual-cost figure.
Raymond believes DeepSeek may not face sales and user-acquisition expenses of the same scale, and quantitative institutions may previously have stockpiled GPUs and retained cash flow. But it still has to support several hundred employees; compute and labor are real costs.
After DeepSeek went viral, local state capital, market-based funds, dollar capital and large companies may all have approached it. 庄明浩 said that, at least as of the program date, it was still bearing annual costs of several hundred million yuan in its own way.
Neither speaker concluded when it might run out of runway. Raymond merely raised the question investors should face directly: if a national champion depends on an individual’s patience and cash flow for an extended period, sustainability itself is a variable that must be monitored.
18. China Saved on Compute but Not on the Revenue Mismatch
The program cited reporting that put OpenAI’s expected 2025 R&D cost at approximately $13.4B, or nearly RMB100B; Zhipu and MiniMax are operating at the billion-yuan level. Raymond used a school-year analogy: OpenAI is like a fifth-grade student and Chinese companies like fourth-graders, but fifth grade costs RMB100B in tuition while fourth grade costs only RMB1B.
Chinese companies may not immediately raise spending to US levels as they advance, but OpenAI will continue moving up the frontier and absorbing massive costs. The expense gap may not be a simple linear relationship.
Domestic chips and compute constraints change the cost structure, but 庄明浩 does not believe lifting the restrictions would take China’s spending directly from 1% of US levels to 1/10. Nvidia GPUs were still used more heavily from 2022 through the first half of 2025, while domestic chips began to see broader use from 2025 onward.
19. Cheap APIs Must Be Validated by Gross Margin, Not Sticker Price
Chinese models are often marketed at a fraction of US peers’ API prices or even less, but financial statements ultimately have to answer who is subsidizing the difference. Raymond asked where the price advantage lands; 庄明浩’s answer was: “In the gross margin of the To B business.”
That business has negative gross margins at Zhipu. Pricing, competitive conditions and business structure can all affect the result, so call volume or market share cannot prove that unit economics work.
MiniMax’s enterprise-model margins may look better in part because price and cost differences across text, voice and video are averaged together. The 2 speakers acknowledged that the prospectus lacks enough granularity to reliably separate the true gross margin of each model category.
20. China’s AI Talent Dividend Remains, but 2026 Could Bring a Sharp Repricing
Raymond annualized R&D payroll expense and divided it by headcount, arriving at roughly RMB810K per Zhipu researcher and RMB940K per MiniMax researcher—less than $150K. 庄明浩’s reaction was: “That fits China’s national conditions, but it really does not fit the current AI competition.”
The US has seen frequent talent offers worth more than $100M, while Chinese teams draw heavily from domestic mathematics and computer-science graduates who move into AI after several years of work. MiniMax also emphasizes that its team is young, so current average pay is not unreasonable.
This may simply be a lagging snapshot: the employees reflected in the statements may have joined in 2024 or 2025. Tencent has reportedly begun aggressively poaching ByteDance talent from 2025, potentially offering 2x pay, while offers for top master’s and PhD graduates are also rising. 庄明浩 expects average cost per employee could change materially in 2026.
21. Zhipu Is Listing for Cash; MiniMax Is Listing for Optionality
Zhipu filed with the China Securities Regulatory Commission in April 2025, preparing for an A-share listing with CICC as its sponsor, but the program said it does not know why the company later switched to Hong Kong. Raymond said changes in the A-share funding pool and issuance cadence may have been factors to weigh, but offered no definitive explanation.
MiniMax has invested at least approximately RMB8B cumulatively. 2025 revenue may reach around $60M, while losses remain around $200M; compute costs were about $100M in the first 9 months, but cash on the balance sheet was approximately $1B, theoretically enough to support the company for years.
MiniMax therefore is not being directly forced by its cash balance in the way Zhipu is. 庄明浩 speculated that with the business rising, Hong Kong sentiment still workable and competitors moving in a cluster, the board may have decided that “it is worth trying” and taken the opportunity to secure a public financing channel.
The program also noted that Chinese capital from state-owned investors, Tencent, Alibaba and others is already involved. The next question is how the public markets will finance the company. For MiniMax and Kimi, the US is not viable, A-shares were never considered, and Hong Kong is effectively the only public-market option.
22. IPO Pricing Is Relative Ranking—and a Gamble on the Window
Raymond’s investment-banking experience is that “there is no optimal answer when it comes to listing.” He once opened a Chinese company roadshow update with, “Syria is at war,” because war raises risk aversion and sends stocks lower, while an IPO is an even higher-risk subset of equities. Events far away can directly determine whether an offering gets done.
He also recalled Bilibili roadshows during trade friction and a sudden snowstorm when he landed, illustrating how difficult the market window is to predict. If Hong Kong encounters any black swan in early 2026, MiniMax may still be able to retreat and prepare again, while Zhipu may have less capacity to absorb the shock.
Sequential listings by comparable companies also create dynamic discounts. If the first company prices at 10x PS, the second may get only 8x unless it is clearly stronger; if the market considers the second stronger, the first may become harder to sell. Daily Fresh, Dingdong, JD.com, Alibaba, Lyft and Uber were cited as examples of this ranking dynamic.
The 2 companies also differ in their sponsor arrangements: MiniMax has CICC and UBS, while Zhipu has only CICC. The program linked this to restrictions preventing US investors from investing in Chinese chip, AI, robotics and biotech projects, and argued that the absence of a US investment bank may limit participation by some dollar institutions.
Biren Technology’s prospectus was overshadowed by the 2 companies. The program also noted that all 4 companies listing that day appeared to break issue price. The speakers believed that if even relatively tangible offerings cannot hold their prices, institutional and retail sentiment could be affected, though the magnitude remains impossible to judge.
Raymond ultimately rejected the idea that a pause is simply “taking a breather”: “You only get one breath for a listing.” Once the company, employees, investors and partners are bound together, a simultaneous loss of morale and market window may make a second attempt difficult. He also mentioned that his own company had filed Hong Kong prospectuses 3 times and seen all 3 lapse.
23. The Final Valuation Buys a Seat at a Table That May Disappear
Traditional PS is difficult to apply directly. SenseTime already had RMB3B-RMB4B of revenue when it listed, yet was still roughly 10 times larger than the 2 companies today. Yunzhisheng, with a market capitalization of around HK$70B on the day of the program, was merely another reference point for sentiment and comparable pricing.
Raymond said that if he were the project analyst, he might use 2027 forward PS to educate investors: first assume rapid revenue growth in 2026 and 2027, then work backward to today’s price. On that basis, OpenAI, with its larger revenue base, actually looks like “the least extreme of the extreme,” while Chinese companies’ PS is not inherently cheap.
Another framework is to ask what an atomic bomb is worth. If the engineering-cost gap between China and the US is thought of as $1B versus $100B, then ask how much it costs to build a model slightly weaker than OpenAI’s. The investor is not buying current profit but a seat at the senior-year table; the company must keep up every year and cannot be held back.
Mistral offers a geopolitical valuation reference. By the program’s account, it is already roughly 2 rounds behind and has never appeared on an SOTA leaderboard, yet is still valued at around $10B, with ASML holding approximately 12%-13%. Europe is willing to pay for a domestic model not controlled by the US or China, offering another valuation angle—but not a direct pricing conclusion for Chinese companies.
These projects are closer to “Star Wars” or the “Manhattan Project.” Early-stage VCs and late-stage or secondary investors judge them differently. Even at a $1B-$2B valuation in 2023, a fund might still decide it had to take another shot and secure a few percentage points of exposure to a leading company, because “if you are not on the list of the biggest companies, you are nothing.”
The program offered no single answer for 杨植麟 and Kimi. In the hypothetical scenarios, remaining independent, secretly filing, waiting, raising capital or being acquired all have substantial rationales. The one clear operating recommendation is to complete audits and other preparations early, keeping the options open. AI founders today have to build products while also “going to Saudi Arabia to meet the royal family and following the US presidential election”—a level of difficulty far beyond the mobile-internet era.