← OMEN

China AI Substitution Monitor

Tracks demand for Chinese open-weight models – DeepSeek, Qwen, GLM (Z.ai), Kimi (Moonshot), MiniMax, MiMo (Xiaomi) – as a leading indicator for US AI equities. Thesis: Chinese open-model adoption ↑ → pricing pressure on proprietary US models ↑ → lower expected inference margins and marginal GPU demand → pressure on US AI infrastructure stocks. Companion to the AI Crash Monitor.

OpenRouter usage: … Hugging Face: … GitHub: … Polymarket: … Equity: … Leaderboards/consumer: snapshot Supply-side: …

Chinese AI Adoption Index – weighted composite, 0–100. Weights follow the source methodology; families without a live or recent source are excluded and weights renormalized.

–/100

ContainedBuildingDisplacing

Substitution Pressure Index (SPI)

–

Chinese-lab tokens ÷ US-lab tokens, OpenRouter trailing week

Chinese share of routed tokens

–

named Chinese labs, trailing week

4-week change

–

share, percentage points

Frontier price gap

–

US flagship output $/M ÷ Chinese flagship

Substitution pressure trend – weekly Chinese vs US lab share of OpenRouter routed tokens (named authors; the platform's "others" bucket is excluded, so Chinese share is a lower bound)

Chinese labs share US labs share Chinese share on Vercel AI Gateway (daily) SPI = CN÷US tokens (lower panel)

OpenRouter usage – by lab – trailing week, live

Tokens routed through OpenRouter only – first-party OpenAI/Anthropic/Gemini API traffic is invisible here, so this measures the price-sensitive, model-agnostic segment where substitution shows up first.

Top models by routed tokens – trailing week, live

#	Model		Tokens/wk	Share

Inference price compression – flagship $/1M tokens, newest model in each line, live from OpenRouter

Model		Input	Output

Hugging Face adoption – 30-day rolling downloads by lab, live

Chinese labs vs US open-weight families (Llama, Gemma, GPT-OSS, Phi, Nemotron, OLMo). Sum of each org's top-50 repos; community quantizations/finetunes not attributed, so both sides are undercounts. Bar length is log-scaled – Qwen's totals would otherwise reduce every other lab to a sliver.

GitHub developer engagement – active flagship repos, live

Repo	Stars	Forks	Δ/day

Δ/day is measured server-side between the updater's daily runs (a browser-local baseline fills in for dev use). Fork counts are the stronger adoption signal.

Leaderboard proximity – snapshot

Best Chinese model on LMArena

–

Elo gap to US leader

–

Chinese models in top 10 / top 20

–

text arena, ordinal ranks

Artificial Analysis index

–

Sources: LMArena's official leaderboard dataset (CC BY 4.0; arena.ai scrape as fallback) and Artificial Analysis (live via its free Data API when the updater has a key, else a dated snapshot). Refreshed by the updater script, values dated in the header.

Router cross-check – Vercel AI Gateway – daily token share, batch

Lab share of text tokens on Vercel AI Gateway – production web apps, a different population than OpenRouter's price-sensitive routing, so it de-biases the headline SPI. When the two routers disagree, the truth is usually in between. Data: “AI Gateway Leaderboard Data” © Vercel, CC BY 4.0 (export cached 24h).

Local adoption – Ollama pulls – lifetime cumulative, batch

Pull counts from the Ollama library – the self-hosted segment that no router or HF download counter sees, and where Chinese open-weight models are disproportionately run. Lifetime totals (a stock, not a flow); the daily rate appears once two refreshes are >20h apart.

Ecosystem gravity – fine-tune trees – batch

Models fine-tuned from each flagship base on Hugging Face (the model-tree count on each base's page). Who builds on your model is a stickier adoption signal than raw downloads – a big tree means training recipes, tooling and habits are already committed to that base.

Lab SDK installs – PyPI weekly – batch

Weekly installs of each lab's first-party Python SDK (pypistats.org). Big caveat: Chinese APIs are OpenAI-compatible, so much Chinese-model usage rides the openai package – the Chinese side is a hard undercount. Watch the trend, not the level.

Compute stock – tracked AI clusters – batch

Country share of H100-equivalent capacity across Epoch AI's GPU-clusters dataset (CC BY 4.0) – the supply-side ceiling under the demand panels above. Read the split as directional, not exact: Epoch estimates it tracks 10–20% of world capacity, and Chinese systems are anonymized with rounded specs upstream, which systematically understates China (Epoch's own analysis puts China nearer 15% of tracked performance). US capacity concentrates in far fewer, far larger clusters.

Notable model releases – batch

Notable-model releases per quarter by lab country, from Epoch AI's models database (CC BY 4.0; notable = state-of-the-art, highly cited, or widely used). Joint US–CN releases fit neither line and are excluded; the current quarter is omitted until complete. Recent quarters get backfilled as Epoch catalogues releases, so the last point drifts up for a few weeks after each refresh.

Chip supply – AI accelerators by designer – batch

Quarterly accelerator output in H100-equivalents (median estimates, wide confidence intervals), US designers (NVIDIA, AMD, Google, Amazon) vs Chinese (Huawei, Cambricon), from Epoch AI's chip-sales dataset (CC BY 4.0). The series ends at the last quarter with estimates for both sides – Chinese estimates lag Nvidia's by a couple of quarters. Domestic designs are not China's whole compute inflow: Epoch separately estimates a median ~660k H100e were smuggled into China through end-2025 (90% CI 290k–1.6M), roughly a third of China's stock – so this panel understates what Chinese labs can actually train and serve on.

Prediction markets – Chinese AI leadership – live Polymarket odds + Kalshi (batch, daily)

Market	History	YES	24h Δ

Polymarket leadership markets resolve on arena.ai rankings; Kalshi is a CFTC-regulated exchange with deeper AI books, fetched by the daily updater (no CORS API). These are the market's own probabilities that Chinese labs take or hold the frontier – the cleanest forward-looking read in this monitor. Rows tagged AUTO were found by searching Polymarket's active markets rather than pinned by id, so the monthly "best Chinese AI model" cohort keeps refreshing as each one resolves; only markets above $5k volume are admitted, which excludes the unnamed placeholder outcomes that sit at 50¢.

US AI equity check – the stocks this pressure would show up in

Prices from the site's shared feed (market-data.json, auto-refreshed server-side). Drawdown is vs the 6-month high; orange sparklines are below it, green at it – red stays reserved for "Chinese labs" everywhere else on this page. Correlation caveat: the AI Crash Monitor's lead-lag study finds no statistically significant lead from prediction-market signals to NVDA/SOXX yet – treat this page as a market-pricing monitor, not a validated forecast.

Why this matters for US AI stocks

Enterprise switching is underway. CNBC (2026-07-07): Coinbase cut AI spend nearly 50% by moving ~1,200 internal agents to GLM-5.2 and Kimi K2.7; Lindy migrated fully off Claude to DeepSeek; Uber capped engineer AI budgets at $1,500/month; Citi reportedly restricted access to the most expensive OpenAI/Anthropic tiers. Chinese models price 60–90% below US flagships for near-frontier quality.

The capex asymmetry is the exposure. US hyperscaler capex was ~$400B in 2025 with >$520B consensus for 2026, vs ~$57B for China's major platforms (UBS). If cheap open-weight models keep compressing inference pricing, the revenue side of that capex – and the margin structure of proprietary US model vendors – is what gets repriced. NVDA's China datacenter revenue already fell from ~$4.6B/yr to ~zero under export controls, so the remaining channel is margin pressure and demand-mix shifts, not lost China sales.

The counter-thesis is live too. 2026 has produced no repeat of the Jan-2025 DeepSeek selloff: NVDA beat earnings through DeepSeek V4's launch window and made new highs after V4-Pro. Chinese labs remain export-control constrained, and US frontier models still hold the #1 spots on quality leaderboards. Watch this page for the divergence, not the conclusion.

Methodology

Chinese AI Adoption Index = Σ(w_i·s_i)/Σw_i over available signal families, each normalized 0–100 against fixed reference ranges: router Chinese token share 0→70% (w=30; mean of the OpenRouter and Vercel-AI-Gateway normalized shares when both are available, OpenRouter alone otherwise), Hugging Face Chinese share of open-weight downloads 0→90% (w=20), GitHub star velocity across the flagship basket 0→300/day (w=15; the basket is rediscovered each run as the two most-starred repos per lab still pushed to within 180 days, so a new release enters it the week it lands), Ollama Chinese share of named-family pulls 0→70% (w=10), Western search/consumer interest share 0→50% (w=10), consumer-app Western top-chart presence 0→100 (basket-mean best top-100 rank, w=10), LMArena proximity 100−Elo-gap, floored at 0 (w=5). Weights follow the source note's suggested scheme, extended with the local-adoption family; excluded families renormalize. SPI = Chinese-lab ÷ US-lab routed tokens (named authors) on OpenRouter, trailing week. Histories: the updater appends one row per UTC day to china-metrics.csv (composite index, SPI, router/HF/Ollama Chinese shares, GitHub star velocity, frontier price gap) – the small history lines across the page read from it, so they are server-durable rather than per-browser; Polymarket odds histories come live from the public CLOB prices-history endpoint. Frontier price gap = median output $/M of US flagship lines (GPT top/mid, Claude Opus/Sonnet, Gemini Pro/Flash) ÷ median of Chinese lines (DeepSeek pro/flash, GLM, Kimi and Kimi Code, MiniMax, Qwen Max). Each line resolves to its newest live model in OpenRouter's catalogue at page load rather than to a pinned version, so a launch or a price cut moves the gap the day it ships and a retired slug surfaces as a named warning instead of quietly leaving the median. Lines are matched on name, never on price: Chinese labs deliberately price their newest flagship below their previous one, so ranking by cost would pick the older model. Supply-side panels (frontier gap, compute stock, chip supply, release cadence) are context, not index inputs – they carry no weight in the adoption index. Frontier gap = best AA Intelligence Index score per lab country, monthly since 2024; lag = months between a US model first reaching today's best Chinese score and the Chinese frontier getting there; near-frontier price ratio = cheapest blended $/M within 3 index points of each side's frontier, US ÷ Chinese.

Caveats

OpenRouter sees only routed traffic – it overstates Chinese share of the total market (first-party US API traffic is invisible) but is the best public real-time proxy for the substitutable segment; the Vercel AI Gateway cross-check (production apps, export cached 24h) partially corrects that bias and currently reads much lower. HF download totals mix model generations and exclude derivative quantizations. Ollama pulls are lifetime cumulative, so they lag shifts; the derived daily rate is the leading part. Fine-tune-tree counts depend on authors declaring base_model metadata. PyPI SDK installs undercount the Chinese side badly (OpenAI-compatible APIs ride the openai package). Leaderboard (LMArena official dataset, arena.ai scrape fallback), search interest (Google Trends), consumer-app chart presence (Apple RSS + Google Play), Vercel, Ollama, Kalshi, fine-tune trees and SDK installs are all refreshed by update-china-data.py; Artificial-Analysis scores refresh live when the updater has an ARTIFICIAL_ANALYSIS_API_KEY, else they remain a dated snapshot, and the Cloudflare-Radar consumer-traffic panel appears only when a CF_RADAR_TOKEN is configured (see freshness strip). Prediction-market prices are thin in places; Kalshi quotes are daily-batch, not live, and the discovered (AUTO) leadership rows rotate as each monthly cohort resolves, so that block is not a fixed panel across visits. The supply-side panels inherit their sources' blind spots: Epoch's cluster dataset covers an estimated 10–20% of world capacity and anonymizes Chinese systems (the split understates China); chip-sales figures are median estimates with wide confidence intervals, count designs rather than China's smuggled-in NVIDIA inflow, and lag ~2 quarters on the Chinese side; notable-model counts backfill for weeks after a release; the AA frontier series is rebuilt from the current catalogue, so delisted models silently leave its history, and it appears only when the updater has an ARTIFICIAL_ANALYSIS_API_KEY. Epoch CSVs are refetched weekly, not per-run. Nothing here is a validated forecast of equity returns – the companion monitor's own lead-lag test is honest about that.

Sources

OpenRouter rankings · Vercel AI Gateway leaderboards (data © Vercel, CC BY 4.0) · Hugging Face Hub API · Ollama library · GitHub API · Polymarket Gamma API · Kalshi trade API · LMArena leaderboard dataset (CC BY 4.0) · Artificial Analysis · Epoch AI data (GPU clusters, chip sales, AI models; CC BY 4.0) · pypistats.org · Cloudflare Radar AI Insights · CNBC 2026-07-07

Built as part of OMEN. Not investment advice.