Research report
Which AI trades are actually one trade.
The AI Supercycle Map · Edition 2, data as of 2 November 2026
Twelve AI bubbles, about 700 names screened, 87 of them mapped into the clusters that genuinely move together. Edition 2 (data as of 2 November 2026) answers one question: which parts of the AI build-out trade as a single thing, and which only look like a theme in a headline.
Look inside before you buy
4,118 words · about 18 min to read · 11 sections · data as of 2 November 2026
Table of contents
- What changed since Edition 1
- What this is
- What this is not
- 1. The method, in one page
- 2. The twelve bubbles, ranked by measured coherence
- The three averages that understate their cluster
- What the ranking actually tells you
- 3. Bubble by bubble: the pairs and the sub-clusters
- 4. Shared names and the ETF overlap trap
- 5. The forward layer: leading signals, re-observed
- 5.1 DRAM / HBM Memory
- 5.2 Semi Equipment / Litho
- 6. Using the map
- 7. What this map cannot do
- 8. Sources
Excerpt
1. The method, in one page
A bubble here is not a valuation judgement. It is a measured statement: these names move together.
The core number is average pairwise correlation (avg_corr): the Pearson correlation of daily returns over the last 252 trading days, computed for every pair of primary members, then averaged. Ten members means 45 pairs; six members means 15. It is the raw correlation: market beta is not removed first. When the whole market falls, as it did in October 2026 (SOXX down 23% and SMH down 20% in the month to 2 November, QA data), raw correlations rise across the board. Read a high number as "these names share a driver", where part of that driver can be the market itself.
How to read it:
- avg_corr above ~0.55: a genuine cluster. One thesis is driving most of the price action. A basket behaves like a single position: its risk is one thesis, however many tickers it holds.
- avg_corr 0.35 to 0.55: a real but loose theme. Sub-clusters exist inside it. Name selection matters more than theme selection.
- avg_corr below ~0.35: a label, not a block. The names are in the same story but not in the same flow. Buying "the theme" here buys idiosyncratic risk you did not intend.
The caveat most people skip: n_pairs. An average of 0.53 across three pairs is close to noise; the same 0.53 across 45 pairs is a fact. Always read avg_corr next to n_pairs. Below 10 pairs, this report flags the average as thin.
One table from the document
2. The twelve bubbles, ranked by measured coherence
| # | Bubble | avg_corr | Ed. 1 | range (min / max) | pairs | Primary members |
|---|---|---|---|---|---|---|
| 1 | Nuclear / SMR | 0.715 | 0.629 | 0.505 / 0.861 | 45 | OKLO, SMR, CCJ, BWXT, LEU, UUUU, NNE, UEC, DNN, LTBR |
| 2 | Networking / Optical | 0.586 | 0.537 | 0.417 / 0.831 | 36 | ANET, ALAB, AAOI, CIEN, COHR, CRDO, FN, LITE, SIVE |
| 3 | Cooling / DC Infra | 0.586 | 0.568 | 0.414 / 0.750 | 21 | VRT, JCI, TT, ETN, EMR, MOD, GNRC |
| 4 | Quantum | 0.578 | 0.572 | 0.134 / 0.936 | 21 | IONQ, QTUM, RGTI, QBTS, ARQQ, QUBT, IBM |
| 5 | Enterprise SaaS AI (thin) | 0.531 | 0.562 | 0.374 / 0.728 | 3 | NOW, SNOW, CRM |
| 6 | Semi Equipment / Litho (data) | 0.511 | 0.574 | 0.178 / 0.906 | 28 | ASML, AMAT, LRCX, ACLS, AEHR, KLAC, ICHR, 3110 |
| 7 | Space / Sat Comms (data) | 0.486 | 0.484 | 0.142 / 0.780 | 28 | RKLB, SPCX, LUNR, KTOS, PL, BKSY, IRDM, SPIR |
| 8 | AI Compute Accelerators | 0.466 | 0.428 | 0.121 / 0.717 | 45 | NVDA, AMD, TSM, ARM, AVGO, SMCI, MRVL, CBRS, QCOM, INTC |
| 9 | Datacenter Power | 0.387 | 0.375 | -0.038 / 0.815 | 36 | CEG, VST, TLN, BE, GEV, AEP, ETR, NVTS, NRG |
| 10 | DRAM / HBM Memory (data) | 0.381 | 0.375 | -0.030 / 0.803 | 15 | HXSCL, MU, RMBS, SNDK, MRAM, WDC |
| 11 | AI-Capex Hyperscalers | 0.322 | 0.316 | 0.136 / 0.618 | 10 | MSFT, AMZN, META, GOOGL, ORCL |
| 12 | Robotics / Automation | 0.294 | 0.254 | -0.251 / 0.623 | 10 | ISRG, CCXI, ABB, ROK, TER |
The preview stops here. The full document is delivered by email after purchase.
Why this matters to a book, not just to a chart
A basket of ten nuclear names at an average pairwise correlation of 0.715 is one position wearing ten tickers. A basket of five hyperscalers at 0.322 is five separate bets. Size them the same way and you are carrying a concentration you never chose. The report gives the measured number for every bubble, with the pair count, the tightest and loosest pair and the sub-clusters inside it, so you can see which averages are facts and which are noise.
The finding the free pages hide
Three published averages understate their cluster because of the data underneath, not the business: a stock split the price history does not adjust for, a broken series for the Seoul-listed SK Hynix line, and foreign or newly listed names. Cleaned, US-listed semiconductor equipment is the second-tightest block on the board (0.706), not the sixth. Nuclear / SMR stays the tightest; AI Compute Accelerators, the trade everyone means by "AI", is still eighth.
Sixteen leading signals, re-observed and dated
For memory and chip equipment, every leading signal was checked again against a dated primary source, mostly SEC filings: eleven re-observed, five marked as not re-observed rather than guessed. Four changed state since May. DRAM spot has gone flat, and the equipment makers now lag the chip index. Every source is listed with its URL.
What is inside
- What changed since Edition 1, in one list, including one correction to the method.
- The method in one page: average pairwise correlation, the pair-count caveat, and the data checks almost nobody makes.
- All twelve bubbles ranked, with the Edition 1 value next to each, the min/max pair range, the pair count and every member ticker.
- The three bubbles whose average is distorted by data problems, with the clean core recomputed.
- Bubble by bubble: the tightest and loosest pair and the sub-clusters inside each one.
- The shared names and the ETF overlap map: which stocks and funds span which bubbles.
- The two signal layers (DRAM / HBM memory, semiconductor equipment): all sixteen signals ordered by lead time, May state next to today's, what was observed, when, and from which source.
- A numbered source list with every URL, and a short section on what the map cannot do.
Buy it if
- You already hold AI names and want to know what your book is really concentrated in.
- You want the thesis measured rather than assumed, and the data behind the measure checked.
- You want the cycle signals that lead memory and semi-equipment, with lead times and dated sources attached.
Skip it if
- You want buy and sell calls. There are none here, by design and by law.
- You want a forecast. Correlation is a backward-looking window; the report says so on its first page.
- You want forward signals for all twelve bubbles. The signal layer covers two of them, and the report is explicit about that.
How delivery works
Pay through Stripe, and the document arrives by email within seconds as a permanent link tied to your address. Read it in the browser or print it to PDF. Nothing to install, nothing recurring: a one-time purchase. Lost the email? Sign in to your library with the same address and it is there.
Included in Pro. See Pro
An educational research publication, not investment advice, trade signals, a managed account, or an advisory service. Nothing in it is a recommendation to buy or sell any security. See our disclosures.