AI Market Intelligence Daily

1. Weekly share markets repriced sharply as Xiaomi cemented its grip on Chinese AI usage. The Xiaomi "above 1.9% this week" contract on Kalshi climbed to 99%, its "above 2.1%" sibling to 96%, and the "top Chinese AI company this month" market now prices Xiaomi at 91% — while rivals collapsed in the same window: DeepSeek's "above 24.1% share" contract fell from 92% to 60%, Qwen's "above 6.7%" from 17% to 1%, and Tencent's "above 3.5%" from 5% to 1%. That combination of one winner consolidating while every competitor's thresholds reprice downward is the signature of a usage-share shift, not a data hiccup, and it points to Xiaomi's assistant-driven phone channel taking meaningful volume in a week when no major model launch made the headlines.

2. Usage volume on the main aggregation routes is running far ahead of where traders expected at the start of the week. Kalshi's OpenRouter contracts for the September 21–27 window moved violently: "above 146T tokens" surged from 8% to 65% and "above 144T" from 23% to 84%, while the top of the ladder ("above 160T") still trades at just 2%, implying the week is tracking toward roughly 150–160 trillion tokens. The cleanest corroborating headline is OpenAI's own Vercel AI Gateway market, which jumped from 32% to 99% on the "above 21% of top-lab spend" threshold — gateway routing through OpenAI models is decisively above that bar. The Moonshot "above 18% of AI Gateway spend" contract, the day's single largest move at 900% (7% to 70%), sits in the same cluster: the gateway-spend series is settling into a new equilibrium, and Moonshot's jump from near-nothing to 70% suggests a sudden, large routing change on its side worth watching in tomorrow's settlement.

3. OpenAI's near-term product and exit markets both ticked up, even as the news cycle stayed negative for the company. "Will OpenAI release a personal AI agent before Oct 1, 2026?" jumped from 13% to 72% (with the 2027 variant at 83%), and "OpenAI IPO by December 31, 2026" crept up from 3% to 4% — but that IPO probability still reads as effectively dead, which is consistent with the day's actual headlines: TechCrunch reported that unsecured OpenAI agents leaked 53 user images to the internet without the lab's knowledge, and that OpenAI's agent swarms have been hammering online databases for months. That is a genuine price-sentiment divergence: the market is pricing an imminent personal-agent release while the press is running agent-security failure stories, so traders are apparently betting the product ships on schedule despite the operational stumbles.

4. The big AI-lab money stories landed quietly in the market prices. Anthropic signed an $11.6 billion, seven-year cloud deal with Akamai and its founders are seeking voting control ahead of an IPO, yet the "Anthropic IPO before 2027" market actually eased from 78% to 74% and Anthropic's weekly share contracts all drifted down (the "above 2.8%" contract fell from 39% to 15%). The market is reading the mega-deal as a cash-flow and margin story rather than an acceleration signal, and the governance news has done nothing to move the exit odds. On the capital side more broadly, Nscale's $3.36 billion pre-IPO convertible and Crusoe's abandonment of its $1.25 billion Boom turbine plan are both infrastructure stories that found little expression in the monitored markets — the "AI bubble burst in 2026" contract held at 11%, so neither the capex retreat nor the new neocloud financing is changing traders' bubble math.

5. The model-leaderboard markets resolved into a Claude-dominated picture as the week closed. Claude holds 74% on "Best AI on Dec 31, 2026" and 93% on "Top Coding AI this week," while the Sep 28 top-model market that had priced claude-fable-5.1-max at 88% crashed to 4% — the field is dispersing across the remaining candidates rather than a rival surging, which typically means the week's ranking data simply didn't confirm a challenger. That dispersion is the one place where price movement likely reflects settlement mechanics and thin late-week trading more than new information: with the week of September 21–28 closing in two days, the leaderboard series is repricing toward its data, and the day's 98 price movers are dominated by these threshold ladders snapping to their final levels rather than by any single news event.