LLM Valuation After IPO: Scarcity Premium, ARR Narratives, and Earnings Quality

    This article introduces a three-phase framework for how large language model (LLM) companies are valued after going public: a supply-scarcity premium, an ARR (annual recurring revenue) growth narrative, and global comparison against earnings quality. It explains how lockup expirations erode the scarcity premium, how to decompose ARR into meaningful quality signals such as deferred revenue, net revenue retention, gross margin, and the Rule of 40, and how to build a unit-token cost model for inference economics. It includes illustrative Python models, a gross-margin sensitivity table showing the limits of trading price cuts for volume, and a practical checklist for judging whether an LLM valuation is sustainable. All figures are based on public information and methodological illustration only and do not constitute investment advice.

    In 2026, the large-model (LLM) industry completed a shift from “private narrative” to “public pricing.” Overseas LLM companies are accelerating toward IPOs, while China’s general-purpose model companies such as Zhipu and MiniMax have entered the public market. Markets now reprice these companies using share prices, market capitalization, revenue quality, and cash flow. All prices, market caps, lockup schedules, index inclusion, and similar data cited in this article come from public reporting, company announcements, index rules, and other public-information sources; they are used only to illustrate the methodology, and the original disclosures should be treated as authoritative. This article is based only on public information and methodological illustration and does not constitute investment advice.

    For technical managers and investors alike, the real questions are not one day’s rally, but: How do lockups affect LLM valuations? How should the ARR metric be decomposed? How do you estimate per-token cost? Where are the limits of trading price cuts for volume? This article breaks the post-IPO valuation process of LLMs into three phases: a supply scarcity premium, an ARR growth narrative, and global comparison with earnings quality. These three are not simple substitutes but a rebalancing of valuation weights—early on, capital structure and free float dominate; in the middle, growth data dominates; and later, fundamentals, revenue quality, and the competitive landscape dominate.

    1. The Three-Phase Valuation Framework at a Glance

    Phase Core pricing variable Basis for capital pricing Valuation metrics Main risk
    Phase 1: Supply scarcity premium Number of investable names, free-float size, lockup schedule Sector coverage, passive index flows Issue price vs. comparable valuation multiple, free-float market cap Lockup expansion, index and rule changes
    Phase 2: ARR growth narrative ARR scale and growth, net revenue retention, gross margin Visibility and durability of growth ARR multiple, Rule of 40 Revenue metric distortion, trading price cuts for volume, model commoditization
    Phase 3: Global comparison and earnings quality Technology generation gap, revenue structure, cash flow Risk-adjusted returns Forward P/E, free cash flow, R&D efficiency Valuation center of gravity shifting lower, the gap narrowing toward global leaders

    2. Phase 1: Supply Scarcity Premium—Capital Can Only Buy “the Sector Itself”

    Scarcity of investable names: capital is forced to buy the whole track

    In the early listing period, only a limited number of standalone LLM companies were listed on the Hong Kong stock market. Globally, companies with foundation models as their core business and already publicly listed are similarly few. For institutional capital, the allocation instruction compresses into one simple action: buy. Managers can neither fully compare the merits of multiple model companies nor use an industry portfolio to diversify single-stock risk, so a single name ends up carrying the allocation demand of the entire sector. When capital has few choices, valuation first reflects “want to buy, but can only buy this one,” rather than the relative merits among companies.

    Extremely small free float: prices are highly sensitive to marginal capital

    Prices are set by marginal trading. In the early listing period, usually only cornerstone investors and a small number of institutions quote in the secondary market, and this free float is tiny relative to total shares. The illustrative function below helps explain pricing elasticity. This function only illustrates pricing elasticity and is not a real pricing formula.

    1. def first_day_premium(
    2. oversubscription: float, # public offering oversubscription multiple, e.g., 1000 means 1000x
    3. free_float_ratio: float, # free-float share as a fraction of total shares, e.g., 0.06 means 6%
    4. demand_index: float, # allocation demand intensity, relative index, 1.0 is neutral
    5. ) -> dict:
    6. """Illustrative pricing-elasticity model only, not a real pricing formula."""
    7. # Higher subscription and smaller free float => stronger IPO sentiment and scarcity pricing
    8. liquidity_pressure = demand_index * (1 + oversubscription / 100) / (free_float_ratio * 100)
    9. estimated_first_day_return = liquidity_pressure / (1 + liquidity_pressure)
    10. return {
    11. "liquidity_pressure": round(liquidity_pressure, 2),
    12. "illustrative_first_day_return": round(estimated_first_day_return, 4),
    13. }
    14. print(first_day_premium(oversubscription=1000, free_float_ratio=0.06, demand_index=1.0))
    15. # Illustrative output: {'liquidity_pressure': 1.83, 'illustrative_first_day_return': 0.6471}
    16. # i.e., under these illustrative parameters, the first-day gain is about 64.7%

    If free float in the early listing period is less than about 6% of total shares, a small amount of incremental buying can move the price significantly. This also explains why first-day market-cap differences do not come entirely from business quality—issue price, free float, capital concentration, and trading structure matter just as much.

    Subscription heat: a direct read of public-market sentiment

    Oversubscription in the public offering is the most direct read of retail and smaller-capital sentiment. If Zhipu is oversubscribed by roughly 1159.46x and MiniMax by roughly 1837.17x, with 420,000 people participating, these two signals correspond respectively to “financing substitution” and “wealth effect.” When primary-market subscription capital far exceeds the allocable amount, winning an allotment itself becomes a scarce resource. Even investors who expect limited gains may use leverage to raise their odds of an allotment, further amplifying early price elasticity. When participants reach the hundreds-of-thousands level, the share price begins to affect more people’s account values, and the self-reinforcing cycle of “rising price—growing buzz—new capital inflow” pushes pricing beyond what fundamentals can explain.

    Institutional view: the premium comes from a combination of “scarcity + lockup”

    According to public reports, UBS Securities analyst Xiong Wei noted that Zhipu’s and MiniMax’s high valuations are partly due to the premium from the scarcity of listed model vendors globally, combined with their not yet entering the lockup-expiry period and their low liquidity. This judgment contains two key conditions: scarcity of investable names and locked-up shares. Scarcity determines “what capital wants to buy”; the lockup period determines “how much is available to buy in the market.” When the two combine, the valuation premium is maximized; if either condition loosens, the premium fades accordingly.

    Event catalysts: passive flows push scarcity to the extreme

    Public reports show that on June 8 the two companies were simultaneously added to the Hang Seng Tech Index and the Hang Seng Composite Index, and Zhipu entered the Stock Connect (Southbound trading) that day. Index-tracking products and Southbound capital brought two layers of incremental demand. First, passive buying ignores fundamentals: funds tracking an index must buy by weight, and this demand is not constrained by valuation, so its impact is especially pronounced on names with already-small free floats. Second, the capital structure changes: Southbound capital is dominated by mainland investors with high attention to the AI theme. As a result, Zhipu touched HKD 2,980 intraday, corresponding to a market cap of about HKD 1.33 trillion; MiniMax touched HKD 1,330 intraday, with a market cap that briefly exceeded HKD 410 billion. In this phase, “whoever’s shares are scarcer” more easily obtains a higher price.

    3. First Test: How Lockup Expirations Squeeze Out the Scarcity Premium

    Two very different “exam papers”

    By public reporting, on July 9 about 153 million MiniMax shares unlocked, 48.9% of total shares; over 80% of pre-IPO shareholders and cornerstone investors stated beforehand that they would hold long-term, yet the stock still closed down 17.98% that day and fell nearly 40% over the next three trading days. The previous day, Zhipu’s unlock in the same batch was only 25.6816 million shares, 5.76% of issued shares; over 60% of cornerstone investors pledged to continue holding, and on the first day after unlock the stock rose instead of falling, up 13%.

    Dimension Zhipu MiniMax Implication for pricing
    Unlock as % of total shares 5.76% 48.9% Free float expands ~8.5x, directly weakening scarcity
    Share of holders pledging to keep holding Over 60% of cornerstone investors Over 80% of pre-IPO and cornerstone shareholders Long-term holding commitments cannot fully offset share expansion
    First-day price after unlock +13% -17.98% The higher the scarcity component, the more sensitive to free-float changes
    Next three trading days — Cumulative ~-40% Selling pressure manifests as sustained price digestion of new shares

    What the market worries about is not the unlock, but “how much the scarcity is worth”

    Long-term holding commitments failed to stop the price decline because investors are not trading “whether old shareholders will sell,” but re-evaluating “how much the scarcity is worth.” Before MiniMax’s unlock, free float was only a small part of total shares, and the price contained a large premium from “being unbuyable.” After the unlock, nearly half of the shares became tradable, and this premium lost the ground it stood on. Zhipu’s unlock was only 5.76% of total shares, so its scarcity structure was basically unchanged and the price was naturally more stable.

    The function below estimates the potential upper bound of sell pressure on the unlock-announcement day. This estimate assumes the announcement-day free float still references prior turnover, and that the holding ratio can be determined immediately after the announcement; in reality, one must also account for index-weight changes, short-selling and hedging strategies, and gradual transfers of old shares.

    1. def unlock_sell_pressure(
    2. unlocked_shares: float, # unlocked shares, in units of 10,000 shares
    3. price: float, # closing price before the announcement, in HKD
    4. avg_turnover: float, # average daily turnover over the last 20 trading days, in units of HKD 100 million
    5. holding_ratio: float, # ratio of unlocking shareholders willing to keep holding, e.g., 0.8 means 80%
    6. ) -> dict:
    7. # Theoretical sellable value = unlocked shares × price × (1 - holding ratio)
    8. sellable_value = unlocked_shares * price * (1 - holding_ratio) / 1e4 # unit: HKD 100 million
    9. # New sell pressure expressed as how many trading days of turnover it equals
    10. days_to_absorb = sellable_value / avg_turnover
    11. return {
    12. "theoretical_sell_pressure": round(sellable_value, 2),
    13. "trading_days_to_absorb": round(days_to_absorb, 1)
    14. }
    15. print(unlock_sell_pressure(unlocked_shares=15300, price=345, avg_turnover=50, holding_ratio=0.8))
    16. # Illustrative output: {'theoretical_sell_pressure': 105.57, 'trading_days_to_absorb': 2.1}

    It should be emphasized that turnover, holding ratio, short-selling hedges, index rebalancing, and gradual transfers of old shares all affect the actual result, and readers should not directly extrapolate the illustrative output into a market trajectory.

    4. Phase 2: The ARR Growth Narrative—After Scarcity Recedes, Capital Needs a New Anchor

    Why ARR becomes the “new anchor” for LLM companies

    After the unlock, the scarcity premium is largely squeezed out, and capital needs a verifiable, sustainable growth anchor. ARR (Annual Recurring Revenue) takes on this role. For LLM companies, ARR mainly has two bases: one is subscription-based, including monthly plans for individual developers and seat- or platform-based subscriptions for enterprises; the other is consumption-based, billed by token usage and annualized over a given period. Compared with traditional software, LLM revenue is more consumption-oriented: the same customer’s monthly API spend fluctuates dynamically with call volume and tokens per request, so ARR is more like a “moving target.”

    ARR’s appeal and blind spots

    ARR replaces scarcity as the core narrative for three reasons: it is repeatable, predictable, and comparable. Revenue comes from ongoing usage and renewals rather than one-time project delivery; combined with net revenue retention and renewal rates, one can project next year’s revenue; and the ARR multiples long established in the SaaS industry provide a unified yardstick for cross-company valuation.

    But ARR also has blind spots: it reflects only growth speed, not revenue quality; it does not reflect gross margin, since in consumption-based revenue inference costs directly erode gross margin; and it is easy to manipulate through metric choices, for example by folding one-time project revenue into the annualized figure.

    Decomposing ARR: from “metric” to “quality”

    Under the consumption model, unit economics determine the true worth of growth:

    1. Gross margin ≈ 1 − (inference and compute cost ÷ related revenue)
    2. Inference cost = input token cost + output token cost + scheduling and operations cost

    Pricing is determined by market competition, while cost is determined by technical capability; only the gap between them is truly sustainable growth. The illustrative code below provides a quick health check of ARR quality.

    1. def arr_quality_check(
    2. arr_growth: float, # YoY ARR growth, e.g., 2.0 means 200%
    3. deferred_revenue_growth: float, # YoY growth of contract liabilities (deferred revenue)
    4. nrr: float, # net revenue retention, 1.2 means 120%
    5. gross_margin: float, # blended gross margin, e.g., 0.45 means 45%
    6. one_off_revenue_ratio: float, # share of one-off/project-based revenue in total revenue
    7. operating_margin: float = 0.15, # operating margin, used for the Rule of 40 illustration
    8. arg: float = 0.4, # Rule of 40 threshold
    9. ) -> dict:
    10. # Signal 1: ARR growth much faster than deferred revenue growth -> watch for early recognition or metric adjustments
    11. arr_deferred_gap = arr_growth - deferred_revenue_growth
    12. # Rule of 40 usually uses revenue growth plus a margin (operating or net), not gross margin.
    13. # Operating margin is used here for illustration; gross margin is not used as a substitute for margin.
    14. rule_of_40 = arr_growth * 100 + operating_margin * 100
    15. signals = []
    16. if arr_deferred_gap > 0.3:
    17. signals.append("ARR and deferred revenue growth diverge; verify revenue recognition basis")
    18. if nrr < 1.1:
    19. signals.append("Net revenue retention is low; growth relies on new customers")
    20. if gross_margin < 0.4:
    21. signals.append("Gross margin is low; watch inference cost optimization progress")
    22. if one_off_revenue_ratio > 0.3:
    23. signals.append("One-off revenue is high; ARR's recurring quality is insufficient")
    24. return {
    25. "arr_deferred_gap": round(arr_deferred_gap, 2),
    26. "rule_of_40_score": round(rule_of_40, 1),
    27. "rule_of_40_passed": rule_of_40 >= arg * 100,
    28. "risk_signals": signals or ["no warning triggered for now"],
    29. }
    30. print(arr_quality_check(arr_growth=2.0, deferred_revenue_growth=1.1, nrr=1.35,
    31. gross_margin=0.55, one_off_revenue_ratio=0.15,
    32. operating_margin=0.15))

    A special note here: the Rule of 40 usually uses revenue growth plus a margin—for example operating margin or net margin—rather than gross margin. Gross margin reflects product-level cost, while margin reflects the result after expenses, R&D, sales, and overall operations. If a company uses gross margin as a Rule of 40 substitute, it can only be treated as illustrative and cannot be taken as the standard basis.

    Four rules of thumb

    • Match between ARR growth and deferred revenue growth. Subscription revenue first shows up as contract liabilities and is later recognized as revenue. If ARR growth is persistently and substantially higher than deferred revenue growth, it often signals metric adjustments, inclusion of one-time revenue, or front-loaded revenue recognition timing.
    • Net revenue retention (NRR). NRR measures the change in current-period revenue from existing customers relative to the prior period, composed of expansion, contraction, and churn. NRR above 100% means the existing customer base itself is growing; NRR persistently below 100% means growth can only rely on continuously acquiring new customers.
    • Customer concentration and scenario dependence. A few top customers contributing more than half of revenue means ARR stability depends on a few decisions; revenue concentrated in a single scenario means that once model capability is commoditized, price wars will directly hit ARR.
    • Rule of 40. This rule of thumb, summarized by the overseas SaaS industry, holds that when revenue growth and margin sum to around 40%, growth and profitability are in basic balance. LLM companies generally front-load R&D and compute investment, so falling short of the line at a given stage is normal, but a clear path between investment and gross margin must be given.

    5. Phase 3: Global Comparison and Earnings Quality

    From “scarce” to “comparable”: the changing frame of reference

    As model companies worldwide progressively enter the public market, the capital frame of reference shifts from “which names are listed in Hong Kong” to “how much are comparable companies worldwide worth.” Comparison dimensions usually include the technology generation gap, revenue structure, customer structure, growth quality, and compliance/data sovereignty. Differences translate into valuation discounts or premiums. The market grants a premium to companies with unique technological barriers or scarce customer resources, and a discount to companies whose growth depends on price wars and whose revenue is highly substitutable.

    Earnings quality: four dimensions beyond ARR

    • Gross margin and its trend: the change in inference cost as a share of revenue directly reflects engineering and algorithm efficiency.
    • Capital expenditure and depreciation: training compute is mostly deployed via self-build or long-term leasing; depreciation policy and capacity utilization determine the “quality” of profit.
    • R&D efficiency: the structure of R&D spend across training, data, and inference optimization, and the model-capability improvement gained per unit of input.
    • Cash flow: whether operating cash flow is positive and whether capex and financing pace are matched. Also watch whether token price cuts and ARR growth move in the same direction; if unit price falls rapidly while usage does not scale up accordingly, the basis for ARR growth is not solid.

    The unit-token economic model: breaking technical cost down to a measurable granularity

    If an LLM company wants its revenue to be “understandable” to the market, the first step is to break inference cost down to a measurable granularity. The illustrative model below applies cache hits only to input tokens and avoids redundant, convoluted notation.

    1. def serve_cost_model(
    2. input_tokens: float, # total input tokens across requests
    3. output_tokens: float, # total output tokens across requests
    4. gpu_hour_cost: float = 25.0, # blended cost per GPU-hour, including depreciation, power, network, ops
    5. tokens_per_gpu_hour: float = 1_500_000, # effective throughput per GPU-hour
    6. cache_hit_ratio: float = 0.3, # input prefix cache hit ratio
    7. batch_efficiency: float = 0.7, # batching efficiency factor
    8. spot_discount: float = 0.6, # discount of flexible/spot resources vs. on-demand cost
    9. ) -> dict:
    10. # Cache hits apply only to input tokens; output token decoding is illustratively scaled to 3.5x input tokens.
    11. effective_tokens = input_tokens * (1 - cache_hit_ratio) + output_tokens * 3.5
    12. gpu_hours = effective_tokens / tokens_per_gpu_hour / batch_efficiency
    13. gpu_cost = gpu_hours * gpu_hour_cost
    14. # Higher flexible/spot resource share => lower unit cost.
    15. flexible_share = 0.3
    16. adjusted_cost = gpu_cost * (1 - flexible_share * (1 - spot_discount))
    17. # Non-compute overhead such as storage, bandwidth, logging and monitoring.
    18. overhead = adjusted_cost * 0.15
    19. total_cost = adjusted_cost + overhead
    20. return {
    21. "effective_compute_tokens": round(effective_tokens),
    22. "gpu_cost_rmb": round(gpu_cost, 4),
    23. "total_inference_cost_rmb": round(total_cost, 4),
    24. }
    25. cost = serve_cost_model(input_tokens=1_000_000, output_tokens=200_000)
    26. price_per_million = 40.0 # pricing example: 40 RMB per million tokens
    27. revenue = price_per_million * (1_000_000 + 200_000) / 1_000_000
    28. print(cost)
    29. print(f"Illustrative gross margin: {(revenue - cost['total_inference_cost_rmb']) / revenue:.1%}")
    30. # Illustrative result: effective compute tokens = 1,400,000; GPU cost ≈ 33.3333 RMB;
    31. # total inference cost ≈ 33.7333 RMB; revenue = 48.0 RMB; illustrative gross margin ≈ 14.3%.

    In this example, effective compute tokens = 1,000,000 × (1 − 0.3) + 200,000 × 3.5 = 1,400,000. Total inference cost is about 33.7333 RMB. Charging 40 RMB per million tokens on 1.2 million tokens yields revenue of about 48 RMB and an illustrative gross margin of about 14.3%. Different models, cache hit rates, batching efficiency, and flexible-resource shares will all change the result significantly.

    General technical paths for cost optimization

    The impact of these techniques on unit economics is not linear but “multiplicatively stacked”:

    1. Model level: model compression and distillation reduce inference compute demand; sparse architectures activate only part of the parameters on demand, raising throughput per unit of compute; routing models by scenario sends simple requests to small models.
    2. Inference-engine level: quantization reduces memory footprint and raises throughput; continuous batching improves GPU utilization; paged attention manages memory fragmentation; prefix caching reuses system prompts and long contexts, and the hit rate has a significant impact on cost.
    3. Architecture level: speculative decoding uses a smaller model to draft and a larger model to verify, reducing large-model calls; compute pooling unifies scheduling across different businesses and workloads; elastic scaling and workload co-location based on metrics such as queue length, latency percentiles, and memory usage let offline batch jobs fill the idle compute between online peaks; and a cost-observability system attributes token consumption and compute cost by tenant, model, and scenario through logging services and monitoring/alerting.

    Gross margin sensitivity: the limits of trading price cuts for volume

    In price competition, the fragility of unit economics is quickly magnified. Assume a blended cost of 12 RMB per million tokens; the gross margin formula is:

    1. Gross margin = (price − blended cost) / price

    Illustrative gross margins at different prices:

    Pricing (RMB / million tokens) Relative to base price (12 RMB) Illustrative gross margin
    6.0 -50% -100%
    9.0 -25% -33.3%
    12.0 base 0%
    24.0 +100% 50%

    The core conclusion: when blended cost approaches price, gross margin approaches zero and quickly turns negative; once blended cost exceeds the selling price, the larger the scale, the larger the loss, and one cannot simply rely on “scale effects to dilute cost.” Therefore, one of the metrics capital focuses on most in Phase 3 is whether price cuts can buy enough call-volume growth to offset the decline in unit gross margin.

    A cost-governance framework for capital markets

    Companies can draw on the cost-governance experience of mature cloud providers, building around three layers: metering, allocation, and an optimization loop. Metering layer: unify the token-billing basis, meter input, output, and cache hits separately, and align with the financial revenue basis to avoid ARR metric distortion. Allocation layer: establish cost centers that allocate compute, storage, and bandwidth costs by business line, product, and customer, identifying high- and low-margin businesses. Optimization layer: build a “cost metric—optimization measure—effect verification” loop, using process metrics such as effective throughput per GPU, cache hit rate, and flexible-resource share to continuously calibrate the unit economic model.

    6. Reader’s Checklist: How to Quickly Judge Whether a Valuation Is Sustainable

    Before the conclusion, use the following five-point checklist to quickly scan the valuation quality of an LLM company:

    1. Check the free float: What is the free-float share as a percentage? Is the free float too small and easily influenced by marginal capital?
    2. Check the unlocks: Are there any large unlocks in the next 12 months? Have the unlocking shareholders’ holding ratio, days of sell-pressure absorption, index rebalancing, and short-selling hedges been factored into the judgment?
    3. Check the ARR basis: What is the mix of subscription vs. consumption? Does deferred revenue growth match ARR? Is one-time project revenue being mistakenly counted as recurring revenue?
    4. Check NRR: Is net revenue retention above 100%? Does growth come from existing-customer expansion or purely from new-customer acquisition?
    5. Check unit gross margin: Are token unit price, inference cost, cache hit rate, batching efficiency, and the gross margin trend clear? Would trading price cuts for volume break unit economics?

    Conclusion: The Pricing Logic Behind Three Migrations

    From the supply scarcity premium, to the ARR growth narrative, to global comparison and earnings quality, the valuation logic of LLM companies has completed three migrations. The underlying pattern can be summarized in three sentences. First, scarcity answers “whether it exists,” not “what it is worth.” The price performance of Zhipu and MiniMax around the unlock shows that the scarcity premium fades quickly with changes in free float. Second, ARR proves “how fast it runs,” not “how steadily it runs.” Only ARR that excludes one-time revenue, is cross-checked against deferred revenue, and is accompanied by a reasonable margin is sustainable growth. Third, earnings quality determines “how far it can run.” Technical strength must ultimately land on revenue quality, gross margin, a profitability path, and the global competitive landscape.

    The takeaway for companies: in the capital market, technical strength can only be priced once it is “measured.” Breaking inference cost down to the token level, and revenue quality down to contract liabilities, net revenue retention, and operating margin, lets the market see a company’s true value. This article is based only on public information and methodological illustration and does not constitute investment advice; share prices, market caps, lockup data, and index data should follow company announcements, exchange disclosures, and index-provider documents.