What investors actually pay for Western tech firms over their Chinese pair.
The founding paper of the index. We pair the leading Western and Chinese firm in each of twelve frontier-tech sectors and compute the price-to-sales premium per pair — capability, revenue, valuation, and the signed premium, with every datum tied to a public filing. The premium becomes a real-time gauge of how investors are pricing US–China decoupling.
The valuation gap between Western and Chinese frontier tech firms is not uniform — it varies by sector, by the regulatory shadow each firm operates under, and by how listing venues price them. This paper pairs the leading Western and Chinese firm in each of twelve frontier-tech sectors and computes the valuation gap (the “Western premium” or “Chinese premium” depending on which side trades higher). The median across the twelve pairs sits at −19.5% (i.e., the median Chinese firm trades about 19.5% higher than its Western pair). Foundation models are decisively Chinese-premium-heavy (median −75.75%, three of four numeric pairs); payments has the cleanest Western premium (Stripe trades +117% above Ant Group, driven by Ant's 2020 IPO halt and lingering regulatory overhang). Each pair is tracked over time, surfacing the discrete repricing events that drive the gap. This is a descriptive paper rather than a predictive one; the value sits in the pair-level structure and the temporal series.
The capability gap is closing. The valuation gap is not.
The price-to-sales premium between Western and Chinese AI firms is the cleanest single read of how investors are pricing US–China decoupling. A narrowing premium prices convergence; a widening premium prices the opposite. We measure it, per pair, on public filings.
OpenAI trades at roughly USD 852 billion against an annualized run-rate of USD 25 billion — a price-to-sales multiple of ~34×.1 Moonshot AI, whose Kimi-K2 family is within striking distance of frontier on multiple public evals, closed a USD 2 billion round in May 2026 at a USD 20 billion valuation — on annualized revenue of ~USD 200 million.2 Boston Dynamics, unprofitable and now Hyundai-controlled, is valued at USD 21–28 billion ahead of a planned US listing on cumulative four-year revenue near USD 285 million.3 Unitree, the Hangzhou humanoid firm whose STAR Market application was accepted 20 March 2026, posted 2025 revenue of RMB 1.708 billion (~USD 240 million), shipped 5,500+ humanoid units, and targets a listing valuation of RMB 42 billion (~USD 6.2 billion).4
In each pair, the Chinese firm posts revenue and capability numbers that are not 1/5 or 1/10 of the Western counterpart's — they are at parity, or close enough that the gap is plausibly closeable in two to three years. The valuation gap is structurally larger than the capability gap, the revenue gap, or the unit-shipment gap. That excess valuation gap is the Western premium.
The premium is both a real-time gauge of how investors read the geopolitical regime and a measurable arbitrage claim — sector-conditional, pair-resolved, and recoverable from filings. This program builds the measurement and ships the public scoreboard.
What exists, what it doesn't do, and where this program slots in.
| Source | Output | Public | Pair-resolved | Forward-looking | Premium quantified |
|---|---|---|---|---|---|
| CB Insights | Subscription dashboards | Yes | No | Mixed | No |
| Crunchbase | Funding events database | Yes | Partial | Mixed | No |
| ITJUZI (IT桔子) | China funding tracker | Paywalled | No | PRC-side only | No |
| McKinsey / BCG | Sector briefs | No | No | Slow cycle | No |
| SCMP / Caixin | Reporting | Yes | No | Episodic | No |
| Sacra / Pitchbook | Private-co. estimates | Partial | Yes (paid) | Pair-level | No |
| Forward Indicators — IV | Pair-level premium scoreboard | Yes | Yes | Yes — quarterly | Yes — per pair |
Sources: CB Insights, Crunchbase, ITJUZI, McKinsey/BCG sector practice, SCMP/Caixin, Sacra, PitchBook public summaries. Compiled May 2026.
What distinguishes Forward Indicators IV from the seven reference programs above: per-pair P/S premium computation against named Chinese and Western counterparts, a public scoreboard with every datum sourced to a regulatory filing or model card, and a quarterly refresh cadence tied to filing events rather than analyst-report schedules. No existing program publishes all three. The dashboards are paywalled, the press is episodic, and the consultancies do not quantify the premium per pair.
Five moving parts. Each is versioned. Each ships.
The pipeline implements the Forward Indicators spine on a single corpus: a matched-pair scorecard, sourced from regulatory filings, scored quarterly, with the premium calculated per pair and aggregated by sector. Each pair carries a capability vector, a revenue normalization, a valuation reference, and a per-pair premium with confidence flag.
Pair selection
Pairs must match on sector taxonomy bucket (frontier foundation model, humanoid robotics, quadruped robotics, AI accelerator silicon, leading-edge foundry, AI public cloud), match on stage (both pre-IPO or both publicly traded, or with a documented cross-stage rationale), carry at least one published quantitative capability score, and have audited or audited-equivalent revenue for at least one side within the last eighteen months. Diversified conglomerates enter on a segment basis (DeepMind, not Alphabet; Cloud Intelligence Group, not Alibaba Group).
Capability scoring
Foundation models report MMLU-Pro, GSM8K or AIME, HumanEval or SWE-Bench, LMArena Elo, and active-parameter count. Robotics reports shipped units, degrees of freedom, peak torque, and a zero-to-three qualitative score per task family. Silicon reports node generation, peak FP16 / FP8 throughput, memory bandwidth, and 12-inch-equivalent wafer capacity. AI cloud reports AI-product revenue run-rate and home-market share. Where multiple variants exist, the strongest released (not gated) model is used; the source eval is recorded.
Revenue normalization
All revenue and valuation figures are converted to USD at the People's Bank of China central parity, HKD–USD, or KRW–USD reference rate on the filing date of the source document. Where Chinese firms report calendar-year revenue and Western firms report on a non-calendar fiscal year, the program reports the most recent disclosed full-year figure for each and notes the period mismatch. Run-rate, ARR, and annualized figures are tagged as such and never mixed with reported full-year revenue in the same column.
Premium calculation
For each pair we compute three numbers. The capability ratio (Western score ÷ Chinese score, mean across dimensions). The revenue ratio (Western USD revenue ÷ Chinese USD revenue, on matched periods). The headline number, the valuation premium:
Premium = (P/SWestern ÷ P/SChinese) − 1
Positive premium means the Western firm trades at a higher price-to-sales multiple than its Chinese counterpart. Where one side has trivially small revenue (DeepSeek pre-funding; Figure AI pre-revenue), the program substitutes a valuation-per-shipped-unit or valuation-per-employee secondary metric and flags the substitution.
Editorial guardrails
Public documents only. Single-source claims are marked. Missing figures are flagged as unverified and source-pending rather than inferred. Pair-symmetric due diligence: when we raise the confidence bar on a Chinese-side number, we raise the same bar on the Western counterpart. Cohort and methodology versions are stamped and changelogged.
Sector-conditional. Robotics carries the premium; foundry does not.
Twelve pairs across foundation models, humanoid robotics (public and private), accelerator silicon, logic foundry, enterprise AI, open-weight models, and payments infrastructure. Each pair shows the firm description, last disclosed valuation, last reported revenue, and a capability anchor on both sides — with the premium summarized as a one-line editorial takeaway.
| Sector | West firm | East firm | Premium | Direction | Confidence |
|---|---|---|---|---|---|
| Humanoid robotics | Boston Dynamics | Unitree | +233% | Western | high |
| Payments | Stripe | Ant Group | +117% | Western | high |
| Foundation models | Anthropic | Zhipu · GLM | −98% (reported) · −95% (ARR) | Chinese | high |
| Foundation models | Mistral | DeepSeek | −82% to −89% | Chinese | medium |
| Foundation models | xAI | Minimax | +70% | Western | medium |
| Foundation models | OpenAI | Moonshot · Kimi | −66% | Chinese | high |
| Semiconductors | NVIDIA | Cambricon | −78% (trailing) · −14% (FY2026 fwd) | Chinese | medium |
| Semiconductors | TSMC | SMIC | +7% | Western | high |
| Enterprise AI | Cohere | 01.AI · Yi (零一万物) | n/c (revenue asymmetric) | Neutral | medium |
| Foundation models | Google DeepMind | DeepSeek | qualitative | Neutral | qualitative |
| Humanoid robotics | Figure AI | UBTECH | >+1,000% (qualitative) | Western | qualitative |
| Humanoid robotics | 1X Technologies | AgiBot (智元) | ≈ par | Neutral | medium |
Sources per row: see dossier § 3. Premium = (P/SW ÷ P/SC) − 1 where P/S is computable on both sides; otherwise qualitative. All currency normalized to USD at filing-date FX.
Median Western premium across the cohort: −19.5%.
Across the twelve documented pairs, eight resolve to a numeric signed premium and four sit outside the like-for-like framework — either revenue-asymmetric (Cohere ↔ 01.AI), structurally non-comparable (DeepMind ↔ DeepSeek), pre-revenue on both sides (1X ↔ AgiBot), or qualitatively unbounded (Figure ↔ UBTECH). The Stripe ↔ Ant Group +117% premium is the cohort's cleanest regulatory-overhang outlier; the median pair sits at −19.5% — just below parity — with an interquartile range of 169.25 percentage points running from Q1 −75.75% (the Hong Kong listing-window foundation-model inversion zone) to Q3 +93.5% (the private-stage Western premium zone). Western and Chinese premium counts split four against four across the numeric cohort, with the remaining four pairs reported qualitatively.
The aggregate −19.5% median masks four very different regimes.
| Sector | Pairs | Median premium | Range | Direction |
|---|---|---|---|---|
Foundation models OpenAI · Anthropic · xAI · Cohere · Mistral · DeepMind ↔ Moonshot · Zhipu · Minimax · 01.AI · DeepSeek | 4 2 reported qualitatively | −75.75% | −96.5% to +70% | Chinese premium dominant (3 of 4 negative) |
Semiconductors / AI silicon NVIDIA · TSMC ↔ Cambricon · SMIC | 2 | −19.5% | −46% to +7% | Mixed |
Humanoid robotics Boston Dynamics · Figure · 1X ↔ Unitree · UBTECH · AgiBot | 1 2 reported qualitatively | +233% | single observation | Western premium dominant |
Payments infrastructure Stripe ↔ Ant Group | 1 | +117% | single observation | Western premium dominant |
Regrouping the cohort into four broad categories shows that the aggregate −19.5% median is the near-cancellation of two opposite regimes, not a generalized state. Foundation models (four numeric pairs) cluster on the Chinese-premium side with a median of −75.75% and the cohort's widest dispersion — 166.5 percentage points from Anthropic and Zhipu (−96.5%) to xAI and Minimax (+70%) — reflecting the Hong Kong rerating on top-tier labs against the Western private-market premium that reappears once the Chinese firm sits at a smaller absolute revenue base. Semiconductors (two numeric pairs) sit at the cohort median by coincidence (−19.5%) but mask a 53-point spread between the NVIDIA and Cambricon trailing inversion and the TSMC and SMIC near-parity foundry anchor. Humanoid robotics produces only one numeric observation (Boston Dynamics and Unitree, +233%) with two pairs reported qualitatively because both sides are pre-revenue or revenue-asymmetric, but the direction is unambiguous: every robotics pair, numeric or qualitative, runs Western-premium positive. Payments contributes a single high-confidence observation (Stripe and Ant Group, +117%) that we treat as the cleanest regulatory-overhang datum rather than a sector trend. The sector view replaces “median near parity” with the operative claim: the foundation-model regime has inverted, the robotics regime has not, and the semiconductor regime is the cohort's only true mixed-signal zone.
Premium trajectories, 2023 → mid-2026.
The six per-pair trajectory sparklines live inside the interactive scoreboard above (§ IV · The cohort) — the sector chips filter both the scatter and the trajectory grid in lockstep.
Six computable pairs carry at least three reference points across the end-2023 to mid-2026 window. Three pairs — DeepMind and DeepSeek, Figure and UBTECH on the Figure side, and 1X and AgiBot — are structurally non-comparable per the methodology above. Anthropic and Zhipu separates after the Hong Kong listing window of late 2025 and early 2026 — the gap widens from −50% at end-2023 to −98% at mid-2026 as the trailing price-to-sales multiple detonates. TSMC and SMIC traces the cohort's cleanest narrowing — +145% to +40% across the same window as SMIC's multiple triples under the export-control regime. xAI and Minimax shows the cleanest revenue-curve maturation, compressing from +1,289% to +70% as both firms scale revenue past the near-zero base.
Reference points end-2023, end-2024, end-2025, mid-2026 · signed premium = (P/SW ÷ P/SC) − 1 · full per-period sourcing in the dossier.
The premium lives in the listing-venue and the staging.
Across the founding cohort, the premium split cleanly by sector. Three pairs ran negative — Moonshot against OpenAI (−66%), Zhipu against Anthropic (−98% reported and −95% on annual run-rate), and NVIDIA against Cambricon (−78% trailing and −14% forward) — because the Hong Kong listing window and Cambricon's post-export-control rerating priced top Chinese AI and accelerator firms above their Western counterparts on trailing revenue. Two pairs sat at or near par— foundry (TSMC and SMIC, +7%) and private humanoid robotics (1X and AgiBot, near par) — because the valuation anchors are exogenous: capacity, margins, and private rounds.
The remaining four pairs ran positive — xAI and Minimax (+70%), Boston Dynamics and Unitree (+233%), Figure and UBTECH (qualitatively north of +1,000%), and a DeepMind and DeepSeek qualitative anchor — concentrating the Western premium in private humanoid robotics and in pairs where the Western side has access to US private capital that the Chinese counterpart does not. The range across the computable pairs runs from +7% to greater than +1,000% on the Western-premium side and −14% to −98%on the inversion side. The two listing-venue mechanisms — the Hong Kong frontier-AI window and the Cambricon post-export-control rerating — produce a separate analytic regime that we report on both trailing and forward bases.
The signed-premium scatter, sector chips, and table that previously sat alongside this finding now live in the interactive scoreboard above (§ IV · The cohort).
What the premium prices in — and what it stops pricing in.
For capital allocators
The premium is a tradable measurement of how decoupling is being priced. A narrowing premium — as seen in the foundation-model pairs after the HKEX rerating — is a price signal that public-market access is the binding constraint, not capability. A widening premium in private humanoid robotics is the market pricing US capital, US customer access, and US-origin foundation models as inputs the Chinese side cannot fully substitute for — even at half the engineering cost.
For policymakers
The premium is the cleanest single read of the cost of decoupling at the firm level.Where the premium is large (private humanoid robotics, >+1,000%), US policy is producing rents for US firms at the expense of efficient capital allocation; where the premium is small or inverted (foundation models, foundry), policy is moving along the grain of the market. Two-sided premium tracking gives the policy community a public scoreboard that survives administrations.
For Chinese founders
The HK listing window has begun to clear the structural discount on top-tier Chinese AI firms. The implied playbook: file early, file in Hong Kong, price aggressively on capability rather than waiting for revenue convergence. The premium reversal in Moonshot and Zhipu is the proof of concept.
For Western founders
The US private-market premium in pre-revenue robotics is a real cushion — but it is the gap that closes fastest when capability convergence becomes legible. Capability parity at a fifth of the valuation is a discount only as long as the buyers and the listing venues stay segmented.
What we will and will not publish.
Public documents only. Every revenue, valuation, and capability number on the scoreboard resolves to an IPO prospectus, an SEC filing, an HKEX or STAR Market filing, a published model card, or an eval report on arXiv. We do not use leaked decks, anonymous sourcing, or non-public secondaries data.
PRC private-market disclosure asymmetry. Chinese private-market disclosure is structurally weaker than US public-market disclosure: shipment counts, ARR, and cap-table composition are typically un-audited until an HKEX or STAR Market prospectus. The program flags every Chinese-side figure sourced from press or analyst reports rather than primary filings, and treats the gap as a measurement-error band rather than a free lunch.
Listing-venue rerating mechanism. Two pairs invert on trailing price-to-sales (Moonshot, Zhipu) and a third inverts on trailing but narrows on forward (Cambricon). In each case the mechanism is the Hong Kong AI listing window of late 2025 and early 2026 or the post-export-control accelerator rerating, not a change in capability. We report trailing and forward multiples where they diverge and never collapse them silently.
Pair-symmetric due diligence.When we raise the confidence bar on a Chinese-side number, we raise the same bar on the Western counterpart. We do not selectively under-cite either side — Figure AI's pre-revenue status earns the same unverified, source-pending flag as a missing AgiBot revenue line.
No-fabrication rule. Where a figure is missing from the public record, the cell is flagged as unverified and source-pending rather than imputed. This is the binding rule for the dossier and for the scoreboard.
Versioned cohort. Pairs and methodology version are stamped. v0.1 is the cut as of 30 May 2026. When a pair is added, removed, or re-graded, the change appears in the changelog with the source that drove it.
Notes
- OpenAI, “Accelerating the next phase of AI”, company blog post, March 2026, announcing the USD 122 billion Series funding at USD 852 billion post-money. Annualized run-rate of USD 25 billion confirmed in The Information and TechCrunch dispatches, February–March 2026.
- Bloomberg, “Kimi chatbot maker Moonshot AI Valued at $20 Billion in Meituan-Led Round”, 7 May 2026; ARR ramp confirmed in TechNode and TechCrunch coverage of the USD 700 million round (Feb 2026) and the USD 2 billion round (May 2026). Round led by Meituan's Long-Z Investments with Tsinghua Capital, China Mobile, CPE Yuanfeng.
- Korea Economic Daily / Seoul Economic Daily / The Korea Herald, January–May 2026, reporting analyst enterprise-value estimates of KRW 30–40 trillion (USD 21–28 billion) ahead of a planned US listing. Cumulative 2022–Q3 2025 revenue KRW 390.7 billion and cumulative losses KRW 1.38 trillion are the Hyundai consolidated disclosure figures.
- Caixin Global, “Unitree Robotics Files for $608 Million STAR Market IPO”, 21 March 2026; Caixin Global, “Unitree Fast-Tracks Shanghai IPO With Target Valuation of $6.2 Billion”, 26 May 2026. Revenue RMB 1.708 billion, adjusted profit RMB 600.1 million, gross margin 60.27%, 5,500+ humanoid units shipped, 32.4% global humanoid unit share, sourced from the Shanghai Stock Exchange STAR Market application file accepted 20 March 2026. Full sourcing for the cohort in projects/markets/RESEARCH.md.
Forward Indicators (2026). “Comparative Tech Firm Premium Scoreboard, v0.” Forward Indicators Working Paper No. IV-01.
DOI: 10.xxxx/fi.wp.iv.01 · pending registration