When Chinese AI hits public markets, the premium changes in days.

From snapshot to trajectory

The founding paper and the sector view both report each pair's premium as a static reading: where it sits today. That framing is insufficient. The static −98% on Anthropic and Zhipu does not communicate that the pair was at −50% at end-2023 and detonated only in the most recent six months under the Hong Kong listing window. A static +40% on TSMC and SMIC does not communicate that the same pair was at +145%thirty months earlier — a 105-point compression that traces a coherent regime change. This piece adds a temporal series to six computable pairs across four reference points (end-2023, end-2024, end-2025, mid-2026) so the premium becomes a trajectory rather than a number. The three cleanest trajectories — Anthropic and Zhipu under the Hong Kong rerating, TSMC and SMIC under the export-control regime, and xAI and Minimax under revenue-curve maturation — each isolate a different mechanism. The trajectory data is, for a descriptive paper, the strongest reading the cohort produces.

Why a snapshot misleads

The static scoreboard reads the cohort as if every pair sat at its current premium in steady state. None of them does. The Hong Kong listing window for Chinese AI labs opened roughly in late 2025 and produced a series of trailing-multiple resets: Moonshot's September 2025 funding, Minimax's January 2026 IPO, Zhipu's February 2026 Hong Kong uplisting, and the May 2026 cluster of Moonshot's Meituan-led round and Zhipu's peak trading day. Each of those events shows up as a step in the relevant pair's trajectory. The static reading at mid-2026 captures the post-event state but loses the structure that put the pair there.

The same problem holds in semiconductors. The TSMC and SMIC pair sits at +40% in mid-2026. Read in isolation, the number suggests a stable Western capability premium. Read as a trajectory from +145% at end-2023 through +141% at end-2024, +60% at end-2025, and +40% at mid-2026, the same number resolves into a monotonically narrowing series — SMIC's price-to-sales multiple triples across the window as domestic-substitution revenue prices the export-control regime. The shape of the series is the substantive finding; the endpoint alone is not.

And the same problem holds in the revenue-curve direction. xAI and Minimax sits at +70% in mid-2026. In isolation, +70% reads as a moderate Western premium. As a trajectory — +1,289% at end-2024, +662% at end-2025, +70% at mid-2026 — it reads as a near-1,200-point compression driven by both firms scaling revenue past the near-zero base. The same final number, three different stories. The temporal series exists to make those stories legible.

Six pairs, four reference points, signed premium on a symmetric-log axis

Six of the twelve documented pairs carry at least three reference points across the end-2023 to mid-2026 window: OpenAI and Moonshot, Anthropic and Zhipu, xAI and Minimax, Boston Dynamics and Unitree, NVIDIA and Cambricon, and TSMC and SMIC. Three further pairs — DeepMind and DeepSeek, Figure and UBTECH, 1X and AgiBot — sit outside the trajectory framework because one side is pre-revenue, not separable from a parent conglomerate, or both. The three newest pairs — Cohere and 01.AI, Mistral and DeepSeek, Stripe and Ant Group — have point-in-time premiums but not yet enough historical reference data for a full trajectory.

Reference points are end-of-period valuation marks paired with the trailing twelve months of disclosed revenue at each cut. The signed premium is the same as in the static scoreboard: the Western price-to-sales multiple divided by the Chinese price-to-sales multiple, minus one. The aggregate plot uses a symmetric-log y-axis to keep both the −98% endpoint of Anthropic and Zhipu and the +1,686% peak of Boston Dynamics and Unitree legible on the same chart without compressing the ±100% range where most of the cohort lives.

Each reference point is sourced to a primary filing: 10-Ks, 6-Ks, S-1s, and Hong Kong Stock Exchange or STAR Market prospectuses. Where a pair's reference point falls between two filings, we use the most recent filing as of the reference date and note the temporal mismatch. End-2023 is null for both OpenAI and Moonshot and xAI and Minimax because neither Chinese firm had disclosed revenue at that cut.

Anthropic and Zhipu — a six-month detonation

The Anthropic and Zhipu pair traces the cleanest single-mechanism trajectory in the cohort. From end-2023 to end-2025, the signed premium drifted modestly from −50% to −37% — a slow narrowing as Zhipu's annual run-rate climbed and Anthropic's private-round multiples settled. Between end-2025 and mid-2026, the premium detonated to −98%— a 61-percentage-point move in six months.

Zhipu's February 2026 Hong Kong uplisting, which peaked on 29 May 2026, reset the trailing price-to-sales multiple by an order of magnitude. Capability did not change; access to public-market capital did.

The mechanism is well isolated. Zhipu's peak Hong Kong valuation on 29 May 2026 was USD 112 billion against FY2025 reported revenue of USD 105 million — a trailing price-to-sales multiple of roughly 1,066 times. Anthropic at the same date: USD 965 billion against USD 45 billion in annual run-rate, a trailing multiple of about 21 times. The pair's premium moves from a regime where both firms traded as private capital-intensive labs with comparable multiples to a regime where one of the two has retail-investor pricing applied to a trailing-revenue base it has not yet matched. The forward run-rate comparison narrows the gap to −95% — a less extreme but still decisive Chinese-premium reading. The trajectory captures the regime transition exactly.

TSMC and SMIC — a monotonic narrowing

The TSMC and SMIC pair traces the cohort's cleanest monotonic narrowing. End-2023: +145%. End-2024: +141%, almost flat. End-2025: +60%. Mid-2026: +40%. A 105-percentage-point compression across thirty months, with the bulk concentrated in the final twelve.

The mechanism is the post-October 2022 export-control regime pricing in on the SMIC side. SMIC's FY2025 revenue grew 16.15% year-on-year to USD 9.33 billion as Chinese hyperscalers and chip designers shifted procurement onto domestic 7-nanometer-class capacity under the Foreign Direct Product Rule scope. The valuation moved with revenue: SMIC's Hong Kong-listed mark in mid-2026 sits at roughly USD 80 billion against revenue growth that has compounded the multiple meaningfully upward over the window. TSMC's price-to-sales has held roughly steady at about 9 times as its revenue grew on N3 and N2 node ramp without a corresponding multiple expansion.

The TSMC and SMIC trajectory is the negative-space companion to the Anthropic and Zhipu trajectory. Where the foundation-model pair shows the Chinese side rerating dramatically against a stable Western base, the foundry pair shows the Chinese side's multiple expanding under a coherent policy regime while the Western side's multiple holds. Different mechanism, similar direction of premium movement, materially different shape of trajectory.

xAI and Minimax — a 1,200-point compression

The xAI and Minimax pair traces the cohort's largest absolute compression. End-2024: +1,289%. End-2025: +662%. Mid-2026: +70%. A near-1,220-point move over eighteen months, driven by both firms scaling revenue past the near-zero base where any small absolute revenue figure produces an enormous price-to-sales multiple.

xAI's 2025 revenue per the IPO filings sat at USD 3.2 billion; the SpaceX acquisition mark in February 2026 lifted the valuation to USD 250 billion, a multiple of about 78 times. Minimax's Q1 through Q3 2025 annualized revenue sat at roughly USD 250 million, with the Hong Kong IPO valuation peaking at USD 11.5 billion in early trading and a multiple of about 46 times. The remaining +70% is the residual Western premium that reappears once the absolute revenue base on the Chinese side is large enough to stop producing artificial multiple inflation.

The xAI and Minimax shape is informative for the rest of the cohort. The end-2024 reading at +1,289% would have looked like a decisive Western premium on a static read; it was, in fact, an artifact of Minimax's near-zero revenue base producing a price-to-sales multiple in the high hundreds. Static premiums on pre-revenue or low-revenue Chinese firms are structurally noisy, and the trajectory data shows exactly how much noise.

OpenAI and Moonshot, NVIDIA and Cambricon, Boston Dynamics and Unitree

OpenAI and Moonshotreads as a noisier version of the Anthropic and Zhipu story: −86% at end-2024, −49% at end-2025, −66% at mid-2026. The non-monotonicity reflects Moonshot's May 2026 Meituan-led round resetting the Chinese-side valuation at a higher level than the end-2025 cut implied, while OpenAI's annual run-rate caught up partially over the same window.

NVIDIA and Cambricontraces the cohort's most stable Chinese-premium reading: −72% at end-2023, −87% at end-2024, −79% at end-2025, −82% at mid-2026. The pair sits in a sustained inversion regime, with Cambricon's price-to-sales multiple consistently exceeding NVIDIA's under the post-October 2022 export-control regime. The narrowness of the band — 15 points peak to trough — is what makes this trajectory informative; the inversion is structural, not transient.

Boston Dynamics and Unitreetraces the cohort's most dramatic Western-premium expansion and subsequent compression: +27% at end-2023, +688% at end-2024, +1,686% at end-2025, +233% at mid-2026. The peak at end-2025 corresponds to Boston Dynamics' pre-listing valuation marks running ahead of Unitree's pre-prospectus disclosure. The compression to +233% by mid-2026 reflects Unitree's STAR Market application disclosing FY2025 revenue at +335% year-on-year, which collapsed the Chinese-side multiple to a level the cohort's static reading captures cleanly.

The trajectory is the finding

The three clean trajectories — Hong Kong rerating, export-control overhang, revenue-curve maturation — each isolate a different mechanism that the static scoreboard could not distinguish. A reader looking at the static cohort sees four Western-premium pairs and four Chinese-premium pairs; a reader looking at the trajectories sees three coherent regime stories and three more whose shape ranges from stable inversion (NVIDIA and Cambricon) to peak-and-compress (Boston Dynamics and Unitree) to noisy step (OpenAI and Moonshot).

For a descriptive paper, the trajectory data is the strongest reading the cohort produces. The static premium at any single date is a function of accidents of disclosure timing, recent funding rounds, and listing-venue mechanics; the trajectory shows what actually moved the pair. Anthropic and Zhipu's six-month detonation under the Hong Kong listing window is the cleanest single example of regime change in the cohort. TSMC and SMIC's monotonic narrowing is the cleanest single example of policy-regime pricing. xAI and Minimax's 1,200-point compression is the cleanest single example of how much static premium readings overstate the signal on low-revenue Chinese firms.

This is a descriptive paper, not a predictive one — the trajectory data does not justify forecasts on where the next reference point will land. But it does justify a stronger version of the founding paper's thesis: the Western premium is not a sector property; it is a regime property, and the regime is set by listing-venue access, policy overhang, and revenue maturation on different timelines for different pairs. Anyone trading on the premium needs to read the trajectory, not the snapshot.