Everyone Knows AI Is the Future. So What Do You Actually Invest In?

In brief

What this research concludes

You can be completely right about AI and completely wrong about the investment. A theme is not a thesis.

What we investigated

The question, and who it looks at

How we tested it

Researched to be refuted, not confirmed

What the evidence says

66 claims survived refutation

The strongest case against this

The episode's own logline does not survive contact with the evidence. 'A theme is not a thesis' is falsified by Amazon and by Apple. The narrower claim — 'entry valuation and vehicle selection determine whether being right about a theme pays' — survives every attack and is the honest version.

History, under test

The strongest historical case

What remains uncertain

What we still do not know

What we refused to publish

15 claims we would not say — and why

1. RefutedX01

   “Coherent reported Q2 FY2026 revenue of $1.7B (+17.5% y/y) with non-GAAP gross margin of 40.2%; Datacenter & Communications grew 33.6% y/y to $1.2B with segment book-to-bill above 4x.”

   Verdict: REFUTED

   Why: Splices two different quarters and cites the wrong filing. Q2 FY2026 (quarter ended 2025-12-31) revenue was $1.69B at 39.0% non-GAAP gross margin. The 40.2% belongs to Q4 FY2026 (quarter ended 2026-06-30), revenue $2.05B, with Datacenter & Communications at $1.62B or 79% of revenue. The cited SEC exhibit is the Q4 release. Neither release contains a $1.2B / +33.6% segment figure, and 'book-to-bill above 4x' appears in NEITHER and is entirely unverified.

   What the research recorded instead: Coherent reported Q4 FY2026 (quarter ended 2026-06-30) revenue of $2.05B, up 34% y/y, with GAAP gross margin 38.5% and non-GAAP 40.2%; Datacenter & Communications contributed $1.62B, 79% of revenue, and full-year FY2026 Datacenter & Communications revenue rose 40% to $5,275M. Source: https://www.coherent.com/news/press-releases/fourth-quarter-and-fiscal-year-2026-results

2. RefutedX02

   “Copper reached a record ~US$14,455/tonne on the LME in August 2026. J.P. Morgan estimates data-center-related copper demand of roughly 475,000 tonnes in 2026 and a global refined-copper deficit of around 330,000 tonnes.”

   Verdict: REFUTED

   Why: The cited IEA commentary is dated 2 March 2026 — five months BEFORE the event described — and says prices 'briefly exceeded USD 14 500 per tonne (intraday) in January 2026'. It places the record in January, not August. The J.P. Morgan figures are not in the source at all: no J.P. Morgan citation, no 475,000-tonne estimate, no 330,000-tonne deficit. The only deficit figure in the source is an IEA projection of 30% by 2035. Neither J.P. Morgan number could be independently confirmed anywhere.

   What the research recorded instead: Copper prices hit record highs in 2026, briefly exceeding US$14,500/tonne intraday on the LME in January 2026, having passed US$12,000/tonne for the first time only in December 2025. The IEA projects the copper market could face a supply deficit of 30% by 2035. Source: https://www.iea.org/commentaries/copper-prices-have-hit-record-highs-but-smelters-face-mounting-strategic-pressures

3. UnverifiableX03

   “Any dark-fibre 'percent lit' statistic — 'less than 5% of boom-era fibre was ever lit', '85–95% remained dark'.”

   Verdict: UNVERIFIABLE

   Why: No primary measurement exists. No FCC, DOE, ITU or carrier dataset supports these percentages; they appear only in secondary commentary. This is the single most tempting and least sourceable number in the dot-com segment.

4. UnverifiableX04

   “Cisco cut 40,000 jobs between 2001 and 2003.”

   Verdict: UNVERIFIABLE

   Why: Cisco's own FY2001 10-K documents a restructuring programme of approximately 6,000 regular employees (about 4,700 terminated by 28 July 2001, plus about 1,500 through attrition). The 40,000 figure comes from secondary reporting and could not be reconciled with any primary filing.

5. UnverifiableX05

   “Cisco's peak market capitalisation on 27 March 2000 was $555 billion (or $569 billion).”

   Verdict: UNVERIFIABLE

   Why: Sources conflict — $555bn, $569bn, and arithmetic from ~6.9bn shares at $80.0625 gives ~$552bn. The exact share count outstanding on that date was not located in a primary filing.

6. UnverifiableX06

   “Any as-traded dollar price for JDS Uniphase's March 2000 peak, including Yahoo's adjusted peak close of $666.81.”

   Verdict: UNVERIFIABLE

   Why: The 1-for-8 reverse split of October 2006 and the 2015 Lumentum spin-off make the adjusted series a price nobody ever paid. Reconstruction from any public adjusted series is not possible.

What to watch next

Dated material, and what would make it stale

Ideas we borrowed, and tested

Thinkers, taken seriously enough to argue with

Evidence & sources

64 sources, by tier

Complete evidence depth

The research desk

• Anthropic is reported to use over one million AWS Trainium2 chips to train and serve Claude, with roughly 1GW of combined Trainium2 and Trainium3 capacity coming online by end-2026, alongside a separate Google Cloud commitment of up to one million TPUs announced 2025-10-23 bringing over a gigawatt online in 2026.

  Verdict: CONFIRMED

  Both halves check out, but against two different primary sources. Google's 2025-10-23 press release (the cited URL) says Anthropic 'will have access to up to one million TPU chips' and 'well over a gigawatt of capacity coming online in 2026' — it contains no mention of Amazon, AWS or Trainium. The Trainium half is confirmed instead by Anthropic's own announcement (anthropic.com/news/anthropic-amazon-compute): 'we currently use over one million Trainium2 chips to train and serve Claude' and 'nearly 1GW total of Trainium2 and Trainium3 capacity coming online by the end of 2026.' Narration note: Anthropic's word is 'nearly 1GW', not 'roughly 1GW' — 'nearly' is the safer read since it means just under. Also cite the Anthropic release, not the Google one, for the Trainium figures.

• Cisco Systems closed at $80.06 on 27 March 2000 — its dot-com peak close — the day it passed Microsoft to become the most valuable public company in the world. Intraday high that day was $82.00.

  Verdict: CONFIRMED

  Yahoo Finance daily OHLC for CSCO on 2000-03-27: open 81.4375, high 82.00, low 79.0625, close 80.0625 ($80.06). I checked every session from 1990 to 2026-08-25 — no close exceeded 80.0625 until Dec 2025 and no high exceeded 82.00 until Feb 2026, so both are genuine dot-com peaks. The Microsoft attribution is asserted by both the cited Barchart piece and CNBC ('the same day that Cisco passed Microsoft to become the most valuable publicly traded company in the world'), and the arithmetic independently supports 27 March rather than 24 March: MSFT fell 6.8% that session (55.84 to 52.03 split-adjusted, i.e. ~$112 to ~$104 as-traded on ~5.28B shares, ~$590B to ~$549B) while CSCO rose to ~$556B on ~6.95B shares — the crossover lands on 27 March.

• Cisco's last stock split was a 2-for-1 with an ex-date of 23 March 2000 — four trading days before the peak. There have been no splits since. Therefore $80.06 on 27 March 2000 is simultaneously the as-traded price and the split-adjusted price; the two numbers are identical.

  Verdict: CORRECTED

  The split facts are right, the day count is not. Yahoo's split event feed for CSCO lists nine splits, the last being 2:1 with ex-date 2000-03-23, and nothing after — so the 'as-traded equals split-adjusted' conclusion holds. But 23 March to 27 March is TWO trading days, not four: the intervening sessions are 24 March (Friday) and 27 March (Monday). Four is the calendar-day gap. Narrating 'four trading days' is a checkable error that a listener with a chart will catch.

• Cisco's post-crash bottom was a closing price of $8.60 on 8 October 2002, with an intraday low of $8.12 the same day. That is −89.3% from the peak close and −90.1% from the intraday peak of $82.00.

  Verdict: CONFIRMED

  Yahoo daily bar for 2002-10-08: open 9.21, high 9.21, low 8.12, close 8.60. Scanning all 6,643 sessions from 2000-03-27 through 2026-08-25, that is both the lowest close and the lowest intraday print of the entire post-peak period, so 'the bottom' is accurate, not merely a local low. Arithmetic checks: 8.60/80.0625 − 1 = −89.26% (rounds to −89.3%); 8.12/82.00 − 1 = −90.10% (−90.1%). CNBC's Dec 2025 coverage independently cites 'a split-adjusted $8.60 on Oct. 8, 2002.'

• Cisco first closed above its March 2000 record on 10 December 2025, at $80.25 — 9,389 days, or 25.7 years, after the peak. It did not exceed the $82.00 intraday high of March 2000 until 3 February 2026 (high $83.25).

  Verdict: CONFIRMED

  Scanning every session after 2000-03-27, the first close above 80.0625 is 2025-12-10 at 80.25 (day range 78.80–80.82) — matching CNBC, Barchart and the Globe and Mail, all of which report the $80.25 close on Dec 10 against 'its previous all-time high of $80.06 reached on March 27, 2000.' Day count: 2000-03-27 to 2025-12-10 is exactly 9,389 days = 25.71 years. The first session with a high above 82.00 is 2026-02-03, high 83.25, close 83.11 — consistent with CNBC's 3 Feb 2026 piece headlined that Cisco 'has finally surpassed its dotcom bubble high.'

• On a total-return basis (dividends reinvested), a buyer of Cisco at the 27 March 2000 close broke even on 24 August 2021 — 21.4 years — more than four years earlier than the price-only breakeven of December 2025.

  Verdict: CONFIRMED

  Using Yahoo's dividend- and split-adjusted close (the series behind the cited URL), the 2000-03-27 base is 51.3333. The first session whose adjusted close reaches or exceeds it is 2021-08-24 at 51.5807 (the prior session is still below). Elapsed: 7,820 days = 21.41 years, so '21.4 years' is right, and 2021-08-24 to 2025-12-10 is 4.30 years, so 'more than four years earlier' is right. One caveat worth knowing even though it does not change the verdict: adjusted-close is a reinvestment proxy that assumes dividends are reinvested at the ex-date close with no tax or commission drag. A different total-return series (Bloomberg, CRSP) could shift the crossing date by days. If narration wants to be bulletproof, say 'August 2021' rather than pinning the exact 24th.

• Cisco paid no dividend at all until 29 March 2011, when it began a $0.06 quarterly payout — eleven years after the peak.

  Verdict: CONFIRMED

  Yahoo's dividend event feed for CSCO has exactly 62 entries and the earliest is 2011-03-29 at $0.06, followed by $0.06 on 2011-07-05, 2011-10-04 and 2012-01-03, then a step up to $0.08 on 2012-04-03 — so 'no dividend at all before then' and 'a $0.06 quarterly payout' both hold, and March 2000 to March 2011 is exactly eleven years. Precision note for the script: 29 March 2011 is the EX-dividend date; the cash actually landed on the 20 April 2011 payment date. 'Began paying' or 'declared its first-ever dividend' is safer phrasing than 'paid' if anyone is likely to check the payment date.

• Cisco's revenue grew almost every year of the lost quarter-century: net sales were $18,928M in FY2000, $40,040M in FY2010, $49,301M in FY2020 and $56,654M in FY2025. Net income went from $2,668M (FY2000) to $10,180M (FY2025).

  Verdict: CORRECTED

  All six figures are exactly right, but the framing is not. Verified: the FY2000 10-K income statement shows NET SALES $18,928 and NET INCOME $2,668 (millions); SEC XBRL gives Revenues $40,040,000,000 for FY2010 (period 2009-07-26 to 2010-07-31), and RevenueFromContractWithCustomerExcludingAssessedTax of $49,301,000,000 for FY2020 and $56,654,000,000 for FY2025, with NetIncomeLoss $10,180,000,000 for FY2025. Where the claim breaks is 'grew almost every year': revenue FELL in 7 of the 25 fiscal years — FY2002 (22,293 to 18,915, a ~15% drop), FY2003, FY2009, FY2014, FY2017, FY2020 and FY2024 (56,998 to 53,803). Seven declines out of 25 is not 'almost every year,' and the FY2002 collapse is the single most rhetorically important one for a dot-com story — glossing it as uninterrupted growth is the kind of thing a listener who pulls the 10-Ks will flag. The cherry-picked decade endpoints happen to all be up years, which is what makes the smooth-growth framing sound plausible.

• Cisco's FY2001 (ended 28 July 2001) revenue ROSE 17.8% to $22,293M — and the company still reported a net LOSS of $1,014M, including a $2.25 billion excess-inventory charge and $1.17 billion of restructuring charges. The restructuring, announced 16 April 2001, cut approximately 6,000 regular employees.

  Verdict: CONFIRMED

  Every element verified verbatim in the cited Exhibit 13 (retrieved from SEC with a compliant user agent; WebFetch gets 403). Selected Financial Data: Net sales $22,293M (FY2001) vs $18,928M (FY2000) = +17.78%, rounds to 17.8%. Net income (loss) $(1,014)M. MD&A: 'we recorded restructuring costs and other special charges of $1.17 billion classified as operating expenses' and 'The excess inventory charge recorded in the third quarter of fiscal 2001 was $2.25 billion.' Announcement date: 'On April 16, 2001, due to macroeconomic and capital spending issues affecting the networking industry, we announced a restructuring program.' Workforce: 'we announced a restructuring program to reduce approximately 6,000 regular employees across all business functions.' One optional nuance the narration may want: the $2.25bn charge was later reduced by a $187M benefit in Q4 FY2001, so the income statement carries $2,249M gross / $2,062M net — the claim as written ('including a $2.25 billion excess-inventory charge') matches the filing's own language and is fine.

• Cisco's revenue actually contracted for two years after FY2001 and did not exceed the FY2001 peak until FY2005: $22,293M (FY2001), $18,915M (FY2002), $18,878M (FY2003), $22,045M (FY2004), $24,801M (FY2005).

  Verdict: CONFIRMED

  All five figures verified. The cited FY2004 Exhibit 13 Selected Financial Data gives, left to right, July 31 2004 / July 26 2003 / July 27 2002 / July 28 2001: $22,045 / $18,878 / $18,915 / $22,293. Source caveat worth noting: a filing dated 2004-09-14 cannot contain the FY2005 number, so I verified $24,801M independently in Cisco's FY2005 10-K Exhibit 13 (https://www.sec.gov/Archives/edgar/data/858877/000119312505187168/dex131.htm), whose five-year table reads: $24,801 / $22,045 / $18,878 / $18,915 / $22,293. The narrative logic also holds: FY2002 and FY2003 both declined, FY2004 ($22,045M) still fell short of the FY2001 peak ($22,293M), and FY2005 was the first year to exceed it.

• From the 27 March 2000 peak to 25 August 2026, Cisco returned about 2.16x with dividends reinvested (a 2.97% annual rate over 26.4 years). The S&P 500 total-return index returned 8.18x over the identical window (8.28% annually).

  Verdict: CONFIRMED

  Independently recomputed from daily price series (Yahoo v8 chart API; the finance.yahoo.com HTML page in the citation is not machine-readable, but it is the same underlying data). CSCO: 2000-03-27 dividend-adjusted close 51.3333 (raw close $80.0625, which is indeed the dot-com peak close), 2026-08-25 close $111.11. Ratio = 2.1645x. Elapsed 9648 days = 26.412 years, CAGR = 2.967%. Matches '2.16x', '2.97%', '26.4 years'. ^SP500TR: 2000-03-27 close 2102.31, 2026-08-25 close 17198.59. Ratio = 8.1808x, CAGR = 8.283%. Matches '8.18x' and '8.28%'. Both legs are exact to the stated precision.

• The 'internet traffic doubles every three months' claim that justified the buildout was false. The NSF-instrumented backbone carried about 15 terabytes per month at end-1994; actual US backbone traffic grew at roughly 100% per YEAR from early 1997. A 3-month doubling would have implied 17-million-fold growth between 1994 and 2000. Level 3 CEO James Crowe cited the 3-to-4-month doubling to analysts at the Q3 2000 results meeting as proof that fibre supply could not meet demand.

  Verdict: CONFIRMED

  Every element appears in the cited Odlyzko paper, several near-verbatim. (a) 'Until the end of 1994, the Internet backbone was funded by the National Science Foundation, and was well instrumented. Hence we know that it carried about 15 TB (terabytes) of traffic each month.' (b) 'Since early 1997, the growth rate in traffic has reverted to 100 percent a year.' (c) 'Hence, from the end of 1994 to the end of 2000, it would have grown by a factor of almost 17 million.' (d) 'During the meeting with financial analysts to discuss the report for the 3rd quarter of 2000, James Crowe, head of Level 3 Communications, cited the doubling of Internet traffic every three or fours months as proof that the supply of fiber in the United States could not possibly meet demand within the next few years.' Crowe was indeed Level 3's CEO; the paper says 'head of'. One fairness note for the script: Odlyzko allows that such growth 'did prevail for a short period during 1995 and 1996' — so 'false' is accurate as applied to the sustained multi-year claim, which is how the claim frames it.

• The Federal Reserve Bank of Richmond documents that FCC chairman-era and telco-executive claims of traffic 'doubling every 100 days' persisted throughout the boom while Coffman and Odlyzko's estimates showed roughly annual doubling; that the Nasdaq telecom index peaked at 1,230.06 on 10 March 2000 and fell 62% by year-end 2000; and that CLEC market capitalisation fell from $86.4bn (1999) to $32.1bn (Feb 2001) to $3.77bn (Feb 2002).

  Verdict: CORRECTED

  Three of the four elements are verbatim correct in Couper, Hejkal & Wolman, 'Boom and Bust in Telecommunications,' Richmond Fed Economic Quarterly 89/4, Fall 2003: 'On March 10, 2000, the Nasdaq telecom index peaked at 1,230.06; by the end of 2000 it had fallen by 62 percent'; 'The market capitalization of CLECs fell 63 percent from $86.4 billion in 1999 to $32.1 billion in February of 2001 and then 88 percent to just $3.77 billion in February of 2002'; and Coffman/Odlyzko 'estimate that the amount of data sent over the Internet has approximately doubled every year since then.' The FCC-chairman attribution is NOT supported. The paper attributes 'doubling every 100 days' to a telco executive only: 'throughout the boom major players outside WorldCom, such as Duane Ackerman, CEO of BellSouth, continued to assert that Internet traffic was doubling every 100 days.' The paper names three FCC chairmen (Dennis Patrick, William Kennard, Michael Powell) but none in connection with any traffic-growth figure — Kennard's quoted Feb 2000 remark is about 'the miracle of the American model for unleashing competition.' The FCC-chairman version of this myth (Reed Hundt, 'In 1999, data traffic was doubling every 90 days') is in the Odlyzko paper cited in claim 4, not in the Richmond Fed paper, and it is 90 days, not 100.

• The decrease in market capitalisation of S&P telecommunications firms alone from 2000 to 2002 was roughly $700 billion — more than 3.5% of the entire value of US corporate equities at the 2000 peak. Telecom employment peaked at about 1.59 million in March 2001 and fell 22% to about 1.30 million by July 2003; announced job cuts exceeded 500,000, nearly a third of peak sector employment.

  Verdict: CONFIRMED

  Verbatim in the cited Richmond Fed paper. Conclusion section: 'The decrease in market capitalization of S&P telecommunications firms alone from 2000 to 2002 was roughly $700 billion, more than 3.5 percent of the entire value of U.S. corporate equities at the stock market peak in 2000.' Employment section: 'Telecommunications industry employment (services plus manufacturing) peaked at approximately 1.59 million workers in March 2001. Employment in telecom-related industries declined 22 percent — an average annual decrease of 8 percent — to about 1.30 million by July 2003.' And: 'Announced figures for job cuts have been even more staggering, and media reports have cited numbers of over 500,000. That is nearly one-third of the sector's total employment at its peak.' Nothing to correct.

• JDS Uniphase's fiscal 2001 (ended 30 June 2001) net sales MORE THAN DOUBLED to $3,232.8 million — and it reported a net loss of $56,121.9 million, including a $50,085.0 million reduction in the carrying value of goodwill and other long-lived assets. Loss per share was $(51.40).

  Verdict: CONFIRMED

  All four figures verified in the cited 10-K, in both the Selected Financial Data and the audited Consolidated Statements of Operations. Net sales $3,232.8M (FY2001) vs $1,430.4M (FY2000) — an increase of 126%, so 'more than doubled' is correct. 'Reduction of goodwill and other long-lived assets' $50,085.0M. Net loss $(56,121.9)M. Basic and diluted loss per share $(51.40). The MD&A separately confirms the impairment 'of approximately $50,085.0 million.'

• Corning's revenue nearly doubled from $3.8bn (1997) to about $7bn (2000) on fibre demand, then fell to roughly $4bn in 2002. It lost $5.5bn in 2001, wrote off nearly all $5bn of goodwill from its 2000 acquisitions, and had laid off 12,000 workers — 28% of its workforce — by mid-2002, with 4,000 more planned.

  Verdict: CORRECTED

  Five of six elements check out; the 2002 revenue figure does not. Confirmed from the Forbes piece: 'nearly doubled Corning's revenue from $3.8 billion in 1997 to $7 billion in 2000'; 'Last year Corning wrote off nearly all of the $5 billion in goodwill from its acquisition binge in 2000, leading to a $5.5 billion loss'; 'Some 12,000 workers, or 28% of its workforce, have been laid off, with another 4,000 to go this year.' Corning's own filings corroborate the revenue and loss figures: the FY2001 10-K five-year table shows net sales of $3,831M (1997) and $7,127M (2000), and the FY2002 10-K shows a 2001 net loss of $(5,498)M. The error: the claim states as fact that revenue 'fell to roughly $4bn in 2002,' but the Forbes article was published 19 June 2002 and the $4bn is a forward-looking projection — 'Total revenue this year will likely drop to $4 billion.' Corning's actual 2002 net sales were $3,164M. Its own FY2002 10-K: 'Consolidated net sales for 2002 were $3.2 billion, a decrease of 48%.' Narrating '$4bn' understates the collapse by about 20% and sources a forecast as an outcome.

• Anthropic, Google and Broadcom announced an expanded partnership for multiple gigawatts of next-generation TPU capacity coming online starting in 2027; Anthropic disclosed run-rate revenue above $30B, up from ~$9B at end-2025.

  Verdict: CONFIRMED

  The cited Anthropic newsroom post (April 6, 2026) states the expanded Google/Broadcom partnership covers 'multiple gigawatts of next-generation TPU capacity' coming online starting in 2027, and that run-rate revenue has surpassed $30 billion, up from roughly $9 billion at the end of 2025. Independently corroborated by CNBC (2026-04-06), Tom's Hardware and DataCenterDynamics, which quantify the deal at 3.5 GW of Google TPU capacity from 2027. Narration tip: the specific figure is 3.5 GW, and the >1,000 customers spending $1M+/yr datapoint is also in the release.

• ASML reported Q2 2026 net sales of €9.3B and gross margin of 54.0%, with €3.8B of EUV system sales, and guided full-year 2026 to €43–45B at 54–56% gross margin. It plans ~65 low-NA EUV shipments in 2026 and is adding ~30% capacity for 2027.

  Verdict: CONFIRMED

  ASML's Q2 2026 press release (2026-07-15) confirms €9.3B total net sales, 54.0% gross margin, €2.9B net income, and FY2026 guidance of €43–45B at 54–56% gross margin. Net system sales were €6.6B of which €3.8B was EUV (57% of system sales, including one High-NA unit). One wording nuance: ASML's CEO statement says it will 'add 30% to our 2026 low NA EUV capacity of around 65 for 2027' — the ~65 is stated as annual low-NA EUV *capacity*, not a shipment plan. Narrate as 'capacity of around 65 low-NA EUV systems, being expanded ~30% for 2027' to stay literally accurate.

• Lam Research reported a record June-quarter (ended 2026-06-28) with revenue of $6.722B and non-GAAP gross margin of 52.0%, guiding the September quarter to $8.10B ± $400M.

  Verdict: CONFIRMED

  Lam's 2026-07-29 release for the quarter ended June 28, 2026 reports revenue of $6,722.24M ($6.72B, +15% q/q, +30% y/y), GAAP gross margin 51.7% and non-GAAP gross margin 52.0%, with CEO Tim Archer describing record revenue, operating margin and EPS. September-quarter guidance is $8.10B ± $400M (non-GAAP GM 52% ±1pt, EPS $2.15 ±$0.15). Exact revenue corroborated at $6,722.24M via the 8-K coverage.

• Arista Networks reported Q2 2026 revenue of just over $3.0B (+37.7% y/y) with gross margin of 63.4%, down from 65.6% a year earlier, and attributed guided margin pressure to anticipated supply-chain cost increases for memory and silicon.

  Verdict: CONFIRMED

  The 2026-08-04 press release confirms Q2 2026 revenue of $3.036B, +37.7% y/y, non-GAAP gross margin 63.4% versus 65.6% in Q2 2025 (GAAP 62.9% vs 65.2%). Caveat on sourcing: the memory/silicon attribution is NOT in the press release body — it comes from management's Q2 2026 earnings-call commentary, where Arista tied the margin decline to end-customer mix plus anticipated supply-chain cost increases for memory and silicon, held the FY2026 gross-margin target at 62–64% including those cost increases, and disclosed multiyear purchase commitments raised to $9.7B. Cite the earnings call, not the release, for that sentence.

• Coherent reported Q2 FY2026 revenue of $1.7B (+17.5% y/y) with non-GAAP gross margin of 40.2%; its Datacenter & Communications segment grew 33.6% y/y to $1.2B with segment book-to-bill above 4x.

  Verdict: REFUTED

  This claim splices two different quarters and cites the wrong filing. Coherent's Q2 FY2026 (quarter ended Dec 31, 2025, reported 2026-02-04) had revenue of $1.69B, +17% y/y (+22% pro forma), GAAP gross margin 36.9% and non-GAAP gross margin 39.0% — NOT 40.2%. The 40.2% non-GAAP figure belongs to Q4 FY2026 (quarter ended June 30, 2026, reported 2026-08-12): revenue $2.05B, +34% y/y, GAAP GM 38.5%, Datacenter & Communications $1.62B or 79% of revenue. The cited SEC exhibit (d128030dex991.htm) is that Q4 FY2026 release, and the claim's date of 2026-08-04 matches neither release. Neither release discloses a Datacenter & Communications figure of $1.2B/+33.6%, and no 'book-to-bill above 4x' appears in either — that number is unverified and should not be narrated. For reference, FY2026 full-year Datacenter & Communications revenue was $5,275M, +40%.

• Vertiv reported Q2 2026 net sales of $3,274M (+24% reported, +18% organic), GAAP operating margin 19.5%, adjusted operating margin 22.6% (up 410bps y/y), and raised full-year 2026 guidance to ~$14.0B sales at ~23.8% adjusted operating margin. The release disclosed no backlog or orders figure.

  Verdict: CONFIRMED

  The 2026-07-29 PR Newswire release confirms Q2 2026 net sales of $3,274.3M, +24% reported and +18% organic; GAAP operating margin 19.5% (+270bps y/y); adjusted operating margin 22.6% versus 18.5% in Q2 2025, i.e. +410bps. Raised FY2026 guidance is $13.8–14.2B net sales (~$14.0B midpoint) at 23.3–24.3% adjusted operating margin (~23.8% midpoint). Verified: the release contains no quantitative backlog or orders figure — it references 'growing pipelines' and capacity expansion qualitatively only.

• GE Vernova reported Q2 2026 revenue of $11.1B (+22%), orders of $24.2B (+88% organically), backlog of $176B and adjusted EBITDA margin of 11.3%. Gas Power equipment backlog rose from 100 GW to 116 GW, and data-center-related orders exceeded $5B year-to-date, more than double the full 2025 total.

  Verdict: CONFIRMED

  GE Vernova's 2026-07-22 Q2 release confirms revenue of $11.1B (+22% total, +12% organic), orders of $24.2B (+88% organic, led by Power and Electrification), total backlog of $176B (up $13.0B sequentially), and adjusted EBITDA margin of 11.3% (+340bps organically, FY guide maintained at 12–14%). Gas Power equipment backlog expanded from 100 GW to 116 GW with a stated path to at least 125 GW by year-end 2026, and data-center-related orders exceeded $5B year-to-date, more than double the full-year 2025 total.

• Amazon reported Q2 2026 revenue of $200.6B, AWS revenue of $42.2B (+36.7% y/y) with a 39% AWS operating margin, cash capex of $53.1B in the quarter and $169B on a trailing-twelve-month basis, and raised full-year 2026 capex guidance to ~$220B from ~$200B — citing higher memory costs among the reasons.

  Verdict: CONFIRMED

  Amazon's Q2 2026 release (2026-07-30) confirms net sales of $200.6B (+20% y/y) and AWS net sales of $42.2B; Andy Jassy's quoted commentary specifies 36.7% y/y AWS growth, 'our fastest growth in 18 quarters.' AWS operating income was $16.6B, a 39.4% margin (the claim's '39%' rounds down but is fair). Cash capex of $53.1B in Q2 is confirmed by the earnings-call transcript ('cash CapEx, which is $53.1 billion in Q2'). The $169B TTM cash capex reconciles exactly with the reported figures: TTM operating cash flow of $161.4B minus TTM free cash flow of −$7.6B = $169.0B (gross TTM purchases of property and equipment were $173.0B). FY2026 capex guidance was raised roughly $20B to about $220B, with higher memory chip costs cited by management (CNBC, Seeking Alpha, DIGITIMES coverage of the same call). One coincidence to avoid conflating on air: $169B is also AWS's annualized revenue run rate this quarter.

• Alphabet reported Q2 2026 revenue of $119.8B (+24% y/y), operating income of $40.8B, Google Cloud revenue of $24.8B (+82% y/y) with $514B backlog, Q2 capex of $44.9B (roughly double the year-ago quarter), and raised full-year 2026 capex guidance to $195–205B from $180–190B. Shares fell on the capex raise.

  Verdict: CONFIRMED

  Verified against Alphabet's own 8-K Exhibit 99.1 (SEC): revenues $119,796M, +24% y/y; operating income $40,770M (+30%, margin 34%); Google Cloud $24,768M, +82% (segment op income $8,814M vs $2,826M). Cash-flow statement shows purchases of property and equipment of $44,924M vs $22,446M in Q2 2025 — exactly 2.0x, so 'roughly double' is precise. The $514B Cloud backlog and the capex guidance move to $195–205B from $180–190B come from the earnings call rather than the press release, but are corroborated consistently across CNBC, MLQ and Motley Fool coverage; CNBC's own headline ('GOOGL stock sinks on 2026 capex hike') confirms the negative share reaction. Nothing in the claim overstates the filing.

• Microsoft reported fiscal Q4 2026 capital expenditures including finance leases of $41B (+69% y/y), with FY2026 Azure revenue exceeding $100B for the first time (+41%) and Q4 Azure growth of 43%. In April 2026 Microsoft called for ~$190B of calendar-2026 capital spending, citing soaring memory prices; it later framed calendar 2026 at ~$175B after extending assumed useful life of data center and office properties to 25 years from 15.

  Verdict: CONFIRMED

  Microsoft's FY26 Q4 press release confirms 'Azure revenue surpassed $100 billion for the first time' and 43% Azure growth in the quarter; Nadella's prepared remarks on the Q4 call give the full-year figure as 'Azure surpassed $100 billion, up 41%.' Capex including finance leases of $41B, +69% y/y, is confirmed by CNBC and Yahoo Finance coverage (the press release separately shows $35.8B of 'additions to property and equipment,' the narrower GAAP line, which is consistent). The April 2026 ~$190B calendar-2026 figure citing memory prices matches CNBC's Q3 FY26 report, and Amy Hood's Q4 commentary confirms the useful-life extension to 25 years from 15 and the resulting ~$175B framing. One nuance a narrator should keep straight: Hood explicitly said the ~$175B is not a spending cut — 'outside of this useful life impact, our calendar year 2026 capex investment expectations remain unchanged'; the reduction is an accounting reclassification as more datacenter leases shift from finance to operating leases. The claim's wording ('later framed calendar 2026 at ~$175B after extending...') is accurate, but don't narrate it as Microsoft pulling back on AI spending.

• Meta reported Q2 2026 revenue of $60.80B (+28% y/y), total costs and expenses of $42.03B (+55% y/y), Q2 capital expenditures of $31.1B, Reality Labs operating loss of ~$4.6B, and raised full-year 2026 capex guidance to $130–145B with total 2026 expenses of $165–169B.

  Verdict: CORRECTED

  Every number checks out exactly against Meta's 8-K Exhibit 99.1 (SEC): revenue $60,801M (+28%), total costs and expenses $42,026M (+55%), capex including principal payments on finance leases $31.08B, Reality Labs operating loss $4,619M, full-year 2026 expenses $165–169B, capex $130–145B. The one thing that is wrong is the verb. Meta did not 'raise' its capex guidance to $130–145B — the release says the range was 'narrowed from our prior outlook of $125–145 billion.' The ceiling did not move; only the floor came up. Same pattern on opex, where Meta said it was 'raising the lower-end of our expense outlook' to absorb $2.4B of legal charges. In a course narrating an AI-capex acceleration story, saying Meta 'raised capex guidance' when the company said 'narrowed' materially overstates the signal.

• Super Micro's fiscal Q4 2026 gross margin was 17.5%, up from 9.5% a year earlier and far above its own prior guidance of 8.2–8.4%, with fiscal 2027 revenue guidance as high as $72B.

  Verdict: CONFIRMED

  All four elements verified in Super Micro's SEC filings. The Q4/FY26 earnings release (8-K Ex-99.1, filed 2026-08-11) states 'Gross margin of 17.5% versus 9.9% in Q3'26 and 9.5% in Q4'25,' with the GAAP reconciliation table showing 17.5% vs 9.5%, and 'For fiscal year 2027, the Company expects net sales in the range of $65.0 billion to $72.0 billion' — so 'as high as $72B' is correct. The 8.2–8.4% prior guidance is confirmed verbatim in Supermicro's separate 2026-07-21 preliminary business update (8-K Ex-99.1): 'GAAP and non-GAAP gross margins are estimated to be in the range of 15% to 17% which is significantly higher than our guidance of 8.2% to 8.4%, primarily due to a favorable customer and product mix.' Worth noting for context if the course uses it: full-year FY26 gross margin was only 10.8% (down from 11.1%), so the 17.5% is a single-quarter mix effect, not a run rate.

• Dell reported Q1 FY2027 ISG revenue of $29B (+181% y/y) with ISG operating income of $3.1B, a 10.5% ISG operating margin; AI-optimized server revenue was $16.13B (+757% y/y), with $24.4B of AI orders booked and a record $51.3B AI backlog. Reporting notes total-company gross margin percentage has compressed materially as AI mix has risen.

  Verdict: CONFIRMED

  Verified against Dell's Q1 FY27 10-Q (SEC, filed 2026-06-09) segment note: ISG net revenue $29,009M vs $10,317M (+181.2%); ISG operating income $3,055M (3,055/29,009 = 10.53%, so 10.5% is right); AI-optimized servers $16,132M vs $1,882M (+757.2%). The $24.4B of AI orders is quoted from Jeff Clarke in the press release. Gross-margin compression is confirmed directly in the 10-Q MD&A table: total gross margin 17.8% of revenue vs 21.1% a year earlier — a 3.3-point decline — and Dell's own outlook language cites 'margin rate pressure resulting from a continuing shift in mix towards our AI-optimized servers offerings.' The $51.3B record AI backlog is the one item NOT in the cited press release or the 10-Q; it came from management on the earnings call ('We exited the quarter with a record $51.3 billion of AI backlog... even after converting $24.4 billion into orders'), and is corroborated across Blocks & Files, Yahoo Finance and Dealroom, which also give the prior-quarter figure of ~$43B. It reconciles arithmetically ($43B + $24.4B − $16.1B ≈ $51.3B). If the course cites the press release as the source for the backlog number, attribute it to the earnings call instead.

• CoreWeave reported Q2 2026 revenue of $2.58B (more than double y/y), a GAAP net loss of $626M (versus $290M a year earlier), interest expense of $640M (up from $267M), principal debt obligations of $35.6B, and revenue backlog of $104.2B as of 2026-06-30.

  Verdict: CONFIRMED

  Verified against CoreWeave's 8-K Ex-99.1 and 10-Q (SEC). Press release: revenue $2,575M vs $1,212M (+112%, i.e. more than double); net loss $626M vs $290M; interest expense, net $640M vs $267M. The 10-Q states plainly: 'As of June 30, 2026, our total indebtedness was $35.6 billion,' confirming the debt figure (balance-sheet carrying values of recourse plus non-recourse debt sum to ~$35.1B, consistent with $35.6B of principal). One small precision note: the press release says revenue backlog was 'approximately $104 billion' as of June 30, 2026, not $104.2B — the extra decimal is not in the primary source, so narrate it as 'about $104 billion.' Also worth flagging for accuracy: CoreWeave's footnote says that backlog excludes more than $25B of net new customer commitments added in early Q3, and the definition is broader than GAAP remaining performance obligations.

• EIA projects total US electricity demand rising from a record 4,195 billion kWh in 2025 to 4,269 billion kWh in 2026 and 4,399 billion kWh in 2027 — roughly 1% growth in 2026 and 3% in 2027, the first four consecutive years of growth since 2007.

  Verdict: CORRECTED

  I pulled the actual July 2026 STEO (eia.gov/outlooks/steo/archives/jul26.pdf, modeling completed July 1, 2026). Table 7a, 'Electricity consumption / Total consumption,' gives annual totals of 4,195 / 4,269 / 4,399 billion kWh for 2025 / 2026 / 2027 — the three levels are exactly right. The growth rate for 2026 is wrong: 4,269 from 4,195 is +1.76%, which rounds to roughly 2%, not 'roughly 1%.' The claim's own numbers contradict its own percentage. 2027 is fine: 4,399 from 4,269 is +3.0%. Separately, neither the word 'consecutive' nor 'since 2007' nor a 'record' characterization of the 2025 figure appears anywhere in the July 2026 STEO — that framing is not supported by the cited source and should not be attributed to it. Note also that the live eia.gov/outlooks/steo/ page now serves a later vintage, so citing the bare URL for these numbers will not reproduce them; cite the July 2026 archive.

• Copper reached a record ~US$14,455/tonne on the LME in August 2026, having first passed US$12,000/tonne only in December 2025. J.P. Morgan estimates data-center-related copper demand of roughly 475,000 tonnes of new installations in 2026 and a global refined-copper deficit of around 330,000 tonnes.

  Verdict: REFUTED

  The cited IEA commentary does not support this claim and contradicts part of it. I fetched the page: it is dated 02 March 2026 — five months before the event the claim describes — and its actual wording is that copper prices surged 'briefly exceeding USD 14 500 per tonne (intraday) in January 2026, having only passed USD 12 000 per tonne for the first time in December 2025.' So the source places the record in January 2026 at ~USD 14,500, not August 2026 at USD 14,455. Only the December 2025 / US$12,000 element is confirmed. The J.P. Morgan figures are simply not in the source: the page contains no J.P. Morgan citation, no 475,000-tonne data-centre demand estimate, and no 330,000-tonne deficit — its only deficit figure is an IEA projection of 'a supply deficit of 30% by 2035.' I could find no independent confirmation of either J.P. Morgan number. A separate commercial aggregator (IndexBox) does report a record US$14,455/tonne on 6 August 2026 (first surpassing US$14,000 in May 2026), so an August record may well be real on a closing-price basis, but that is a weak secondary source that conflicts with the IEA's January intraday level, and it is not the source cited. Do not narrate this as written.

• Campbell and Turner's daily index of British railway share prices, based at 1,000 in January 1843, peaked at 2,017 on 9 August 1845 and fell to a low of 672 on 16 April 1850 — a peak-to-trough decline of 66.7%. By December 1850 the index was still 14.3% BELOW its January 1843 starting level, i.e. the eight-year holder of railway shares lost money outright.

  Verdict: CORRECTED

  Every NUMBER is exact — verified verbatim in the cited MPRA paper (Campbell & Turner, "'The Greatest Bubble in History': Stock Prices during the British Railway Mania", MPRA 21820, p.11): 'The index peaked at 2,017 on 9th August 1845 ... reaching a low of 672 on 16th April 1850. Overall, the index fell by 14.3 per cent between January 1843 and December 1850, and from peak to trough, prices fell by 66.7 per cent.' Summary table also confirms Max 2017.4 / Min 672.1. BUT the concluding inference is not supported and must not be narrated: the paper's own figure notes state 'Capital gains for each company are weighted by market capitalization to produce a daily market index' — it is a PRICE index that EXCLUDES dividends. The paper elsewhere notes established railways were 'increasing their dividends substantially' during the boom before cutting them in the downturn. A 14.3% capital loss over eight years does not establish that holders 'lost money outright'; total return is not computed in this paper. Separately, be aware a rival index (used by the Cambridge Group's railway atlas) puts the peak at 1,984 on 8 Aug 1845 and the nadir at 673 on 19 Apr 1850 — that is Campbell & Turner's sub-index of railways established BEFORE 1843, not a contradiction, but do not mix the two series.

• The British railway network the mania financed was real and enduring: route mileage went from about 1,775 miles (1840) to about 6,890 miles (1850) to 15,537 miles open for traffic by 1870, while annual passenger journeys rose from 28 million (1844) to 288 million (1870). In 1846 alone Parliament passed roughly 272 railway Acts authorising about 9,500 miles of new line.

  Verdict: CORRECTED

  The cited parliament.uk 'Fire and steam' page supports NONE of these figures and contradicts one of them. Retrieved full text of that page gives only: '8,590 miles of railway were authorised in 1845-7' and 'In 1846, more than 700 railway Bills were introduced' and 'By 1910 Great Britain had 20,000 miles of railway.' So '9,500 miles authorised in 1846 alone' conflicts with the cited source's 8,590 miles across three years, and '272 Acts' is not in it. Wikipedia's Railway Mania article currently reads '263 Acts of Parliament for setting up new railway companies were passed, with the proposed routes totalling 9,500 miles' — 263, not 272. VERIFIED: the passenger figures are exact, from UK Parliament's 'Plight train' page: 'The 28m journeys taken in 1844 had increased to 288m by 1870.' NOT VERIFIED: 1,775 miles (1840) — Grace's Guide, citing Bradshaw's Railway Companion 1840, gives 1,646 miles, and other standard accounts give roughly 1,500; the 1,775 figure traces only to a railway-enthusiast forum post. 6,890 miles (1850) — unsourced, and UK Parliament's own 'Plight train' page says Great Britain had 'less than 5,000 miles' in 1848. 15,537 miles (1870) — could not be found on the cited page or any authoritative source; it circulates without attribution. The mileage series may be a UK-including-Ireland series from the Railway Year Book 1915, but I could not confirm that.

• The American version repeated twice. After Jay Cooke & Co failed on 18 September 1873 over Northern Pacific railroad bonds, 89 of the country's 364 railroads went bankrupt and by 1877 about 20% of US rail mileage was in receivership. In the twelve months to June 1894, following the Panic of 1893, over 125 railroads went into receivership; ultimately roughly one quarter of the nation's rail mileage passed into receivers' hands.

  Verdict: CONFIRMED

  Two load-bearing figures verified against the cited source and a second source. PBS American Experience (the cited URL): 'the banking firm of Jay Cooke and Company, a firm heavily invested in railroad construction, closed its doors on September 18, 1873' and 'A startling 89 of the country's 364 railroads crashed into bankruptcy.' Wikipedia's Panic of 1873 confirms the 18 September bankruptcy date; Wikipedia's Long Depression corroborates the 89 railroads. For 1893-94, Wikipedia's Panic of 1893 states 'one-fourth of all rail mileage went into receivership,' matching the 'roughly one quarter' claim. CAVEATS the narrator should know: (a) the '20% of US rail mileage in receivership by 1877' figure circulates widely but I could not trace it to a primary or scholarly source — it is not in the cited PBS page, Wikipedia's Panic of 1873, or Long Depression; (b) 'over 125 railroads in the twelve months to June 1894' is a soft floor and I could not confirm the exact count for that specific twelve-month window. The harder, better-documented number for the same moment is that as of 30 June 1894, 192 US railroads covering 40,819 miles — about a quarter of the network — were in receivership. Prefer that figure if you want a citable one.

• Cisco's own Visual Networking Index put global IP traffic at 122 exabytes per month in 2017 (1.5 zettabytes/year), rising to a projected 255 EB/month in 2020 and 396 EB/month in 2022 (4.8 ZB/year). Against the ~15 terabytes per month the NSF-instrumented backbone carried at end-1994, the technology forecast of the 1990s was directionally vindicated many times over.

  Verdict: CORRECTED

  Two numbers are off. (1) The 2020 figure is 254, not 255. Table 8 of the cited Cisco VNI Forecast and Trends 2017-2022 white paper gives 'Total IP traffic 122 / 156 / 201 / 254 / 319 / 396' EB per month for 2017-2022. (2) The NSFNET figure is 16.3, not ~15. Odlyzko's Table 1 ('Traffic on Internet backbones in U.S. ... estimated traffic in terabytes during December of that year') gives 1994 = 16.3 TB/month, and states the 1990-94 data 'is taken from the trustworthy statistics for NSFNET, the original backbone funded by NSF.' CONFIRMED exactly: the 2017 and 2022 figures, quoted verbatim from the white paper's executive summary — 'Annual global IP traffic will reach 4.8 ZB per year by 2022, or 396 Exabytes (EB) per month. In 2017, the annual run rate for global IP traffic was 1.5 ZB per year, or 122 EB per month.' Note also that 255/254 EB for 2020 was a FORECAST made in 2018, not an observed outcome — say 'projected', as the claim does.

• Nortel Networks filed for creditor protection on 14 January 2009. At its July 2000 peak it was reported at about C$367 billion in market value and represented over a third of the value of Toronto's TSE 300 index; in June 2001 it announced a quarterly loss of about $19 billion.

  Verdict: CORRECTED

  Three of four elements check out; the market-value figure does not survive scrutiny. CONFIRMED from the cited Canadian Encyclopedia article: 'Nortel's extended and painful period of corporate downsizing culminated in the company filing for bankruptcy protection on 14 January 2009' and 'in July 2000, Nortel's common shares reached a historic high of $124.50 on the Toronto Stock Exchange (TSE) and represented over 35 per cent of the value of Toronto's TSE 300 index' — so 'over a third' is right, and 'over 35 per cent' is the sharper number. CONFIRMED elsewhere: on 15 June 2001 Nortel warned it expected a second-quarter loss of $19.2 billion (it reported $19.4 billion on 19 July 2001) — so 'about $19 billion' is fair, but say 'warned of' or 'announced it expected', since the loss was reported in July. NOT CONFIRMED: 'about C$367 billion.' That figure appears nowhere in the cited Canadian Encyclopedia article, which gives no dollar market-cap figure at all; the only sourcing I could find for C$367bn is a low-quality commercial blog. Wikipedia's Nortel article instead states 'Nortel's market capitalization fell from C$398 billion in September 2000 to less than C$5 billion in August 2002.' The peak market cap is genuinely contested (roughly C$366-398bn depending on the date and share count used), so pin the claim to what the encyclopedia actually says rather than to a specific dollar figure.

• NVIDIA's most recent REPORTED quarter is Q1 FY2027 (quarter ended April 26, 2026, reported May 20, 2026): revenue $81,615M (+85% y/y, +20% q/q); Data Center revenue $75,246M (+92% y/y); GAAP gross margin 74.9%; GAAP diluted EPS $2.39; non-GAAP diluted EPS $1.87. Under the new market-platform reporting, Hyperscale was $37,869M and 'AI Clouds, Industrial & Enterprise' (ACIE) was $37,377M — i.e. hyperscalers are now only ~50% of Data Center revenue.

  Verdict: CONFIRMED

  Every figure verified against the cited SEC-filed press release (EX-99.1 to the 8-K filed 2026-05-20, accession 0001045810-26-000051) and the accompanying CFO Commentary (EX-99.2). Press release: 'reported record revenue for the first quarter ended April 26, 2026, of $81.6 billion, up 20% from the previous quarter and up 85% from a year ago', GAAP/non-GAAP gross margins '74.9% and 75.0%', GAAP/non-GAAP diluted EPS '$2.39 and $1.87'; summary table shows Revenue $81,615M. CFO Commentary's 'Revenue by Market Platform' table gives Data Center $75,246M (+21% q/q, +92% y/y), Hyperscale $37,869M, 'AI Clouds, Industrial, & Enterprise' $37,377M, Edge Computing $6,369M. The ~50% split is NVIDIA's own words: 'Hyperscale revenue increased sequentially and remained at approximately 50% of Data Center revenue, while the remaining 50% came from a continued diversification of customers.' (37,869/75,246 = 50.3%.) 'Most recent REPORTED quarter' also verified: I pulled NVIDIA's full EDGAR submissions index and there is no Item 2.02 (results of operations) 8-K after 2026-05-20 as of today, 26 August 2026. WARNING FOR THE COURSE: this is about to go stale — NVIDIA reported Q2 FY2026 on 27 August 2025, so Q2 FY2027 results are likely days away. Date-stamp the narration.

• NVIDIA guided Q2 FY2027 revenue to $91.0 billion ±2%, with GAAP/non-GAAP gross margins of 74.9%/75.0% ±50bp, and stated explicitly: 'We are not assuming any Data Center compute revenue from China in our outlook.'

  Verdict: CONFIRMED

  Verified verbatim in the cited CFO Commentary (EX-99.2, accession 0001045810-26-000051): 'Outlook for the second quarter of fiscal 2027 is as follows: Revenue is expected to be $91.0 billion, plus or minus 2%. We are not assuming any Data Center compute revenue from China in our outlook. GAAP and non-GAAP gross margins are expected to be 74.9% and 75.0%, respectively, plus or minus 50 basis points.' The quotation is word-for-word correct including the first-person phrasing (the press release carries the third-person variant, 'NVIDIA is not assuming any Data Center compute revenue from China in its outlook' — quote the CFO Commentary if you use 'we'). Supporting context worth adding: the same commentary states 'No shipments of Data Center Hopper products to China occurred during the quarter, compared with $4.6 billion in the first quarter of fiscal year 2026', so the zero-China assumption reflects an already-realised zero, not a new prospective haircut.

• NVIDIA's current shipping architecture generation is Blackwell — specifically 'Blackwell 300' products, which drove the Q1 FY2027 Data Center ramp, with NVIDIA stating 'our Blackwell architecture remains the majority of our revenue.' The next generation, Vera Rubin (including the Vera CPU and BlueField-4 STX), was ANNOUNCED in Q1 FY2027 but is not yet the revenue driver.

  Verdict: CONFIRMED

  Both quotations verified verbatim in the cited Form 10-Q for the quarter ended April 26, 2026 (accession 0001045810-26-000052): 'Data Center revenue was $75.2 billion, up 92% from a year ago and up 21% sequentially, driven by the ramp of our Blackwell 300 products and demand for our InfiniBand, Spectrum-X Ethernet, and NVLink solutions', and 'Gross margin was approximately flat sequentially as our Blackwell architecture remains the majority of our revenue.' Both sentences also appear in the CFO Commentary. The Vera Rubin point is confirmed by the Q1 FY2027 press release Data Center highlights: 'Announced the NVIDIA Vera Rubin platform, including the NVIDIA Vera CPU, the world's first processor purpose-built for agentic AI, and NVIDIA BlueField-4 STX, accelerated storage infrastructure for agentic AI factories.' One nuance for accuracy: Vera Rubin was more than a slideware announcement in the quarter — NVIDIA also disclosed 'new NVIDIA Vera Rubin-powered A5X instances' with Google Cloud — but the claim's substance, that Blackwell (not Rubin) is the majority of revenue, is exactly what NVIDIA states.

• NVIDIA reported Q1 fiscal 2027 revenue of $81.6B (+85% y/y), Data Center revenue of $75.2B (+92% y/y), GAAP gross margin 74.9% / non-GAAP 75.0%, and GAAP net income of $58.3B. It guided Q2 FY27 to ~$91.0B revenue at ~75% gross margin.

  Verdict: CONFIRMED

  Every figure matches NVIDIA's Q1 FY27 press release (EX-99.1 to the 8-K filed 2026-05-20, SEC accession 0001045810-26-000051): revenue $81,615M up 85% y/y; record Data Center revenue $75.2B up 92% y/y (compute $60.4B +77%, networking $14.8B +199%); GAAP GM 74.9% / non-GAAP 75.0%; GAAP net income $58,321M (+211% y/y). Outlook: 'Revenue is expected to be $91.0 billion, plus or minus 2%' with GAAP/non-GAAP gross margins of 74.9%/75.0% ±50bp. One narration-worthy caveat: the outlook explicitly assumes zero Data Center compute revenue from China.

• NVIDIA's Q1 FY2027 10-Q discloses that three direct customers each exceeded 10% of total revenue, at 21%, 17% and 16% — roughly 54% of revenue from three buyers.

  Verdict: CONFIRMED

  Verified verbatim in the 10-Q (nvda-20260426.htm): 'Direct Customers – For the first quarter of fiscal year 2027, three direct customers represented 21%, 17%, and 16% of total revenue, all of which was primarily attributable to the Compute & Networking segment.' 21+17+16 = 54%. Prior-year comparative was two customers at 16% and 14%. Note for accuracy on air: these are direct customers (OEMs/ODMs/system integrators/CSPs that buy from NVIDIA), not necessarily the end buyers of the compute — the same 10-Q separately discloses accounts-receivable concentration of 30%/18%/16%.

• Micron reported fiscal Q3 2026 (quarter ended 2026-05-28) revenue of $41.46B versus $23.86B the prior quarter, with gross margin of 84.9%, and guided fiscal Q4 to ~$50B revenue at ~86% gross margin. DRAM was $31.3B, ~76% of revenue, with ASPs up in the low-60s percent range q/q.

  Verdict: CONFIRMED

  Press release (EX-99.1, filed 2026-06-24) confirms revenue $41,456M vs $23,860M prior quarter, quarter ended May 28 2026, and FQ4-26 guidance of '$50.0 billion ± $1.0 billion' at 'Approximately 86%' gross margin. DRAM detail confirmed in Micron's own FQ3-26 prepared remarks: 'Fiscal Q3 DRAM revenue was a record $31.3 billion, up 343% year over year, and represented 76% of total revenue... Prices increased in the low-60s percentage range.' One precision point: 84.9% is the NON-GAAP gross margin; GAAP gross margin was 84.6%. Say 'non-GAAP' or use 84.6% if the script frames it as a reported GAAP figure.

• SK hynix reported Q2 2026 revenue of KRW 79.32 trillion and operating profit of KRW 60.54 trillion — a 76% operating margin, an all-time high — driven by HBM, AI-server DRAM and eSSD. HBM4 entered mass shipment in the quarter.

  Verdict: CONFIRMED

  SK hynix's own newsroom release (July 29, 2026) states revenues of 79.3187 trillion won and operating profit of 60.5426 trillion won 'with an operating margin of 76%... marking an all-time high quarterly performance,' attributing it to 'high-value-added products, including HBM, DRAM for AI servers, and eSSD.' HBM4 confirmed: 'The company began mass shipments of HBM4 in the second quarter and will ramp up production in the second half of the year.' Y/Y comparison for context: Q2 2025 revenue 22.232T won, OP 9.2129T won (+257% / +557%).

• Samsung Electronics' Device Solutions (semiconductor) division posted Q2 2026 revenue of KRW 127.5 trillion and operating profit of KRW 89.2 trillion, an operating margin of roughly 70%, while the mobile business swung to an operating loss.

  Verdict: CONFIRMED

  Samsung Global Newsroom release confirms: 'The DS Division posted KRW 127.5 trillion in consolidated revenue and KRW 89.2 trillion in operating profit for the second quarter' (89.2/127.5 = 70.0%), against group totals of KRW 171.5T revenue / KRW 89.5T operating profit. The mobile loss is confirmed verbatim: 'The MX and Networks Businesses posted KRW 33.2 trillion in consolidated revenue and reported an operating loss of KRW 0.7 trillion for the second quarter.' Two small notes: the release is dated July 30, 2026 (the claim metadata says July 29), and the reported loss line is MX *and Networks* combined, so 'the mobile business' is a slight simplification.

• TSMC reported Q2 2026 revenue of US$40.20B (NT$1,270.38B), gross margin 67.7% and operating margin 60.3% — both records. HPC was 66% of revenue; advanced nodes (7nm and below) were 77% of wafer revenue. Full-year 2026 capex was guided to $60–64B.

  Verdict: CONFIRMED

  All figures check against TSMC's 2Q26 earnings release and call transcript: 'consolidated revenue of NT$1,270.38 billion'; 'second quarter revenue was $40.20 billion'; 'Gross margin for the quarter was 67.7%, operating margin was 60.3%'; 'Advanced technologies, defined as 7-nanometer and more advanced technologies, accounted for 77% of total wafer revenue'; on the call, 'HPC increased 20% quarter over quarter to account for 66% of our second-quarter revenue'; and CFO Wendell Huang: 'decided to raise our full-year 2026 capital budget to be between USD60 billion and USD64 billion.' Caveat on wording only: neither the release nor the call calls the margins 'records' — GM rose 1.5pts and OM 2.2pts sequentially, which makes them highs in this series, but 'both records' is your inference, not TSMC's language. Safer phrasing: 'the highest margins TSMC has posted.'

• TSMC management stated that advanced packaging capacity is so tight it is limiting customers' growth, and is scaling CoWoS reticle sizes (5.5-reticle in production planning, 14-reticle targeted for 2028).

  Verdict: CORRECTED

  The first half is confirmed verbatim — CEO C.C. Wei on the 2Q26 call: 'our packaging capacity is so tight that now it limits my customers' growth.' The reticle specifics are not supported. The cited Tech Times article contains zero mentions of 'reticle' (I searched the full text). The only reticle discussion on the call is an analyst noting TSMC 'previously already announced 14 times reticle CoWoS roadmap,' and IR's Jeff Su saying 'we roadmap to even larger than 14x reticle size with CoWoS.' There is no 5.5-reticle figure anywhere in the quarter's materials, and no 2028 date is attached to the 14x roadmap — 14x is the already-announced baseline TSMC says it will exceed, not a 2028 target. Do not narrate the parenthetical as stated.

• Broadcom reported Q2 FY2026 AI semiconductor revenue of $10.8B, up 143% y/y, guided Q3 FY26 AI revenue to ~$16B, and guided full fiscal 2026 AI semiconductor revenue to ~$56B (roughly +180% y/y), with fiscal 2027 AI revenue guided in excess of $100B.

  Verdict: CONFIRMED

  The first two elements are verbatim in the press release (EX-99.1 to the 8-K filed 2026-06-03, SEC accession 0001730168-26-000051), quoting Hock Tan: 'Q2 semiconductor revenue from AI of $10.8 billion grew 143% year-over-year... in Q3 we expect semiconductor revenue from AI to grow over 200 percent year-over-year to $16.0 billion.' Total Q2 revenue was $22,187M (+48%). The $56B and $100B figures are NOT in the press release — they came from the earnings call, and I confirmed both against contemporaneous coverage of that call: 'AI semiconductor revenue growth of 180% year over year to $56 billion' for FY26 and '$100 billion AI revenue guidance for fiscal 2027, which would be about 78% year-over-year growth.' A Motley Fool piece on the same call adds that management chose to 'reiterate rather than raise' the FY27 target, and it is a stated goal ('over $100 billion by fiscal 2027') rather than formal quarterly guidance — prefer 'target' over 'guided' for the FY27 number.

• Corning's split-adjusted closing low was $1.10 on 8 October 2002 — −99.0% from its $113.33 close on 1 September 2000. It did not close above the 2000 high again until 6 February 2026, 25.4 years later.

  Verdict: CONFIRMED

  Verified against Yahoo Finance daily OHLC pulled directly (split-adjusted for the Oct-2000 3:1 split). 1 Sep 2000 close = $113.3333 (also the max close of 2000-2002 and that day's intraday high). 8 Oct 2002 close = $1.10, the minimum close AND minimum intraday low of the window; 1.10/113.3333 − 1 = −99.03%. First close above $113.33 thereafter = 6 Feb 2026 at $122.16 (prior closes: 5 Feb $112.79, 3 Feb $112.79, 27 Jan $109.74 — all below). Elapsed 1 Sep 2000 → 6 Feb 2026 = 25.43 years, so '25.4 years' is right. Note for narration: this is a price-only statement; on a total-return (dividend-reinvested) basis breakeven came earlier.

• Corning's 2025 results were records: GAAP net sales $15.63bn (2024: $13.118bn), with Optical Communications sales of $6.274bn, up 35% year on year, plus a multiyear agreement with Meta worth up to $6 billion to support US data-centre buildout.

  Verdict: CONFIRMED

  Verified against the primary SEC document, not just the IR page: Corning 8-K filed 28 Jan 2026 (accession 0000024741-26-000034), exhibit glw-20251231xex99xq42025.htm. FY2025 GAAP net sales $15,629M vs FY2024 $13,118M (+19%); Optical Communications FY2025 net sales $6,274M vs FY2024 $4,657M, stated as +35% Y/Y; release headlines 'Corning and Meta Announce Multiyear, Up to $6-Billion Agreement' for next-generation US data centers. One narration caution: $15.63bn is the GAAP figure — Corning's own headline 'record' number is CORE sales of ~$16.4bn (+13%), so do not mix the two in the same sentence.

• WorldCom filed for Chapter 11 on the night of Sunday 21 July 2002, listing over $107 billion in assets — the largest corporate bankruptcy in US history at the time, far exceeding Enron's the previous December. It had admitted in late June to failing to account for some $3.8 billion of expenses over five quarters (the figure later grew), employed 60,000 people and had laid off nearly 17,000 the month before.

  Verdict: CONFIRMED

  Every element matches the cited PBS NewsHour report of 22 July 2002: filed Sunday night, 'over $107 billion in assets', far exceeding Enron's December filing, admission in late June of failing to account for $3.8 billion in expenses over the last five quarters, 60,000 workers, nearly 17,000 laid off the previous month. 21 July 2002 was indeed a Sunday. The parenthetical '(the figure later grew)' is correct — the restatement ultimately reached roughly $11bn.

• Global Crossing filed for Chapter 11 on 28 January 2002, the fourth-largest US bankruptcy to that date, after spending roughly $15 billion in about five years building fibre-optic networks. It emerged in 2003 with Singapore Technologies Telemedia taking 61.5% for $250 million, and was finally acquired by Level 3 Communications in a stock deal valued at about $3 billion including $1.1 billion of assumed net debt, completed October 2011.

  Verdict: CONFIRMED

  Verified: (a) filing date — Global Crossing's own 10-K (SEC accession 0001193125-05-052933) states 'We filed for bankruptcy protection on January 28, 2002'; (b) fourth-largest at that date — its $30.185bn of assets ranked behind only Enron ($63.4bn, Dec 2001), Texaco ($35.9bn, 1987) and Financial Corp. of America ($33.9bn, 1988), ahead of PG&E ($29.8bn); (c) emergence and STT stake — the same 10-K states the reorganization value 'was determined pursuant to our Plan of Reorganization and ST Telemedia's $250 million equity investment for 61.5% ownership', with the plan effective 9 December 2003; (d) Level 3 — TechCrunch (11 Apr 2011) gives ~$3bn all-stock at $23.04/share including ~$1.1bn of assumed net debt, and the deal closed 3 October 2011. CAVEAT — the one figure I could NOT source independently is 'roughly $15 billion in about five years'. It is a widely repeated press round number, and SEC filings put cumulative network capex in the same ballpark (2001 alone was $2.643bn; PP&E net peaked near $12bn as originally reported), but I found no primary source stating it. Keep it hedged ('roughly', 'about'), or attribute it, rather than presenting it as a precise figure.

• Amazon's peak close was $106.69 as traded on 10 December 1999 (intraday high $113.00 on 9 December), equivalent to $5.334 after the June 2022 20-for-1 split. Its trough close was $5.97 as traded on 28 September 2001, with an intraday low of $5.51 on 1 October 2001 — a drawdown of −94.4% on closes and −95.1% intraday. It first closed above the 1999 record on 23 October 2009, 9.9 years later.

  Verdict: CONFIRMED

  Verified against Yahoo Finance daily OHLC pulled directly. Max close of the era = 10 Dec 1999 at $5.334375 split-adjusted → ×20 = $106.6875 ≈ $106.69. Max intraday high = 9 Dec 1999 at $5.65 → $113.00 (10 Dec's high was only $112.00). Min close of 2001-02 = 28 Sep 2001 at $0.2985 → $5.97. Min intraday low = 1 Oct 2001 at $0.2755 → $5.51. Drawdowns: 5.97/106.69 − 1 = −94.40%; 5.51/113.00 − 1 = −95.12%. First close above $5.334375 = 23 Oct 2009 at $5.9245 (21 and 22 Oct were $4.67). Elapsed 10 Dec 1999 → 23 Oct 2009 = 9.87 years. The June 2022 20-for-1 split is correct (effective 6 June 2022).

• Amazon's revenue grew straight through the crash — $1,639.8M (1999), $2,762.0M (2000), $3,122.4M (2001) — and reached $716,924M in 2025. A buyer at the December 1999 peak had made about 49x on a total-return basis by 25 August 2026.

  Verdict: CONFIRMED

  Revenue verified in the cited FY2001 10-K itself (SEC accession 0001032210-02-000059): 'Net sales $3,122,433 / $2,761,983 / $1,639,839' (thousands) for 2001/2000/1999 — the claim's rounding is correct. FY2025 net sales of $716,924M verified via SEC XBRL company-concept data for CIK 1018724 (CY2025 frame, 10-K). Multiple: AMZN closed $261.06 on 25 Aug 2026; 261.06 / 5.334375 = 48.94x, so 'about 49x' is right, and because Amazon has never paid a dividend, total return equals price return here.

• Microsoft closed at $59.5625 split-adjusted on 27 December 1999 (as traded, $119.13, before the February 2003 2-for-1 split); intraday high $59.97 on 30 December 1999. Its closing low over the following decade was $15.15 on 9 March 2009, −74.6% from the peak. It did not close above the 1999 record until 21 October 2016 — 16.8 years — with a total-return breakeven on 16 July 2014, 14.6 years.

  Verdict: CONFIRMED

  Verified against Yahoo Finance daily OHLC plus adjusted close, pulled directly. Max 1999 close = 27 Dec 1999 at $59.5625 split-adjusted → ×2 = $119.125 (the claim's $119.13). Max 1999 intraday high = 30 Dec 1999 at $59.96875 ≈ $59.97. Min close 2000-2009 = 9 Mar 2009 at $15.15; 15.15/59.5625 − 1 = −74.56%, i.e. −74.6%. First close above $59.5625 = 21 Oct 2016 at $59.66 (20 Oct was $57.25) — 16.82 years, i.e. 16.8. First adjusted close (dividends reinvested) above the 27 Dec 1999 adjusted close of 36.1547 = 16 Jul 2014 at 36.8728 — 14.55 years, i.e. 14.6. The February 2003 2-for-1 split is the only split between the two dates, so the as-traded/adjusted relationship is exactly 2x as stated.

• Microsoft's revenue grew from $22,956M in FY2000 to $77,849M in FY2013 (3.4x) and net income from $9,421M to $21,863M (2.3x). Yet the total return from the 27 December 1999 peak close through 31 December 2013 was −14.7%.

  Verdict: CONFIRMED

  FY2000 figures verified in Microsoft's FY2000 10-K (SEC accession 0001032210-00-001961): five-year selected data shows Revenue $22,956 and Net income $9,421 for FY2000. FY2013 figures verified in the cited FY2013 10-K (accession 0001193125-13-310206): 'Revenue $77,849 ... Net income $21,863'. Ratios: 77,849/22,956 = 3.39x (3.4x) and 21,863/9,421 = 2.32x (2.3x). Total return computed from Yahoo adjusted closes: 30.8429 (31 Dec 2013) / 36.1547 (27 Dec 1999) − 1 = −14.69%, i.e. −14.7%.

Community

Talk it through

Now train it

Take it further