This article discusses valuation methods for AI companies at IPO. It does not refer to any specific listed company or investment target. All figures are hypothetical values for ease of calculation and do not represent actual offering information. When reading, refer to the prospectus and regulatory disclosures.

1. What Typically Drives a Valuation Correction

The market’s valuation of large-model companies over the past two years has been built mainly on growth expectations. When investors shift to a stricter review framework, valuations move toward cash flow and revenue quality. There are three common review dimensions.

First is the speed of revenue realization. Subscription and enterprise API revenue can grow quickly, but if valuation reaches trillion-dollar scale, the price-to-sales ratio will still be very high. Investors will ask whether this revenue can convert into stable profit in the coming years.

Second is compute cost. Model training and inference require continuous capital expenditure, and free cash flow may be negative for a long period. The speed at which the cost curve declines directly determines whether valuation can be absorbed by business performance.

Third is the competitive landscape. The speed at which open-source models catch up can compress the pricing room for closed-source models. Whether leading companies can maintain their technical moat is the hardest variable to quantify in valuation models.

The weight of these three factors depends on revenue growth, gross margin, and unit compute cost. If revenue growth continues to outpace cost growth, the magnitude of a valuation correction usually becomes smaller.

During the roadshow, the buy-side typically uses more conservative assumptions than the sell-side to calculate valuation, and final pricing often falls between the two sides’ expectations. This is a common phenomenon in IPO pricing.

2. The Role of Cornerstone Investment

Cornerstone investors are institutional investors that sign subscription agreements with the company before the offering, usually subscribing to a certain number of shares at the final offer price. The agreement generally includes a lock-up period. The length of the lock-up period varies by exchange and agreement terms and should be based on prospectus disclosure.

Cornerstone investment mainly has three functions.

1. Support the offering size: During market volatility, subscribed volume can provide a floor and reduce the risk of offering failure.
2. Signal industry judgment: If subscribers come from upstream segments such as chips, storage, and cloud platforms, the market will view this as a supply chain endorsement of the company’s long-term demand.
3. Narrow pricing disagreement: Large subscriptions reduce the uncertain public-market portion, making pricing more likely to move closer to one end of the offering range.

The cornerstone subscription share can be observed with a simple ratio. Let cornerstone subscription amount be C and public offering amount be P. The cornerstone ratio is C/(C+P). The higher the ratio, the lower the offering outcome’s reliance on public-market sentiment. This is only an observation metric; it cannot directly imply the probability of a successful offering.

3. Three Valuation Metrics in a Prospectus

Valuation figures in a prospectus can differ significantly depending on the metric used. The following hypothetical case illustrates this. Assume a company’s offer price is $70 per share, post-offering basic shares are 10 billion, and employee options, RSUs, and convertible securities total 800 million shares on an as-converted basis.

Metric	Calculation	Hypothetical case result	Use case
Basic market cap (offering valuation)	Offer price × post-offering basic shares	$700 billion	Measures market pricing at listing
Fully diluted valuation	Offer price × (basic shares + potential shares)	$756 billion	Measures the value of all potential equity
Private placement subscription valuation	Subscription price × post-offering basic shares	Depends on subscription price	Reflects the price industrial capital is willing to pay

Note that new shares issued multiplied by the offer price gives gross proceeds, not valuation. The two are often confused. In actual fully diluted valuation calculations, options are usually converted under the treasury stock method; specific figures should follow the prospectus.

When the subscription price equals the offer price, the private placement subscription valuation is the same as basic market cap. However, the subscriber’s identity and lock-up conditions give this price additional reference meaning.

The situation of selling shareholders should also be viewed separately. Existing shareholders sell existing shares in the offering, and the proceeds belong to the sellers, not the company. This does not increase company cash or change post-offering total shares. Early shareholders participating in the sale of existing shares effectively lock in part of their gains early. Readers should distinguish this from the company’s fundraising amount.

4. Implications for Technical Practitioners and Enterprise Users

4.1 Compute-Layer Demand Is Relatively Observable

Valuations at the model layer can be highly debated, but demand at the compute and infrastructure layer is easier to observe through procurement data and capacity announcements. For teams planning to build or rent large-model capabilities, near-term compute supply may remain tight. Engineering practices around GPU scheduling, batching, and inference caching still have direct value.

4.2 Enterprise Selection Checklist

When selecting a provider, check the following items one by one:

• Complete end-to-end validation on your own business evaluation set; use public leaderboards only as an auxiliary reference
• Calculate true cost based on token billing, including context length, cache hit rate, and batch processing ratio
• Confirm the level of support for function calling, structured output, long context, and streaming responses
• Check service-level commitments (SLA), rate-limiting policies, and handling of burst traffic
• Confirm compliance capabilities such as data isolation, training data usage terms, and audit logs
• Assess migration cost: whether prompts, evaluation sets, and API wrappers can be reused with another provider

4.3 Bottom-Up Cost Estimation

The example below uses hypothetical unit prices per 1,000 tokens in USD.

def monthly_cost(queries_per_day, days, avg_input_tokens, avg_output_tokens,
                 input_price_per_1k, output_price_per_1k):
    """Prices are per 1,000 tokens; return total cost for the specified period."""