Large Model Vendors Race Toward IPO: Valuation, Commercialization, and AGI Ambitions
A leading independent Chinese LLM vendor’s IPO prospectus offers a rare financial lens on the sector, reframing valuation around revenue, costs, and cash flow. This article breaks down its private-deployment and MaaS revenue models, explains the cost structure behind widening losses, and examines the transition from model demos to production-grade deployment as vendors prepare for public markets and AGI.
I. A Prospectus Lifts the Veil on Industry Valuation
For a long time, public understanding of China’s leading LLM startups was shaped mainly by funding announcements and launch events. Valuation figures, revenue mix, and loss scale remained obscured. Until recently, a leading independent LLM vendor in China passed its listing hearing and disclosed a prospectus through public channels, giving the first clear look: based on its latest funding round, the company was valued at approximately RMB 24.3 billion.
The significance of this number is not whether it is high or low, but that it serves as an anchor. Before this, primary-market valuation of LLM companies relied more on “narrative pricing”: model leaderboard rankings, team background, and funding-round heat. The prospectus pulls the narrative back to financial statements: revenue, cost, gross margin, R&D expenses, and cash flow. As valuation moves from “story” to “numbers,” the sector’s valuation logic shifts from “who has the more impressive technology” to “who can turn technology into a sustainable commercial loop faster.”
The timing matters. Three years after generative AI shook the industry, the noise of the “hundred-model battle” has receded, and the players left at the table have completed a first round of elimination. These survivors have proven that their technology innovation is not inferior to international peers, and now they are sprinting toward capital markets—a typical signal of transition from adolescence to adulthood.
II. Commercialization Capability: Breaking Down Revenue Structure and Growth Curve
According to the prospectus, the vendor’s core commercialization revenue comes from two major segments, which are also the two most representative business models among current Chinese LLM vendors:
1. On-Premises Deployment (Private Delivery)
For B2B customers, the vendor provides private deployment of LLMs, delivering the model and inference services into the customer’s own data center or dedicated environment. This business has several characteristics:
- Large ticket size: usually project-based, covering model licensing, custom fine-tuning, system integration, and O&M services;
- Strong stickiness: once deployed, customers become deeply bound on data assets and workflows, making replacement costly;
- Broad industry coverage: mainly internet technology, public services, telecom, traditional enterprises, consumer electronics, retail, media, and consulting. Public-service and telecom customers have rigid requirements that data not leave their domain, making private solutions a natural fit.
The prospectus shows that nearly 85% of the vendor’s revenue comes from on-premises deployment. This explains a key reality in China’s LLM commercialization: under the dual constraints of data compliance and industry access, private delivery remains the primary monetization model in the B2B market.
2. Cloud Deployment (MaaS Services)
Through a MaaS (Model-as-a-Service) platform, the vendor offers model API access and token-based invocation services to developers. This is the more mainstream model internationally, billed by usage, with controllable marginal costs and the scale elasticity typical of cloud services. Its revenue model can be expressed in simplified pseudocode:
def maas_revenue(input_tokens, output_tokens, unit_price):# Input and output are priced separately; output unit price is usually highercost = input_tokens * unit_price["input"] \+ output_tokens * unit_price["output"]return cost# Monthly bill per customer ≈ daily call volume × 30 × tokens per call
The ceiling of the MaaS model depends on two variables: the scale of the developer ecosystem and the pace at which per-token inference costs fall. Every order-of-magnitude decline in inference cost opens new usage volume for previously uneconomic scenarios, such as long-document analysis and multi-turn agent tasks.
3. Three Observations on Growth Data
From 2022 to 2024, the vendor’s revenue grew from RMB 60 million to RMB 310 million, a CAGR of about 130%. In the latest two first-half periods, revenue grew more than 300% year over year. Based on the latest annual revenue, its share of the enterprise LLM market reached 6.6%, ranking first among independent LLM vendors. Three observations deserve elaboration:
- Base effect: early revenue was extremely low, so high growth is inevitable. The real question is whether absolute revenue can cross the scale threshold in 2025–2026;
- Heavy revenue mix: 85% dependence on private delivery means growth is constrained to a considerable extent by project delivery cycles and labor input, giving it less scale elasticity than a pure MaaS model;
- Overseas increment: in the latest reporting period, the overseas revenue share rose rapidly to about 12%, mainly from Southeast Asia. Against the rise of “sovereign LLM” demand, Chinese large-model technology achieved systematic overseas expansion for the first time, and the long-term opportunity in this business may be larger than in the domestic market.
III. Losses and Cash Burn: How to Read an LLM Company’s Balance Sheet
The prospectus also exposes the other side of the coin: losses far larger than revenue growth, with R&D expenses continuing to rise sharply. Measured by mature software company standards, this is clearly not a passing report card. But using a mature company’s yardstick on an AGI company less than six years old is itself a mismatch.
Understanding this mismatch requires returning to the LLM cost structure:
- Upfront training costs: each generational upgrade of a foundation model means compute investment on the order of hundreds of thousands of GPU card-hours, and the commercial return cannot be determined when the investment is made;
- Inference costs expand linearly with scale: the faster revenue grows, the more tokens are consumed, and GPU procurement and rental expenses rise in tandem, suppressing gross margin in the short term;
- Rigid talent costs: compensation packages for top algorithm researchers sit at the highest tier in the industry and cannot be diluted through scale.
Put another way, current LLM losses are not a signal of business failure; they are the result of capitalizing future capacity in advance. The real question is where the burn curve ends. The industry consensus is that unit economics turn positive only when the slope of declining inference costs exceeds the slope of revenue growth. Before that point, financing capability and cash flow management are part of the core competitiveness.
IV. From Demos to Deployment: The Real Test of the Transition
The market’s expectation for this generation of LLM companies is no longer benchmark scores but the proof of a complete proposition: how to convert novel, demo-oriented model technology into practical, adapted, end-to-end deployment capability. At least three hurdles stand on this path:
- Engineering hurdle: between model capability and production environments lie a series of engineering problems—context management, retrieval augmentation, tool calling, hallucination suppression, and evaluation systems. Enterprise customers are not buying model weights; they are buying a complete solution of model + engineering + services;
- Scenario hurdle: general capability does not equal scenario value. Vertical know-how, private data governance, and business-process restructuring are key to realizing value. This also explains why private-deployment revenue remains so high;
- Cost hurdle: customers are willing to pay for solutions that are an order of magnitude cheaper than manual work and a dimension smarter than traditional software. Pricing room is determined jointly by inference cost and replacement cost.
For developers, this transition period means an opportunity window: application-layer engineering capabilities around models—Agent orchestration, RAG optimization, evaluation, and observability—are becoming a new skill premium. General categories offered by cloud providers, such as object storage, message queues, container platforms, logging, and monitoring and alerting, form the standard infrastructure foundation for these applications.
V. Conclusion: A Bet That Must Convince Everyone
Before the widely believed future of AGI arrives, independent LLM vendors can only prepare as thoroughly as possible. That includes a series of complex strategic tradeoffs: self-developed technology versus ecosystem partnership, domestic priority versus overseas expansion, and deep private deployment versus betting on MaaS. An IPO essentially lays these internal tradeoffs open before the public market for scrutiny by all stakeholders.
For the industry, the first company to “reach shore” provides a reference valuation coordinate system and a commercialization sample. For developers and enterprise users, entry into capital markets means the leading vendors’ service stability and long-term investment commitments gain stronger endorsement. Change often happens suddenly and exponentially at a certain moment; before that, whether a company can survive the long investment period will determine who remains at the table.
