AI Market Intelligence Daily

1. The GPU compute markets on Kalshi repriced sharply higher across the board, and the move lines up cleanly with this week's supply-side headlines. The "H100 SXM price up in next week?" contract jumped from 16% to 97%, the H100-above-$2.75-today leg went from 29% to 99%, the above-$3 leg more than quintupled from 2% to 18%, and A100 strikes climbed in the same direction (above $1.23 on Aug 31 from 4% to 35%, above $1.50 on Sep 4 from 1% to 7%). This matches the news flow: Amazon reportedly tripled its order of Nvidia chips on surging demand, Anthropic signed a $45B compute deal with Nscale, and Nvidia is closing in on a Hugging Face acquisition that would extend its reach into model distribution. The market is pricing a genuine tightening of the older H100/A100 fleets, not a broad uniform shortage.

2. The Anthropic Mythos release calendar moved in one direction, and the internal structure now looks consistent. The "Mythos before Oct 1, 2026" contract surged 500% from 2% to 12%, the Dec 1 leg doubled from 11% to 22%, and the Nov 1 leg rose from 6% to 14%, all while the Sep 1 leg sat at just 5% and the Jan 1, 2027 contract held at 13%. That ordering (Sep < Oct < Nov < Dec) is coherent, and it suggests traders are shifting the expected release window toward late fall rather than doubting a release at all. The timing coincides with reports that Anthropic's August risk report describes an unreleased internal model that is already more capable than Mythos 5, which makes a compressed Q4 release window plausible.

3. The "best AI on December 31, 2026" ladder on Kalshi is consolidating around Anthropic, and the move is not headline-driven. Claude holds 68% of the year-end best-model odds while ChatGPT sits at 16%, Grok at 9%, and Qwen at 1% (up 17% relatively from 1%, still negligible). None of today's 15 news stories is a benchmark release, which makes the slow drift toward Claude look like private information or positioning rather than a news reaction. It is consistent with the Mythos term-structure move above and with the "top AI model this month" contract for claude-opus-5-max, which nearly tripled from 39% to 91% in 24 hours while the claude-opus-5-high leg collapsed from 44% to 6% — traders are concentrating on a single Claude variant rather than the family.

4. OpenAI's markets moved in a way that does not line up cleanly with its news coverage, which is worth flagging as a divergence. On Polymarket, "Will OpenAI IPO by December 31, 2026?" fell 9% from 16% to 14% and "GPT-6 by December 31" rose only 3 points to 65%, yet on Kalshi the "GPT-6 before Oct 1" leg rose 22% from 18% to 22%. The OpenAI news flow this week is mostly about ads on ChatGPT's free tiers in India, a Meta executive leaving for OpenAI, and its appearance at TechCrunch Disrupt — nothing that should move IPO or model-release odds materially. The modest uptick in the October GPT-6 leg against a flat-to-down December contract on the other venue suggests the two markets are being traded by different books rather than reacting to a shared signal.

5. The usage-share and token-throughput markets point to the same underlying story as the compute repricing: demand is outrunning supply, but the mix is shifting. The OpenRouter total-tokens contract above 122T fell 82% from 73% to 13%, meaning traders expect August's cumulative token volume to come in well below the high end of the range, while OpenAI's "above 15.5% share" contract sits at 99% and the "above 19.5%" leg is down to 19% (from 21%). DeepSeek's and Tencent's weekly share legs also slipped. The most coherent read is that total AI inference volume is growing but growth is decelerating at the margin, which is exactly the regime in which older H100/A100 capacity tightens and prices rise — the same regime the GPU contracts are pricing in.