Are Americans losing faith in AI while the Chinese keep building?

AI sentiment is asymmetric. Capex is the proof.

US AI media sentiment peaked in the eighteen months between the November 2022 ChatGPT launch and the May 2024 GPT-4o keynote. By April 2026 the rolling 13-week index prints near zero on the v0.1 seed. The DeepSeek-V3 release of 26 December 2024 and DeepSeek-R1 of 20 January 2025 broke the “Chinese AI is two generations behind” frame in mainstream US coverage inside of four weeks. Five weeks later the TD Cowen note on Microsoft's ~2 GW of canceled lease commitments converted a one-firm capability story into a sector-wide infrastructure story.¹

In the same window, Chinese AI sentiment moved the other way. The AI+ initiative was formalized in the March 2024 Government Work Report and extended in 2026; Alibaba committed ~USD 53 B in three-year AI+cloud capex on its FY25 Q3 call; ByteDance's 2025 AI capex was reported by the Financial Times at ~USD 20 B; the East-Data-West-Compute phase-II buildout added another ~USD 12 B in state-direction commitments.² A major Chinese AI capex, fundraise, or model launch has landed every four to eight weeks across 2025 and 2026 H1. The US side, in the same period, has netted ~4.75 GW of canceled, delayed, and scaled-back data-center capacity at the hyperscaler level.

When Chinese AI sentiment is rising and US AI sentiment is normalizing, the structural assumption that US frontier labs dominate becomes contestable. That asymmetric move is the program's leading signal.The remainder of this paper documents the cancellation record (§ II), runs the symmetric capex tracker (§ III), shows the sentiment series (§ IV), logs the live prediction (§ V), and specifies the methodology (§ VI).

Five moving parts. Each is versioned. Each ships.

The pipeline implements the Forward Indicators spine on a single corpus split four ways: US AI media sentiment, Chinese AI media sentiment, US AI DC disclosures, and Chinese AI capex announcements. v0.1 ships a deterministic stub scorer over hand-built seed; v0.2 swaps the stub for a calibrated LLM scorer over the live corpus.

Per-article LLM scoring with structured-output prompts returns a sentiment_score on [−1, +1] and a coverage confidence on [0, 1]. Anchored to a hand-coded calibration set of 120 articles (60 US, 60 CN) labeled across 2024-Q1 → 2026-Q2. The 13-week rolling mean is the headline series; window length is the standard equity-research convention.

Every cancellation, delay, or scale-back enters as a structured DCEvent with capacity in MW (rounded to 50 MW in v0.1) and a status verb. Quarterly review pulls from hyperscaler earnings transcripts, local utility interconnect filings, TD Cowen + SemiAnalysis sector notes, and major dispatches from Bloomberg / WSJ / The Information.

Symmetric: CapexAnnouncement with category (model launch / fundraise / state capex / provincial / talent), actor, USD-m figure, and source. Sources: Caixin + 36Kr + Xinhua + MIIT and MOST press releases + the published 14- and 15-FYP fascicles cross-referenced with Brief II.

Three baselines ship alongside the LLM model: a last-observed-value baseline (the t-13 reading), a random walk, and a feature-only logistic regression over the six v0.1 features. The headline metric is Brier score relative to baseline; calibration-curve reliability is the secondary metric.

Brier score, Brier lift over baseline, calibration-curve reliability, and coverage of the 80% credible interval. v0.2 ships a held-out evaluation against the 120-article calibration set and a ten-fold cross-validation report.

What we will and will not publish.

Public sources only. Every DC event and every capex announcement on the published scoreboard resolves to a hyperscaler earnings transcript, a local interconnect filing, a press-of-record dispatch (TD Cowen, Bloomberg, WSJ, FT, The Information, Caixin, 36Kr), or a published model card / arXiv tech report.

Capacity rounded to 50 MW. v0.1 rounds every DC capacity figure to the nearest 50 MW. Triangulating across multiple overlapping press accounts of the same lease portfolio would otherwise introduce false precision.

Reported vs confirmed. Rows resting on secondary financial-press reporting rather than primary disclosure are tagged [reported] and do not carry full weight in the bubble-normalization composite.

No fabrication. The TD Cowen ~2 GW Microsoft headline is the founding citation; it is reported across multiple outlets but is not yet tied to a primary Microsoft disclosure. We do not infer additional capacity figures. The corpus is the source of truth.

Score cap. The v0.1 bubble-normalization composite is clamped at 0.85 by construction to avoid false-precision overconfidence on a stub scorer.