Load-bearing claims in this essay passed 42 groups × 3 adversarial votes (126 votes: 13 groups clean, 29 with calibration corrections, 0 overturned), with contradiction-search and methods-audit seats added for 3 single-source empirics (6 verdicts: one "single-source" number independently rebuilt from five SEC filings, one vendor-reported figure vetoed as load-bearing, one primary source recovered but with one of its numbers restricted). All figures as of July 2026; capex and revenue move quarterly — mind the timestamp.
In September 2023, Sequoia's David Cahn published "AI's $200B Question." The algorithm fits on a napkin: take Nvidia's run-rate data center revenue, multiply by 2 (GPUs are roughly half the total cost of ownership of a data center), multiply by 2 again (assume end users need a 50% gross margin) — and you get the lifetime revenue this year's GPUs must earn to pay back: $200 billion. Subtract his generous assumptions for big tech's incremental AI revenue, and a $125B+ annual hole remains.
In June 2024 the same algorithm produced "AI's $600B Question," with the hole growing to roughly $500 billion. On July 8, 2026, came "AI's $1.5T Question": a single year's capex now implies roughly $1.5 trillion in required lifetime revenue, roughly $3 trillion cumulatively. Mind the yardstick: $200B→$600B→$1.5T is the gross requirement series; $125B→$500B is the net hole after subtracting revenue, and Cahn published no net-hole number in 2026 — reading $1.5T as "the hole tripled again" is a common misreading. Two things matter more. Cahn never calls it a bubble — his words: "Speculative frenzies are part of technology, and so they are not something to be feared." And in 2026 he turned the criticism on himself: the GPU-only math "will increasingly underestimate" future revenue requirements, because it omits TPUs/ASICs and increasingly expensive memory.
An arithmetic problem that compounds for three years naturally leads everyone to the same analogy: the telecom and fiber bubble of 1999. Both sides use it — bears say the capex curve rhymes, bulls say "this time the revenue is real." This essay does three things: gets 1999 right first (it is misremembered to a degree most people don't suspect); lays out this cycle's ledger with its yardsticks straightened; then runs the analogy through a physical — where it holds, where it breaks, and what both sides pretend not to see.
How much did US telecom capex actually burn? Circulating figures run from "$500 billion" to "$1 trillion." Lay the yardsticks out and a ladder appears.
The cleanest primary source is the Richmond Fed's 2003 post-mortem (Couper, Hejkal & Wolman): US communications equipment investment (in 1996 dollars) grew from about $62 billion a year in 1996Q1 to over $135 billion a year by 2000Q4 — nearly 18% average annual growth; its share of total private investment approached 7% in 2000 and fell back to 4.8% by the end of 2002. Telecom structures investment (the civil works of laying fiber) jumped $9 billion in 1999 alone. Then the collapse: equipment investment bottomed in 2001Q4 below $93 billion — 69% of its level a year earlier.
The "$500 billion" is Om Malik's tally of new bonds issued by telecom companies in 1996-2001 — financing raised, not physical capital deployed. The "$1 trillion" is a wide-yardstick narrative that stacks debt, acquisitions, and equipment with double counting. At the narrow end, Odlyzko offered the most restrained estimate: rebuilding America's long-haul fiber infrastructure would cost roughly $150-300 billion in total, of which the "wasteful and wasted" portion came to about $100 billion — roughly 0.8-1% of 2000 US GDP (the ratio is our derivation, not his). Same bubble, different yardsticks, a 5-10x spread. The lesson transfers directly: for any "AI capex vs 1999" comparison, first ask what the numerator measures.
The most famous "fact" of 1999: internet traffic doubles every 100 days. Level 3's CEO James Crowe used it to argue America's fiber supply could not possibly meet demand; former FCC chairman Reed Hundt put it in a book; WorldCom was among its most energetic distributors.
Andrew Odlyzko of AT&T Labs debunked it in 2000, in real time: actual US backbone traffic doubled roughly annually (70-150% a year), not every three to four months (which would mean 700-1,500% a year). Were "doubling every 100 days" true, traffic would have had to grow roughly 17 million-fold from end-1994 to end-2000 — physically impossible. He supplied the utilization mechanism too: on the SWITCH transatlantic link, from May 1996 to October 2000, traffic grew about 87% a year while capacity grew about 144% — supply persistently outran demand, and utilization fell structurally. His verdict was on the record at the time: "The myth ... is dangerous ... It surely helped inflate the current bubble in optical networking stocks."
While we're calibrating, fix another circulating number: "only 2.7% of fiber was lit." That precise value has no clean primary source — 2%, 2.7%, 5%, and 11% each have different origins with clashing definitions (lit strands vs carried traffic; all fiber vs newly laid). The honest statement: utilization sat somewhere between single digits and about a tenth, with the direction anchored by capacity growth far outrunning traffic growth.
The other signature of the 1999-2002 bubble era: where the demand story fell short, accounting stepped in. WorldCom overstated assets/pre-tax income by more than $11 billion in total — of which the piece that maps directly onto "disguising operating expense as capex," the line-cost capitalization, was about $3.5 billion (calling the full $11B "capitalization" is a common threefold exaggeration). It filed for bankruptcy on July 21, 2002, listing $107 billion in assets and $41 billion in debt — then the largest in US history. Qwest was found by the SEC to have fraudulently recognized over $3.8 billion in revenue in 1999-2002. Global Crossing inflated revenue through capacity-swap round-tripping and went bankrupt in January 2002.
Vendor financing by equipment makers was the bubble's financing signature. The hardest primary source is Lucent's FY2000 10-K: credit and guarantee commitments to customers capped at about $8.1 billion, with about $2.1 billion actually outstanding — roughly $700 million of exposure to a single customer, Winstar, written off in full, and about $3.5 billion of customer-financing bad debt provisioned over two years. Cisco's FY2001 doubtful-accounts provision was about $268 million (the circulating "$900M write-off" misreads a balance-sheet reserve balance as a P&L charge). From 2000, about 47 CLECs went bankrupt or exited the market (ALTS's count); the roughly 300 CLECs worth about $87 billion in market value in 1999 shrank to about 200 worth about $4 billion by 2002. Note Lucent's structure: committed $8.1B vs disbursed $2.1B — announced amounts running 4x ahead of deployed ones. Keep that trap in mind; it returns below.
The collapse in full: roughly $2 trillion of telecom market value erased and about half a million jobs lost (Paul Starr's 2002 accounting); FCC chairman Powell testified to the Senate in July 2002 that the industry owed about a trillion dollars, "much of which will never be repaid." Moody's global speculative-grade default rate (issuer-weighted) peaked around 10% in 2001 and eased to about 8.4% in 2002 — a whole-market yardstick; a clean telecom-only default series still doesn't exist.
Push the lens back another eighty years and Odlyzko's archaeology of the 1840s British Railway Mania yields the two sharpest, most transferable conclusions.
First: manias are not information failures — they are reliable counter-evidence being ignored. In the peak year, 1847, Britain invested about £44 million in railways in a single year — against a national budget of about £50 million; 1846 saw a record 4,538 miles of railway authorized. Reliable quantitative evidence existed at the time showing demand could not support expected returns (investors expected railway revenues around £60 million by 1850 or so; the 1852 actual was £15 million) — but the technology narrative overrode the arithmetic. Odlyzko calls it a "collective hallucination."
Second: infrastructure surviving ≠ investors getting paid. The railway network survived and eventually prospered (railway revenue was 6.0% of British GDP in 1905); but the railway share index fell about 60% from its July 1845 peak to its October 1849 trough, and partly-paid households were wiped out by capital calls. Fiber likewise: dark fiber was indeed picked up cheap by the Googles, and collapsed bandwidth prices did make Netflix, YouTube, and AWS's cost environment possible — none of which saved the telecom shareholders of 2000. One more honest note: "bubbles leave productive residue" remains a qualitative narrative — no one has published a serious quantification converting the fiber glut into "$X of later savings." It can be a historical footnote; it cannot be an investment thesis.
The big four's calendar-2026 capex guidance sums to roughly $710 billion at company-stated midpoints: Microsoft ~$190B (including ~$25B from memory and component inflation), Alphabet $180-190B (with 2027 guided to "significantly increase"), Amazon ~$200B, Meta $125-145B. For scale: Amazon's full-year 2025 capex was $131.8B, and Meta raised its own range from $115-135B to $125-145B within a year.
Three yardstick details belong in the text, not a footnote. One, the yardsticks differ: Amazon reports cash capex while Meta/Microsoft include finance leases — the sum is an order-of-magnitude figure, not an audited total. Two, headlines understate commitments: Meta's contractual commitments jumped $107 billion in a single quarter — multi-year cloud and infrastructure agreements never touch current-period capex. Three, inflation content: Microsoft says ~$25B of its $190B is component price increases — that part of capex growth is inflationary and buys no additional compute.
Start by admitting the yardstick chaos. Microsoft says its "AI business ARR" exceeds $37 billion, up 123%; AWS says its "AI revenue run rate" passed $15 billion in the AI wave's first three years; Google doesn't break out AI revenue at all, offering only Google Cloud as a whole ($20B/quarter, +63%) and backlog ($462B). These are all vendor-defined, unaudited annualized metrics with incompatible numerators; they cannot be summed across vendors.
On the model-company side: OpenAI reached an annualized run rate of about $25 billion by the end of January 2026, with internal guidance of roughly $30 billion in full-year 2026 revenue and a ~$14 billion loss; its audited 2025 books (leaked June 2026) show $34 billion of total spending and a $38.5 billion net loss — roughly 7.6x 2024's. Anthropic's self-reported run rate jumped from $14 billion in February 2026 to $47 billion by mid-May. And a double-counting trap: OpenAI paid Microsoft/Azure about $17 billion in 2025 (audited yardstick) — money that also shows up inside Microsoft's AI revenue; add hyperscaler AI revenue to model-company revenue and the same compute dollars get counted twice. (Correcting a circulating version: "OpenAI pays Azure $13B" is a mix-up — $13B is Microsoft's cumulative investment commitment to OpenAI, the opposite direction.)
Take the most aggressive yardsticks available and visible "AI revenue" still sums to tens of billions — against $710 billion of single-year capex. An order of magnitude. That is the factual floor under every gap calculation.
Cahn's algorithm is dissected above. The other figures routinely cited alongside it use different denominators, and the juxtaposition itself is the error:
The central methodological split beneath the whole fight is a denominator question: align revenue against full capex, or against annual depreciation? The Cahn family uses a single year's full capex, and the gap explodes. Critics point out capex is depreciated over 3-6 years; use annual depreciation as the denominator (roughly a third to a sixth of full capex) and the gap shrinks dramatically. The strongest version of the rebuttal — "revenue already exceeds depreciation" — depends on a $1.19-per-dollar-of-depreciation figure that could not be traced to any primary source and cannot bear weight. But the methodological split itself is real: the bubble callers and their critics are frequently not arguing about facts. They are arguing about denominators.
This capex cycle is already a macro variable. Harvard's Jason Furman, using BEA data: investment in information-processing equipment and software is only about 4% of US GDP, yet accounted for 92% of GDP growth in the first half of 2025; strip those categories out and H1 2025 annualized growth was 0.1%. He supplied his own hedge: absent the AI boom, lower rates and electricity prices "could maybe make up about half" of it. Yardstick note: BEA's "information processing equipment and software" is wider than data centers, so 92% is an upper bound. The number cuts both ways, exactly as in 1999: it is evidence that AI is carrying US growth — and it means that when capex decelerates, growth is instantly exposed. Telecom equipment's near-7% share of private investment in 2000 is the same seat, occupied by a new tenant.
The most authoritative characterization of the financing shift comes from the Bank for International Settlements. BIS Bulletin 120 (January 2026) puts the conclusion in its title: "Financing the AI boom: from cash flows to debt." The hard numbers: private credit loans outstanding to AI-related companies grew from near zero to over $200 billion (almost 8% of total private credit loan volumes), with BIS projecting $300-600 billion by 2030; funds originated over $40 billion of AI loans in 2025 alone (about $3 billion in 2010). The sharpest sentence is about pricing: spreads on AI private credit are nearly identical to non-AI (about 6.2 vs 6.1 percentage points) — "either lenders may be underestimating the risks of AI investments … or equity markets may be overestimating the future cash flows AI could generate." At least one of those two markets is wrong.
The same bulletin has a cooling side: BIS puts the AI investment rise at around 1% of US GDP — similar to the shale boom, and about half the rise in IT investment during the dot-com era — and assesses macro-financial stability risks as "appear moderate." Not yet 1999 in scale; converging on 1999 in structure — that is the most accurate one-line summary of the BIS yardstick.
Why the forced turn to debt? Operating cash flow no longer covers the burn. The big four's 2026 capex is on track to consume about 90% of operating cash flow (about 65% in 2025); Alphabet's FY2025 free cash flow of $73.3 billion is projected by analysts to fall to about $8 billion in 2026; Amazon's FY2025 FCF fell about 71% year-over-year to $11.2 billion, with 2026 projected negative (note: the projections are analyst forecasts, not results). Oracle has already crossed the line: FY2026 (ending May 2026) free cash flow of −$23.7 billion, $43 billion borrowed during the year; reported capex $55.7 billion, with FY2027 guidance of about $70 billion in net cash outlay.
Hence a financing toolkit 1999 never had:
The closest thing this cycle has to vendor financing is the circular arrangements between suppliers and customers. Deal by deal, with yardsticks:
Against 1999: Lucent was "committed $8.1B, disbursed $2.1B"; today it is "announced $100B, definitive agreement unsigned." The structure is isomorphic — supplier balance sheets subsidizing customer demand, with announced amounts manufacturing demand signals. The differences must be recorded just as faithfully: NVIDIA is mostly making minority equity investments rather than lending customers money to buy its goods, a different legal structure; and the funders (Oracle aside) are hyperscalers with real profits, not the never-profitable CLECs of 1999. The off-balance-sheet total now has a hard number too: the five hyperscalers (including Oracle) carry about $662 billion of data center leases signed but not yet commenced — a figure first attributed to Moody's via media, which this essay's verification rebuilt independently from the five companies' SEC filings ($661.7B, a mix of operating and finance leases). Note it is an undiscounted whole-lease-term total (typically 10-15 years) and must not be compared in the same layer as annual capex or on-balance-sheet debt.
Grade the moment on Minsky's three-stage scale (the original definitions, Levy Institute Working Paper 74): hedge (cash flows cover principal and interest) → speculative (interest covered, principal rolled over) → Ponzi (both depend on selling assets or new borrowing). The big four moving from "capex at 65% of OCF" to "90% plus bond issuance plus off-balance-sheet SPVs" is a textbook hedge→speculative slide; Oracle (negative FCF, debt-funded expansion, a single-customer bet) sits deep in speculative territory. Analyst Paul Kedrosky, using the same framework, puts the overall judgment as "We're now seeing elevated leverage and speculative financing in AI infrastructure" — but he attaches no Ponzi label, and writing the present as "already Ponzi" is a distortion. The framework's value is that it supplies observable deterioration criteria: the next stage begins when refinancing windows become a condition of survival and new debt mainly retires old debt.
Here the 1999 analogy meets its most interesting two-way stress: buried fiber lasts 20 years. GPUs?
The undisputed part first (all from 10-Ks): between 2020 and 2024, the hyperscalers collectively stretched server depreciation from 3-4 years to 5-6. Amazon: 3→4 (2020) →5 (2022) →6 years (2024); Microsoft: 4→6 years (effective FY2023); Alphabet: 4→6 years (effective January 2023); Meta: stepwise to 5.5 years (effective January 2025). The profit impact is also disclosed: Microsoft's FY2023 operating income higher by about $3.7 billion; Alphabet's FY2023 depreciation lower by about $3.9 billion; Meta's 2025 depreciation expected about $2.9 billion lower. Longer lives materially lift reported profits — no one disputes this; the dispute is whether it is justified.
Then the 2025 reversal signal: Amazon's SEC filings moved a subset of servers and networking equipment from six years back to five (effective January 1, 2025), with the stated reason — "increased pace of technology development, particularly in the area of artificial intelligence and machine learning" — an anticipated ~$0.7 billion reduction in 2025 operating income, plus a $920 million accelerated-depreciation charge in Q4 2024 for early retirements. That is a company's own filings conceding that AI shortens equipment life — the depreciation skeptics' only audit-grade, internally generated confirmation. (For completeness: the same 10-K extended heavy equipment from 10 to 13 years; the net effect on operating income was still positive.)
Short seller Michael Burry's accusation (from November 2025): lengthening depreciation is "one of the more common frauds of the modern era"; he estimates $176 billion of understated depreciation across 2026-2028, with Oracle's 2028 earnings overstated about 27% and Meta's about 21%; he puts true GPU economic life at 2-3 years. His line: "I am not claiming Nvidia is Enron. It is clearly Cisco." Required labeling: the underlying arithmetic sits behind a paywall, CNBC says it cannot independently verify the numbers — accusation-grade — and NVIDIA circulated a seven-page memo rebutting him point by point.
The bulls' counter-evidence is "old cards don't die": the A100 (launched 2020) was still actively renting in mid-2026 at roughly $1.09-2.04/hr on specialist platforms (multi-source confirmed; hyperscaler clouds up to $4-5/hr) — a six-year-old part with a live secondary market, in direct conflict with "2-3 years to zero." But this evidence supports only the qualitative point: "still rentable" is not "rental margins ever recovered the purchase price," and no primary-source unit economics exist for the latter. The other frequently cited number — CoreWeave's 2022-vintage H100s re-booked at ~95% of original price on contract expiry — was vetoed as load-bearing by this essay's methods-audit seat: single source, maximally interested self-report (CoreWeave's valuation depends on GPU residuals), n=1, "original price" undefined. It may be quoted only as a company statement with its interests labeled, not as fact.
Jensen Huang's "When Blackwell starts shipping in volume, you couldn't give Hoppers away" (GTC 2025) is half-quoted by both sides. The full context is frontier inference workloads — his same-session addendum was "There are circumstances where Hopper is fine. Not many." Note the dismissal actually targets the inference market the bulls rely on; reading it as "training only, inference unharmed" is an over-charitable gloss.
The depreciation question damages the 1999 analogy in both directions. Against the bears: 1999's fiber was a 20-year asset that could sit dark until a Google arrived; if GPUs are truly 3-5-year assets, any glut clears faster and more painfully — the bust would not look like 2001's. Against the bulls: nobody in 1999 could stretch a bubble three extra years by editing depreciation schedules; if the 5-6-year convention is eventually falsified by product cadence (NVIDIA is now on an annual cycle), the profit give-back arrives retroactively and all at once. Chancellor lists "depreciation stretched from roughly 3-3.5 years to roughly 6-6.5" among his foremost earnings warning signs, and not without reason.
1999 was "built and unused" (capacity growing 144-165% against traffic's 87-88%); 2026's testimony is "built and still short." Microsoft CFO Amy Hood (FY26Q1 call): "I have been short now for many quarters. I thought we were going to catch up. We are not." Pichai (April 2026): "We are compute constrained in the near term. ... our cloud revenue would have been higher if you were able to meet the demand." Nadella locates the bottleneck precisely at power: "It's not a supply issue of chips; it's actually the fact that I don't have warm shells to plug into" — chips sitting in inventory for want of powered shells.
Revenue really is materializing: Google Cloud +63%; AWS's fastest growth in fifteen quarters; Microsoft's AI ARR more than doubling in a year; Anthropic's run rate up 3.4x in three months. NVIDIA's Q3 FY2026 revenue was $57.0 billion (+62%, primary 8-K), and its CFO put Blackwell+Rubin revenue visibility through end-2026 at about $500 billion (visibility — not backlog, and not realized revenue). All of this differs substantively from the 1999 CLECs' build-first-find-revenue-later spiral.
The demand evidence is produced almost entirely by the supply side. "Sold out" comes from CFOs defending capex; token growth (Google's monthly volume up ~7x in a year to 3.2 quadrillion) is an all-products yardstick including caching and internal calls — Google itself pushes AI Overviews into every search — and cannot proxy for revenue or compute occupancy; the $500 billion of visibility comes from the largest seller. Third-party measurement is almost absent: no credible independent GPU utilization data exists, and rental pricing's two tracks (spot down roughly two-thirds to three-quarters over three years, while 1-year contract prices rose from ~$1.70 at end-2024 to ~$2.65 by mid-2026 — the latter from SemiAnalysis's paywalled data via secondary accounts) are compatible with both "normal generational discounting" and "localized glut." No single indicator adjudicates.
The honest summary: the demand-reality evidence is directionally consistent but positionally monolithic. The lesson of 1999 is precisely that "doubling every 100 days" was also directionally consistent and positionally monolithic — everyone saying it was selling fiber or bandwidth — and the one person with independent data was right. What this cycle lacks is its Odlyzko: an independent measurer of utilization and demand.
Perez (technological revolutions and financial capital): installation (frenzied build-out) → crash → deployment (golden age) is the narrative AI bulls quote most ("the crash is the doorway to the golden age"). Two under-known facts: Perez herself concedes in the 2002 book that the phase-dating is circular ("the phases have been dated taking the occurrence of crises into account," p.79); and her own 2024 position is that AI belongs inside the ICT revolution that began a half-century ago — not a new revolution. "AI is in the frenzy phase of a new revolution" is her citers' extrapolation, not her claim.
GPT diffusion economics (Bresnahan & Trajtenberg 1995; David 1990, the dynamo paper): general-purpose technologies pay back late because complementary investment — organizational redesign, process change — takes decades. This is the bulls' theoretical root for "no returns yet ≠ bubble." Its problem is Perez's problem: "returns haven't arrived because it's early" cannot be falsified in advance, indefinitely.
Janeway (productive bubbles): bubbles come in two kinds — those that leave productive assets after bursting (railways, grids, fiber — "no one tore up the railroad tracks ... The infrastructure remained") and those that don't (housing, 2004-07). But note that in November 2025 Janeway did not certify AI as productive — he framed it as an empirical question answerable only after the fact ("What is the value-creating potential of LLMs?"), flagging the risk that token output commoditizes into a low-margin business; a truly productive bubble is confirmable only years after the frenzy cools and passes through the trough of disillusionment.
Chancellor (capital cycle): high returns attract capital → oversupply → returns destroyed. Of the five, this one is observable in advance: capex growth, entry rates, and depreciation policy are all measurable. Chancellor's 2026 judgment: AI's ratio of hype to proven efficacy is the most extreme he has seen; and he explicitly rejects the productive-bubble consolation — "irresponsible to argue that bubbles are good for society"; misallocation slows growth rather than speeding it.
Minsky (financial instability): applied in 3.4 above. The other operational framework: financing structures can be classified and tracked.
The meta-conclusion: the first three frameworks explain powerfully after the fact and falsify weakly before it — at the level of theory, whether this is 1999 cannot be decided in advance. The criteria with real ex-ante bite come in exactly two families: observable supply-side variables (the capital cycle) and financing-structure classification (Minsky). Which dictates the shape of this essay's closing section: every testable claim is pinned to observables.
Like 1999 (structural layer): the demand narrative is produced by the supply side with no independent audit ("doubling every 100 days" ↔ today's token counts and visibility yardsticks); a modern variant of vendor financing (announced ≫ deployed; supplier balance sheets subsidizing customers); financing shifting from cash flow to debt, off-balance-sheet and structured (BIS's official characterization); accounting conventions as a first-order profit variable (the depreciation fight ↔ the era's capitalization fraud — different in degree, same family of mechanism); macro concentration (4% of investment driving 92% of growth ↔ telecom's near-7% of private investment).
Unlike 1999 (substantive layer): the funders are the strongest balance sheets in corporate history, not unprofitable CLECs (BIS and Allianz both note this); revenue is materializing fast and supply-constrained ($37B ARR +123%; "short now for many quarters" ↔ capacity forever outrunning traffic); the scale relative to GDP is about half the dot-com IT rise (BIS: ~1%); GPUs are 3-6-year assets where fiber was 20-year — the clearing dynamics differ; and there is no proven 1999-style accounting fraud — the depreciation fight is a yardstick dispute, not an established fraud.
What both sides pretend not to see: bulls skip the railway/fiber iron law — infrastructure surviving ≠ investors getting paid — and the downside rigidity of 90% OCF consumption plus off-balance-sheet guarantees; bears skip Cahn's own admission that his denominator is muddled, the arithmetic that shrinks the gap under a depreciation yardstick, and BIS's ~1%-of-GDP cooling note. And one number both sides quote wrong: Allianz's finding that AI capex-revenue growth divergence runs about 46%, "exceeding the 32% of the 2001 telecom cycle" — this essay traced the original report; the 32% comparator has zero methodological support anywhere in it, and Allianz itself calls the telecom analogy "imperfect," in a report whose stance is "war-proof for now." Using "46>32" to prove "already worse than 1999" mistakes a rhetorical anchor for a measurement.
The one-line verdict (as of July 2026): not yet 1999 in scale, better than 1999 in demand quality — but converging on late-1999 in financing structure at visible speed; and the way the demand narrative is produced — supplier-manufactured, independently unmeasured — is the deepest structural resemblance of all. Not a yes/no; a set of monitorable sliding variables.
Ordered by evidence strength, each with an observable test.
Load-bearing claims passed three-vote adversarial verification plus dual-seat audits for single-source empirics; fidelity of quotation is not truth of content, and evidence grades are labeled claim by claim. Related research on this site: AI Code Review: Cure for the Verification Bottleneck, or Turtles All the Way Down? (the verification bottleneck), The "95% of AI Pilots Fail" Physical (the yardstick ladder for enterprise AI returns), and The '70% of Transformations Fail' Autopsy (zombie-statistics methodology).