AI Model Valuations Shift to Performance-Driven Metrics as China's Vendors Lose Narrative Premium
The AI large-model market is moving from narrative-driven valuation to financial verification. This article examines how public-market pricing, revenue structure, cost pressure, and API usage economics are reshaping valuations for global AI labs and Chinese model vendors, and why revenue quality, cost structure, customer retention, and unit economics will determine the next phase of differentiation.
From Stories to Numbers: A Paradigm Shift in AI Valuation
Over the past two years, valuation growth in the global AI large-model sector has been almost decoupled from financial performance. The market was willing to pay a premium for the imagination of “intelligence emergence,” and capital chased metrics closely tied to technology narratives: parameter scale, multimodal capabilities, and reasoning performance. As leading labs begin public listing procedures and disclose detailed financials, however, the pricing mechanism is undergoing a fundamental shift: the valuation anchor is moving from “future possibilities” to “currently verifiable financial performance.”
The core marker of this shift is greater pricing transparency. When one leading AI lab gave a roadshow price range of $70 to $78 per share, the implied market capitalization was approximately $916 billion to $1.02 trillion. Including reserved portions such as employee options and restricted shares, the fully diluted valuation was approximately $1.4 trillion. At the midpoint of the range, the implied market capitalization was about $968.4 billion, almost flat with the $965 billion post-money valuation from its previous private financing round. This suggests that public-market investors did not grant the same “growth expectation premium” as the primary market, but instead scrutinized the fundamentals of its existing business more cautiously.
This valuation “convergence” is not an isolated event. Another leading AI lab reached a valuation on the order of $1.25 trillion in primary-market financing, but raised only $11 billion, below previous market expectations. Both leading labs have entered the trillion-dollar valuation range, yet market pricing has become more conservative. Behind this is a shift in investor consensus: as model capability gaps narrow and technology narratives become more homogeneous, financial numbers become the only yardstick for distinguishing company quality.
A Full Review of Revenue Structure and Profitability
Greater financial transparency allows outsiders, for the first time, to systematically examine the actual operating conditions of these high-valuation AI companies. Take one leading lab as an example: it recognized $33.24 billion in revenue in the first nine months, including about $17 billion in the third quarter, indicating strong growth momentum. The key signal, however, lies in the revenue structure: about 80% of first-nine-month revenue came from usage-based API calls. This means its business model is highly dependent on actual usage by enterprise customers, rather than the “book certainty” brought by prepaid contracts or subscriptions.
On profitability, although the company reported $4.08 billion in adjusted operating profit, it still had an operating loss of about $1.26 billion under U.S. GAAP. The difference mainly comes from stock-based compensation costs and complex depreciation arrangements for compute infrastructure. More noteworthy is that the company disclosed contractual obligations and other commitments totaling about $900.9 billion, almost all related to compute and infrastructure. This means that in the coming years, a large amount of the company’s cash flow will be tied up in fixed-asset investment, and the possibility of positive free cash flow in the near term is limited.
The financial data of another leading lab presents different problems. Its third-quarter revenue was about $10 billion, up nearly 50% quarter over quarter, still a bright growth rate. But its adjusted operating margin was about -109%, an improvement from the previous quarter, while operating losses continued to expand. This shows that even with rapid revenue growth, the costs of model training and inference are still significantly eroding profit margins.
A common trend emerges from the data of these two companies: the revenue ceiling for AI large-model companies has not yet appeared, but the cost curve is similarly steep. Whether valuations can remain at the trillion-dollar scale ultimately depends on how quickly the gap between these two variables narrows.
China’s Model Vendors: The Narrative Tide Recedes
As the financial data of leading global labs becomes more transparent, the “narrative premium” that Chinese AI model vendors once relied on is also fading. In the summer, domestic vendors experienced a dense wave of model releases. But based on a sample of eight major vendors, with September 1 as the dividing point, model release events fell from 10 to 7 across two equal 60-day windows before and after that date. A slower release cadence is not necessarily bad in itself—it may mean vendors are shifting from a “parameter race” to “application refinement”—but the market tends to interpret it as an inevitable response to the change in valuation logic.
During the “narrative support” phase, the valuation logic of Chinese AI model vendors was driven almost entirely by two factors: first, “tracking mapping” of technology progress by leading overseas labs—that is, after overseas companies upgraded their models, expectations for domestic substitution in the Chinese market rose in tandem; second, policy support for the AI industry. Together, these formed an “expectation-discounting” model: the market was willing to pay a premium for technology capabilities that had not yet been commercialized, because it believed those capabilities would convert into revenue within a few years.
But once the financial data of leading overseas labs became public, the market had a reference frame. Investors could estimate more precisely how long a cycle, how much capital expenditure, and how high customer acquisition costs would be required for an AI company to move from model R&D to scaled API call revenue under a similar technology route. This “benchmark valuation method” replaced the “imagination valuation method,” meaning Chinese AI model vendors must present their own revenue growth curves, customer retention rates, and unit economics to support valuations.
Back to Business Fundamentals: From Model Capability to Revenue Conversion Efficiency
The shift in valuation logic imposes stricter requirements on Chinese AI model vendors. In the narrative era, model evaluation scores, leaderboard rankings, and parameter counts were core KPIs for financing. In the performance era, the market is truly focused on three types of financial metrics:
First, revenue quality. That is, whether revenue comes from real enterprise-level API calls or from “contract stacking” through prepayments. The share of usage-based revenue is a key observation point—a high share means the business model has been tested by the real market, while a low share may hide risks such as channel stuffing or related-party transactions.
Second, cost structure. The costs of large-model companies mainly consist of three parts: compute procurement and depreciation, R&D staff compensation, and inference infrastructure operations. The first two are rigid for a considerable period, so improvements in inference efficiency—supporting larger-scale concurrent requests on the same compute—become the key variable determining gross margin. Industry technology solutions are shifting from “a single model serving a single task” to “multi-model mixed scheduling,” routing requests of different complexity to models of different sizes to reduce per-token inference costs.
Third, customer retention and expansion. API business revenue growth depends on two things: natural growth in usage by existing customers, through model capability upgrades and context window expansion, and the speed of new customer acquisition. As model capabilities converge, customer decision factors are shifting from “single-model performance” to “platform comprehensive capability”—including private data deployment options, compatibility with existing cloud infrastructure, and seamless upgrade support for existing applications during model iterations.
Adaptation: Chinese Vendors Shift Their Strategic Focus
Facing the shift in valuation logic, the strategic focus of Chinese AI model vendors is also shifting in observable ways. At the technology level, a lower release frequency does not mean R&D has stalled; rather, it reflects a shift from “refreshing public leaderboards” to “improving usability in real scenarios.” Specific signs include greater emphasis on the deployment results of long-text processing in vertical industries, stronger emphasis on the stability of tool calling and structured output, and performance optimization under cost constraints.
At the commercialization level, vendors are repositioning their roles in the value chain. The business model of externally supplying models is diverging into two directions. The first is “infrastructure-style” pricing for developer communities—using usage-based token pricing as the core competitiveness and pursuing marginal cost declines under scale effects. The second is deep customization for enterprise customers—using private deployment, industry fine-tuning, and data security compliance as core selling points and maximizing per-customer value. The former benchmarks the usage-based pricing model of leading overseas labs; the latter is closer to the sales logic of traditional enterprise software services.
At the capital strategy level, Chinese vendors are also adjusting financing pace and valuation expectations. Under pressure from the transparent valuation pricing of leading overseas labs, financing logic anchored to “overseas valuation comparables” is no longer viable. The new generation of financing narratives places greater emphasis on unit economics—using customer lifetime value, customer acquisition cost, gross margin, and revenue retention to justify valuations.
Conclusion: A Long-Term Rewrite of the Valuation Framework
Overall, the valuation framework for AI large-model companies is undergoing a long-term reshaping from “scientific project” to “commercial entity.” The IPO process of leading overseas labs provides a benchmark coordinate system for the entire industry: the market can read from it the speed of inference cost decline, the linear trend of revenue growth, and the payback cycle of capital expenditure. Once this information is public, it becomes the first reference for all investors evaluating similar companies.
For Chinese AI model vendors, this shift is both a challenge and an opportunity. The challenge is that bidding farewell to the narrative premium means financing becomes harder and the valuation growth curve becomes flatter. The opportunity is that once the shift from “telling stories” to “delivering performance” is completed, companies with real technology accumulation, commercialization capability, and financial discipline will stand out from homogeneous competition and gain more sustainable capital support. It is foreseeable that in the next 12 to 18 months, valuation divergence among Chinese AI model vendors will intensify markedly: companies with clear revenue paths will enjoy a liquidity premium, while companies still relying on concept-driven narratives may face valuation correction pressure.
The shift in valuation logic is not a cyclical fluctuation but a structural rebuild. When the financial data of all major participants is on the table, AI model valuations will ultimately return to a systematic assessment of core variables—revenue, cost, growth, and risk. The “narrative era” of this industry is ending, and the “performance era” has just begun.
