Within 48 hours, Zhipu and MiniMax completed their listings on the HKEX, elevating 'domestic large models' from a technical spectacle in public discourse to a scrutiny of financial statements and cash flows. While capital markets are certainly interested in AGI, they are more concerned with whether you are in the business of 'models' or 'making money with models.'

From publicly disclosed information, this pair of 'giants' appear to be two fish swimming in the same river: both operating within a triangle of high growth, heavy investment, and persistent losses; the bulk of their R&D expenses goes towards computing power, which acts like an unavoidable tax; while they have substantial cash on hand, it burns away just as quickly. The difference lies in their swimming styles: Zhipu resembles a B2B company focused on 'project-based + delivery,' with a longer chain of trust from state-owned assets and government-enterprise clients; MiniMax, on the other hand, is more akin to a C2C company centered on 'product-based + distribution,' relying on application go overseas (overseas expansion) to acquire users and revenue, then using capital to further amplify its imagination.

Listing is not graduation; it's a make-up exam. The Hong Kong stock market is willing to price large model companies, but only if they can clearly articulate two things: first, whether revenue growth can penetrate the fog of 'subsidies, marketing, and one-time deliveries'; second, whether model capabilities can be translated into replicable, scalable, and profitable products/platforms. Otherwise, the louder the bell rings, the more it serves as a reminder to the industry—the story has reached this point, and it's time to reconcile the accounts.

Zhipu's path is quite typical: first, gain trust, then scale up. Trust comes from two levels: one is its 'legitimate' technical and organizational background, and the other is a delivery approach more suited to government-enterprise needs—localized deployment, exclusive models, and industry-specific solutions. Its revenue structure naturally resembles that of the pre-SaaS era of 'enterprise software': large contract values, seemingly respectable gross margins, payment collection cycles tied to acceptance milestones, and growth that resembles steps rather than curves. The historical revenue and periodic losses disclosed by Zhipu in its prospectus largely reflect this 'project elasticity': growth can be rapid, but its continuity depends on the rhythm of orders and deliveries.

The advantage of this model is its high certainty: B2B clients are willing to pay for compliance, security, and controllability, especially in sectors like finance, government affairs, and energy, where 'usability' and 'manageability' often outweigh 'the strongest.' Consequently, Zhipu's narrative leans more towards technology and capability stacks: models, platforms, toolchains, and privatization capabilities—this is the language it uses to secure budgets from clients and to command a 'long-term premium' from capital.

However, behind this certainty lies another ceiling: project-based models fear having 'per capita output' locked in. The better you sell, the more your delivery team expands; the more you emphasize customization, the harder it is to achieve scale effects of standardized products. More realistically, the government-enterprise market inherently has 'budget cycles' and 'procurement cycles,' and when the macro environment tightens, growth can be put on pause. In other words, Zhipu today more closely resembles an 'AI-era enterprise software and computing infrastructure provider,' and the common fate of enterprise software is to either transform delivery into platform capabilities or to forever circle in the triangle of expanding personnel, projects, and management.

MiniMax, on the other hand, has taken the opposite direction: first, acquire users, then force out commercialization. Its core assets are not contracts but rather its product matrix and user base—applications like Talkie/Xingye and Hailuo, which in the context of overseas expansion, resemble 'AI-native content platforms': using generative capabilities to create experiences, retaining users with those experiences, monetizing through payments and advertising, and then using data and cash flows to fuel model iterations. Public disclosures show that MiniMax's revenue primarily comes from overseas, and its commercialization strategies are closer to C2C tactics like subscriptions, in-app purchases, and advertising.

The key to the C2C route lies in 'growth continuity': you don't need to wait for acceptance, nor for budget approvals; as long as the product hits the mark, it can spread exponentially. But C2C is also more ruthless: it's a distribution business. User acquisition costs, marketing, channels, platform rules, and compliance risks all directly feed into the profit and loss statement. The faster you grow, the more variable costs of computing power and bandwidth follow you like a shadow; if you stop marketing, growth may immediately slow; if you market too aggressively, gross margins and losses can rapidly deteriorate. Consequently, we see that MiniMax's cost structure more closely resembles that of an internet application company: high periodic sales and marketing expenses, significant pressure on gross margins—this is not an operational 'mistake' but rather the cost of its chosen model.

When viewed together, these two companies are actually answering the same question: should the value of large models be reflected in 'selling capabilities' or 'selling experiences'? The former relies on trust and delivery, while the latter on products and distribution; the former more easily secures institutional budgets, while the latter more readily captures capital's imagination. The difference lies not in idealism but in accounting categories.

The financial statements of large model companies share a nearly ruthless commonality: within R&D expenses, computing power accounts for an extremely high proportion; and R&D expenses often constitute the bulk of period expenses. In other words, the more you resemble a 'legitimate large model company,' the more you resemble a 'high-intensity GPU-burning company.' This is not a problem unique to domestic companies but rather an industry structure in the early stages of generative AI commercialization: both training and inference incur hard costs, and it's difficult to fully offset them through efficiency improvements in the short term.

This also explains why, despite similar high growth, the market exhibits two different sentiments: for B2B, the market worries about 'scale without economy'—the more you deliver, the more personnel you need, the more accounts receivable you accumulate, and the tighter your cash flow becomes; for C2C, the market worries about 'growth not being free'—the more users you have, the higher your inference costs rise, the harder it is to stop marketing, and the thinner your gross margins become.

Consequently, after listing, the true focus should no longer be on 'how many times revenue has multiplied this year,' but rather on two more fundamental tables: unit economic models and cash flow structures.

For Zhipu, the key is not just having a more comfortable gross margin but also being able to transform delivery from 'labor-intensive' to 'platform-intensive.' The most intuitive signals include: whether the proportion of cloud API revenue can continuously rise, whether the proportion of customized solutions per client can decrease, whether the structure of accounts receivable and contract liabilities can improve, and whether R&D investment can be translated into reusable industry components rather than starting from scratch each time. The coexistence of revenue growth and expanding losses presented in its disclosures essentially reflects a typical phase where 'scale is amplifying, but platformization hasn't fully caught up yet.'

For MiniMax, the core contradiction resembles the age-old proposition of 'platform versus content companies': you can use products to drive growth, but you must make that growth cheaper and more durable. A high proportion of overseas revenue is certainly good—dollar-denominated income, a larger market, and more mature payment habits; but overseas also means heavier compliance costs, stronger platform dependencies, and more frequent rule changes. More importantly, C2C AI applications easily fall into a 'homogenization of features + competition for emotional value': when the gap in model capabilities is quickly closed, what often determines success or failure is not parameters but rather character design, community relationships, content supply, distribution efficiency, and payment design. In other words, on the surface, MiniMax is selling AI, but at its core, it's more like creating 'new forms of content and companionship.' Once this type of business succeeds, the ceiling is very high; but when it doesn't, losses can be very real.

Both companies also face a common 'mid-term trap': as the industry transitions from 'training-driven' to 'inference-driven,' the cost war will become even more brutal. Training is a one-time heavy asset investment, while inference is an ongoing variable cost; training can be resolved through financing, but inference must be resolved through commercialization. You'll see many companies talking about 'efficiency improvements, MoE, and long contexts' at the model level, but what capital markets care more about is—can you reduce inference costs to a sufficiently low level and achieve gross margins under the same user scale and call volume? Otherwise, the larger the scale, the faster the losses, and what is termed 'growth' can actually become a curse.

Therefore, placing the listings of Zhipu and MiniMax within the industry coordinate system, they more closely resemble stress tests of two 'viable paths': one proving whether independent large model companies can grow into platform-based enterprise service providers in China's government-enterprise market; the other proving whether Chinese teams can create AI-native application ecosystems with paying capabilities in the global C2C market.

Neither proof is easy, but at least it shifts the question from 'whose model is stronger' to 'whose business is more viable.'

Many view this 'dual listing' as a highlight moment for domestic large models, but a more accurate description is that it's the first time domestic large models have been forced to explain themselves in 'financial language.' Over the past two years, the industry has relied on technological breakthroughs and financing to sustain itself; over the next two years, it will need to survive through product pricing, cost efficiency, and cash flow discipline.

Zhipu's challenge is to transform 'trustworthy delivery' into a 'replicable platform,' making growth less dependent on manpower and customization; MiniMax's challenge is to evolve 'hit applications' into a 'sustainable ecosystem,' making growth less reliant on marketing and luck. One must move from projects to platforms, the other from products to systems. While their routes may seem opposite, they will ultimately converge in the same river: whoever can bear more stable demand at lower inference costs and turn model capabilities into a sustainable commercial structure will be able to use listing as an amplifier rather than a life-extending pill.

Capital markets have already granted them an entry ticket; the next assignments they must submit are far from romantic: how to improve gross margins, stabilize cash flows, and achieve growth that doesn't become 'more expensive as it grows.' The story of large models hasn't ended, but the way the story is told has changed—from 'what I can do' to 'what I can earn.' This may well be the greatest significance of the HKEX bell.

Written by: Lyu Xuemei

Edited by: Dahai

(Image sources: Network, removal upon infringement claim)