Large Model IPO Breakdown: Why 16.3 Billion in Half-Year Revenue Still Comes with 80.7 Billion in Losses

1. Start with the Conclusion: High Revenue Does Not Mean Profitability

An AI company disclosed in its IPO materials that it generated approximately RMB 16.3 billion in revenue for the half-year period, while recording a loss of approximately RMB 80.7 billion. Looking at these two numbers alone tends to produce two misjudgments:

First, that revenue is already large enough that the company should be entering a profitable phase; second, that a loss far exceeding revenue proves the business model cannot work.

In reality, the financial statements of large model companies are typically shaped by three factors at once: the pace of revenue recognition, the intensity of infrastructure investment, and non-cash accounting items. Especially during rapid expansion, the income statement may look worse than the underlying operating condition—but that does not mean all losses stem from day-to-day operations.

A more reasonable approach is to separate revenue, gross profit, operating loss, net loss, and cash flow, and to distinguish three questions:

1. Does the business itself have genuine paying demand?
2. How much compute and serving cost is consumed for each unit of revenue?
3. How much of the loss is cash expenditure, and how much is accounting recognition?

Only after completing these three layers of analysis can you judge whether the company is investing ahead of future growth or using heavy costs to sustain an unsustainable revenue scale.

2. RMB 16.3 Billion in Revenue: First Confirm Where It Comes From

Revenue for large model companies typically comes from several business lines:

1. Model invocation and inference services

Companies provide text generation, code generation, content moderation, knowledge Q&A, and other capabilities to developers or other organizations through APIs, charging by call volume, input/output characters, compute time, or subscription packages.

This revenue grows quickly and has a low barrier to use, but it carries an obvious problem: every call may incur compute cost. If pricing is below unit inference cost, larger revenue may actually widen losses.

2. Enterprise subscriptions and professional services

Enterprise customers may pay by number of accounts, usage quota, feature modules, or annual contracts. This revenue is usually more predictable than consumer revenue, but customers often require private deployment, data isolation, permission management, audit capabilities, and industry adaptation—raising delivery costs accordingly.

3. Cloud model hosting and compute services

Some companies deploy models on container platforms, dedicated compute clusters, or within enterprise environments, charging for hosting, operations, and resource usage. Revenue from this line may be substantial, but hardware depreciation, network fees, storage costs, and operations staffing grow in tandem.

4. Customized solutions

These include knowledge base Q&A, intelligent customer service, coding assistance, process automation, and data analytics. Revenue from such projects does not necessarily represent standardized product revenue, since it may include consulting, implementation, integration, and ongoing maintenance fees.

Therefore, to judge the quality of RMB 16.3 billion in revenue, look beyond the total and examine:

• The ratio of subscription revenue to one-time project revenue;
• The ratio of standardized product revenue to custom delivery revenue;
• The share of revenue from the top five customers;
• Whether contract liabilities continue to increase;
• Whether accounts receivable are growing too quickly;
• Whether revenue depends on a small number of large customers;
• Whether revenue recognition keeps pace with actual cash collection.

If revenue comes mainly from standardized subscriptions and stable model invocations, business quality is generally higher. If a large share comes from one-time delivery, long-term receivables, or concentrated large-customer procurement, the sustainability of future revenue warrants caution.

3. Why Inference Costs Become the Core Pressure on the Income Statement

The most underestimated cost for large model companies is not office space or ordinary R&D salaries—it is the compute resources required to run models.

The cost of a single model request can be roughly understood as:

Cost per call
= input compute cost
+ output compute cost
+ storage and network cost
+ scheduling and system operations cost

Output compute cost is often higher than input compute cost, because the model must continuously generate content. The longer the answer, the more complex the context, and the higher the concurrency, the more pronounced the resource consumption.

Key factors affecting inference cost include:

• Model parameter scale;
• Context length;
• Output length;
• Concurrency;
• Request peaks;
• Hardware utilization;
• Batch processing capability;
• Cache hit rate;
• Model compression and quantization effectiveness;
• Service-level requirements.

If a company actively lowers prices to capture market share, it may see “revenue growth, gross margin decline.” Especially when prices fall faster than unit costs, economies of scale do not automatically produce profitability—they amplify losses instead.

This is also why “cheaper models” keep appearing across the industry. Low prices do not simply mean weaker technical capability; multiple engineering approaches may be behind them:

1. Using smaller models for simple tasks;
2. Routing complex requests to stronger models;
3. Caching results for common questions;
4. Compressing input text;
5. Adopting more efficient inference scheduling;
6. Setting different service levels for different customers;
7. Separating real-time tasks from offline tasks.

For enterprise customers, what really matters is not “how much per million characters” but the total cost per business task. For example, how many model calls a customer service ticket requires, the average output length, the failure retry rate, and whether human review is needed—these factors together determine the final cost.

4. What Different Types of Items Might the RMB 80.7 Billion Loss Include

A “net loss of RMB 80.7 billion” does not necessarily come entirely from the model business itself. At minimum, break it down into the following items.

1. R&D expenses

Training foundation models requires substantial compute resources, data processing capability, and engineering teams. Training costs include not only compute equipment rental but also data cleaning, experiment iteration, evaluation systems, model safety, and infrastructure development.

If a company is in a technology race, it may train multiple versions simultaneously, and some models may never reach production. These expenditures appear as expenses in the current period, but their goal is to secure future technical capability and market position.

2. Depreciation of servers and compute equipment

When a company purchases equipment itself, it does not expense the entire outlay in the current period but depreciates it over the estimated useful life. For training and inference clusters, equipment refreshes quickly, and older equipment may depreciate technologically at a rapid pace.

Depreciation does not necessarily represent current cash outflow, but it continuously weighs on profit margins. Companies must monitor:

• Original equipment value;
• Depreciation period;
• Equipment utilization;
• Whether assets face impairment risk;
• Whether new equipment can be absorbed by revenue growth.

3. Share-based compensation

Technology companies commonly use equity or options to attract R&D and management talent. Share-based compensation usually does not require direct cash payment in the current period but must be recognized as an expense under accounting rules.

As a result, a company may show “a large net loss while operating cash flow has not deteriorated to the same degree.” Share-based compensation is not a fictitious cost—it represents the economic cost of future shareholder dilution—but it must be distinguished from cash expenditure in analysis.

4. Fair value changes of financing instruments

If a company has issued convertible instruments, preferred interests, or financing arrangements with complex terms, changes in market valuation may alter the carrying value of liabilities or equity instruments.

Such losses are often strongly accounting-driven and may actually increase book losses when company valuation rises. The reason is that certain financing instruments must be settled at a higher value in the future, requiring the company to recognize corresponding fair value changes.

This is a typical source of “the higher the valuation, the larger the book loss.”

5. M&A, asset impairment, and one-time charges

During expansion, a company may incur business integration, asset restructuring, intangible asset impairment, legal fees, and listing preparation costs. These items do not necessarily indicate ongoing operating problems but can significantly affect net profit in a single reporting period.

Therefore, to judge whether a loss is serious, expenses should be categorized as:

Ongoing operating costs
+ R&D investment
+ Asset depreciation
+ Share-based compensation
+ Fair value changes
+ One-time items

Different categories of expenses have completely different implications for future profitability.

5. Beyond the Income Statement, Cash Flow Matters More

To judge whether a large model company is healthy, cash flow is often more informative than net profit.

Three statements deserve close attention:

1. Income statement: examine the profit structure

The income statement answers “how much revenue and expense the company recognized during the accounting period.” It reflects gross margin, sales expense ratio, R&D expense ratio, and net profit margin, but is easily affected by depreciation, share-based compensation, and fair value changes.

2. Cash flow statement: check whether money actually moves

Operating cash flow helps determine whether customers actually pay. If revenue grows rapidly but operating cash flow remains persistently negative, it may mean:

• Accounts receivable are growing too quickly;
• Large customers have long payment cycles;
• Contracts have not been fully delivered;
• The company offered large discounts to acquire customers;
• Prepaid compute resources consumed significant cash.

Investing cash flow mainly reflects equipment purchases, data center construction, and long-term asset investment. For model companies, persistently high investing cash flow is not surprising, but it is necessary to judge whether these investments translate into higher resource utilization and revenue scale.

3. Balance sheet: check whether expansion depends on financing

If a company’s cash and cash equivalents come mainly from multiple funding rounds rather than operating activities, the company is still in a capital-driven stage. At this point, monitor:

• How many months of runway the cash reserves can support;
• Whether quarterly cash burn is expanding;
• Whether short-term debt is rising quickly;
• Whether capital expenditures continue to climb;
• Whether there are large contractual payment obligations.

A commonly used metric is cash runway:

Cash runway (months)
= available cash ÷ average monthly net cash outflow

If a company holds substantial cash but monthly cash consumption is expanding rapidly, book cash does not equal a long-term margin of safety.

6. Low-Price Model Competition Changes the Business Model Itself

When low-price models appear in the industry, the impact is not limited to the selling price of a single product—it may change how profit is distributed across the entire market.

Low-price competition typically produces four outcomes.

First, model capabilities gradually become commoditized

When differences in accuracy, response speed, and context capability narrow across services, customers focus more on price, stability, data security, and integration difficulty. The premium attached to the model itself may decline.

Second, cost control matters more than parameter scale

Future competition may not be about who has the largest model, but about who can complete the same business task at lower cost. Inference scheduling, model routing, caching, compression, and resource utilization all directly affect gross margin.

Third, application-layer value becomes more critical

If foundation model prices continue to fall, the profit margin from simply reselling model invocations will be squeezed. Companies need to embed model capabilities into business processes—such as customer service, R&D, operations, risk control, and knowledge management—creating value through process efficiency gains.

Fourth, revenue growth and profit growth may decouple

Low-price strategies can rapidly expand call volume but do not necessarily improve profit. Only when scale growth brings higher hardware utilization, lower unit inference cost, and stronger customer retention will revenue growth gradually translate into profit growth.

7. How Technical Teams and Enterprise Customers Should Evaluate Such Services

For users, comparing model unit prices alone is insufficient—a complete unit economics model is needed.

The following evaluation framework can be used:

Cost per business task
= model invocation fee
+ data processing fee
+ storage and network fee
+ human review fee
+ failure retry fee
+ system operations fee

Also evaluate the following metrics:

• First response time;
• Peak concurrency capability;
• Output quality stability;
• Success rate for long-text tasks;
• Data isolation and permission control;
• Log audit capability;
• Failure recovery capability;
• Service migration difficulty;
• Dependence on a single vendor.

In architecture design, a tiered strategy can be adopted:

1. Use lightweight models for simple classification, summarization, and format conversion tasks;
2. Use high-capability models for complex reasoning and high-value tasks;
3. Enable caching for high-frequency repeated requests;
4. Use batch processing for low-urgency tasks;
5. Retain backup service paths for critical business;
6. Continuously track the real cost of each task type through monitoring systems.

The goal is not to blindly pursue the lowest unit price but to reduce the total cost of ownership across the entire business chain.

8. How to Judge Whether a “Huge Loss” Is Acceptable

It is not unusual for large model companies to incur losses in their early stages, but losses must be explainable as controllable investment rather than unbounded consumption.

Four questions can serve as a starting point:

1. Is gross margin improving

If unit inference cost keeps falling as revenue grows, economies of scale are forming. Conversely, if the higher the revenue, the lower the gross margin, beware of buying scale with low prices.

2. Are customers retained

The number of new customers matters less than renewal rate, depth of usage, and customer expansion revenue. A genuinely healthy business model should move customers from trial use toward core business usage.

3. Does R&D investment translate into product capability

High R&D spending is not inherently alarming; the key is whether it brings lower cost, higher stability, stronger security, or greater willingness to pay. If R&D investment only produces parameter growth over time without improving business metrics, investment efficiency deserves scrutiny.

4. Is cash burn predictable

A company can lose money in stages, but cash consumption should be explainable and predictable by management. If financing needs are raised every quarter, the business model has not yet formed a stable closed loop.

Conclusion: Do Not Be Misled by Revenue and Loss Figures Alone

Half-year revenue of RMB 16.3 billion and a loss of RMB 80.7 billion form a striking contrast on the surface, but these numbers alone cannot directly prove success or failure.

What really matters is breaking them down:

• Whether revenue is real, stable, and sustainable;
• Whether model invocations carry positive gross margin;
• Whether R&D and infrastructure investment are generating economies of scale;
• How much of the loss consists of non-cash items;
• Whether operating cash flow continues to deteriorate;
• Whether low-price competition will further compress margins;
• Whether the company can convert model capability into replicable business value.

For investors, avoid looking only at revenue growth or net loss scale; for technical teams, avoid comparing only model invocation unit prices; for enterprise customers, focus more on total cost and long-term portability across the complete business process.

The core question for the large model industry is ultimately not “can the model be sold,” but whether it can use controllable compute costs to continuously deliver business outcomes that customers are willing to pay for over the long term.