From DowJo — where this research becomes practice. Train your judgment before you risk your money.
Course Research · Episode 10
Why Did Six Different Stocks Move at the Same Time?
The evidence-led takeaway
They were never six different stocks. They were one position.
Ask Jo about this researchJo reads this note before answering. Sign in to ask Jo with this evidence attached — you come straight back here.
In brief
What this research concludes
They were never six different stocks. They were one position. On 27 January 2025 a chip designer, a maker of data-centre cooling gear, a turbine builder, an electricity generator, a fuel-cell company and a reactor designer with 2.8 million dollars of revenue on file all fell between 17 and 30 percent — on a session when the equal-weighted S&P 500 closed HIGHER, the Dow closed up 0.65 percent, and about three quarters of the cap-weighted index fall was one of the six. Strip out the market and the standard style factors and what is left over — the part that is supposed to be each company’s own private business — arrives in all six at once, four to eleven standard deviations below zero, while four regulated utilities sit three to four and a half ABOVE it on the same afternoon under the same ten-basis-point fall in the ten-year yield. The edge of the move ran straight through the utility sector, between companies that sell power at market prices and companies that earn a regulated return on invested capital. What the six shared was not an industry, a supply chain or a customer. It was one assumption about how much computing the world was going to buy — and the number that repriced it had been public for 26 days.
The judgment skill this hands over
When several things move at once, do not start by hunting for the connection between them. Find the EDGE of the move — the nearest comparable thing that did NOT move — and name what is on each side of it. That edge is the exposure. Then ask whether the thing on the other side of it is something a company has FILED, or something the market has ASSUMED.
What we investigated
The question, and who it looks at
This is episode 10 of Investing in the AI Era, a 14-episode course. The film tells the story; this note is the research behind it, kept inspectable.
Who and what this looks at
How we tested it
Researched to be refuted, not confirmed
Candidate claims went to an independent pass instructed to refute them against primary sources. Survivors became evidence; casualties became the refused list below.
The method, in the research team's own words
BOUNDED BY DESIGN, AND PRE-REGISTERED. The quantitative core was fixed in writing before any data existed: course/research/AI10-lanes/L0-comovement-preregistration.md names the cohorts, the windows, every statistic, six falsifiers and the declared limitations, and was COMMITTED TO GIT AT 8e4d0a0 BEFORE the first price was pulled. Section 0 of that file separates what was already known to the thread (the 27 January single-session moves, inherited from Episode 6's completed research, which selected those tickers months earlier for an electricity-constraint episode that never used them) from what was blind (every correlation, every loading, every base rate, the lead/lag result, the revenue test and the filing evidence). Claiming a clean blind pre-registration would have been the exact failure the file exists to prevent. The computation is one committed script, tools/ai10-comovement.mjs, run once. Two independent price vendors were pulled and cross-checked: they agree on every event-session figure to 0.000 percentage points, and they reproduce AI06's independently recorded numbers exactly. Factors are the Kenneth French daily five-factor file, used as published. Fundamentals are SEC XBRL company facts, most recently filed version of each period. Treasury yields are the US Treasury's own daily par curve. TWO DEFECTS WERE FOUND IN THE COMPUTATION ITSELF, BEFORE ANY RESULT WAS READ. The Gauss-Jordan back-substitution indexed the row rather than the matrix and returned NaN, silently nulling every factor residual; and the revenue series took the first XBRL tag with enough rows, which returned NVIDIA's 2017-2020 series as if it were current, because a filer can switch revenue tags mid-history. Around that, a bounded research wave of nine lanes with the agent count fixed in advance. Seven landed: the event dated precisely, the six companies' own business models from their filings, the counter-case at full strength, mechanical flow explanations tested against the SEC's Form N-PORT census and FINRA short-sale files, the measurement literature, the historical regime case, and the attributed intellectual sources. Two lanes (the aftermath, and a dedicated falsifier hunt) were lost to session limits and agent stalls; their subject matter is substantially covered elsewhere — the aftermath by the counter-case lane's hyperscaler capex work (F44), the falsifier hunt by the deterministic breadth, Treasury-yield and mechanical-explanation results. The campaign was HARVESTED AND STOPPED rather than restarted, under the studio's standing rule that research is a budget. The adversarial pass that did complete changed the film twice. It caught an arithmetic error in the sentence the episode is built around — the gap between the cost figure's publication and the session is 31 days, not the 26 the draft narrated, because 26 is the gap to a different paper — and it argued, correctly, that 'the number that repriced the market' is an interpretation of trader intent that no document in the record supports. The film now says what the record supports. VERDICT VOCABULARY: SUPPORTED / CORRECTED / MISLEADING / UNSUPPORTED / DO-NOT-NARRATE, per docs/RESEARCH.md. The broad adjudicated candidate universe, including everything evaluated and rejected, is preserved separately in course/research/AI10-research-shortlist.json.
Pre-registered before any data was pulled
The firms, metrics and window for the quantitative work were fixed in writing before a single number was requested, so the result could not be cherry-picked after the fact. Every pre-registered case is reported regardless of direction. The full pre-registration is in the research desk.
What the evidence says
41 claims survived refutation
Each carries its source, the date the thing happened and the date it was said — two different facts — and its limits, stated by the research team rather than left for you to discover.
- FactResearchF01
On Monday 27 January 2025 six US-listed companies selling six different products all fell between 17 and 30 per cent in one session: NVDA -16.97%, VRT -29.88%, GEV -21.52%, VST -28.27%, BE -24.79%, SMR -27.53%.
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: Vendor-redistributed consolidated-tape data, not an exchange filing. Two independent vendors agree on every event-session figure to 0.000 percentage points; the audit trail is course/research/AI10-lanes/prices-daily.csv.
Research note
The film states the range, not each figure. The chart draws all six.
- FactPrimary sourceF02
The S&P 500 fell 1.458% that session (6101.24 to 6012.28) and the S&P 500 ETF fell 1.41%.
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: FRED and Cboe both redistribute S&P Dow Jones Indices data rather than being the index owner; the two agree exactly.
- FactPrimary sourceF03
Of the thirty Dow Jones Industrial Average constituents as of 24 January 2025, 22 closed higher and 8 closed lower on 27 January 2025; the median constituent gained 1.9%.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: THE DJIA THIRTY IS THE ONLY MECHANICALLY DEFINED LARGE-CAP CONSTITUENT LIST THIS COMPUTATION HOLDS. It is not the S&P 500 and must never be narrated as "the market".
- FactPrimary sourceF04
Only 2 of those thirty fell more than 5% — NVIDIA at -16.97% and Cisco at -5.06%. Johnson & Johnson gained 4.13%, Procter & Gamble 3.38% and Apple 3.18%.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
- InterpretationResearchF05
The Fama-French five-factor model decomposes a stock's excess return onto an overall market factor plus factors for firm size, book-to-market equity, operating profitability and investment. Daily factor values are published by the Kenneth R. French Data Library and are built from the CRSP universe.
Event n/a · Published Feb 1993
Limits: French warns that the full factor history is reconstructed monthly and can change when CRSP revises its database. The 2015 paper itself reports that HML becomes redundant for average returns once RMW and CMA are added.
- FactPrimary sourceF06
Feeding the session's factor values through each name's own 2023-24 five-factor loadings predicts NVIDIA at about -5.8%, Vertiv at about -3.9% and a small RISE for Bloom Energy, against actual moves of -16.97%, -29.88% and -24.79%. Market and style factors account for 31% of NVIDIA's move and between -1% and 17% of the other five.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Loadings estimated before the event and applied across the window boundary; if a name’s true loadings moved, the residual carries both the surprise and the drift. The claim made is only that the ordinary relationship does not explain the session.
- FactPrimary sourceF07
Run on the S&P 500 ETF itself, the same decomposition explains 108% of the session move with a pre-event R-squared of 0.995 — the method’s own internal control.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
- FactResearchF08
The Utilities Select Sector SPDR fell -2.31% that session.
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: Vendor-redistributed consolidated-tape data, not an exchange filing. Two independent vendors agree on every event-session figure to 0.000 percentage points; the audit trail is course/research/AI10-lanes/prices-daily.csv.
Show the remaining 33 pieces of evidence
- FactResearchF09
On the same session Vistra fell -28.27%, Constellation Energy -20.85% and Talen Energy -21.59%.
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: Vendor-redistributed consolidated-tape data, not an exchange filing. Two independent vendors agree on every event-session figure to 0.000 percentage points; the audit trail is course/research/AI10-lanes/prices-daily.csv.
- FactResearchF10
On that same session four large regulated electric utilities closed HIGHER: American Electric Power +4.53%, Dominion Energy +4.21%, Southern Company +3.62% and Exelon +3.35%.
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: Vendor-redistributed consolidated-tape data, not an exchange filing. Two independent vendors agree on every event-session figure to 0.000 percentage points; the audit trail is course/research/AI10-lanes/prices-daily.csv.
Research note
Independently recorded by AI06 research (claim H284) months earlier and reproduced here exactly from two vendors.
- FactPrimary sourceF11
The 10-year US Treasury constant-maturity yield fell 10 basis points on 27 January 2025, from 4.63% to 4.53%.
Source: US Department of the Treasury, daily par yield curve rates (opens home.treasury.gov)
Event 27 Jan 2025 · Published 27 Jan 2025
Limits: Close-of-business quotations, not intraday. This is the honest competing explanation for the regulated utilities rising, and the film says so.
- InterpretationPrimary sourceF12
The line separating the fallers from the risers on 27 January 2025 ran between companies whose earnings depend on market power prices and companies whose earnings are set by a regulator on invested capital — not between industries or sectors.
Source: Interpretation of the same-session comparison, supported by the filing language in F31 and F32
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: NOT A CONTROLLED EXPERIMENT. Regulated and merchant generators differ in more ways than one, and two of the four control names carry some market-priced generation of their own (F33-adjacent, recorded in do_not_narrate X09). What the same session does establish is narrower: whatever moved the fallers was not the market, not the sector label and not the yield move, because both columns had all three.
- InterpretationPrimary sourceF13
Across the nine pairs of the six where at least eight overlapping quarters exist, mean price-return correlation is 0.4741 and mean year-over-year revenue-growth correlation is 0.1927.
Source: tools/ai10-comovement.mjs; revenues from SEC XBRL company facts (data.sec.gov), most recently filed version of each period
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: STATISTICALLY THIN BY CONSTRUCTION: eight to twenty-six overlapping quarters per pair. This is INTERPRETATION, not proof, and the frame says the word. Pairs with fewer than eight overlapping quarters were excluded in the pre-registration, which is why NVIDIA — whose fiscal quarters end in January, April, July and October — appears in none of them.
- InterpretationPrimary sourceF14
GE Vernova and Vistra: price-return correlation +0.61 over 2023-2026; year-over-year revenue-growth correlation -0.05 across eight overlapping quarters.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Eight quarters. INTERPRETATION. The pairing is the scene’s hinge because the divergence is largest, and it is drawn alongside the eight pairs that are less dramatic, including two where revenue correlation EXCEEDS price correlation.
- FactPrimary sourceF15
On 27 January 2025 the most recent revenue NuScale Power had on file was $2,821,000 for the nine months ended 30 September 2024, filed 8 November 2024. Its FY2024 figure of $37.0m was not filed until 3 March 2025 and the market could not have known it that Monday.
Event period ended 2024-09-30 · Published 8 Nov 2024
Research note
POINT-IN-TIME. The nine-month figure is what was knowable; the annual figure is hindsight. The narration says "the nine months to September" so no listener can infer an annual number.
- FactPrimary sourceF16
Mean pairwise correlation of the six: 0.3921 across 2023-24, 0.6214 through 2025, 0.4294 in 2026 to date.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: See F35: a correlation measured over a more volatile window is biased upward even when the underlying correlation is constant. The base-rate line (F36) is the control for that.
- FactPrimary sourceF17
The four regulated utilities move together at 0.7, 0.7057 and 0.6763 across the same three windows — higher than the six reach in any of them — and their 2023-24 correlation is still 0.6698 after the market and all five style factors are removed, against 0.0244 for thirty ordinary large caps. Their mean single-factor R-squared is only 0.0302.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
Research note
This is the episode’s control and it outranks its result: a high correlation number, on its own, says nothing about a story.
- UnknownPrimary sourceF19
No statistic in this episode establishes what caused the 27 January 2025 session. The attribution to the DeepSeek coverage is contemporaneous press consensus, not a measurement.
Source: Stated limitation of a single-session event study with no counterfactual
Event 27 Jan 2025 · Published 5 Sep 2026
Limits: This appears in the film in the narration, at full size, not as a footnote.
- FactPrimary sourceF20
DeepSeek released the R1 model on 20 January 2025; the model repository was created at 03:46 UTC that day and the company's release page carries the same date.
Event 20 Jan 2025 · Published 20 Jan 2025
Limits: Repository creation time can slightly precede the public announcement.
- FactPrimary sourceF22
The $5.576 million figure appears in the DeepSeek-V3 Technical Report, where it is derived by assuming a rental price of $2 per H800 GPU-hour applied to 2.788 million GPU-hours for one training run. It is an arithmetic product of GPU-hours and an assumed market rate, not a disclosed expenditure.
Source: DeepSeek-AI, "DeepSeek-V3 Technical Report", arXiv:2412.19437, Section 1 (opens arxiv.org)
Event training completed before 2024-12-26 · Published 27 Dec 2024
Limits: Excludes, by the paper’s own words, prior research and ablation experiments; makes no claim about hardware capital cost, salaries, data acquisition or inference serving.
- FactPrimary sourceF23
The cost figure was published on 27 December 2024, twenty-six days before the session that repriced it, and it is not in the R1 paper at all — the R1 paper (arXiv:2501.12948, 22 January 2025) contains no training-cost figure, no dollar amount and no GPU-hour total.
Event 27 Jan 2025 · Published 27 Dec 2024
Limits: Sequence is not causation. That the figure was public for 26 days without a market reaction is consistent with slow diffusion, with the app-store surge as the trigger, and with pure coordination. The film says the FACT did not change and the ATTENTION did, which is what the dated record supports.
- FactPrimary sourceF24
The exclusion clause sits in the sentence immediately after the figure and is word-for-word identical in both arXiv versions of the report, so it cannot be described as a later clarification added under scrutiny.
Source: arXiv:2412.19437 v1 and v2, full-text comparison (opens arxiv.org)
Event n/a · Published v1 2024-12-27; v2 2025-02-18
Research note
Quoted in the source: "Note that the aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data."
- FactResearchF25
Weighting every S&P 500 constituent equally instead of by market capitalisation, the index closed HIGHER on 27 January 2025, at +0.07%.
Source: Invesco S&P 500 Equal Weight ETF (RSP) daily close (opens query1.finance.yahoo.com)
Event 27 Jan 2025 · Published retrieved 2026-09-05
Limits: RSP is a fund tracking the equal-weight index rather than the index itself; the two can differ by a few basis points.
- FactPrimary sourceF26
The Dow Jones Industrial Average closed HIGHER on 27 January 2025, up 0.651% (44,424.25 to 44,713.58).
Source: FRED series DJIA, sourced from S&P Dow Jones Indices (opens fred.stlouisfed.org)
Event 27 Jan 2025 · Published series maintained continuously
Limits: The DJIA is price-weighted and holds only 30 names, so its rise is partly a composition artifact — which is the point being made, not a caveat against it.
- FactPrimary sourceF27
NVIDIA's index weight multiplied by its own fall accounts for about three quarters of the S&P 500's decline that session: 6.599% x -16.97% = -1.12 percentage points of a -1.458% index move.
Event 27 Jan 2025 · Published holdings as of 2024-12-31
Limits: Weight is taken at the prior month-end, not intraday on the session; the arithmetic is a first-order attribution, not an index-provider decomposition.
- InterpretationResearchF28
For large US stocks, systematic influences explain only about a fifth of a typical daily return: Roll's measured average adjusted R-squared is approximately 0.20 with daily data, and a variance decomposition attributes roughly 13% of a typical stock's volatility to market-wide movement and about 77% to firm-specific movement.
Event n/a · Published Jul 1988
Limits: Roll’s exact figures could not be re-verified against the article text within this campaign and rest on the widely cited summary; the CLMX decomposition was verified against the NBER working paper. The film states the magnitude loosely ("about a fifth") for this reason.
Research note
This is what makes the episode’s real finding the CO-ARRIVAL of residuals rather than their size.
- FactPrimary sourceF29
Standardised by each name's own 2023-24 five-factor residual standard deviation, the six residuals on 27 January 2025 are NuScale -4.29, NVIDIA -5.75, Bloom Energy -7.15, GE Vernova -8.84, Vertiv -10.94 and Vistra -11.48 standard deviations. Residuals are idiosyncratic by construction and are supposed to be independent across companies.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
Research note
THE EPISODE’S CENTRAL EVIDENCE. Not the size of any one residual, which is unremarkable on a single day, but six of them arriving on one session with the same sign.
- FactPrimary sourceF30
The same test on the four regulated utilities returns +3.08 to +4.35 standard deviations — large and POSITIVE on the session the six were four to eleven standard deviations negative. Constellation and Talen, both merchant, return -8.35 and -9.20.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
- FactPrimary sourceF31
Vistra states in its own annual report: "We are not guaranteed any rate of return on capital investments in our businesses", and that its revenues and cash flows depend in large part upon wholesale market prices. The sentence is identical in the FY2024 filing that was current on the session.
Source: Vistra Corp., Form 10-K, Item 1A Risk Factors (opens sec.gov)
Event fiscal years 2024 and 2025 · Published FY2024 and FY2025 filings
- FactPrimary sourceF32
American Electric Power states that state commissions set its retail rates "on a cost-of-service basis, which is designed to allow a utility an opportunity to recover its cost" and earn a reasonable return on the investment used to provide service. Southern discloses Georgia Power's retail return on equity set at 10.50%, and Dominion a legislated authorized return with 85-100% of over-earnings credited back to customers.
Event fiscal year 2025 · Published FY2025 filings
Research note
Terminology differs by company and state: AEP says "cost-of-service basis" and never "authorized ROE"; a search for one phrase misses the evidence.
- FactPrimary sourceF34
NuScale Power's own annual report states that it has not commercialised or sold its power modules and has not entered a binding contract with a customer to deploy one, and that all revenue generated to date arises from engineering and licensing fees and services.
Source: NuScale Power Corporation, Form 10-K FY2024, Item 1A Risk Factors and MD&A (opens sec.gov)
Event fiscal year 2024 · Published 3 Mar 2025
Limits: This filing post-dates the session by five weeks. The point-in-time revenue figure the film speaks is F15, from the 10-Q filed 8 November 2024; the never-sold-a-reactor statement is a standing description of the company, present in the FY2023 filing as well.
- InterpretationPrimary sourceF35
A correlation measured on a subsample selected by large or volatile observations is biased UPWARD even when the true correlation is constant. A Federal Reserve paper gives the closed form and worked numbers: with a true correlation of 0.50, conditioning on the outer 10% of one variable's distribution yields a measured 0.771. Forbes and Rigobon show that correcting for this removes most apparent contagion.
Event n/a · Published Dec 1997
Research note
Longin & Solnik (2001) and Ang & Chen (2002) correct for the bias and still find genuine asymmetry, concentrated in falling markets rather than in volatility as such. The film states the caveat and then points at the base-rate control rather than pretending the caveat away.
- FactPrimary sourceF36
Measured the same way over the same three windows, thirty ordinary large caps move together at 0.1987, 0.2979 and 0.118. Their 2023-24 to 2025 rise of about a tenth is the control for the volatility-conditioning bias; the six rose by more than twice as much.
Source: tools/ai10-comovement.mjs, computed from the committed price and factor files
Event 27 Jan 2025 · Published computed 2026-09-05
Limits: Cohorts, windows, statistics and falsifiers were fixed in course/research/AI10-lanes/L0-comovement-preregistration.md and committed to git at 8e4d0a0 BEFORE any data was pulled or this script was written.
Research note
NVIDIA is itself a DJIA constituent from November 2024; the figure is computed both with and without it and the two differ by less than 0.01.
- FactPrimary sourceF37
China's official manufacturing PMI for January 2025 printed at 49.1 against a consensus of about 50.1, released on 27 January Beijing time — the evening of 26 January in New York, and therefore before the US session.
Event 27 Jan 2025 · Published same-day coverage 2025-01-27; NBS English page posted 2025-02-10
Limits: The release timing rests on contemporaneous reporting because the NBS English page carries a February posting date. It is recorded here as a competing overnight macro negative that the episode cannot rule out, which is exactly how the film uses it.
- FactPrimary sourceF38
Bloom Energy announced on 14 November 2024 — ten weeks before the session — a procurement agreement with American Electric Power for up to 1 gigawatt of fuel cells, with an initial order of 100 megawatts. On 27 January 2025 Bloom Energy fell 24.79% and American Electric Power rose 4.53%.
Event 14 Nov 2024 · Published 14 Nov 2024
Limits: A company announcement, not an SEC filing: Bloom filed no 8-K between 20 December 2024 and 27 February 2025, so the agreement was public but not in a periodic report on the session. "Up to 1 GW" is a ceiling with an initial 100 MW order, not a firm gigawatt. The two moves are a same-session fact, not a causal claim about the contract.
Research note
THE SHARPEST SINGLE FACT IN THE EPISODE. Buyer and seller of the same equipment, one signed agreement, opposite sides of the line, 29 percentage points apart on one afternoon.
- FactPrimary sourceF39
NVIDIA disclosed before the session that Data Center was $30,771m of $35,082m of quarterly revenue (87.7%) and that cloud service providers were approximately 50% of Data Center revenue — roughly 44% of total revenue from a handful of cloud buyers. Its FY2025 10-K discloses three direct customers at 12%, 11% and 11% of total revenue.
Event fiscal 2025 · Published 20 Nov 2024
Limits: The 10-K post-dates the session; the concentration disclosure was in the prior years’ filings in the same form. Direct customers are anonymised as A, B and C.
- FactPrimary sourceF40
Vertiv's own SEC-filed earnings releases attribute its order growth to the data-centre market — trailing-twelve-month organic orders up about 37% (October 2024) and about 30% (February 2025), both stated by the company as driven by the hyperscale and colocation data centre market — and its backlog grew 29.9% during 2024, from $5,526.7m to $7,178.8m.
Event fiscal 2024 · Published 23 Oct 2024
Limits: Vertiv discloses NO numeric customer-concentration figure and NO share of revenue from data centres; its concentration language is qualitative only. The attribution is the company’s own characterisation of order growth, not an audited revenue split.
- FactPrimary sourceF41
GE Vernova's Q4 2024 earnings release and slide deck, both published 22 January 2025 — five days before the session — contain zero mentions of data centres, AI or hyperscalers, and its FY2024 10-K mentions data centres once, in an IT-systems context. Its Power segment equipment backlog FELL during 2024, from $13,636m to $12,461m, with gas turbine gigawatts shipped down from 13.8 to 11.9.
Event fiscal 2024 · Published 22 Jan 2025
Limits: Absence in a press release is not absence of exposure. What it establishes is that the company itself was not, in its own results communication five days before, presenting itself as a data-centre story.
Research note
Bounded not-found stated precisely: zero occurrences in those two named documents, searched for "data cent", "artificial intelligence" and "hyperscal". On the earnings CALL the same day the chief executive did attribute 2024 gas orders to data-centre demand only PARTIALLY, and gave no figure.
- FactPrimary sourceF42
Vistra's only named contracts with the large cloud buyers, as of the session, were approximately 605 megawatts of SOLAR power purchase agreements — 200 MW with Amazon and 405 MW with Microsoft — against a generation fleet of roughly 41,000 megawatts. Its FY2024 10-K described nuclear and gas large-load deals as still 'in various discussions'.
Source: Vistra Corp. Form 8-K exhibit 99.1 (Q3 2024 earnings release) and Form 10-K FY2024 (opens sec.gov)
Event fiscal 2024 · Published 7 Nov 2024
Limits: The 10-K post-dates the session by a month; the solar PPAs were disclosed in November 2024 and were public on the day.
- FactPrimary sourceF43
Across the FY2024 annual reports of all six companies, no single large cloud buyer is named as a customer of more than one of them.
Event fiscal 2024 · Published FY2024 filings
Limits: A BOUNDED NOT-FOUND across six named documents, not a claim about the world. NVIDIA anonymises its customers as A, B and C, so a shared customer could exist and be undisclosed; Vertiv names Microsoft, AWS and Google Cloud only as examples defining a market segment, not as its own customers.
- FactPrimary sourceF44
In the eleven days after 27 January 2025 the four largest buyers of AI data-centre capacity all reported, and none cut capital-expenditure plans: Meta guided to $60-65bn for 2025 against $39.2bn in 2024, Alphabet to about $75bn against $52.5bn, Microsoft stated it was on track to invest about $80bn in FY2025, and Amazon's Q4 capital investments of $26.3bn were described by management as reasonably representative of a 2025 run rate.
Event fiscal 2024 results and 2025 guidance · Published 29 Jan 2025
Limits: Guidance is an intention, not a commitment, and capital-expenditure definitions differ between these filers (Meta includes finance-lease principal; Amazon reports purchases of property and equipment). Microsoft’s ~$80bn figure comes from a company blog post, not a filing. This establishes that the assumption was not abandoned by the buyers. It establishes NOTHING about whether the assumption is correct, and the film says so in the next sentence.
What surprised us
Where the simple story did not survive
Logged by the research lanes as they worked, before anything was written. The full log is in the research desk.
- 1
The counter-case is TRUE at one end of the six and collapses at the other. GE Vernova's own Q4 2024 results release — five days before the session — mentions data centres zero times, and its Power equipment backlog FELL over 2024.
- 2
Bloom Energy had agreed on 14 November 2024 to supply American Electric Power with up to a gigawatt of fuel cells. On 27 January 2025 the seller fell 24.79% and the buyer rose 4.53%. Buyer and seller of the same equipment, opposite sides of the line, on one afternoon.
- 3
The cost figure that travelled with the coverage was published 31 days earlier, in a different paper, about a different model. The R1 paper contains no cost figure at all.
- 4
The equal-weighted S&P 500 closed HIGHER that session and the Dow closed up 0.65%. About three quarters of the cap-weighted index fall is one constituent.
- 5
The four "boring" regulated utilities are the most tightly co-moving group in the entire study — 0.70 in every window, and 0.67 after removing the market and all five style factors, against 0.024 for thirty ordinary large caps.
The strongest case against this
The evidence against our own conclusion
Carried at full strength, before the conclusion — not as a footnote.
Claim: The six moved together because they share a real economic driver: one capital-expenditure budget at a handful of cloud buyers is the demand for all of them, so the market was right to reprice them as one.
Carried in film: S05, in Dow's own voice, at full strength and before it is answered.
Strongest support: NVIDIA disclosed Data Center at 87.7% of quarterly revenue with cloud service providers about half of that, and three direct customers at 12/11/11% of FY2025 revenue. Vertiv's own filed earnings releases attribute roughly 30-37% trailing organic order growth to the hyperscale and colocation data-centre market, with backlog up 29.9% to $7.18bn.
Where it collapses: GE Vernova's own Q4 2024 results release and slide deck, published five days before the session, mention data centres zero times, and its Power equipment backlog FELL over 2024. Vistra's only named contracts with those buyers were about 605MW of solar against a roughly 41,000MW fleet. NuScale had one Class 1 customer, in Romania. And across all six FY2024 annual reports, no single cloud buyer is named as a customer of more than one of them.
Verdict: PARTIALLY SUPPORTED. The economic link is real, disclosed and quantified at one end of the six and effectively absent at the other — which is itself evidence that what the whole group shared on the day was an expectation rather than a customer.
History, under test
The strongest historical case
What the past licenses — and, stated just as plainly, what it does not.
Note: Lane 6 landed 44 adjudicated candidates on the historical regime case and they are preserved in the shortlist. NONE reached the film: the episode's own evidence turned out stronger than any analogy, and a historical parallel would have cost roughly a minute for a weaker version of a point the 27 January cross-section already makes. Recorded here so the next episode that wants it does not re-research it.
Status: EVALUATED, NOT USED — see course/research/AI10-research-shortlist.json
What would change our mind
The conclusion is wrong if…
Six falsifiers were written down before the data was pulled (pre-registration §8). Two were tested and did not fire: the regulated utilities did not fall with the merchant ones (F10), and the mechanical flow explanations are present but too small by orders of magnitude (X07, X08). One fired PARTIALLY and improved the film: the counter-case that the link is genuinely economic is TRUE at the NVIDIA and Vertiv end and collapses at the GE Vernova, Vistra and NuScale end (F39-F43), so the episode carries both halves. What would still overturn the reading: if the six's factor-residual correlation before the event were no higher than an ordinary large-cap base rate, the day created the co-movement rather than revealing it; if a shared named customer appeared across several of the six in later filings, the shared driver is contractual rather than expectational; and if the residual separation between the merchant and regulated columns on 27 January 2025 ever fails to reproduce, the central evidence is gone. src/data/ai10-evidence.ts::assertEvidence() throws on the last of these at render time.
What remains uncertain
What we still do not know
Stated by the research team, in full, rather than smoothed over.
- No statistic here establishes what caused the session. The attribution to the DeepSeek coverage is contemporaneous consensus. Competing overnight events exist and are dated: China's January manufacturing PMI at 49.1 against a consensus near 50.1, published 09:30 Beijing time on 27 January, which is the evening of the 26th in New York; a Colombia tariff threat announced and withdrawn on the 26th; and the Bank of Japan's 24 January rate rise.
- Six names is a small sample and every statistic computed on it has wide error bars.
- The revenue co-movement test rests on eight to twenty-six overlapping quarters per pair and is INTERPRETATION.
- Factor loadings estimated in 2023-24 are applied to an event in 2025; if a name's true loadings moved, the residual carries both the surprise and the drift.
- RESIDUALS DEPEND ON THE FACTOR SET AND THIS IS DEMONSTRABLE: Kenneth French's three-factor and five-factor daily files carry SMB values with OPPOSITE SIGNS on 27 January 2025 itself (-0.65 against +0.30). A three-factor decomposition of the same session would not give the same residuals.
- A correlation measured over a more volatile window is biased upward even when the underlying correlation is constant (F35). The base-rate line is the control for that; it is not a correction.
- The control cohort chosen as an ordinary-large-cap base rate was NOT unrelated either: the six defensive names show a factor-residual correlation of 0.171 against 0.024 for the mechanical DJIA base rate, because they share a rate sensitivity the market factor does not capture. A defect in the control design, reported as one — and the episode's own point arriving uninvited.
- Two of the four control utilities are not purely regulated: Southern Power sells at market-based rates and Dominion's Contracted Energy segment earns market prices. Exelon is the clean case.
- Whether the assumption that repriced on 27 January 2025 was correct is not settled by anything in this episode. The four largest buyers did not cut their plans in the eleven days after (F44), which says the assumption was not abandoned and nothing about whether it was right.
What we refused to publish
16 claims we would not say — and why
The do-not-narrate list. Some are popular; some are true but unproven; some are simply not this note’s to make. Each refusal is enforced in production, not just recorded.
- RefutedX01
“DeepSeek trained a frontier model for $5.6 million.”
Verdict: REFUTED AS STATED
Why: The figure is an assumed rental price ($2/GPU-hour) times one training run’s GPU-hours, and the paper’s own next sentence excludes prior research and ablation experiments. It is not a total cost and DeepSeek never presented it as one. The film says "a comparable one had been trained for about five and a half million dollars" and immediately dates it and bounds it.
- UnverifiableX02
“DeepSeek actually spent about $1.6bn on servers and $944m on operations.”
Verdict: UNVERIFIABLE
Why: Commercial-research estimates published after the session by a firm selling research to the semiconductor industry. Not audited, not disclosed, not independently reproducible. Interesting, and not narratable.
- UnresolvableX03
“DeepSeek caused the 27 January 2025 selloff.”
Verdict: UNRESOLVABLE — MUST BE STATED AS UNKNOWN
Why: One session has no counterfactual, and at least four other real events sit in the window: a Colombia tariff threat announced and withdrawn on the 26th, China’s January manufacturing PMI missing consensus overnight, the Bank of Japan’s 24 January rate rise, and a US new-home-sales print during the session. The film narrates the attribution as contemporaneous consensus and says out loud that agreement is not measurement.
- Partly unverifiableX04
“The DeepSeek app reached number one on the US App Store on 27 January 2025.”
Verdict: PARTIALLY UNVERIFIABLE
Why: The best-sourced date is Sunday 26 January, from app-analytics firm Appfigures via TechCrunch; 27 January is the PUBLICATION date of the reporting, not the event date. Apple publishes no dated historical chart archive, so no Tier-1 confirmation exists. The film says "a Chinese lab released a model" and does not use the chart claim at all.
- RefutedX05
“Trump’s chip-tariff pledge and his "wake-up call" remark moved the market that day.”
Verdict: REFUTED
Why: Both were delivered at the House Republican retreat on the EVENING of 27 January, at least ninety minutes after the 16:00 ET close. They cannot have caused the session. They are also Episode 11 territory.
- RefutedX06
“It was a bad day for the stock market.”
Verdict: REFUTED
Why: The equal-weighted S&P 500 closed higher, the Dow closed up 0.65%, 22 of its 30 constituents rose, and the median Dow constituent gained 1.9%. About three quarters of the cap-weighted index’s fall is one constituent.
Show the remaining 10 refused claims
- RefutedX07
“Thematic ETF flows or index membership explain the co-movement.”
Verdict: REFUTED
Why: Tested against the SEC’s own Form N-PORT census: no widely held fund held all six, GICS classification actively split them across three sectors, every identifiable AI/nuclear/clean-energy/data-centre themed fund held about $1.0bn of the six combined against roughly $112bn traded that day, and those funds generally held MORE shares after the event than before. Present, and too small by orders of magnitude to be sufficient.
- RefutedX08
“Short sellers drove the fall.”
Verdict: REFUTED
Why: Short volume as a share of FINRA-reported volume FELL on 27 January in five of the six names against the prior three sessions, and the 15-to-31 January short-interest change has no common sign across the six. The day’s selling was disproportionately long selling.
- Partly unverifiableX09
“The four control companies are purely regulated utilities.”
Verdict: PARTIALLY REFUTED
Why: Exelon is: its own 10-K states it does not generate the electricity it delivers and passes procurement costs through without mark-up. Southern is not — Southern Power sells at market-based rates — and neither is Dominion, whose Contracted Energy segment earns market prices. The film says "companies whose rates a regulator sets" rather than "purely regulated", and the frame carries the caveat.
- RefutedX10
“GE Vernova is an AI company.”
Verdict: REFUTED
Why: Two-thirds of its Power segment revenue is SERVICES on an installed base of roughly 7,000 gas turbines, not new equipment. The film never characterises any of the six as an AI company; it describes what each one sells.
- UnverifiableX11
“Vertiv earns a stated share of its revenue from data centres.”
Verdict: UNVERIFIABLE
Why: Vertiv’s 10-K discloses revenue by geography and by product-versus-service and does NOT disclose a data-centre share or any customer concentration. A bounded not-found in that filer’s own filings, not a claim about the world.
- RefutedX12
“Twenty-two of the thirty companies in the market rose that day.”
Verdict: REFUTED AS STATED
Why: The thirty is the Dow Jones Industrial Average, the only mechanically defined constituent list this computation holds. Narrating it as "the market" would be the bounded-result-to-universal-claim error this course teaches people to catch. Every frame naming it names the universe.
- UnresolvableX13
“This shows the AI trade is a bubble / that these companies are overvalued.”
Verdict: UNRESOLVABLE — DO NOT NARRATE
Why: Bubble dynamics, the capital cycle, overbuild and valuation levels are Episode 12. Nothing in this episode measures value, and no statistic here can support a valuation claim.
- UnresolvableX14
“Export controls, the AI diffusion rule or China policy explain the repricing.”
Verdict: UNRESOLVABLE — DO NOT NARRATE
Why: Episode 11 territory. The DeepSeek release appears here only as the dated event that repriced a shared assumption.
- UnresolvableX15
“Any statement about what a viewer should buy, sell, hold or avoid.”
Verdict: UNRESOLVABLE — DO NOT NARRATE
Why: DowJo trains judgment, not trade calls. No buy/sell/hold language, no price targets, no signals, no copy-trading framing.
- RefutedX16
“The residual analysis proves what caused the move.”
Verdict: REFUTED
Why: A large residual means the model does not explain the observation. It does not name a cause. The film uses the residuals to rule things OUT and to show that ten supposedly independent surprises arrived together — which is a statement about structure, not about causation.
What to watch next
Dated material, and what would make it stale
The six's trailing co-movement, currently 0.43 in 2026 to date against 0.62 through 2025.
5 Sep 2026Status: live
What changes: If the group re-converges above its 2025 level, the film's 'they came back apart' sentence needs re-measuring. Re-run the script; the chart is derived, not typed.
Whether a shared named customer appears across several of the six in later annual reports.
5 Sep 2026Status: not found as of the FY2024 filings
What changes: A named overlap would move the counter-case from partially supported toward supported, and the film would need to say so.
Ideas we borrowed, and tested
Thinkers, taken seriously enough to argue with
Claim → author → evidence → counter-argument → historical test → current relevance. Never doctrine.
Eugene Fama and Kenneth French
Co-movement between securities is exposure to a common factor, not a relationship between the companies.
- Evidence:
- Common risk factors in the returns on stocks and bonds, JFE 33(1) 1993; A five-factor asset pricing model, JFE 116(1) 2015.
- Counter-argument:
- Roll's critique (JFE 1977): any factor decomposition is a joint test of the phenomenon and the proxy, and the true market portfolio is not observable. Harvey, Liu and Zhu (RFS 2016) argue that hundreds of mined factors mean a new one should clear a t-statistic above 3.
- Historical test:
- The five-factor model's own authors report that HML becomes redundant for average returns once profitability and investment are added — the factor set is not settled.
- Current relevance:
- The episode uses the decomposition only to SUBTRACT what is ordinary, never to price anything.
Boyer, Gibson and Loretan; Forbes and Rigobon
Measured correlation rises in stressed subsamples even when the true correlation is constant.
- Evidence:
- Federal Reserve IFDP 597 gives the closed form and worked numbers (true 0.50 measures 0.771 in the outer 10%); Forbes & Rigobon, Journal of Finance 57(5) 2002, find almost no increase in unconditional correlations after correcting for it.
- Counter-argument:
- Longin & Solnik (JF 2001) and Ang & Chen (JFE 2002) correct for the bias and still find genuine asymmetry — concentrated in falling markets rather than in volatility as such.
- Historical test:
- Applied here to the episode's own statistic: part of the 0.39 to 0.62 rise is this artefact, and the film says so on screen.
- Current relevance:
- It is the reason the base-rate line is drawn at all.
Richard Roll; Campbell, Lettau, Malkiel and Xu
A typical large stock's daily return is mostly its own business; systematic influences explain about a fifth.
- Evidence:
- Roll, 'R2', Journal of Finance 43(3) 1988; CLMX, NBER WP 7590 / Journal of Finance 2001, attributing roughly 13% of a typical stock's volatility to the market and about 77% to the firm.
- Counter-argument:
- Roll's own figures could not be re-verified against the article text in this campaign and rest on the widely cited summary; the film states the magnitude loosely for that reason.
- Historical test:
- It is what turns the episode's finding from 'the factors don't explain it' — unremarkable — into 'ten supposedly independent surprises arrived together', which is not.
- Current relevance:
- The single idea that makes the residual test mean anything.
Evidence & sources
29 sources, by tier
Tier 1 is primary and authoritative — filings, regulators, official statistics. Journalism and books are attributed ingredients, never proof by reputation.
- ResearchYahoo Finance chart API (consolidated tape), cross-checked against stockanalysis.com (opens query1.finance.yahoo.com)supports F01, F08, F09, F10
- Primary sourceFRED series SP500, sourced from S&P Dow Jones Indices, cross-checked against Cboe SPX daily history; ETF from the vendor price file (opens fred.stlouisfed.org)supports F02
- Primary sourcetools/ai10-comovement.mjs, computed from the committed price and factor filessupports F03, F04, F06, F07, F14, F16, F17, F29, F30, F36
- ResearchFama & French, "Common risk factors in the returns on stocks and bonds", JFE 33(1) 1993, 3-56; "A five-factor asset pricing model", JFE 116(1) 2015, 1-22; Kenneth R. French Data Library (opens mba.tuck.dartmouth.edu)supports F05
- Primary sourceUS Department of the Treasury, daily par yield curve rates (opens home.treasury.gov)supports F11
- Primary sourceInterpretation of the same-session comparison, supported by the filing language in F31 and F32supports F12
- Primary sourcetools/ai10-comovement.mjs; revenues from SEC XBRL company facts (data.sec.gov), most recently filed version of each periodsupports F13
- Primary sourceNuScale Power Corporation Form 10-Q, accession 0001822966-24-000144, and Form 10-K, accession 0001822966-25-000024 (opens sec.gov)supports F15
- Primary sourceStated limitation of a single-session event study with no counterfactualsupports F19
- Primary sourceHugging Face repository metadata for deepseek-ai/DeepSeek-R1; DeepSeek "DeepSeek-R1 Release" page (opens huggingface.co)supports F20
- Primary sourceDeepSeek-AI, "DeepSeek-V3 Technical Report", arXiv:2412.19437, Section 1 (opens arxiv.org)supports F22
- Primary sourcearXiv:2412.19437 (27 Dec 2024) and arXiv:2501.12948v1 (22 Jan 2025), full-text scans (opens arxiv.org)supports F23
- Primary sourcearXiv:2412.19437 v1 and v2, full-text comparison (opens arxiv.org)supports F24
- ResearchInvesco S&P 500 Equal Weight ETF (RSP) daily close (opens query1.finance.yahoo.com)supports F25
- Primary sourceFRED series DJIA, sourced from S&P Dow Jones Indices (opens fred.stlouisfed.org)supports F26
- Primary sourceIndex weight from SEC Form N-PORT holdings of the iShares Core S&P 500 ETF as of 31 December 2024; arithmetic applied to the session return (opens sec.gov)supports F27
- ResearchRichard Roll, "R2", Journal of Finance 43(3), 1988, 541-566; Campbell, Lettau, Malkiel & Xu, "Have Individual Stocks Become More Volatile?", NBER WP 7590 / Journal of Finance 2001 (opens nber.org)supports F28
- Primary sourceVistra Corp., Form 10-K, Item 1A Risk Factors (opens sec.gov)supports F31
- Primary sourceAmerican Electric Power, The Southern Company and Dominion Energy Forms 10-K, Item 1 Business (opens sec.gov)supports F32
- Primary sourceNuScale Power Corporation, Form 10-K FY2024, Item 1A Risk Factors and MD&A (opens sec.gov)supports F34
- Primary sourceBoyer, Gibson & Loretan, "Pitfalls in Tests for Changes in Correlations", Federal Reserve IFDP 597; Forbes & Rigobon, "No Contagion, Only Interdependence", Journal of Finance 57(5), 2002 (opens federalreserve.gov)supports F35
- Primary sourceNational Bureau of Statistics of China, Purchasing Managers’ Index for January 2025 (opens stats.gov.cn)supports F37
- Primary sourceBloom Energy press release, "Bloom Energy Announces Gigawatt Fuel Cell Procurement Agreement with AEP"; session moves from the committed price file (opens bloomenergy.com)supports F38
- Primary sourceNVIDIA CFO Commentary on Q3 FY2025 results and NVIDIA Form 10-K for the fiscal year ended 26 January 2025 (opens investor.nvidia.com)supports F39
- Primary sourceVertiv Holdings Co Forms 8-K exhibit 99.1 (Q3 and Q4 2024 earnings releases) and Form 10-K FY2024 (opens sec.gov)supports F40
- Primary sourceGE Vernova Inc. Form 8-K exhibit 99.1 filed 22 January 2025, the accompanying presentation, and Form 10-K FY2024 (opens sec.gov)supports F41
- Primary sourceVistra Corp. Form 8-K exhibit 99.1 (Q3 2024 earnings release) and Form 10-K FY2024 (opens sec.gov)supports F42
- Primary sourceSynthesis of the FY2024 Forms 10-K of NVIDIA, Vertiv, GE Vernova, Vistra, Bloom Energy and NuScale Power (opens sec.gov)supports F43
- Primary sourceMeta, Alphabet and Amazon Forms 8-K exhibit 99.1 (Q4 2024 results); Microsoft official blog, 3 January 2025, and Form 10-Q for the quarter ended 31 December 2024 (opens sec.gov)supports F44
Complete evidence depth
The research desk
Everything the research evaluated before it was distilled — including what it rejected, and why.
The research desk
Before the evidence above was distilled, the research evaluated 224 candidate claims across 0 lanes, rejected 16 with a recorded reason, and logged 8 surprises. A rejected claim with a reason is the most reusable thing research produces — the desk keeps all of them.
- laneDeepSeek trained a frontier model for $5.6 million. (Id: X01; Verdict: REFUTED AS STATED; Reason: The figure is an assumed rental price ($2/GPU-hour) times one training run’s GPU-hours, and the paper’s own next sentence excludes prior research and ablation…
- laneDeepSeek actually spent about $1.6bn on servers and $944m on operations. (Id: X02; Verdict: UNVERIFIABLE; Reason: Commercial-research estimates published after the session by a firm selling research to the semiconductor industry. Not audited, not…
Save this research and keep following the story — you come straight back to this desk.
Community
Talk it through
Discuss this research
Members only · in 📡 Intelligence Watch
Challenge the argument, bring evidence, or ask what would change our mind. This note is discussed in 📡 Intelligence Watch, a DowJo room where members compare reads. Filings, news events, and what they actually mean — the discussion room for the intelligence feed.
Read and reply, then keep following the story — you come straight back to this note.
Education-first ground rules apply in every room: mechanisms and evidence, never calls.
Now train it
Take it further
Reading is where judgment starts. Practice is where DowJo measures it — and remembers what to train next.
Train it
Can you spot what the market already priced in?
Five reps on today's tape. Commit a call before the market grades it.
Train itTrain it
Think you can spot it before the reveal?
Train it on a chart: an unseen setup, revealed bar by bar, graded on your process.
Sign in to train this setupLearn the concepts underneath
Join the discussion
Challenge the argument, bring evidence, or ask what would change our mind — in the open, with other members.
MentorAsk Jo what could invalidate this thesis.
Jo reads this note before answering. Sign in to ask, with this evidence attached — you come straight back here.
Ask Jo with this evidence attachedProvenance — where this note comes from
This page is a deterministic projection of canonical research artifacts. It adds presentation and discovery; it never adds, removes or softens a finding. Question taken from the title; thesis from the packet. Projected 21 Sep 2026 by dowjo-research-projector v1.1.0 from origin commit c83c8797ab79.
| Artifact | Path (origin-relative) | Identity | Version |
|---|---|---|---|
| Episode registry | course/episodes.json | sha256 4686fb04ef23… | — |
| Evidence packet | course/research/AI10-evidence-packet.json | sha256 289c3e6a67aa… | compiled 2026-09-06 · rev 1 |
| Research shortlist | course/research/AI10-research-shortlist.json | sha256 a18c5f45956d… | — |
| Approved script | course/scripts/AI10-the-edge-of-the-move.json | sha256 8f34fce59f5e… | v 1 |
| Episode manifest | course/manifest/AI10.manifest.json | sha256 aeccfaa22ae5… | — |
| Approval record | renders/approved/AI10-FINAL.json | sha256 bf0ba6de18c6… | compiled 2026-09-13 |
| Pre-registration | course/research/AI10-lanes/L0-comovement-preregistration.md | sha256 ace88b7669ba… | — |
| Poster still | qa/stills-AI10/RESEARCH-POSTER.png | sha256 74eff69de839… | — |
Approved master (AI10-final02-corrected.mp4): sha256 633ec6694d920c9cc0ffa80b8a05f98f9d26744a00b8d58b58e76640b481b7ec
Commercial disclosure: none. No affiliate relationship, review copy, sponsorship or publisher relationship applies to this note.
For education only. Not financial advice. No buy, sell or hold recommendations, ever.
Related research
That looks interesting too
Investing in the AI Era · Episode 6
AI Eats Electricity — and Then the Grid Fights Back
"Electricity" is not one constraint.
Shares AI Infrastructure, Artificial Intelligence, Markets, Investing
Investing in the AI Era · Episode 1
Everyone Knows AI Is the Future. So What Do You Actually Invest In?
You can be completely right about AI and completely wrong about the investment.
Shares Artificial Intelligence, AI Infrastructure, Markets
Investing in the AI Era · Episode 14
The Market Becomes Your Curriculum
The thesis of the theses: under every film sat the same two questions, asked until they no longer needed the subject — compared with what, and as of when?
Shares Artificial Intelligence, Markets, Investing