The study covers 7–13 September 2026 (UTC). The market-cap-weighted return of the Large Language Models segment was +21.82%, yet only 3 of its 12 assets rose. The index-level result does not describe the typical asset in the sample.
Venice Token (VVV) held 85.50% of the segment's starting market capitalization and gained 26.25%. Its contribution was +22.44 percentage points, exceeding the segment's overall +21.82% because the other 11 coins contributed −0.63 points in total. Excluding VVV, their own start-cap-weighted return was −4.31%. This demonstrates concentration, not a proven cause of the move.
LLM: performance, breadth, and concentration
- Weighted return: +21.82%. Median return of a typical asset in this sample: −2.31%.
- Breadth: 3 of 12 gained, or 25%. The other 9 declined.
- Top-1 weight: VVV held 85.50% of starting market capitalization. Top-1 contribution: +22.44 percentage points.
- Ex-Top1 return: −4.31% for the other 11 coins using their own starting weights.
The weighted measure tracks the start-cap-adjusted aggregate. Breadth counts gainers, the median describes the typical asset in the sample, and the leader's contribution measures dependence on one coin.
Segment membership and underlying data
The LLM membership was frozen in evidence pack 2da8cff8e757ade9, generated on 18 September 2026 at 23:56 UTC. All 12 eligible assets appear below. The first column gives the ticker and CMC ID. Weight uses starting market capitalization; contribution equals weight × return. Display values are rounded; the pack retains full precision.
| Asset | Start market cap, $ | Weight, % | Return, % | Contribution, pp |
|---|---|---|---|---|
| VVV (35509) | 818,249,427.30 | 85.50 | +26.25 | +22.442 |
| KAITO (35763) | 75,635,703.50 | 7.90 | −3.40 | −0.268 |
| CGPT (23756) | 20,133,603.05 | 2.10 | −7.83 | −0.165 |
| MIRA (38495) | 15,779,524.24 | 1.65 | −0.32 | −0.005 |
| ASM (6069) | 7,933,202.19 | 0.83 | −1.23 | −0.010 |
| ALU (9637) | 5,834,923.70 | 0.61 | −17.56 | −0.107 |
| PAAL (27178) | 5,797,715.37 | 0.61 | −8.32 | −0.050 |
| PALM (28567) | 2,490,414.76 | 0.26 | −4.32 | −0.011 |
| HGPT (25419) | 1,989,739.81 | 0.21 | −12.92 | −0.027 |
| HEART (15584) | 1,853,663.92 | 0.19 | −1.00 | −0.002 |
| ARC (13051) | 863,626.97 | 0.09 | +9.94 | +0.009 |
| VAI (8479) | 504,355.07 | 0.05 | +21.60 | +0.011 |
| Total | 957,065,899.88 | 100.00 | +21.82 weighted | +21.816 |
The full sample: performance and breadth
Of 166 published coins with sufficient history, 33 rose. The −5.31% median describes a typical asset in this full sample without capitalization weighting. It was below the −3.53% weighted return. This is a different sample from the LLM segment, and its figures must not be substituted for LLM metrics.
Weighted return, median return, and the share of advancing coins use one scale here, but they measure different properties and should be read together.
AI outperformed BTC, but this was not a broad rally
AI-sector capitalization fell 2.85%, compared with a 3.86% decline for BTC and 3.51% for the total market. AI therefore outperformed BTC by 1.01 percentage points and the total market by 0.66 points.
This compares percentage changes in different benchmarks: AI-sector capitalization, BTC price, and total-market capitalization. It indicates relative strength, not an investable portfolio return or a proven cause. At the same time, 133 of the 166 coins declined.
The start of the week is rebased to 100. The chart compares relative paths, not absolute market sizes.
Other anomalies this week
Separate from the main LLM finding, Generative AI returned a weighted +17.33% while only 3 of 18 coins rose. VVV held 72.39% of the starting weight and contributed +19.00 percentage points. The other 17 coins returned −6.07% without it.
Segments overlap: one coin may belong to several groups. This chart does not show breadth or explain causes.
The strongest weekly regime change occurred in ZKML & Cryptographic AI. Its weighted return was +34.52% in the previous week and then fell to −15.32%. The change between the two periods was −49.84 percentage points.
All four assets in the segment declined: FLock.io (FLOCK) −10.84%, Nillion (NIL) −21.51%, Mind Network (FHE) −18.38%, and Marlin (POND) −9.20%. These figures do not prove a shared cause, but they show that the prior acceleration failed to broaden into the following week.
What to watch next week
- whether the share of advancing coins recovers toward at least half of the sample
- whether median and weighted returns begin to converge
- whether AI keeps outperforming BTC without relying on one large asset
- whether LLM and Generative AI breadth improves after removing the capitalization leader
- whether the ZKML segment continues lower after its sharp reversal
Method and limitations
The study covers 7–13 September 2026 (UTC). Coin returns use the 6 and 13 September daily closes as boundary observations for that week. Segment membership and the published sample were frozen in the pack generated on 18 September at 23:56 UTC. This is a saved membership version, not a claim that taxonomy labels were reconstructed as of 7 September.
The 12 LLM coins use CoinMarketCap daily snapshots (cmc_latest in the pack); BTC and market comparisons use internal market snapshots (internal_rollup). See the Data Catalog for source details. Coin return = (end price / start price − 1) × 100. Weighted return = sum(start market cap × return) / sum(start market caps). Breadth is the share of positive returns. Median is the midpoint of the ordered return series. Ex-Top1 recalculates after excluding the largest starting-cap asset. Differences versus BTC and the total market are in percentage points. Taxonomy segments overlap and cannot be added together.
Source data pack: 2da8cff8e757ade9.