Kalshi AI — Mispricing Audit

2026-07-14 · methodology v3.1 (conviction-weighted, EV-agnostic sizing; resolver-priced probabilities) · $1,000 book · 45-day horizon · picks count toward the public track record and are cleared for auto-copy
TL;DR. The AI category is efficiently priced this week. Of ~375 active AI-theme markets, 20 candidates survived to diligence, and after checking each against its actual resolver — the Ornn GPU-compute index (whose raw feed I pulled directly), the Arena leaderboards, and Datacurve's DeepSWE benchmark — exactly one pick cleared every screen: a small MEDIUM-conviction NO on RTX 5090 compute exceeding $0.75 by Friday ($69 at risk, 6.9% of capital). 93.1% of the book is cash, deliberately. Seventeen diligenced rejects are logged below and shadow-tracked to settlement.

1 · How this was researched

Mode: theme-match. The Kalshi mirror has no AI category, so I selected by keyword across all categories (trading_events.title, trading_markets.title, trading_markets.rules), searching: openai, anthropic, claude, fable, gpt, gemini, grok, xai, nvidia, deepseek, llama, mistral, agi, chatbot, llm, superintelligence, chatgpt, sam altman, "ai" (word-bounded), lmarena, artificialanalysis. That returned ~375 active markets closing within 45 days (by Aug 28), spanning five families: NVIDIA GPU compute-price ladders (Ornn index), Arena leaderboard markets, model-release-date markets, the Anthropic export-control saga, and macro-AI markets (IPO announcements, Challenger job-cut attribution).

Stage-1 mechanical screen (volume24h ≥ 1,500; spread ≤ 5¢; 5¢ < price < 95¢; no mention-markets; no already-published picks) left 8 survivors; I extended diligence to 12 near-misses (v24 500–1,500 or one leg of a diligenced event) so the reject log would be meaningful. Stage-2/2.5 diligence pulled the rules text (primary + secondary) for every candidate, profiled the named resolver, and — where the resolver publishes data — fetched the resolver's own feed directly: I queried Ornn's dashboard API (api.ornnai.com/api/gpu/<type>/index-history, 91 daily prints per GPU), the Arena text and Math leaderboards, the LM Code Arena (WebDev) board, and the Datacurve DeepSWE table. Fourteen-day Kalshi candlestick histories were pulled for every candidate to catch already-eaten edges.

Per v3, ev_pct/edge_cents are recorded for calibration but did not set position sizes; sizing is by conviction tier and entry band only. Existing live picks (19 tickers, including KXTOPMODEL-26JUL31-CLAUF and KXMATHAI-26JUL31-GEMI in this category) were excluded from the candidate set up front.

2 · Markets reviewed

TickerMarketYes bid/askv24Verdict
KXRTX5090WS-26JUL17-0.750RTX 5090 > $0.75 on Jul 174 / 72,550PICK — BUY NO @ 95
KXH200MS-26JUL-4.000H200 July avg > $4.0065 / 66162WATCHLIST (volume screen)
KXLLM1-26JUL31-AAnthropic top-ranked LLM Jul 3193 / 947,066PASS — fairly priced
KXTOPMODEL-26JUL31-CLAUTopus-4-6-thinking top model Jul 3175 / 805,204REJECT — resolver unobservable
KXTOPMODEL-26JUL20-CLAUTopus-4-6-thinking top model Jul 2079 / 837,421REJECT — resolver unobservable
KXTECHRANKLISTAICODE-26JUL20-CHATChatGPT #1 Code Arena Jul 2059 / 623,263REJECT — 1-Elo coin flip
KXTECHRANKLISTAICODE-26JUL20-CLAUClaude #1 Code Arena Jul 2035 / 361,942REJECT — coin-flip band
KXCODEAI-26JUL31-CHATChatGPT #1 DeepSWE Jul 3179 / 85653REJECT — edge eaten Jul 10
KXMATHAI-26JUL31-CLAUClaude #1 Math Arena Jul 3183 / 85694REJECT — no edge, unstable
KXGEMINI-GEMI35P-26JUL24Gemini 3.5 Pro before Jul 2434 / 351,433REJECT — coin-flip band
KXGEMINI-GEMI35P-26JUL31Gemini 3.5 Pro before Jul 3170 / 74339REJECT — negative edge
KXRTX5090MS-26JUL-0.500RTX 5090 July avg > $0.5079 / 84524PASS — fairly priced
KXRTX5090WS-26JUL17-0.500RTX 5090 > $0.50 on Jul 1788 / 891,056PASS — fairly priced
KXB200MS-26JUL-6.00B200 July avg > $6.0076 / 866REJECT — 10¢ spread
KXB200WS-26JUL17-7.000B200 > $7.00 on Jul 1741 / 42218REJECT — coin-flip band
KXA100WS-26JUL24-1.500A100 > $1.50 on Jul 241 / 41,566REJECT — overpriced tail
KXANTHROPICRESCIND-26JUN-26AUG01DoD rescinds Anthropic designation before Aug 120 / 270REJECT — illiquid
KXCHAICUTS-26AUG06-T1AI #1 job-cut reason (July Challenger)43 / 480REJECT — illiquid
KXLLM1-26JUL20-ANTHAnthropic best AI, week of Jul 2098 / 9918,049SCREEN-CUT — fully priced
KXFABLEDISABLED-27-26JUL31Fable 5 US access disabled again before Jul 311 / 4718SCREEN-CUT — fully priced

3 · The pick

Pick 1 — KXRTX5090WS-26JUL17-0.750 · BUY NO @ 0.95 limit · MEDIUM CONVICTION

"Will the NVIDIA RTX 5090 compute per hour price be above $0.75 at 4 PM ET on Jul 17?" — closes 2026-07-17 20:00 UTC (3 days).

Current market
YES 4 / 7 → NO ≈ 93–96; last YES trade 7¢
My probability (NO)
98% vs market-implied ≈ 95–96% at my entry
Conviction / band
MEDIUM · deep-favorite band (>90¢ effective entry)
Size
73 contracts @ 95¢ = $69.35 at risk (6.9% of capital; MEDIUM cap is 7%)
Numbers (record only)
edge +3¢/contract · EV +3.2% · max payout $73.00 · EV ≈ +$2.19. These did not set the size — the tier cap did.
Cluster
ornn-gpu-index

Resolver profile (Stage 2.5). Resolves solely on the value of NVIDIA RTX 5090 compute per hour as reported by Ornn (dashboard.ornnai.com), USD index, at 4 PM ET Jul 17; rules: "If no data is published… the most recently available published data will be used" — a missing print resolves off the last value ($0.51), which favors NO. Priced risks: −1.5 pts complex-wide melt-up contagion (see tail risk), −0.5 pts feed methodology change/revision before expiry. Evidence-based ~100% → model 98%.

Thesis. I pulled the resolver's own feed (91 daily prints). RTX 5090 is at $0.51 (Jul 13) and has printed 0.50–0.51 for eight consecutive days; July's range is 0.50–0.53. YES needs +48% in 3 days from a standing start. The only comparable move in the dataset (+55% over 3 days, May 28–30) came at the climax of a two-week parabola that had already doubled the index — there is no precedent for this move out of a flat regime. The current datacenter squeeze (B200 +35% in two weeks, H200 +13%) has left consumer-card compute untouched.

Evidence.

Tail risks (cleanest way to lose). The B200 shortage cascades down-stack and retail renters bid up RTX 5090 this week. Mitigant: in the May episode the RTX 5090 spike lagged H200's peak by ~13 days; a repeat lands after Jul 17 expiry. Second tail: an Ornn methodology change mid-week (unpriceable beyond the −0.5 pt deduction; revisions after expiration are excluded by rule).

Liquidity / entry context (recorded at pick time). YES book: bid 4¢ × 1, 3¢ × 5, 2¢ × 84; NO resting bids 94¢ × 163, 92¢ × 693. Contracts fillable at or below my 95¢ NO limit right now: 0 — this is a maker order (a 95¢ NO bid = offering YES at 5¢ into a market that traded 2,550 contracts at 4–8¢ today). v24 = 2,550; OI = 6,758. Execution: rest at 95¢, do not cross above 96¢; cancel any unfilled remainder at Jul 16 close.

Top watchlist (not a pick) — KXH200MS-26JUL-4.000 · would-be BUY YES @ 66 · MEDIUM

The largest modeled edge on the board, cut by the Stage-1 volume screen (v24 = 162), logged as a shadow-tracked reject. "H200 July average > $4.00": 13 of 31 days are already published averaging 4.151, so NO requires the remaining 18 days to average below $3.891 — an immediate, sustained ~11% drop from today's $4.39 against a rising tape. My model: 82% vs 66¢ ask (+16¢). It is not free money — H200 crashed −45% in ten days after the May 17 spike, and June averaged ~3.6 — but the risk-reward at 66¢ is real. Orderbook depth at pick time: 218 bid @ 65¢ / 264 offered @ 66¢ (tight 1¢ spread despite thin volume). Trigger: if v24 clears 1,500 while YES ≤ 70¢, this becomes a MEDIUM pick in the next run.

4 · Recommended $1,000 portfolio

#TickerActionLimitContractsCostMax payoutConvictionBandCluster
1KXRTX5090WS-26JUL17-0.750BUY NO95¢73$69.35$73.00MEDIUM>90¢ deep favoriteornn-gpu-index
Cash held$930.65 (93.1%)Cash is a position; the category is efficient this week.

Cluster exposure (cap: 15% of capital per cluster)

ClusterCost% of capitalCapStatus
ornn-gpu-index$69.356.9%15% ($150)OK

All Ornn-resolved GPU markets are treated as one cluster despite partially offsetting payoffs, because they share a single resolver (feed outage/methodology tail). Had the H200 watchlist item been deployed, the combined cluster would have been trimmed to the $150 cap.

Conviction exposure

TierPer-pick capDeployedPicks
HIGH$150$00
MEDIUM$70$69.351
LOW$30$00

Risk profile

Execution notes

5 · What I rejected and why

Every Stage-2 survivor below is machine-logged in picks.json and shadow-tracked to settlement — if the rejects outperform the pick, the screen itself is the finding. Verdicts against the resolver, not the headline.

Cut by the Stage-3.5 entry-band screen (the v3 discipline, visible)

Cut on resolver risk (Stage 2.5)

Cut as fairly priced (no edge to defend)

Cut on liquidity / spread / already-moved

6 · Sources

Data: Kalshi read-only mirror + Kalshi public trade API (prices in cents; books as of 2026-07-14 ~15:05 UTC) + resolver feeds cited above. All probabilities are subjective estimates; binary contracts resolve to $0 or $1 and the full cost of any contract can be lost. Rejects listed here are machine-logged and shadow-tracked to settlement as a test of the selection screen. This report is research, not financial advice. Methodology v3.1: conviction-tier and entry-band sizing; EV recorded but never used for sizing; 15% per-cluster cap; resolver risks priced as explicit point deductions.