42 — The Overhyped Analysis

42 — The Overhyped Analysis

All 50 records classified overhyped · 25 sectors · research date 2026-09-15

0. Ranking method, and why score alone is the wrong instrument

composite_score cannot rank hype, for a structural reason. The score is capped at 40 + 60E, so a claim with a thin evidence base is already penalised — and that penalty compresses precisely the records this document is about. Ranking the fifty by score would put T-20-20 (67.8) and T-11-20 (70.8) at the top and T-17-19 (28.7) at the bottom, which inverts the finding: the creator-economy market-size statistic is the least traceable number in the database and the net-zero target count is one of the best-evidenced records in it.

Primary metric — the substantiation gap. For each record:

Prominence  P = mean(velocity, capital, strategic_importance)
Substantiation S = mean(adoption, revenue)
GAP = P − S

P measures how much attention, money and perceived strategic weight the claim carries. S measures whether anyone is actually using it and whether money is coming back. The gap between them is the thing being reported: claims that have captured capital and attention without capturing customers or revenue.

Rejected alternative, and why. The obvious metric — prominence minus evidence quality — is contaminated. Our own research produces high evidence_quality precisely when we can falsify a claim well: humanoids score 5/5 on evidence quality and source diversity because Unitree's prospectus and Agility's S-4 exist. Ranking on that metric would rank humanoids 30th out of 50, which is absurd. Good sceptical research must not lower a hype score. Both rankings are published below so the reader can audit the choice.

One important consequence of the chosen metric, visible at the bottom of the table: four records have a negative gap. They are not attention-without-adoption claims at all. They are measurement critiques — cases where the underlying activity is large and real but the metric everyone quotes measures the wrong thing. That is a genuinely different species of hype and §4 treats it separately.

1. The top 20 by substantiation gap

1 · T-13-19 — Humanoid robots doing useful paid work at scale

Sector 13 · score 51.1 · E=1.00 · GAP 3.33 (P 4.33, S 1.0) · high confidence, triangulated

• The claim as it circulates. Humanoids are entering industrial deployment; "deployment trackers" report thousands of units in the field; the category carries roughly $100bn+ of aggregate private and public valuation (Figure AI at $39bn in Sept 2025, Unitree peaking near $66bn, XPeng's Dogotix at $6.3bn).
• What the evidence supports. Every damaging number comes from a prospectus, an SEC filing or the industry's own trade body. Unitree, the global volume leader, shipped 5,632 humanoids cumulatively 2023–2025 with under 10% of 2025 revenue from industrial applications; the G1 sells at $13,500, a research-and-entertainment price point. Agility Robotics — the leading US deployment story — disclosed $1.8m of 2025 revenue against a $140m operating loss in its S-4. Tesla had zero Optimus units doing useful work as of January 2026 against a promise of 10,000 in 2025. IFR itself states mass adoption "will not happen within the near- and medium-term future."
• Where the gap comes from. A definitional gap, not a factual dispute. Trackers count announced pilots, letters of intent, demonstration units and units produced as "deployments." Filings count revenue. Almost the entire public disagreement is this.
• Who benefits. Component and actuator suppliers; GPU and simulation vendors; early investors who achieved liquidity in 2026; every private holder marked against a peak comparable.
• What would have to become true. One audited, multi-hundred-unit paying industrial deployment with disclosed revenue. The category is not disproven; it is undisclosed. Unit figures for Figure, Apptronik, 1X and AGIBOT are entirely undisclosed, so the true installed base is unknown, not known to be small.

2 · T-08-20 — Orbital data centres as near-term AI infrastructure

Sector 08 · 44.3 · E=0.50 · GAP 2.67 · low confidence

• Claim. Compute is moving to orbit; Starcloud raised $170m at a $1.1bn valuation (2026-03-30) and the thesis is used to justify space-sector valuations generally.
• Evidence. Starcloud-1 was a 130-pound satellite carrying a single Nvidia H100, launched November 2025; Starcloud-2, with a B200, was still pending. The company's own CEO projects favourable economics in three to five years and mainstream adoption "about a decade out." Microsoft's president: "We're keeping our feet on the ground."
• Gap source. A founder's own technical estimate is years beyond the narrative built around his funding round. That divergence is the record.
• Beneficiaries. Launch providers; radiation-hardened electronics suppliers; funds marking up early positions.
• What would make it right. Heat rejection in vacuum and radiation-induced error rates solved at rack scale, plus Starship cadence at a price that beats terrestrial capacity — which is itself scaling far faster. The market has already begun repricing: SPCX closed at $125.33 on 2026-08-04, below IPO price and 44% off its June peak.

3 · T-06-20 — Reshoring pledges treated as delivered manufacturing capacity

Sector 06 · 52.1 · E=0.40 · GAP 2.67

• Claim. $480bn+ of announced US pharmaceutical manufacturing investment, 22 sites, ~44,000 jobs, reported as though it were capacity.
• Evidence. The pledges run four to ten years and no public figure exists for how much is under construction. The one independent cross-check points the other way: a Council on Foreign Relations analysis observed that equipment-supplier share prices had largely not moved, and that even a conservative 15%-of-capital-to-equipment assumption — implying $75bn+ of orders — had not materialised.
• Gap source. A pledge is directly convertible into tariff relief: Commerce-approved onshoring plans cut the Section 232 rate to 20%, and to 0% combined with an MFN agreement. Announcing large is costless relative to the alternative.
• Beneficiaries. Companies receiving tariff relief for announcements; political sponsors; consultancies sizing the opportunity. Losers: states granting incentives against unbuilt capacity, and equipment suppliers who modelled order books on the pledges.
• What would make it right. State permit and groundbreaking records showing conversion. The gap will not become visible until exemptions come up for renewal.

4 · T-13-20 — "Every industrial company will become a robotics company": generalist manipulation declared solved

Sector 13 · 51.8 · E=0.60 · GAP 2.67

• Claim. Physical AI has reached the real world; top benchmark placement; success on new tasks "more than twice as often" as competitors; learning from a single demonstration.
• Evidence. All vendor-stated, none independently replicated in this research. No disclosed production deployment reports task success rate, cycle time or MTBF for a foundation-model-driven robot. Global industrial robot installations were flat at ~542,000 units in 2024 after 541,302 in 2023. If generalist manipulation were production-ready, that number and the programming-intensive cost structure would already be moving.
• Gap source. Announcements describe toolchain and simulation integration, which is real, and are received as describing autonomous task generalisation, which is not. The conflation is in the reception as much as in the claim.
• What would make it right. ISO 10218 / ISO TS 15066 presume deterministic, analysable behaviour; no harmonised standard yet certifies a learned policy. A conformity-assessment route is a precondition, not a detail.

5 · T-05-19 — SMRs as a near-term answer to data-centre power

Sector 05 · 62.0 · E=0.70 · GAP 2.67 · high confidence

• Claim. Small modular reactors will power the AI buildout; hyperscaler partnerships are announced regularly.
• Evidence. Zero commercial SMR megawatt-hours have been delivered anywhere in the US. The genuine 2026 milestones are early-stage: TerraPower's first-ever NRC construction permit for a commercial non-LWR (March 2026, groundbreaking April); Kairos's Hermes 2 groundbreaking (April 2026). DOE's pilot targeted three designs reaching criticality by 2026-07-04 and had three of eleven projects with a Final Documented Safety Analysis as of May 2026. DOE's actual near-term plan — 2.5 GW by 2027, 5 GW by 2029 — relies entirely on uprates, licence renewals and restarts, and the $80bn federal nuclear partnership is built on AP1000 large reactors. Nuclear's share of US generation is flat at 18% in both 2026 and 2027.
• Gap source. Timeline arithmetic. A construction permit in 2026 implies operation in the early-to-mid 2030s against load growth forecast for 2027–2030.
• The tell, stated in the record itself: capital scores 4 and revenue scores 1. The divergence between those two numbers is the whole story.
• What would make it right. The precedent nobody cites is NuScale/UAMPS, cancelled in 2023 at ~$89/MWh. HALEU fuel supply is unresolved for several designs.

6 · T-10-20 — Solid-state batteries "entering mass production"

Sector 10 · 37.7 · E=0.60 · GAP 2.67 (adoption 0, revenue 0)

• Claim. Imminent since roughly 2017, repeatedly "within 2–4 years."
• Evidence. Pilot lines and sample cells; SNE Research's top-ten global installations for Jan–Jul 2026 are entirely conventional lithium-ion chemistries and no solid-state manufacturer appears. The sharpest evidence is fiscal rather than technical: China's new lithium-battery consumption tax (2% from 2026-09-01, 4% from 2027-09-01) exempts sodium-ion, solid-state and fuel cells through 2028-12-31. A government does not exempt a product that is generating meaningful tax base.
• Gap source. Manufacturer timelines are strategic communications aimed at investors and partners. When a corporate timeline and a fiscal forecast disagree, the one with money attached is the better estimator.
• Cost of believing it. The actual cost revolution happened in LFP at $81/kWh against NMC's $128/kWh — a chemistry Western analysis largely dismissed until 2023. OEMs that deferred conventional cost work while waiting are the losers.

7 · T-10-19 — "Robotaxis are scaling globally"

Sector 10 · 62.9 · E=1.00 · GAP 2.50 · high confidence, triangulated

• Claim. "14 cities," "28 cities," "global robotaxi leadership."
• Evidence — regulator-reported miles. A peer-reviewed analysis of CPUC filings covering California's first ~1,000 days of commercial service (Aug 2023 – Dec 2025) found 13,790,147 trips and 86,269,177 vehicle miles in total — about 0.003% of one year of US light-duty VMT. 46.4% of those vehicle miles carried no passenger; occupancy averaged 1.4; the passenger-onboard share plateaued at 55–57% from mid-2025, i.e. the deadhead problem stopped improving. Waymo's 220.6m cumulative rider-only miles through March 2026 sit in five metros (Phoenix 80.6m, SF Bay 67.1m, LA 51.8m, Austin 15.8m, Atlanta 5.4m). No operator anywhere publishes revenue per vehicle-hour, remote-assistance ratio or contribution margin.
• Gap source. "City counts are announcement-driven and grow cheaply; miles are regulator-reported and grow expensively." The sector's public narrative uses the cheap metric. Compounding it: the operator with the loudest claims operates primarily where disclosure is weakest — California requires quarterly CPUC and DMV reporting, Texas does not.
• The honest version. A real commercial service in roughly five metro clusters worldwide, growing fast from a very small base, with no published positive contribution margin. That is a genuine achievement and is separately recorded as T-10-09 (74.8) — the same underlying facts support a high-scoring emerging_signal record and an overhyped record, because the claim being tested is different.

8 · T-01-19 — AGI declarations as a capability claim

Sector 01 · 52.0 · E=0.80 · GAP 2.33 · high confidence

• Claim. On 2026-09-03 OpenAI released GPT-6 Astra and its president closed the briefing with "Welcome to the AGI era."
• Evidence. Independent measurement the same month: Artificial Analysis v4.3 scores GPT-6 Astra at 53, tied with Claude Fable 5.1; it loses 64% to 20% on Epoch's MirrorCode. OpenAI itself conceded the model is "harder to monitor" in evasion tests and delayed its cybersecurity capability release.
• Gap source. The claim has no falsifiable definition. It is recorded as marketing under the claim-type ladder and the benchmarks are recorded as fact.
• Who benefits. Frontier labs raising capital (AGI framing supports Anthropic's $965bn and OpenAI's $730bn marks) and, asymmetrically, safety-advocacy organisations whose case the framing strengthens — the Sanders/Casar Ban Artificial Superintelligence Act was introduced the same day.
• What would make it right. An agreed operational definition and an independently administered test. Neither exists.

9 · T-06-19 — AI-compressed drug discovery timelines

Sector 06 · 45.3 · E=0.70 · GAP 2.33 · high confidence

• Claim. AI compresses ten-year discovery to eighteen months.
• Evidence. As of 2026-09-15: zero AI-originated molecules approved anywhere; exactly one in Phase III (rentosertib, single country, on an endpoint that has defeated most entrants in IPF, with Phase 2a data whose primary endpoint was safety). Recursion discontinued five programmes in 2025 including a Phase II asset showing no efficacy signal. Generate Biomedicines priced its 2026 IPO below its last private valuation and traded below issue.
• Gap source. A category error. The 12–18-month figure is a preclinical-candidate-nomination metric presented as a development-timeline metric. Discovery was never the long pole; clinical attrition is, and no platform has shown it changes Phase II or III success rates.
• What would make it right. A positive Phase III readout for rentosertib, or any platform demonstrating a shift in late-stage success rates.

10 · T-03-19 — "Quantum advantage" and logical-qubit counts as proxies for usefulness

Sector 03 · 46.4 · E=0.70 · GAP 2.33

• Claim. Vendors compete on qubit counts, quantum volume and "verifiable advantage" demonstrations.
• Evidence. Best publicly reported logical-qubit results in 2026 span roughly 12 (Infleqtion) to 96 (academic neutral-atom), against IBM's own target of 200 logical qubits and 100 million gates in 2029 for the first large-scale fault-tolerant machine. That is roughly two orders of magnitude in gate count. Advantage demonstrations are on sampling and echo-type problems chosen because they are hard classically, not because anyone needs them solved.
• Gap source. "Logical qubit" is not a standardised term — it is used for memory-only encoded qubits, non-universal gate sets and full fault-tolerant computation, which differ by orders of magnitude in difficulty. Headline counts are not comparable across vendors.
• Second-order damage. The marketing directly drives post-quantum-cryptography policy deadlines in sector 04 (see T-04-20 at rank 30).
• What would make it right. qLDPC overhead reduction (T-03-14) could compress the timeline faster than sceptics expect. The hype being wrong today does not make the technology wrong.

11 · T-08-19 — Golden Dome framed as an immediate contract bonanza

Sector 08 · 59.5 · E=0.60 · GAP 2.33

• Claim. Retail financial media described the programme as beginning "with a massive USD 151 billion contract."
• Evidence. $151bn is an IDIQ ceiling shared by more than 2,100 awardees with no funding obligated at the base contract level — MDA said so explicitly. The programme office's actual FY2026 money was $22.5bn of a $24.4bn allocation, did not arrive until March 2026, and was 90% obligated by 2026-08-20. As of this research date no FY2027 appropriations bill had been enacted, and Gen. Guetlein warned that without continued funding after 30 September the programme "will stop with everything we've already built and delivered."
• Gap source. Conflation of a contract ceiling with an obligation. Roughly sevenfold against the programme office's actual funds, and far more against obligations to any single company.
• Losers. Suppliers who staffed up against the ceiling number; investors who bought defence names on the headline.
• What would make it right. Nothing about the technology — a successful intercept test is claimed and the target operational date is summer 2028. What would have to change is the number in circulation, plus an enacted FY2027 appropriation.

12 · T-11-19 — "The IPO window has reopened"

Sector 11 · 66.4 · E=0.80 · GAP 2.17 · high confidence

• Claim. Record aggregate exit value means a broad reopening.
• Evidence. Four companies account for 93.5% of 2026 venture exit value. 44 US VC-backed IPOs had priced by late July, against fewer than 50 annually every year since 2022 — while 950+ private companies are valued at $1bn+, so under 5% of the eligible population listed. Aftermarket performance is weak: SpaceX −30.3% from debut, Cerebras −34.7%, against a Renaissance IPO Index up 12.7%.
• Gap source. Value and count diverge. Morgan Stanley argues the opposite case on sector breadth and both can be true: broad by sector count, extremely narrow by value.
• Why it matters most. If the window is narrow, the 2021-vintage overhang does not clear, DPI does not recover for the median fund, and the whole liquidity-engineering apparatus stays load-bearing. This is the single most consequential misreading available in the sector.
• What would make it right. IPO counts can inflect quickly on two or three mega-listings; Anthropic's and OpenAI's listing status was unconfirmed at this research date.

13 · T-25-19 — Agentic AI disintermediating online travel agencies

Sector 25 · 42.9 · E=0.60 · GAP 2.17 · low confidence

• Claim. AI agents are displacing OTAs as the travel discovery and booking layer.
• Evidence. Booking Holdings' Q2 2026 earnings release does not mention AI agents or agentic booking at all, and marketing expense moved from 4.6% to 4.7% of gross bookings — ten basis points. Direct channel mix held in the mid-fifties percent of room nights; room nights grew 5% to 325m. OpenAI named Booking.com and Expedia as launch partners (2025-10-06) and published no travel booking volumes.
• Gap source. If discovery were migrating at the claimed pace, the first observable symptom would be customer-acquisition economics, which are disclosed quarterly and have barely moved.
• Precedent being ignored. 1996–2005 produced re-intermediation, not disintermediation; and the fourteen-year "NDC will bypass the GDS" story shows how slowly travel distribution actually changes. The genuine measured disintermediation is in GDS air bookings (T-25-10), not through AI assistants.

14 · T-12-19 — Agentic shopping as an imminent replacement for search-led commerce

Sector 12 · 43.2 · E=0.50 · GAP 2.17

• Claim. AI agents are about to intermediate a large share of retail transactions.
• Evidence. Twelve months after Google announced AP2 with 60+ partners and OpenAI and Stripe launched the Agentic Commerce Protocol, not one participant — OpenAI, Google, Stripe, Mastercard, American Express, PayPal or Adyen — has published a transaction count, GMV figure or active-merchant number. The only quantified retail-side evidence is Adobe's AI-referral data, which reports conversion rates while withholding the denominator, and whose growth decayed from 693% to 138% year over year in six months. At launch, Instant Checkout supported single-item purchases only.
• Gap source. Infrastructure and partnership announcements are cheap to make and generate coverage; volume disclosure is voluntary and unflattering if small.
• The distinction that resolves it. AI-assisted discovery is demonstrably real. Agent-completed checkout is demonstrably unmeasured. Conflating the two is the sector's most common analytical error. A single credible volume disclosure would move this record from overhyped to emerging_signal immediately.

15 · T-15-20 — Fully generative games and autonomous AI NPCs

Sector 15 · 45.2 · E=0.60 · GAP 2.17

• Claim. Generative models will produce dynamic NPCs and eventually whole games on demand.
• Evidence. Steam's own disclosure data shows ~60% of AI disclosures concern visual assets, with audio, localisation and writing following — the least architecturally consequential uses. No commercially significant shipped title is built around generative NPCs. The named shipped 2026 uses are asset-layer: EA's synthetic commentary in NHL 27 (2026-09-10) and Level-5's promotional art, for which the CEO apologised. Engine-level agentic tooling launched only in September 2026, so systems-level integration has effectively no shipped track record.
• Gap source. A three-year disconfirming series that is never cited: Nvidia's ACE has been demonstrated at CES every year since 2023 without converting into a shipped commercial title.
• The cost of the conflation. Overestimating how fast production costs will fall and underestimating near-term labour impact, which is concentrated and real in exactly the asset roles being substituted.

16 · T-18-20 — Announced data-centre pipeline capacity treated as a demand measure

Sector 18 · 58.8 · E=0.70 · GAP 2.00

• Claim. JLL reports more than 66 GW of North American data-centre capacity under construction — noting it exceeds Germany's entire electricity requirement.
• Evidence. CBRE, counting capacity physically under construction in primary markets, reports 7,481 MW. Roughly ninefold. ConstructConnect tracks 136 started projects worth $84.1bn — a project-level count far smaller than announcement-based capacity implies. And data-centre investment sales totalled only $1.7bn in H1 2026, so there is almost no arm's-length transaction evidence validating any capacity-derived valuation.
• Gap source. Scope and stage definition, and neither firm publishes its inclusion rule. Corroboration from sector 05: ERCOT suspended Batch Zero against 474 GW of requests (~90% data centres) saying its own forecast was probably too high; Exelon disclosed it expects 22% of its 65 GW pipeline to materialise. A 22% materialisation rate on 66 GW describes a very different market.
• Fairness. The contracted portion is genuinely large and genuinely leased — 80.4% preleasing on 7,481 MW is real. The overhype is specifically in the multiple of that figure quoted as equivalent. It is also possible the critique is wrong in the other direction: JLL's definition may be more complete rather than more inflated.
• Where loss would land. Debt underwritten against pipeline assumptions rather than signed leases.

17 · T-24-19 — "AI music is taking over streaming"

Sector 24 · 49.7 · E=0.50 · GAP 2.00

• Claim. Deezer reported over 50% of daily new uploads at peak in June 2026 — roughly 90,000 tracks a day — were fully AI-generated.
• Evidence. From the same disclosure: fully AI-generated music is 1–3% of total streams. And Deezer identified up to 85% of streams on fully AI-generated tracks as fraudulent in 2025 and demonetised them, against platform-wide fraud of 8%. Net of fraud, genuine human demand is plausibly in the low tenths of a percent of streams. Global recorded music grew 6.4% to $31.7bn; US mid-year 2026 revenue grew 6.9%; Spotify reached 300m premium subscribers with no disclosed impact.
• Gap source. A supply-side statistic read as a demand-side statistic. The two numbers come from the same press release and only one of them is ever quoted. Suno's $5.4bn valuation reflects investor expectation, not disclosed revenue — its Series D announcement contains no revenue, ARR or subscriber figure.
• Why the distinction is operationally decisive. The correct response to a supply flood — metadata, detection, fraud enforcement, identity verification — is entirely different from, and much cheaper than, the response to demand substitution.
• The strongest argument against our own classification, recorded as a risk on the record: substitution would appear first in low-attention functional genres (background, ambient, production, fitness) that no published series separates.

18 · T-03-20 — Intel Foundry's external-customer turnaround treated as settled

Sector 03 · 61.8 · E=0.40 · GAP 2.00

• Claim. Intel Foundry has turned the corner on 18A high-volume manufacturing and Panther Lake shipping.
• Evidence. 18A yields remain below profitable levels and are not expected to reach target cost thresholds until end-2026 at the earliest. 14A rests on two prospective customers whose decisions run from H2 2026 into H1 2027, and Lip-Bu Tan has said Intel may slow or cancel 14A without external foundry revenue. Nova Lake is reported as likely to slip to 2027. Meanwhile TSMC holds 67.7% gross margins and is raising capex to $60–64bn.
• Gap source. National-champion framing and equity positioning substituting for disclosed customer commitments. Two prospects are not a customer base.
• Why the asymmetry matters. Dismissing a genuine turnaround is as costly as believing a false one — Intel Foundry is the only plausible non-Taiwanese source of leading-edge logic at scale. A single named 14A customer commitment would invalidate this scepticism overnight. E=0.4 and single_source are recorded honestly; this assessment rests on March 2026 reporting and H2 2026 decisions may already have changed it.

19 · T-24-20 — Collective or statutory licensing for AI training as the imminent settlement

Sector 24 · 46.4 · E=0.80 · GAP 2.00 (adoption 0, revenue 0)

• Claim. AI training will be resolved by a collecting society or a statutory remuneration right, on the model of mechanical or performance rights.
• Evidence. Every primary source points the other way. IFPI explicitly asks the EU to reject statutory remuneration rights, compulsory licensing and mandatory collective management. The US Copyright Office's Part 3 report examined both and recommended neither, and remains a pre-publication draft sixteen months on; Part 2 recommended no legislative change at all. The UK consultation closed 2025-02-25 and had produced no outcome by 2026-03-19. Every actual deal — Udio-UMG, Udio-WMG, Suno-WMG, Suno-BMG, Suno-Believe, KLAY with all three majors — is bilateral, confidential and per-catalogue.
• Gap source. The largest rightsholders, who would be the principal beneficiaries of a levy, are lobbying against one, because bilateral leverage pays them more.
• Who is harmed by the gap persisting. Independent artists, songwriters, mid-list authors and local publishers — for whom the difference between a collective and a bilateral regime is the difference between being paid and not being paid — and who are planning for a mechanism that will not arrive.
• What would make it right. A broad fair-use ruling for developers in the OpenAI MDL would destroy bilateral leverage, at which point the same trade bodies would demand a statutory right. Briefing completes in early November 2026.

20 · T-21-20 — "Institutional adoption" evidenced by announcements, TVL and holder counts

Sector 21 · 54.0 · E=0.90 · GAP 2.00 (adoption deliberately scored 0)

• Claim. Institutional adoption of digital assets, evidenced by partnerships, total value locked, and holder counts.
• Evidence. One tracker reported total RWA holders rising 111.23% in thirty days — a doubling in a month, which is a distribution event or a methodology change, not adoption. The same publisher reports "represented asset value" of $361.91bn alongside "distributed asset value" of $38.89bn for the same market — a ninefold spread between two definitions that commentary routinely collapses into one number. Three independent trackers disagree by 4% on the sector's single most-cited figure. And the SEC's Gotbit judgment (LR-26598, proposed final judgment 2026-07-28, parallel criminal guilty plea) establishes that wash trading to "create the false impression of market interest" is a prosecuted practice, not a theoretical risk.
• Gap source. The sector's standard metrics are structurally vulnerable to manufacture: holder counts are addresses and cost a fraction of a cent to create; TVL is a price times a quantity reported as a quantity; "partnership" is not a revenue event.
• The operative rule this record establishes. A claim in this sector counts only if it appears in an audited filing, a regulator's document, a fund's daily NAV, or posted margin.

2. The full ranking, 21–50

Alternative ranking. Under the rejected prominence-minus-evidence-quality metric, the top five would be T-04-19 (+2.33), T-03-20 (+2.00), T-06-20 (+1.67), T-08-19 (+1.33), T-08-20 (+1.17). The two rankings agree on T-06-20, T-08-19 and T-08-20 and disagree sharply on the humanoid and robotaxi records, for the reason set out in §0.

3. What the fifty have in common: eight structural mechanisms

3.1 Vendor-captured evidence bases

The evidence for a claim exists only where the claimant produced it. The diagnostic is a deliberate search that returns nothing neutral — and the absence is the finding, not a gap in research effort.

• T-04-19: a search for autonomous-SOC efficacy returned "exclusively vendor blogs, vendor-sponsored buyer's guides and consultancy content marketing." Evidence quality scored 1; E=0.2, the lowest in the database.
• T-07-19: a dedicated search for independently measured bank AI outcomes returned "only vendor ROI marketing pages and content-farm 'definitive guide' articles."
• T-09-19: "search results for industrial AI adoption statistics are almost entirely Tier-C content farms citing unsourced downtime-reduction percentages. That the credible number had to come from the Census and a Reserve Bank, not from the trade press, is itself the finding."
• T-06-19, T-22-20, T-16-20, T-13-20: same pattern.

Across the 50, 46 evidence items are labelled marketing — 2% of all claims in the database, but concentrated here. The editorial rule that a marketing claim may never be rendered without its label is the primary defence.

3.2 Untraceable market-size figures

A number circulates with no nameable modeller, methodology or date.

• T-17-19: "no creator-economy market-size figure encountered in this research cycle could be traced to a named methodology or modeller."
• T-21-20: one publisher reports $361.91bn and $38.89bn for the same market under two definitions; three trackers disagree 4% on stablecoin market cap with no published reconciliation.
• T-17-20: the traceable figure is ~$120m; the circulating figures are "multi-billion 'relational AI' market estimates" that met Tier-C content-farm criteria on sight.
• Sub-mechanism — date laundering: searching for YouTube's payout figure in September 2026 surfaces a February 2024 article stating "$70bn over three years" as if current.
• Sub-mechanism — rounded cumulative disclosures cannot be differenced: YouTube's "$100bn over four years" was reported at the same $100bn a year earlier, so no annual growth series can be derived from it.

3.3 Deadline-driven marketing on regulations that then slip

The most-marketed compliance deadline of 2026 did not bind.

• T-04-20: PQC spending is driven by EO 14412's dated 2030/2031 obligations, not by an updated threat estimate, but is sold using the threat framing. Sequencing follows the driver: a mandate calls for a cryptographic inventory; an imminent break calls for re-keying twenty-year-confidentiality data first. Those are different programmes.
• T-01-16 / T-02-20 / T-09-19 / T-23-20 / T-14-19: the EU AI Act's high-risk employment obligations moved 2026-08-02 → 2027-12-02, a ~16-month slip, and appear as a stale premise in five separate overhyped records.
• Precedents in the same shape: T-16-18 (third-party cookies: 2022 → 2023 → 2024 → 2025 → abandoned) and T-19-09 (FSMA 204: 2026-01-20 → no enforcement before 2028-07-20).

3.4 Announcement-versus-delivery conflation

• T-06-20: $480bn pledged, construction unknown.
• T-18-20: 66 GW announced, 7,481 MW building.
• T-08-19: a $151bn ceiling against $22.5bn in hand.
• T-12-19: 60+ protocol partners, zero transactions disclosed.
• T-03-20: two prospective customers described as a customer base.

3.5 Demonstration-versus-deployment conflation

The gap between what is technically possible and what is economically deployed is where most technology forecasting fails.

• T-13-19 / T-13-20: videos and benchmarks against $1.8m of disclosed revenue and flat 542,000-unit installations.
• T-10-19: city counts against regulator-reported miles.
• T-08-20: one H100 in orbit against a $1.1bn valuation.
• T-15-20: three consecutive years of CES NPC demonstrations with no shipped title.
• T-03-19: sampling-problem advantage demonstrations against a 200-logical-qubit requirement.

3.6 Supply-side statistics read as demand-side statistics (new — found in five sectors)

• T-24-19: upload share (>50%) read as listening share (1–3%, most of it fraudulent).
• T-21-19: raw on-chain transfer volume read as payment volume.
• T-20-20: target counts read as emissions outcomes.
• T-25-20: stated intent read as booked volume.
• T-12-19: protocol adoption read as consumer adoption.

3.7 Advocacy models quoted as forecasts (new)

• T-07-20: the ICBA's $1.3tn deposit-flight figure is a conditional worst case produced to win a specific rulemaking argument about affiliate yield — "not a forecast but a negotiating position" — and is quoted without its provenance.
• T-11-20: distributor-generated demand claims ("aggregate demand exceeding $1 trillion for pre-IPO AI exposure") recorded as marketing.

3.8 The missing denominator (new)

A rate is published without its base.

• T-12-19: Adobe reports AI-referral conversion rates and withholds the denominator.
• T-07-19: use-case counts and employee-access percentages are inputs reported in the register reserved for results.
• T-22-20: usage (54% of teens) reported as efficacy.
• T-16-19: 61% assistant adoption reported as traffic loss.

4. Overhyped is not the same as false — the four species

This is the most important section for fair use of the classification. The fifty records divide into four groups requiring four different responses.

Species A — Real trend, inflated timeline (19 records)

The mechanism is genuine, the science or engineering is advancing, and the date is wrong. These should be kept on a watchlist with a named falsifier, not dismissed.

T-13-19 · T-13-20 · T-08-20 · T-05-19 · T-10-20 · T-10-19 · T-03-19 · T-06-19 · T-15-20 · T-12-19 · T-25-19 · T-20-19 · T-02-20 · T-04-19 · T-22-20 · T-23-20 · T-17-20 · T-16-20 · T-24-19

The record that proves the point: T-10-19 and T-10-09 describe the same facts. One is overhyped (the claim "scaling globally") and one is a 74.8-scoring emerging_signal (multi-metro paid driverless networks, ~500,000 weekly Waymo rides). The trend is real; the adverb is not.

Species B — Number error (7 records)

The underlying activity is real and often large; a specific figure in circulation is wrong by a stated multiple. These are the easiest to fix and the most immediately actionable.

Species C — Measurement critique (9 records)

The activity is real and often healthy; the metric everyone quotes measures the wrong thing. These are the negative-gap records at the bottom of the table, and they are among the best-evidenced in the database (mean E = 0.80).

T-20-20 (target counts ≠ emissions) · T-25-20 (intent ≠ bookings) · T-14-20 (revenue ≠ admissions) · T-09-20 (employment ≠ output) · T-19-20 (aggregate channel data contradicts the narrative) · T-12-20 (network launch ≠ media business) · T-01-20 (national ≠ regional) · T-09-19 (vendor survey ≠ Census) · T-22-19 (relative within-occupation ≠ absolute unemployment)

These require substituting a better metric, not abandoning the subject. Several also cut against a fashionable pessimism: T-19-20 finds restaurants beating grocery, T-12-18 finds the retail apocalypse cooling, T-02-19 finds no measurable SaaS displacement.

Species D — Empirically contradicted by disclosed data (15 records)

A named, checkable primary disclosure points the other way. These are the records where the strongest language is warranted.

T-02-19 (every incumbent forward metric grew double digits) · T-07-20 (deposits grew 0.8% for an eighth quarter) · T-16-19 (the named advertiser says budgets are static) · T-19-19 (Coca-Cola +5%, Nestlé positive, neither mentions GLP-1) · T-23-19 (eleven issuers, zero quantification) · T-16-20 (academic parity; flat agency margins) · T-15-19 (10:1 loss ratio) · T-05-20 (four of eleven units not running; order vacated) · T-24-20 (every primary source opposes) · T-03-20 (yields below profitable; 14A contingent) · T-07-19 (two earnings cycles, zero attributable figures) · T-18-19 (11.91% delinquency; −66bp cap-rate spread) · T-01-19 (tied at 53; losing 64–20) · T-04-20 (estimates span two decades) · T-14-19 (the contract prices substitution above hiring)

Where we may be wrong, recorded explicitly

Six records carry a stated risk that the sceptical reading is itself the error:

• T-03-20 (Intel Foundry). "Being wrong here is asymmetric — dismissing a genuine turnaround is as costly as believing a false one." One named 14A customer inverts it.
• T-24-19 (AI music). Substitution would appear first in functional genres that no series measures. "Absence of evidence over an 18-month window is weak evidence of absence."
• T-13-19 (humanoids). Unit figures for Figure, Apptronik, 1X and AGIBOT are entirely undisclosed — the installed base is unknown, not small.
• T-19-19 (GLP-1 food). "The correct position is unverified, not refuted." Companies have an obvious incentive not to attribute weakness to a drug.
• T-17-19 (creator economy). "The search budget was exhausted early and a fuller search might locate a defensible methodology. This record should be revisited at full search budget."
• T-18-20 (data-centre pipeline). JLL's definition may be more complete rather than more inflated; neither firm publishes its rule.

Sector-discovery caveat: sectors 13–18 completed with 5–11 searches each. Nine of the fifty overhyped records sit in that band (T-13-19, T-13-20, T-14-19, T-14-20, T-15-19, T-15-20, T-16-19, T-16-20, T-17-19, T-17-20, T-18-19, T-18-20). Their Tier-A verification is sound — the humanoid case rests on a prospectus and an S-4 — but their discovery is narrow, which matters most for T-17-19, where the central claim is an absence.

5. A reusable hype-detection checklist

Derived from the fifty. Each item is a test with a pass condition, in the order a working analyst should apply them. Three or more failures is a hype classification.

Tier 1 — Provenance (apply first, costs nothing)

1. Name the modeller. Can the figure be traced to a named organisation, a stated methodology and a date? Fails: T-17-19, T-17-20, T-07-18. If no modeller can be named, stop — the figure is unusable regardless of how often it appears.
2. Check the date of every search result before using it. Stale statistics recirculate without recency signalling. Fails: T-17-19 (a Feb 2024 figure surfacing as current in Sept 2026).
3. Is every locatable source produced or sponsored by a seller? Fails: T-04-19, T-07-19, T-09-19, T-16-20. A deliberate search returning only vendor content is itself the finding.
4. Is the claim marketing wearing the clothes of fact? A vendor claim about its own product or market is marketing and may never be rendered without the label. Fails: T-01-19, T-13-20, T-11-20.

Tier 2 — Arithmetic (the highest-yield tests)

5. Ceiling or obligation? Gross or net? Represented or distributed? Fails: T-08-19 (7x), T-21-20 (9x), T-18-20 (9x).
6. Supply-side or demand-side? Uploads, transfers, targets, intents and protocol partners are supply. Streams, payments, emissions, bookings and transactions are demand. Fails: T-24-19, T-21-19, T-20-20, T-25-20, T-12-19.
7. Where is the denominator? A conversion rate without a base, a use-case count without a cost saving, a usage figure without an outcome. Fails: T-12-19, T-07-19, T-22-20.
8. Aggregate or median? Demand the count and the distribution beside the sum. Fails: T-11-19, T-12-20.
9. Does the number move with price? TVL, market cap, GMV and "revenue recovery" are price × quantity reported as quantity. Fails: T-21-17, T-14-20.

Tier 3 — Delivery

10. Announcement or delivery? Ask for put-in-place spending, groundbreakings, equipment orders, regulator-reported volumes — not press releases. Fails: T-06-20, T-18-20, T-03-20.
11. Demonstration or deployment? Ask for task success rate, cycle time, MTBF, contribution margin, revenue per unit. Fails: T-13-19, T-13-20, T-10-19, T-15-20.
12. Does the capital score exceed the revenue score by two or more? The single most portable numeric tell in this database. Fails: T-05-19 (4 vs 1), T-13-19 (5 vs 1), T-10-19 (5 vs 1), T-03-19 (4 vs 1), T-08-20 (4 vs 1).
13. If the efficiency claim were true, where would it show up in audited accounts? Then look there. Fails: T-16-20 (agency margins flat), T-07-19 (no efficiency-ratio delta), T-09-19 (segment growth traced to data centres, not factories).

Tier 4 — Regulation and time

14. Which of the four regulatory stages is this — proposed, enacted, in force, enforced? Vendor marketing routinely collapses them. Fails: T-04-20, and five records leaning on the deferred EU AI Act deadline.
15. What is the base rate of slippage for this class of deadline? In this research cycle the EU AI Act slipped 16 months, FSMA 204 slipped 30, and cookie deprecation slipped four times and was abandoned. Discount accordingly.
16. Is the claim conditional on a rule that has not been made? Fails: T-07-20 (the $1.3tn figure is contingent on affiliate-yield treatment that is explicitly unresolved).
17. Is the timeline from an interested party, and is even their timeline later than the narrative? Fails: T-08-20 (the founder says three to five years; the coverage implies now).

Tier 5 — Disconfirmation discipline

18. Name the falsifier before forming the view. If no observation could change the conclusion, it is a prediction, not a forecast.
19. Search for the disconfirming series specifically. Regulator-reported miles, Census put-in-place, CPUC filings, BLS/BTOS, FDIC deposits, audited segment revenue. In every top-20 record the disconfirming series was public and free.
20. Has this claim been made before, and what happened? SOAR made the autonomous-SOC promise in 2017 (T-04-19). NuScale/UAMPS was cancelled in 2023 (T-05-19). Solid-state has been "2–4 years away" since 2017 (T-10-20). Nvidia ACE has been demonstrated at CES every year since 2023 (T-15-20). The precedent is almost never cited by the promoter.
21. State how you could be wrong. Six of the fifty records carry an explicit risk that the sceptical reading is the error. A hype critique without one is itself a hype position.

6. Closing observation

The single most striking finding across the fifty is that the disconfirming evidence was almost always free, public and one search away. CPUC filings, the Agility S-4, the Unitree prospectus, Census BTOS, the Booking Holdings 10-Q, FDIC deposit data, EIA's STEO, Deezer's own press release, Meta's segment disclosure, Marsh's rate index, the Columbia ad study. Nothing in this document required privileged access.

The gap between prominence and evidence therefore is not an information gap. It is an attention allocation gap — and it persists because, in every one of the fifty cases, the parties with the resources to close it are the parties who benefit from it staying open.

More research from the library · Explore trend records