Priced on Revenue, Not on Revenue Quality

Introduction

Our stance. We prefer exposure to the highest-quality-ARR names in the cohort (Anthropic on the enterprise/API mix and margin inflection; OpenAI kept Neutral on franchise reach pending net-revenue clarity) and are cautious on thin-float, triple-digit-multiple names (Zhipu, MiniMax post-lockup; Moonshot above 100x P/S on current ARR). Because most of the universe is private or newly listed, this is a relative-value and watchlist stance, not a set of price targets; the concrete expressions and monitoring triggers are in §6.

Source: Anthropic Series G and Series H announcements and 2Q26 results (valuation, run-rate revenue); OpenAI round and financials as reported by Bloomberg, FT and The Information; HKEX filings and market data (Zhipu 2513.HK, MiniMax 0100.HK); Moonshot financing per Bloomberg and SCMP. ARR is reported run-rate, not audited; private valuations are last-round marks; multiples are the author's calculations. *China P/S reflects a thin-float peak that corrected post-lockup, shown to illustrate ordering, not as a clean comparable. As of 24-Aug-2026.

Mortise view. Read the last two columns together. The ordering of the multiple column runs opposite to the ordering of the quality column: the strongest revenue-quality signal (Anthropic) carries the lowest multiple; the weakest (sub-$0.5bn revenue at triple-digit P/S) carries the highest. That inversion is the trade.

Related read

Crude Oil Through the Cycles: how supply-regime shifts re-anchor a price center; a template for reasoning about structural rather than cyclical repricing.

The Efficiency Revolution in Snacking: monetising supply-chain efficiency; a companion study in how business-model structure, not headline growth, drives durable economics.

Mortise Capital is a student-led research platform. This report is for educational and informational purposes only and does not constitute investment advice, an offer or a solicitation. It is fair, clear and not misleading to the best of our ability; figures are reported or estimated as marked and are not audited. See the Disclosure Appendix.

Section 1: The cohort and the revenue-quality question

This report reasons from one house question, namely what is the investable implication and is it already priced, and moves through the standard chain: a sector event (the 2026-27 listing wave), a transmission channel (audited disclosure plus real free float), company exposure (seven independent labs), market pricing (the multiple map) and expression (a relative-value stance). We first define the object being priced.

1.1 Scope: who is in the cohort

Bottom line. The universe is the set of independent frontier-model developers that monetise the model directly and are listed or filing to list, and it excludes the big-tech captives whose model revenue is bundled and unobservable.

The cohort is Anthropic and OpenAI in the US, and Zhipu, MiniMax, Moonshot/Kimi, DeepSeek and StepFun in China. Big-tech models (Gemini, Grok, Doubao and the Alibaba/Tencent families) enter only as competitive and demand backdrop: their model revenue is embedded inside cloud and advertising P&Ls and cannot be compared to a standalone ARR figure. The cohort matters precisely because, for the first time, most of its members are about to report numbers a public investor can audit and trade.

1.2 A revenue-quality taxonomy

Bottom line. Not all AI revenue is the same asset. We rank four revenue types by durability and margin; this ranking, not headline growth, is what the market is failing to price.

A dollar of enterprise API revenue and a dollar of promotional consumer-subscription revenue command very different multiples in any mature software framework, yet the cohort is largely valued on undifferentiated run-rate. We grade revenue on four axes: gross margin, net revenue retention, switching cost and cyclicality. Enterprise API tokens score highest: they are usage-metered, embedded in customer workflows and expand with the customer (Anthropic's reported net revenue retention is exceptionally high). Enterprise seat contracts score next. Consumer subscriptions are lower quality: churn-prone, promotion-sensitive and diluted by large free-user bases that drag gross margin. Captive-bundled revenue is unobservable and therefore uninvestable as a standalone.

Source: Gross margin and net revenue retention per SemiAnalysis and company disclosures; revenue mix per reported financials and company statements. Quality grades are the author's qualitative framework, not a rating-agency scale; NRR at the leader is a reported, unaudited figure.

Mortise view. The taxonomy is the spine of the report. If the market priced revenue by quality grade, the ordering in Exhibit 1 would run A, B, C from cheap to dear. It runs the other way. §2 explains why the operating economics support the grades; §3 measures the mispricing.

Section 2: Industry economics and the two clocks

2.1 Clock one: inference-cost deflation lifts everyone

Bottom line. Cost-to-serve is falling fast enough to turn the cohort's unit economics positive; this is a genuine, cohort-wide tailwind and the part of the story consensus has largely absorbed.

The clearest single ratio in the sector is compute cost per dollar of revenue. At Anthropic it reportedly fell from about $0.71 in 1Q26 to roughly $0.56 in 2Q26, a ~21% decline in one quarter, and blended gross margin has moved from roughly -94% in 2024 into the 44-70% range in 2026 depending on how amortised training cost is treated, with API-only gross margin quoted above 80% by SemiAnalysis. This deflation is driven by model-efficiency gains and infrastructure scale, not price rises, which is what makes it durable rather than a one-off. Crucially it applies to the whole cohort: the open-weight Chinese labs and DeepSeek in particular are pushing the cost floor down for everyone.

2.2 Clock two: the operating line diverges by business model

Bottom line. The same gross-margin tailwind produces opposite operating results depending on revenue mix; this second clock is what consensus is under-weighting.

Anthropic reported its first profitable quarter in 2Q26, roughly $11.5bn of revenue with positive adjusted operating income, and SemiAnalysis projects about $1bn of GAAP EBIT (a ~6% margin) in 3Q26. OpenAI, over the same window, ran a non-GAAP operating margin near -122% in 1Q26 (an operating loss around $7bn on ~$5.7bn of quarterly revenue) and booked a heavy operating loss on ~$13.1bn of revenue in 2025. The gap is not primarily capability or top-line growth, both of which are scaling fast; it is mix. Anthropic monetises high-margin enterprise API and coding workloads, while a consumer base near 900m weekly users dilutes the economics at the peer. The China cohort sits further left again: audited filings show R&D running three to four times revenue (Zhipu spent RMB3.18bn of R&D against RMB724m revenue; MiniMax spent ~$253m against a fraction of that in revenue).

2.3 Model quality has converged, so it cannot explain the multiple gap

Bottom line. At the frontier, capability is now near-parity; the surviving differentiator is monetisation, which is exactly what the multiple map ignores.

On Artificial Analysis' intelligence index the leading families cluster tightly, at Claude Fable 5 60, GPT-5.6 Sol 59 and Kimi K3 57, with Kimi reportedly ahead on some coding and agentic tasks. When three labs on three cost structures and two continents score within three points of each other, capability is no longer the variable that should drive an order-of-magnitude spread in revenue multiples; business-model quality should. The market is instead paying up for growth optics and scarce float.

Source: Anthropic 2Q26 results; SemiAnalysis (inference cost and EBIT estimates); Artificial Analysis (model intelligence index); Zhipu and MiniMax HKEX filings (R&D and revenue). 2026 figures reported or estimated; margins vary by amortisation treatment.

Mortise view. Two clocks means the bull and bear can both cite “margins are improving” and talk past each other. Gross margin is rising everywhere (clock one, largely priced); the operating result diverges by mix (clock two, under-priced). Our edge is insisting on clock two.

Section 3: The inverted multiple map

3.1 What the market pays per dollar of ARR

Bottom line. On like-for-like run-rate figures, the quality leader is the cheapest name in the cohort, the opposite of what an efficient market would price.

Anthropic's May Series H set a $965bn post-money against a run-rate that has since reached ~$65bn, about 15x reported ARR, with the cohort's best margins, fastest growth and only positive operating quarter. OpenAI's March round set $852bn against a run-rate near $40bn (~21x) on weaker near-term economics. The China names, priced on scarce float, carry triple-digit price-to-sales on sub-$0.5bn revenue: Moonshot's reported ~$31.5bn against ~$0.3bn June ARR is above 100x, and exceeded 150x on its spring run-rate. Ranked by multiple, the cohort orders almost exactly inversely to revenue quality.

3.2 Why the ordering is inverted

Bottom line. The market is pricing three observable factors, namely growth rate, narrative and float scarcity, and under-pricing the one it cannot yet audit: revenue durability.

Growth optics reward the fastest headline ramp regardless of margin content. Narrative rewards the consumer brand with 900m weekly users over the enterprise vendor most users have never opened. Float scarcity mechanically inflates price where only a few percent of shares trade; Zhipu listed with a ~2.7% float and promptly ran +500%. None of the three is revenue quality. Our variant perception is that as audited disclosure and real float arrive, the first three fade and the fourth, durability, becomes visible and priced.

3.3 The gross-versus-net asterisk

Bottom line. Before trading the inversion, we must neutralise the one accounting difference that could explain it away, and that is not fully possible until the S-1 lands.

Anthropic recognises cloud-reseller revenue (via AWS, Google, Microsoft) on a gross basis, counting total end-customer spend as revenue and booking partner payouts as cost, while OpenAI reports core revenue on a net basis. Gross recognition inflates Anthropic's top line and therefore compresses its price-to-ARR relative to a net-reporting peer. This is the honest counter to our own thesis: part of Anthropic's apparent cheapness may be an accounting artifact rather than a mispricing. We do not believe it explains the full gap, since the margin and retention differences are real and independent of the gross/net choice, but it is material enough that we anchor the trade to the S-1, which must reconcile gross and net, rather than to today's headline multiple.

Source: Underlying data per Exhibits 1 to 3 (company announcements, HKEX filings, SemiAnalysis, Artificial Analysis). Grades and gap labels are the author's judgement. *“Too cheap” is conditional on the gross-versus-net reconciliation in §3.3 and §5.

Mortise view. The gap column is the thesis in one view. We hold conviction on the ordering (A-grade revenue should not be the cheapest in a cohort) while keeping the downside legible: the Anthropic label is explicitly conditioned on the S-1 reconciliation, and the China labels rest on the more robust observation that sub-$0.5bn revenue cannot support triple-digit multiples through a lockup.

Section 4: The listing wave as the forcing function

4.1 Why now: the calendar

Bottom line. The catalyst is dated. Three of the cohort's largest names are moving to public markets inside a roughly twelve-to-fifteen-month window, and each listing is a disclosure-and-float event.

Anthropic filed a confidential S-1 on 1 June 2026 and is targeting a Nasdaq debut as early as October 2026, with press reports of a ~$2tn ambition against guided full-year revenue of $100-120bn. Moonshot/Kimi is unwinding its offshore VIE structure and is reported to be preparing a Hong Kong filing by year-end for a late-2026/early-2027 listing at a valuation above $30bn. OpenAI has filed confidentially but is now widely expected to list in 2027, holding out for a price above $1tn. StepFun is a smaller HK candidate. The timing is what makes the idea actionable now rather than a standing observation: a mispricing with no forcing event can persist indefinitely; this one has a calendar.

4.2 The live demonstration in Hong Kong

Bottom line. We need not assume the repricing mechanism; the Hong Kong tape has already demonstrated it twice.

Zhipu (2513.HK) and MiniMax (0100.HK) listed back-to-back on 8-9 January 2026. On thin floats they ran to roughly +500% and +470% above offer, and their combined market cap briefly topped HK$300bn, exceeding JD.com. Two forces then hit in sequence. Lockups expired in July 2026, adding sellable supply to names whose prices had been set by a sliver of float; and the 17 July Kimi K3 launch triggered single-session drops of ~24-30% in Zhipu and ~16-18% in MiniMax. This is the transmission channel in compressed form: when float widens and a genuine competitive datapoint arrives, prices set on scarcity and narrative revert toward fundamentals. The cohort faces the same two forces as it lists.

Source: Anthropic and OpenAI announcements and filings; HKEX; company and press reports (Bloomberg, SCMP, 36Kr, TechCrunch). Dates are reported targets and subject to market conditions.

Mortise view. Anthropic's IPO is the single highest-information event in the calendar because its prospectus must reconcile gross and net revenue, the one disclosure that either confirms the mispricing or collapses our thesis (§5). We would rather trade around that print than ahead of it.

Section 5: Risks and thesis breakers

Bottom line. The view has one dominant breaker (accounting) and several second-order risks; each has a stated threshold that would flip or weaken it.

• Gross-versus-net accounting (primary breaker). If Anthropic's S-1 restates revenue to a net basis and the cut is large, with our threshold set at a reduction of more than ~30% that lifts its true multiple toward the peer, the “cheapest equals best quality” inversion narrows to noise and the Preferred stance is wrong. This is why the trade is anchored to the S-1.

• Profit reversal. Anthropic itself flags that full-year profitability is uncertain given the H2 compute and training ramp, and critics argue the 2Q26 profit was flattered by a ramp-up discount on a large compute contract. Threshold: a return to negative operating margin in 3Q26 would weaken the durability argument, though not the retention or mix argument.

• China ARR genuinely accelerating. The bear-for-us case on the China names is that the multiples are not absurd because ARR is compounding fast (Kimi went $0.1bn to $0.3bn in three months). Threshold: if Moonshot lists with run-rate ARR above ~$0.6bn, roughly a doubling from June, the implied multiple compresses toward ~50x, the valuation reads as an aggressive but defensible growth multiple rather than a bubble, and the Cautious stance should be dropped.

• The IPO window closes. A market drawdown or an adverse regulatory or legal event (for example around corporate restructurings) could defer the listings. The mispricing can then persist for several quarters: the thesis is not wrong, but the catalyst is postponed and the horizon extends.

• Model commoditisation. Tencent Research characterises the LLM API market as monopolistic competition with lower-than-expected entry barriers; open-weight leaders (DeepSeek, Kimi) keep pushing the price floor down. If capability parity turns into a price war, gross margins compress cohort-wide and the quality premium we are trading shrinks for everyone.

• Concentration, key-person and geopolitical risk. Founder dependence, compute-supply concentration and US/China export-control and listing-venue risk all sit on the cohort. These widen the distribution of outcomes rather than pointing in one direction, and argue for expressing the view in relative-value rather than outright form.

Source: Scenario probabilities are the author's subjective estimates for illustration, not forecasts; underlying facts per Exhibits 1 to 5.

Section 6: Trade expression and what to monitor

Bottom line. Because most of the universe is private, the view is expressed as a relative-value stance plus a disciplined watchlist, not price targets, with an explicit no-trade on names we cannot access.

The clean expression of “quality is mispriced relative to level” is a pair: constructive on the highest-quality-ARR exposure, cautious on the thinnest-float, highest-multiple names. Today the only directly tradeable legs are the two Hong Kong listings (2513.HK, 0100.HK), which serve as the live short-quality gauge; the high-quality long leg becomes expressible at Anthropic's IPO, conditional on the S-1 reconciliation. Until then the private names are watchlist, not position. We would scale the long leg only if Anthropic prices at or below its private mark on a net-revenue basis, and we would treat any China name listing with ARR below ~$0.6bn at triple-digit P/S as confirming the cautious leg.

Source: Signal sources as noted: SEC and HKEX filings (Tier 1); SemiAnalysis and Artificial Analysis (Tier 2/3). Trigger thresholds are the author's decision rules.

Mortise view. The discipline that matters most is patience: the highest-information event (Anthropic's S-1) is also the one that could break the thesis. The correct expression today is therefore a small, tradeable short-quality gauge in Hong Kong plus a fully specified long-leg plan that triggers only on disclosure: conviction with a defined exit, not a binary bet.

Disclosure Appendix

Analyst Certification

I, Baining Zhang, hereby certify that all of the views expressed in this report accurately reflect my personal views about the subject companies and their securities, and that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this report. This report was prepared for educational and internal research purposes within Mortise Capital's Industry Research function and has not yet completed the house peer-review stage; it should be treated as a working draft pending review.

Nature of this report

Mortise Capital is a student-led research and educational platform. This material is for educational and informational purposes only and does not constitute investment advice, a recommendation, an offer or a solicitation to buy or sell any security. Nothing here is personal advice. “Stance” labels (Preferred / Neutral / Cautious / Watch) are relative-value research opinions on private or newly listed entities and are not price targets or ratings on tradeable instruments. Most of the universe is not accessible to public investors.

Data quality and uncertainty. Private-company revenue figures are reported or leaked run-rate ARR, not audited financial statements, and different companies compute revenue on different bases (notably gross versus net revenue recognition), which limits comparability. Single-sourced or contested figures, above all Anthropic's first quarterly operating profit and its gross-basis revenue, are flagged in the text and are the top validation priorities. Forward figures marked with “E”, “~” or “projected” are estimates. Probabilities in Exhibit 6 are subjective. Past performance is not indicative of future results.

Conflicts of interest. Mortise Capital and its members may hold, or in future hold, positions in securities referenced. Investors should consider this report as only one factor in any decision and seek professional advice appropriate to their circumstances. Before any external distribution or fundraising use, appropriate professional and regulatory guidance should be obtained.

Originality. This note is an original Mortise Capital synthesis. It cites third-party data and reporting for facts but expresses all analysis, structure, framing and exhibits in the firm's own words and design; it does not reproduce third-party wording, layouts or chart designs.

Sources and attribution

Tier 1 (Primary). Anthropic Series G and Series H announcements, 2Q26 results, and confidential S-1 (1 Jun-26); OpenAI confidential S-1; HKEX filings and IPO prospectuses for Zhipu (2513.HK) and MiniMax (0100.HK); company financial disclosures (R&D and revenue).

Tier 2 (Professional data). Artificial Analysis (model intelligence and price); SemiAnalysis (inference economics, EBIT projections); exchange data for the HK-listed pair.

Tier 3 (Credible reporting / sell-side). TechCrunch, Bloomberg, Reuters, Financial Times, CNBC, Forbes, South China Morning Post, 36Kr, BigGo; CSC, Global Large Model Research Framework (Jun-26); Morgan Stanley, Global and China AI Semiconductors; Tencent Research Institute.

Tier 4 (Triangulation only, flagged in text). Sacra, Value Add VC, FutureSearch, EBC and similar trackers, used to cross-check wire reporting and never as sole support for a load-bearing number.

Figures are as of 24 August 2026 and were cross-checked across at least two independent sources where the source tier was below primary. Multiples and growth rates are the author's calculations from the reported figures cited above; Mortise Capital analysis is the interpretation of these sources, not itself a data source.