Chinese AI demand is not falling. It is leaving DeepSeek.
On October 26, 2026 the STAR 50 closed at 1,315.50, 41.7% below its July 1 intraday high of 2,255.25. The coverage read that as China’s AI bubble bursting. Measured against search demand in four Western markets, the index and the demand are not describing the same thing — and neither of them is describing “Chinese AI.”
Read this first. This study rests on one instrument. A second — a different operator, a different collection method, a different population — was registered before any data moved and was not run, because the endpoint would not answer. A replication planned is not a replication run. What follows is a finding about Google Ads search volume, not yet a finding about Chinese AI.
Strip DeepSeek out and Chinese AI demand grew 72% in the US, 59% in the UK, 90% in Germany and 213% in Japan, year over year.
DeepSeek peaked in 2025-01 — eighteen months before the index did — and is down 80% in the US and 87% in the UK and Germany since. None of that decline begins on July 1, 2026. What the selloff coincided with was a fall already two-thirds complete.
The decision this is for
A multi-year AI contract sitting unsigned, and whether to execute it now or hold a quarter or two. The buyer is a VP of Engineering, a Head of Data or a CIO. They are contested by a CFO who reads a 41.7% selloff in Chinese AI equity as evidence that AI pricing is about to fall, and that waiting is free.
Opposite answers lead to opposite actions, and both are expensive. If demand is falling with the equity, waiting is rational and the buyer gains leverage. If demand is flat or rising while equity falls, waiting costs a quarter of capability against competitors who did not wait, and the expected discount never arrives.
What the vendors actually did
Same month, year over year — 2025-08 against 2026-08 — which is how a series with an annual shape is read. Two terms were cut by the contamination screen before any of this was computed: baichuan and mistral, both carrying pre-2022 volume above a tenth of their 2026 peak, which means they were matching something that existed before the vendor did.
| Vendor | US | UK | Germany | Japan |
|---|---|---|---|---|
| kimi ai | +309% | +404% | +410% | +831% |
| glm ai | +307% | +181% | +177% | +529% |
| qwen | +50% | +22% | +83% | +173% |
| minimax ai | −33% | −45% | −46% | +122% |
| ernie bot | −46% | −64% | −64% | −57% |
| deepseek | +0% | −33% | −33% | +22% |
Four of the six moved in the direction the headline says they did not. Kimi, GLM and Qwen grew in every market measured. DeepSeek was flat in the US, fell a third in the UK and Germany, and grew in Japan.
Why this was nearly missed
| Market | Chinese arm | Western arm |
|---|---|---|
| US | deepseek — 90% | chatgpt — 91% |
| UK | deepseek — 86% | chatgpt — 91% |
| Germany | deepseek — 78% | chatgpt — 91% |
| Japan | deepseek — 82% | chatgpt — 92% |
An arm total measures one brand and calls it a category. On the totals, Chinese AI “fell” 16–19% in the UK and Germany. That figure was DeepSeek’s own decline, with three vendors growing three- to ninefold underneath it. The headline and the data pointed in opposite directions, and only the headline would have been published.
So the rule this study now carries: before any group total is read, report the top constituent’s share of it. Above half, the total is a proxy for one name and is labelled as one.
The West is rotating the same way
| Vendor | US | UK | Germany | Japan |
|---|---|---|---|---|
| claude ai | +397% | +309% | +512% | +233% |
| gemini ai | +173% | +232% | +510% | +307% |
| anthropic | +124% | +172% | +398% | +124% |
| chatgpt | +23% | +22% | +22% | +22% |
| grok ai | +22% | +49% | +49% | +50% |
| copilot ai | −18% | +49% | +0% | +0% |
| perplexity | −33% | −33% | −45% | −55% |
| openai | −45% | −45% | −33% | −33% |
| llama ai | +0% | −19% | +0% | +50% |
ChatGPT grew to within a point of the same figure in all four markets. Claude, Gemini and Anthropic multiplied. And OpenAI and Perplexity fell in every market measured. The incumbent brand term in each bloc is the one losing ground — DeepSeek on one side, OpenAI on the other. Same shape, both sides.
What this means for the contract
The evidence does not support waiting. Western AI demand grew in all four markets year over year, with no break at the July 1 boundary. There is no demand-side cooling to wait for, and no mechanism here by which the selloff delivers the buyer a discount.
The sharper question is vendor selection, not exposure. A stack anchored to DeepSeek is anchored to a vendor whose own demand peaked in January 2025 and has fallen 70–87% since, while three Chinese competitors grew. That is a risk inside the Chinese bloc rather than a reason to leave it.
The scorecard
Sixteen priors were registered before retrieval. Ten were scoreable on this instrument; five held. Six remain unresolved, five of them because the second instrument was never run.
| # | Prior | Result | Note |
|---|---|---|---|
| A4 | A Chinese vendor term falls more than 40% from its peak | HIT | Three did — but from a January 2025 peak, not from July 2026. A hit for a reason unrelated to the registered cause. |
| A5 | Western demand is up, same month year over year | HIT | Up in all four markets. |
| D1 | The exposed move is not uniform across markets | HIT | — |
| D2 | Japan's exposed arm falls least | HIT | Thin on the arm totals, decisive year over year: Japan is the only market where Chinese AI demand grew. |
| F2 | The release effect is visible in the leader's own series | HIT | And uninformative. The term moved one band rung, on a term with two distinct values all year. See the limits. |
| A3 | The exposed-to-unexposed ratio falls after July 1 | MISS | Two markets up, two down. No direction. |
| F3 | That ratio sits below its first-half average | MISS | Same split. |
| F1 | At least three of twenty terms fail the contamination screen | MISS | Two did. |
| D3 | Category terms are usable in fewer than four markets | MISS | They were usable in all four and too small to measure in all four, which is a different thing and the prior did not say it. |
| C2 | Divergence is the surviving explanation | MISS | See below. |
And none of the four registered explanations was right. Contagion, containment, divergence and “nothing measurable” were all rejected. The surviving explanation — rotation between vendors inside each bloc — was never written down, because every candidate had been phrased as a total going up or down. A candidate set written in levels cannot return a composition change, which means every prior built on it could only be wrong, never right for the right reason.
Both scorecards are in the calibration ledger.
What this cannot tell you
One instrument. Stated at the top and repeated here because it is the limit that matters most. A second instrument sharing no operator, no collection method and no population was registered and not run. Until it is, this is a finding about Google Ads search volume.
Three releases sit inside the window. V4-Flash on July 31, 2026 · V4-Pro on August 13, 2026 · V4.1-Flash on September 10, 2026. All of them push the Chinese arm up and none of them are sentiment, which is why the levels in that arm are not read as demand for China and the comparison is made vendor by vendor instead.
The instrument has a floor, and the biggest term sits on it. Google bands and rounds its figures. Across all of 2026, deepseek took exactly 2 distinct values in the US — while glm ai took 6. The term carrying most of the Chinese arm is the least sensitive one in it.
The window ends 2026-08. The current month is incomplete and five of eighty series ended a month early, so every comparison here runs only over the window present in every series used. A missing month is missing, never zero.
Search volume measures interest, not purchase. A developer evaluating a model and a buyer signing a contract are not the same person, and this instrument cannot tell them apart.
How it was built
Two passes. The first seeded the term sets per market and produced no measurement; the second measured. The first pass broke the second one’s design before it ran, which is what it was for.
The original control arm was a category phrase rather than vendor names. It differed from the treatment on two axes at once and ran 130 to 259 times smaller depending on the market — no market could fill a ten-keyword row above 500 searches a month. It was rebuilt as Western vendor names, differing on exposure alone, before any measurement.
Terms were seeded from the data, never written in advance. Every vendor launched after 2022, so the 2019–2021 months arriving in the same request became the contamination screen at no extra cost: a term whose pre-2022 maximum exceeds a tenth of its 2026 peak is matching something else.
The brief, its priors, its falsifiers and both findings records were written before the data they describe. Every change made after registration is declared in the brief with the date and whether it was made before or after seeing data.
The next one
This study is re-run when the window extends past 2026-08, and the second instrument is run if its endpoint becomes reachable. Both outcomes publish, including the one where the replication disagrees.
This is one question of twenty
This study reads one rotation, on one surface, on one day. The AI visibility Decision Brief answers the six decisions the seat funding this work is already arguing about — twenty questions chosen by measured demand, across four assistants and the AI Overview, with what was predicted before anyone looked.
See what the brief covers$495 standard, $995 custom. Pre-made and delivered on payment. If the decision turns out not to be decidable from evidence, it is refunded.