The −19.5% headline hides two opposite tech-pricing regimes.

A headline median that hides everything important

The twelve-pair cohort returns an aggregate median signed premium of −19.5% across the eight numeric pairs. Read literally, that number says the typical Chinese frontier-tech firm trades at roughly twenty percent above its Western counterpart on a price-to-sales basis — a near-parity reading that mildly favors the Chinese side. Read literally, that number is misleading. Disaggregating the cohort into four sector buckets — foundation models, semiconductors, humanoid robotics, and payments — shows that the aggregate is the near-cancellation of two structurally opposite regimes. Foundation models cluster on the Chinese-premium side at a median of −75.75%; humanoid robotics is unambiguously Western-premium-positive at +233%; semiconductors are the cohort's only true mixed-signal zone; payments contributes a single high-confidence regulatory-overhang datum, Stripe and Ant Group at +117%. The operative claim that emerges from the sector view is not “Chinese firms trade slightly above Western counterparts.” It is “the foundation-model regime has inverted, the robotics regime has not, and semiconductors split between accelerator rerating and foundry parity.”

Why a single median is the wrong summary

The eight numeric pairs split four against four. The Western-premium-positive group: Boston Dynamics and Unitree (+233%), Stripe and Ant Group (+117%), xAI and Minimax (+70%), TSMC and SMIC (+7%). The Chinese-premium-negative group: Anthropic and Zhipu (−96.5%), Mistral and DeepSeek (roughly −85%), NVIDIA and Cambricon (−46% blended), OpenAI and Moonshot (−66%). When two sub-distributions live on opposite sides of zero with comparable magnitudes, the median collapses to near-zero by construction. The median in that case is not a measurement of the typical pair — it is a measurement of how nearly the two regimes cancel.

A reader looking at −19.5% in isolation would conclude that the Chinese side carries a mild premium. A reader looking at the four sector buckets would conclude that where you sit in the cohort determines which side you trade on, and the magnitude of the gap is large in both directions. Those two readings produce different portfolio decisions, different policy reads, and different research priorities. The sector view is the one the paper actually wants its readers to anchor on.

The widest dispersion, the inverted center

Four numeric pairs sit in the foundation-models bucket: OpenAI and Moonshot (−66%), Anthropic and Zhipu (−96.5% midpoint), xAI and Minimax (+70%), and Mistral and DeepSeek (roughly −85% midpoint). Two further pairs sit here without a numeric reading: Cohere and 01.AI is revenue-asymmetric on the Chinese side; DeepMind and DeepSeek is structurally non-comparable because Alphabet does not separate the DeepMind segment in its financials.

The numeric subset has a median of −75.75% with a range of 166.5 percentage points— the cohort's widest dispersion by a wide margin. Three of the four numeric pairs sit on the Chinese-premium side, and they sit there decisively: the typical foundation-model pair is not slightly Chinese-premium-positive, it is heavily so. The single positive observation in the bucket, xAI and Minimax at +70%, sits in the bucket precisely because Minimax is the smallest of the Chinese frontier labs by trailing revenue, which puts both firms' absolute revenue bases close enough that the Western-side premium reappears.

The Hong Kong listing window of late 2025 and early 2026 reset the foundation-model regime. The Chinese side now trades at a premium on trailing revenue because the public-market access constraint just relaxed.

The mechanism behind the −75.75% median is well isolated. Moonshot, Zhipu, and Minimax all touched Hong Kong Stock Exchange listing events between September 2025 and May 2026. Each event reset the trailing price-to-sales multiple on the Chinese side by an order of magnitude. The Western counterparts — OpenAI, Anthropic, and xAI — remained private throughout and continued to be priced on forward annual run-rate. The premium gap is therefore largely an artifact of listing-venue asymmetry: trailing price-to-sales on a publicly traded firm against forward run-rate on a private firm produces structurally different multiples even when the underlying capability and revenue are comparable.

The only true mixed-signal zone

Two numeric pairs sit in semiconductors: NVIDIA and Cambricon (−46% blended, −78% on trailing, −14% on FY2026 forward) and TSMC and SMIC (+7%). The bucket median is −19.5%, which coincides with the aggregate cohort median by chance; the underlying range is −46% to +7%, a 53-percentage-point spread.

The two pairs are differently mechanism-bound. The NVIDIA and Cambricon inversion is a domestic-substitution rerating: Cambricon's first-half 2025 revenue grew 4,348% year-on-year as Chinese hyperscalers shifted procurement onto domestic accelerator silicon under the post-October 2022 export-control regime, and the trailing price-to-sales multiple detonated alongside. The forward FY2026 figure narrows the inversion to −14% as consensus revenue catches up. The TSMC and SMIC near-parity at +7% is the structural counter-pair: foundry valuations are anchored by capacity, gross margin, and node generation rather than by listing-venue narrative, and the two firms' multiples sit within a few percent of each other despite TSMC being two nodes ahead.

Reading the semiconductors bucket as −19.5% would smuggle in two completely different stories: a Chinese domestic-substitution rerating in accelerator silicon, and a structural Western capability premium in logic foundry that has held basically constant. The sector view keeps those stories separate.

Western-premium positive everywhere it can be read

Three pairs sit in humanoid robotics: Boston Dynamics and Unitree at +233% (the cohort's largest numeric Western premium); Figure and UBTECH, qualitatively north of +1,000% but not computable on a like-for-like basis per editorial guardrails; and 1X and AgiBot, near parity on absolute valuation but also not computable because both sides are pre-revenue. Only one of the three produces a clean numeric reading, but the direction across all three is unambiguous.

Boston Dynamics is valued at USD 21 to 28 billion (midpoint USD 24.5 billion) on cumulative four-year revenue of USD 285 million. Unitree is targeting a STAR Market listing valuation of USD 6.2 billion on FY2025 revenue of USD 240 million, with more than 5,500 humanoid units shipped. The Western firm prints at a price-to-sales multiple of 86 times; the Chinese firm at 25.8 times. The Western firm carries roughly 3.3 times the Chinese firm's multiple despite shipping fewer humanoid units in the same year. Every robotics pair, numeric or qualitative, runs Western-premium positive; the smallest premium in the bucket (1X and AgiBot at roughly par) still has 1X at 1.74 to 4.8 times AgiBot on absolute valuation. The premium does not invert when both sides are pre-IPO; it just shrinks toward par.

A single high-confidence regulatory-overhang datum

Payments contributes one pair: Stripe and Ant Group, +117%. The cohort treats this as the cleanest regulatory-overhang reading rather than a sector trend, because the sector contains essentially no other Western-Chinese pair where both sides are at comparable scale and stage. The mechanism — Ant's November 2020 IPO halt, the subsequent forced restructuring, and the institutionalized markdown visible in Alibaba's 6-K disclosures — is exhaustively documented in the cohort-expansion piece. The bucket median is the single observation.

From “near-parity median” to four operative claims

The sector view replaces a single misleading summary statistic with four operative claims that each support a different downstream decision:

One.Foundation models are now decisively Chinese-premium-heavy on trailing price-to-sales, driven by the Hong Kong listing window. The relevant trade is not “Chinese frontier AI is cheap” (it isn't, anymore); it is “the trailing multiple is the wrong yardstick — forward run-rate reverses the sign in three of the four pairs.”

Two.Humanoid robotics is unambiguously Western-premium-positive, with the magnitude widening at the private-stage tier. The relevant trade is “the premium is the market pricing US capital, US customer access, and US-origin foundation models as inputs the Chinese side cannot fully substitute for.” The premium will close fastest when one of those three inputs becomes substitutable.

Three.Semiconductors are the cohort's only true mixed-signal zone. Accelerator silicon has inverted under domestic-substitution rerating; logic foundry has not. The relevant read is bucket-internal: NVIDIA and Cambricon and TSMC and SMIC are not the same story.

Four.Payments contributes a single high-confidence regulatory-overhang datum that should be read as such — not extrapolated to a sector trend, but used as the cohort's reference point for what regulatory overhang costs at scale.