From 19 August to 7 October, Bitcoin gained 30.87%, yet most tokens failed to beat it. We compared 155 projects, separated them into subsectors, reconstructed product announcements and exchange listings, and checked what happened before and after the largest moves. The evidence tells a more selective story than a broad “AI season.”
A glance at the largest tokens could suggest that the whole sector joined the same rally. NEAR more than tripled from our starting point, FET and VVV comfortably beat Bitcoin, and several small projects returned hundreds of percent.
That impression does not survive a closer look.
We used 19 August 2026 as a common starting point for every AI Coin Map project with a usable price history. Of 166 published projects, 155 qualified for the main ranking.
Bitcoin returned 30.87%. The typical token in the sample, measured by the median, returned only 18.62%. Just 69 of 155 tokens beat Bitcoin, or 44.5% of the sample. Another 44 finished the period with a negative return.
Our central finding: the sector rose, but there was no uniform rally. Prices repriced individual projects and parts of the infrastructure stack very differently.
The weighted basket rallied, while the typical token lagged Bitcoin
The gap between the median and the market-cap-weighted result makes that selectivity clear. For 116 projects with usable starting capitalization, excluding questionable supply-proxy values, the fixed-weight price index gained 89.22%. The median across all 155 ranked tokens was just +18.62%.
These are different samples. Within the same 116-asset weighted basket, the median was +36.04%, 75.86% of assets rose, and 51.72% beat Bitcoin.
TAO had the largest starting weight: Top-1 Weight was 20.20% and Top-1 Contribution was +11.17 percentage points. Removing TAO and recalculating the weights of the remaining 115 assets produced an Ex-Top1 Return of +97.81%. This basket's rise was therefore not simply the result of its largest constituent.
Strong returns were unevenly distributed. A number of large assets and several exceptionally strong groups lifted the weighted result.
Only 69 of 155 AI tokens beat Bitcoin. The chart shows the top 20 and bottom 15 returns. The middle 120 are omitted. The sample contains currently published projects and excludes 11 invalid histories.
“The sector went up” tells us little about the return of any particular holding. Getting the broad narrative right could still have meant underperforming Bitcoin.
The winners: NEAR's 215% gain and even larger small-cap moves
Small projects led the absolute return ranking. NEURAL gained 363.20%, LNQ 329.88%, CLORE 323.10%, AGRS 307.30%, ARC 277.08%, and NATIX 268.22%.
One of the most unusual results belonged to a much larger asset: NEAR gained 215.35%.
With an ending market capitalization of approximately $6.55 billion, NEAR ranked ninth among all 155 tokens and beat Bitcoin by 184.48 percentage points. That is a small-cap-style return on a multibillion-dollar starting valuation.
Size alone does not capture the risk either. Despite its 363% gain, NEURAL suffered a sampled maximum drawdown of roughly 60%. AGRS ended with reported 24-hour volume of about $1,500, and ARC with about $2,700.
A 215% rise in NEAR and a 300% rise in a tiny token are not interchangeable market outcomes.
The surprise: agents were not the strongest subsector
Autonomous agents attracted considerable attention in 2025–2026. It would have been easy to expect that category to lead the next rally. Our sample points elsewhere.
| Subsector | Median return |
|---|---|
| ZKML & Cryptographic AI | +115.49% |
| Large Language Models | +86.38% |
| Decentralized Compute / GPU | +71.91% |
| Inference / Model Serving | +60.48% |
| On-chain / Verifiable AI | +55.29% |
| Model Training | +38.82% |
| Privacy / Secure AI | +38.71% |
| Federated Learning | +28.93% |
| Data Marketplaces | +20.20% |
| AI Agents & Autonomous Platforms | +12.77% |
| Agentic AI | +12.28% |
The categories overlap, and the ZKML group contains only four projects. Their results cannot be added together, and a tiny sample should not be treated as a representative investable index.
Even with those limitations, the contrast is substantial. In decentralized compute / GPU, 24 of 30 projects beat Bitcoin, or 80%. In AI Agents & Autonomous Platforms, the share was just 31.48%.
Subsector returns differ sharply. ZKML median excess versus BTC was +84.62 percentage points, but the sample contains only four projects. Categories overlap. Membership was frozen at extraction, not reconstructed for the starting date.
While agents dominated much of the conversation, stronger typical returns came from compute, model serving, LLMs, privacy, and verifiable AI.
This is evidence of relative price performance, not a map of capital flows. We measured prices, not transfers between wallets.
Rotation was a sequence, not a single move
Breaking the period into weekly intervals reveals another layer. From 19 to 26 August, LLMs, compute, and inference led among the larger categories. On 2–9 September, LLMs and inference led again, with model-training networks moving closer.
The sharpest acceleration came on 16–23 September:
- Inference / Model Serving: median +39.37%.
- Model Training: +29.00%.
- LLMs: +28.48%.
The following week, leadership shifted toward on-chain / verifiable AI and decentralized compute.
No subsector led throughout the period. The heatmap shows median excess returns versus BTC across successive intervals. The final interval spans 7.25 days, versus seven days for the others. Categories overlap.
“Rotation” is useful here, provided we are clear about what it means. We did not observe money being transferred directly from agents to compute. We observed changing relative strength across overlapping groups.
No subsector remained the best performer throughout the seven intervals.
NEAR: a multibillion-dollar token gained 215%
NEAR deserves a closer look. Since 19 August:
- Return: +215.35%.
- Excess return versus BTC: +184.48 percentage points.
- Ending market cap: approximately $6.55 billion.
- Sampled maximum drawdown: approximately −12.64%.
NEAR was the result that first pushed us to investigate the events behind this rotation. During the study window, the project had developments in confidential inference, NEAR AI Cloud, private AI integrations, and OpenRouter.
At 17:24:17 UTC on 16 September, the nearcore 2.14.0-rc.1 pre-release appeared with post-quantum cryptography components. A pre-release is not evidence that those components were activated on mainnet.
A tempting explanation would be: “NEAR announced quantum protection, and the market bought NEAR.” The chronology is less tidy.
In the three days before the exact release timestamp, NEAR was already outperforming BTC by 6.67 percentage points. Relative performance after the release was:
- One day: +20.82 percentage points versus BTC.
- Three days: +39.12 percentage points.
- Seven days: +65.16 percentage points.
NEAR was already rising before nearcore 2.14.0-rc.1. The pre-release coincided with a faster trend, not an established causal effect. This pre-release does not establish mainnet activation. The median is not a tradable index. BTC has a missing 28 September anchor, with no interpolation.
The post-quantum pre-release did not start the entire NEAR move. It coincided with a sharp acceleration of an existing trend.
Two days later came the NEAR AI Cloud integration with SayGm. Before our main window, NEAR had already introduced staking for AI compute and developed confidential AI.
Rather than one decisive catalyst, there were several parallel developments: infrastructure, privacy, agents, token utility, Intents, and protocol engineering. Our data cannot assign a share of NEAR's 215% gain to each one.
TAO did not lose, but it was no longer the only obvious sector bet
Against NEAR's result, it is easy to say that Bittensor / TAO fell behind. That misses an important distinction.
TAO gained 55.29%, beating Bitcoin by 24.42 percentage points. It outperformed the market benchmark used in this study.
What changed was the comparison set: NEAR gained 215%, VVV 91%, FET 89%, and PHA 225%, while dozens of smaller tokens doubled or more. In that setting, a 55% gain can appear unremarkable.
One hypothesis is that Bittensor is becoming harder to value as a single token. Dynamic TAO and subnet economics create additional ways to allocate capital within the system.
We cannot verify the claim that capital moved from TAO into subnet tokens. That would require historical alpha-token prices, AMM reserves, staking flows, and subnet emissions.
The narrower conclusion is supported: TAO remained a strong performer, but it was no longer the only obvious way to gain exposure to the sector.
ICP, FET, and RENDER did not move as one group
Other large tokens also beat Bitcoin, with different degrees of strength:
ICP had a confirmed DFINITY–UNDP partnership involving sovereign cloud and decentralized AI pilots. We did not find a persistent positive price effect versus BTC immediately after that announcement.
RENDER fits the stronger compute / GPU group, but our data cannot separate new AI workload from broader Render Network demand.
FET became particularly interesting once we added exchange events to the analysis.
A listing is not a universal explanation for a rally
We separately reconstructed exchange events from 1 August to 7 October. The search found 13 confirmed clusters of new market access, including spot, perpetual futures, and other futures products.
Those events did not explain most of the rally.
For NEAR, ICP, RENDER, PHA, POND, RLC, AIOZ, SENT, NEURAL, LNQ, AGRS, NATIX, PIN, and other strong performers, we found no confirmed new major exchange catalyst that explained the entire trajectory.
There were counterexamples too. CLORE gained roughly 323% even though HTX removed its spot trading during the period. DSYNC gained roughly 179% while MEXC closed its perpetual contract.
Exchange access matters. “New listing → higher price” does not work as a general causal model.
VVV: the first wave came before OKX spot
Venice Token / VVV returned 91.30% over the full period, beating BTC by 60.43 percentage points.
Venice combines an AI product and API with token burns and a scheduled reduction in emissions. But we could not test “more users → more revenue → higher token price,” because a comparable daily usage and revenue series for August–September was unavailable.
The sequence we could verify was different. From 2 to 9 September, VVV gained 57.97% while Bitcoin gained only 2.28%. OKX scheduled VVV/USDT spot trading for 15 September at 14:00 UTC.
VVV subsequently outperformed BTC by approximately 14.39 percentage points over three days and 28.67 over seven days.
OKX expanded access to an asset that had already risen sharply. Better market depth remains a hypothesis, because the study has no historical order-book series. The first wave preceded the listing.
FLOCK: a small-cap case with two very different OKX outcomes
We initially approached FLock.io / FLOCK as an interesting small-cap research candidate. Its performance was strong, but it was not the top-ranked winner:
- Return: +121.16%.
- Excess return versus BTC: +90.28 percentage points.
- Overall rank: 18th.
- Low-cap rank: 13th.
- Ending market capitalization: approximately $30.1 million.
FLock had developments involving Red Hat / Open Cluster Management, research, FOMO, model tokens, and new models. Neither the Red Hat announcement nor the research publications adequately explained the price path.
The exchange chronology added a useful contrast. From 2 to 9 September, FLOCK gained 87.27% versus BTC's 2.28%, an advantage of 85.00 percentage points.
OKX scheduled the ordinary FLOCK/USDT perpetual contract for 12 September at 10:00 UTC. The claim that this contract caused the entire rally fails the chronology test: FLOCK was already rising sharply beforehand.
After the perpetual launch, FLOCK's excess return versus BTC was:
- One day: +38.00 percentage points.
- Three days: +15.72 percentage points.
- Seven days: +14.15 percentage points.
OKX then scheduled FLOCK X-Perp for 23 September at 08:00 UTC. This is an expiring futures product, not a second ordinary perpetual contract. The subsequent relative returns had the opposite sign:
- Three days: −8.21 percentage points versus BTC.
- Seven days: −20.69 percentage points.
FLOCK's strong 16–23 September interval also ended eight hours before the scheduled X-Perp opening.
FLOCK gained 87.27% on 2–9 September, before the 12 September OKX perpetual opening. Relative performance weakened after X-Perp. X-Perp is an expiring futures product, not an ordinary perpetual. Product and exchange events overlap. BTC has a missing 28 September anchor.
This is a useful example because it separates an appealing market story from an observed sequence: the token price rises, a major venue expands access, and relative performance changes again after a contract opens.
That sequence does not establish causality. Without trade-level evidence, we cannot label the different outcomes as stronger demand or profit-taking.
FET: derivatives access arrived closer to the start of a stronger move
FET's chronology was different. OKX scheduled its ordinary FET perpetual contract for 5 September at 03:30 UTC. Relative performance before that point was more modest. Over the following three days, FET beat BTC by approximately 16.94 percentage points.
FET X-Perp followed two days later, so the two events cannot be treated as independent experiments. Even so, this was one of the few cases in which new derivatives access coincided with the early part of a stronger move.
That does not prove a listing effect. It does explain why exchange chronology belongs in the analysis.
The biggest percentages were often the hardest to explain
A return ranking alone makes NEURAL, LNQ, CLORE, AGRS, and ARC look like the main winners. Each gained hundreds of percent. Yet these were also cases where headline performance and market quality diverged most visibly.
NEURAL
- Return: +363.20%.
- Sampled drawdown: approximately −60%.
- Ending reported 24-hour volume: approximately $74,000.
AGRS
- Return: +307.30%.
- Ending reported 24-hour volume: approximately $1,500.
ARC
- Return: +277.08%.
- Ending reported 24-hour volume: approximately $2,700.
For many such moves, we could not reconstruct a major confirmed product or exchange catalyst. That does not establish manipulation. It means the percentages should not be read as equivalent evidence.
A 300% gain in a deep market and a 300% gain with a few thousand dollars of reported daily turnover carry different execution risks.
NEURAL, AGRS, and ARC show large gains alongside low reported volume. Returns, size, and execution liquidity are different properties. Volume is measured at the endpoint, after the rally. Possible supply-proxy capitalization is flagged in the data and does not establish circulating float.
The small-cap ranking is a list for further investigation, not a list of the “best coins.”
Product progress did not guarantee a rising token
The sample also contained counterexamples on the product side. Allora released:
Yet ALLO fell 16.68% across the full period and lagged BTC by almost 48 percentage points.
Acurast continued engineering work, published technical updates, and introduced AI demos and API access. Its ACU token barely participated in the broader rally.
Project development and token returns are not the same thing. Supply, unlocks, liquidity, token economics, exchange access, existing valuation, speculative demand, and whether usage creates demand for the token can all matter.
What did the market reward?
We found no single announcement, listing type, or universal metric that explained the winners. Several patterns were visible:
- Compute, inference, LLMs, and verifiable AI had substantially stronger typical returns than the broad agents category.
- Large assets could also produce exceptional relative strength. NEAR was the clearest example.
- Listings often followed the first strong move, particularly in FLOCK and VVV.
- Extreme small-cap returns alone said little about project quality or execution conditions.
- Technical releases did not automatically lead to token repricing.
What to watch in the next phase
If these performance patterns persist, useful research signals will require more than “AI” in a project's description. Watch for several forms of evidence at the same time:
Actual usage
Inference requests, GPU workloads, active agents, data usage, and API calls.
A link between usage and the token
Burns, fees, staking, required settlement, and a clear mechanism for value capture.
Wider market access
Major spot listings, perpetual contracts, and Korean markets, with a check for whether the price move began beforehand.
Supply
Emissions, unlocks, buybacks, burns, and changes in circulating supply.
Liquidity
Not just 24-hour volume, but order-book depth, spreads, and the ability to execute a meaningful position.
Relative strength
A token can rise 20% and still underperform if Bitcoin rises 40% over the same interval.
Small caps: a research question, not a promise of the next 10×
FLOCK is one of the more interesting small-cap cases in this study, but not the only one. NATIX, LNQ, CLORE, PIN, ORAI, and other compute or data projects also merit investigation.
PHA belongs in a separate discussion: its $59.84 million ending capitalization exceeded this sample's low-cap threshold.
Potential upside and uncertainty coexist here. Some leaders have already multiplied in price, some have weak liquidity, and some still have no established explanation for their move.
The useful question is not “Which token will do the next 10×?” It is “Which smaller project already shows the characteristics associated with larger outperformers, without a comparable valuation?” That question still requires evidence rather than a list of supposed gems.
The conclusion: one narrative, very different price outcomes
From 19 August to 7 October, the surface story was straightforward: Bitcoin rose, interest in the sector strengthened, and many tokens rallied.
Underneath, most tokens failed to beat Bitcoin. Subsector leadership changed across intervals. Compute, inference, LLMs, and verifiable AI were stronger than the broad agents group. NEAR demonstrated exceptional relative performance at multibillion-dollar scale.
TAO did not fail, but more alternatives appeared alongside it. VVV and FLOCK had their first strong moves before the major listings we identified. Some of the most spectacular small-cap winners were also the hardest to explain fundamentally.
The central result is not a recommendation for one token. It is that the market did not trade as a single bet on artificial intelligence. Compute, models, inference, privacy, data, and agents had markedly different price outcomes.
Further opportunities may emerge where that repricing is incomplete. A project's name or an AI label is not enough to identify them.
How we calculated the results
The study runs from 19 August 2026 at 00:00 UTC to 7 October 2026 at 06:00 UTC.
Prices come from CMC snapshots in the existing server-side history. If an intraday starting observation was unavailable, the preceding day's exchange UTC close was used. The anchor tolerance is 30 minutes. Actual observation timestamps are retained in the download, and ending observations are no later than 06:00 UTC.
Return = 100 × (ending price / starting price − 1). Excess return versus BTC = token return − BTC return over the same interval, expressed in percentage points. It is not a formal estimate of financial alpha.
The weighted return is the sum of individual returns multiplied by their shares of starting capitalization. Weights remain fixed, with no rebalancing. The median describes the typical token. Breadth is the share with a return above zero: 111 of 155, or 71.61%. Beating-BTC breadth uses a positive return difference versus BTC instead.
The median is the main subsector comparison because extreme individual gains have less influence on it. The groups overlap.
Low-cap means an ending capitalization below $49.14 million, the 75th percentile of positive capitalizations in the initial export. This is a sample-relative threshold, not a universal classification.
Membership and taxonomy were frozen on 7 October, not reconstructed for 19 August. The study therefore has a current-project selection limitation. Of 166 cards, three were excluded for missing starting anchors and eight for price jumps with zero reported volume that require validation.
The exchange search is not a complete catalog of every venue. Trading times are official scheduled openings, and the first actual trade was not independently verified. Announcement time and scheduled trading time were kept separate. Price behavior before and after each event was compared with BTC, not treated as proof of causality.
Market-cap changes are not capital inflows, because capitalization also depends on token supply. Reported 24-hour volume does not measure order-book depth or execution quality. AI Score was not used.
Source data and weight calculations, CSV · Chart data, events, and sources, JSON. Server snapshot version: e5f2cddc18db760e. The full SHA-256 is retained in the download.