Data governance note · Internal research infrastructure
Where the numbers come from
A layer-by-layer source register for the AGI infrastructure stack. Every figure that enters the Stack to AGI report should be traceable to one of the sources below, tagged by tier, dated, and marked as fact, estimate, or opinion. This document exists so that the analysis can be audited rather than trusted.
01 — Sourcing protocol
Rules before sources
A list of good sources does not by itself produce reliable analysis. What produces reliable analysis is a set of rules about how sources are used, applied consistently and visibly. These eight rules govern every figure in the report.
- Every number carries a source, a tier, and a date.Format: value [Source, tier, as-of date]. A number without an as-of date is unusable in a cyclical sector, because the reader cannot tell whether it describes the current state or a state that has since reversed.
- Primary beats secondary, always.If a figure appears in a news article, find the filing, report, or dataset the article drew it from and cite that instead. Trade press is an index to primary sources, not a source.
- Company-disclosed operating metrics outrank third-party estimates for that company.TSMC's own capex guidance beats any analyst estimate of TSMC's capex. The exception is where the company has an incentive to shade the number — order backlog definitions, "AI revenue" segmentation, and announced-capacity claims all require independent corroboration.
- Announced is not contracted; contracted is not built; built is not energised.This distinction is the single largest source of error in AI infrastructure analysis. Track each project at its actual stage and never aggregate across stages into one headline number.
- Two independent sources for any figure that carries a conclusion.Independent means not sharing an upstream source. Three outlets reprinting the same consultancy press release is one source, not three.
- Record the disagreement rather than resolving it silently.Where credible sources conflict, present the range and say which one the analysis uses and why. Silent resolution destroys the reader's ability to audit.
- Estimates are labelled and their derivation is shown.Any figure derived rather than sourced is marked ESTIMATE with the arithmetic exposed. Derived figures should never be re-cited later as though they were facts.
- Absence is reported.Where no reliable source exists, write "Data unavailable from accessible sources" and say what would be needed. Do not substitute a plausible number. Several genuinely important variables in this stack — transformer lead times, precision reducer capacity, high-purity quartz volumes — fall into this category.
On paywalled sources. Several of the highest-quality sources in this register are expensive subscription services (SemiAnalysis, Dell'Oro, TrendForce, Benchmark Mineral Intelligence, Wood Mackenzie, BNEF). They are listed because they are the correct source, not because access is assumed. Where a paid source is the only reliable one, the register says so — that is itself a finding about how confident the analysis can be.
02 — Reliability tiers
How sources are ranked
Every source in this register carries a tier. The tier reflects verifiability and incentive structure, not brand reputation or how frequently a source is quoted in the press.
Tier 1
Statutory and primary
Filed under legal obligation, or produced by a statistical agency under a published methodology. SEC filings, USGS, EIA, FERC, IEA, central banks, court and regulatory dockets. Errors are possible; deliberate misstatement carries legal consequence.
Tier 2
Industry standard-setter
Trade bodies and specialist datasets whose numbers the industry itself uses for planning. SEMI, WSTS, IFR, Uptime Institute, MLCommons, Epoch AI. Methodologies are published; coverage may be partial.
Tier 3
Commercial research
Paid analyst houses with real primary collection. Useful and often the only source, but methodology is proprietary, revisions are frequent, and the client base creates incentives. Cite as estimate, never as fact.
Tier 4
Press and company narrative
Trade press, press releases, conference keynotes, investor-day slides. Use to detect that something happened and to locate the primary document. Do not use as the source of a number.
03 — Layer source map
Sources by stack layer
Ordered from the material foundation upward. Each layer lists the sources that should carry the analytical weight, what specifically to take from each, and the caveat that matters most. The last column is the one to read first.
L0Critical materials, mining and refining
Chokepoint: refining, not reserves
The governing question at this layer is separation and refining capacity by country, not tonnage in the ground. Sources must therefore report refined output and processing share, which most mining datasets do not. Export-control status is a live policy variable and belongs to primary government publication, never to press summary.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| USGS Mineral Commodity Summariespubs.usgs.gov/periodicals/mcs2026/ | T1 free | Annual · early Feb | World production, reserves, import reliance, five-year salient statistics for 90 commodities, plus a policy-events narrative per commodity. The anchor dataset for gallium, germanium, rare earths, graphite, quartz, rhenium, hafnium. | Reports production, and only patchily reports refining and separation capacity — which is where the leverage sits. Data lag roughly one year. Gallium data is in kilograms, not tonnes. |
| USGS Mineral Industry Surveys | T1 free | Monthly / quarterly | Higher-frequency movement between annual MCS editions; useful for detecting a shortage before it reaches the press. | Coverage is uneven by commodity; several critical by-product metals have no monthly series. |
| IEA Global Critical Minerals Outlook | T1 free | Annual · ~May | Refining and processing concentration by country, project pipeline, price and investment analysis. The best free source specifically on the refining chokepoint. | Framed around energy transition demand; AI-stack demand is not separately modelled. |
| China MOFCOM / GACCMinistry of Commerce announcements; customs export data | T1 free | Ad hoc; customs monthly | The authoritative text of export control measures and their scope, effective dates and expiry. Monthly customs data shows actual export volumes of gallium, germanium, graphite, rare earths and magnets — the empirical test of whether controls bind. | Chinese-language primary; English summaries frequently misstate scope. The gap between a control's announcement and its licensing practice is where the real signal lies. The November 2026 suspension expiry is a scheduled risk date and must be checked against MOFCOM directly. |
| EU Raw Materials Information System (JRC) | T1 free | Continuous | EU-side dependency mapping and Critical Raw Materials Act implementation, including permitting timelines for European projects. | Europe-centric; global figures largely re-derived from USGS and industry sources. |
| Adamas Intelligence | T3 paid | Quarterly | Rare earth magnet demand and NdFeB supply chain — the best specialist coverage of the dysprosium/terbium dependency. | Small house; single-analyst risk. Treat as estimate. |
| Benchmark Mineral Intelligence · Fastmarkets · Argus · SMM | T3 paid | Daily / weekly | Assessed prices for materials with no exchange-traded benchmark: rare earth oxides, high-purity gallium, spherical graphite, spodumene. | Assessed prices are survey-derived, not transacted. In thin markets the assessment can move on very few trades. Different agencies publish materially different prices for the same material. |
| Producer filingsMP Materials, Lynas, Wacker, Aurubis, Neo Performance | T1 free | Quarterly | Actual separated-oxide and refined output, realised prices, capacity expansion timing and capex. The most reliable read on whether Western refining capacity is genuinely being built. | Company-defined capacity metrics ("nameplate", "run-rate") are not comparable across producers. |
Metrics to maintain from this layer: refined output share by country (gallium, germanium, dysprosium, terbium, graphite) · export licence approval rates · assessed price series for NdFeB inputs · Western separation capacity commissioned vs announced · status and expiry of Chinese export controls.
L1Energy, power and grid
Moat durability: strong
The most source-rich layer in the stack and the one where secondary reporting is least reliable, because interconnection queues, announced capacity and contracted capacity are routinely conflated. Grid operator filings are the primary evidence; everything else is commentary.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| IEA Electricity (annual) and Key Questions on Energy and AIiea.org/reports | T1 free | Feb; AI report ~annual | Global and regional data-centre electricity demand with a published methodology and a stated base case. The 2026 update puts data-centre demand at roughly 485 TWh in 2025 doubling to roughly 950 TWh by 2030, with AI-focused facilities growing far faster than the aggregate. | Relies on Omdia accelerator-shipment inputs, so the projection inherits that supply forecast. Base case only; the report itself flags meaningful post-2030 upside. Country granularity is coarse. |
| US EIAForms 860 & 923, Hourly Grid Monitor, STEO, AEO | T1 free | 860 annual · 923 monthly · Grid Monitor hourly · STEO monthly | Generator-level inventory, planned additions and retirements with expected in-service dates; actual generation and fuel consumption; near-real-time balancing-authority demand. The definitive US supply-side dataset. | Planned additions in Form 860 are developer-reported and systematically optimistic on timing. Grid Monitor data is preliminary and revised. |
| RTO/ISO primary filingsERCOT large-load queue & CDR; PJM load forecast and capacity auction results; MISO, SPP, CAISO equivalents | T1 free | ERCOT monthly · PJM load forecast Jan · auctions annual | The single best evidence on data-centre load growth. ERCOT publishes large-load interconnection requests separately from generation, which no other operator does as cleanly; the IEA notes the large-load connection queue grew more than 3.5x between 2024 and 2026. | Queue volume is not demand. Developers file duplicate and speculative requests across operators for the same project. Use signed interconnection agreements and posted financial security as the filter, and treat raw queue totals as an upper bound only. |
| Lawrence Berkeley National Laboratory"Queued Up" interconnection reports; US data centre energy usage report | T2 free | Annual · ~Apr | Standardised cross-ISO queue analysis with completion rates by technology — the correction factor that converts queue volume into expected build. Historic completion rates have run well below 30%. | Publication lags the underlying data by several months. |
| NERC Long-Term Reliability Assessment | T1 free | Annual · Dec | Reserve margin adequacy by region under stated load-growth assumptions. The clearest statement of where power scarcity becomes a binding constraint on siting. | Assessment, not forecast; assumptions are conservative by design. |
| FERC dockets and Form 1 | T1 free | Continuous | Co-location and behind-the-meter nuclear arrangements, large-load tariff proceedings, transmission cost allocation. The regulatory outcome that determines whether a given PPA structure is legal. | Docket text is dense and slow; outcomes often turn on procedural detail rather than headline ruling. |
| Equipment maker disclosuresGE Vernova, Siemens Energy, Hitachi Energy, ABB, Schneider, Eaton, Vertiv | T1 free | Quarterly | Orders, backlog, book-to-bill and capacity expansion for turbines, transformers, switchgear and power distribution. The best available proxy for equipment lead times, which are otherwise unpublished. | Backlog definitions differ by company and are not comparable. Order intake includes options and framework agreements at some issuers and not others. Read the definition in the notes before comparing. |
| Nuclear-specific: US NRC, DOE HALEU programme, WNA | T1 free | Continuous | Licensing status, uprate approvals, restart applications, ITAAC progress and HALEU availability — the hard gate on the SMR portion of the power thesis. | Licensing milestones are necessary but far from sufficient predictors of in-service dates. |
| BNEF · Wood Mackenzie · Utility Dive · Heatmap | T3/T4 | Continuous | PPA pricing, deal detection, policy tracking. Useful for finding the docket or 8-K to read. | Never the source of a number that reaches the report. |
Metrics to maintain: data-centre TWh by region · large-load queue MW with signed IA filter · reserve margins in PJM, ERCOT, MISO · turbine and transformer backlog and book-to-bill · announced nuclear PPA capacity vs licensed capacity · HALEU milestones.
L2Semiconductor manufacturing and capital equipment
Moat durability: very strong
Unusually well served by trade-body statistics with long consistent histories. The correct discipline here is to work from bookings and billings rather than revenue, because revenue at this layer lags the order cycle by four to six quarters and will mislead on turning points.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| SEMIBillings report; World Fab Forecast; silicon shipment statistics | T2 free/paid | Billings monthly · Fab Forecast quarterly | North American equipment billings (monthly, free); global fab construction, equipment spending and capacity by node and region (World Fab Forecast, paid). The standard reference for the industry's own planning. | Billings is a three-month average and is revised. World Fab Forecast is expensive and its capacity figures are modelled, not surveyed. |
| WSTS via SIA | T2 free | Monthly · plus semi-annual forecast | Global semiconductor billings by region and product category, the longest consistent industry series. | Categories predate the AI accelerator era and map poorly onto it. Three-month moving average obscures inflection. |
| SEAJ (Japan) | T2 free | Monthly | Japanese equipment billings — a useful independent cross-check on SEMI's North American series, and the better read on deposition, etch and test tooling. | Japanese-language primary release. |
| Equipment maker filingsASML, Applied Materials, Lam, KLA, Tokyo Electron, ASM International | T1 free | Quarterly | ASML's bookings, EUV and High-NA unit shipments, and China revenue share are the highest-information disclosures in the entire stack. Deferred revenue and backlog give twelve-to-eighteen-month forward visibility. | ASML stopped guiding bookings in some periods; check what is actually disclosed rather than assuming continuity. Systems recognised vs shipped differ. |
| TSMC monthly revenue and quarterly report | T1 free | Revenue ~10th monthly · results quarterly | The highest-frequency reliable indicator of leading-edge foundry demand anywhere. Quarterly call gives capex guidance, node revenue mix, and CoWoS advanced-packaging capacity commentary. | Monthly revenue is unaudited and mixes nodes and customers. Advanced-packaging capacity is described qualitatively; precise CoWoS wafer numbers are not disclosed and any specific figure circulating publicly is a third-party estimate. |
| US BIS Federal Register rules; Japan METI; Dutch export regulations | T1 free | Ad hoc | The actual legal text of equipment and chip export controls, entity list additions and licence policy. Scope is decided in the definitions section, not the press release. | Press coverage of these rules is unreliable on scope with striking consistency. Read the rule. |
| TechInsights | T3 paid | Continuous | Teardowns, die analysis, process node verification and manufacturing cost models — the only way to verify claimed node capability independently. | Very expensive. Cost models are reconstructions. |
Metrics to maintain: ASML bookings and EUV/High-NA units · SEMI billings and SEAJ cross-check · TSMC monthly revenue growth and capex guidance · WFE spend by region · China share of equipment revenue · export control scope changes.
L3Compute silicon: accelerators, memory, networking, packaging
Moat durability: moderate, eroding
The layer with the widest gap between what is publicly claimed and what is publicly verifiable. Accelerator shipment volumes, HBM allocation and packaging capacity are all central to the thesis and none is disclosed directly. The register therefore leans on customs data, memory contract pricing and independent estimation.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| Vendor filings and callsNvidia, AMD, Broadcom, Marvell, Micron, SK hynix, Samsung | T1 free | Quarterly | Data-centre segment revenue, gross margin, inventory, purchase commitments and supply-agreement disclosures. Purchase obligations and prepayments in the notes are the most underused forward indicator at this layer. | Customer concentration is disclosed as unnamed percentages. "AI revenue" definitions differ by vendor and are not comparable. Customer warrant arrangements are in the equity footnotes, not the release, and materially affect per-share economics. |
| Epoch AI data hubepoch.ai/data — AI Data Centers, Chip Sales, Hardware, Models | T2 free | Continuous · updated to Sep 2026 | Independent, methodologically documented estimates of deployed AI compute built from satellite imagery, permits and disclosed data. The AI Data Centers database currently covers 86 sites, roughly 13.9M H100-equivalents and 13.3 GW of IT power, with an estimated 46% coverage of global deployed capacity. Downloadable CSVs with stated confidence intervals. | Estimates, and presented as such — coverage confidence interval on the 46% figure runs from 26% to 79%. US-weighted. The older GPU Clusters database is deprecated and should not be mixed with the newer project-level data. |
| TrendForce | T3 paid | Monthly / quarterly | DRAM and NAND contract prices, HBM capacity allocation and supplier share, foundry revenue ranking. The de facto memory price benchmark. | Contract price assessments are survey-based. HBM allocation figures are estimates and have been revised substantially. Widely quoted as though Tier 1 — it is not. |
| Taiwan MOF export orders; Korea 20-day export data | T1 free | Monthly · Korea 20-day ~21st | Government customs data on electronics and semiconductor exports. Korea's 20-day series is among the earliest hard macro reads on memory and logic demand available anywhere. | Aggregated categories; strong price/volume mixing effects, particularly during memory price swings. |
| MLCommons / MLPerf | T2 free | ~2–3x per year | Audited, like-for-like training and inference throughput across accelerators. The only vendor-neutral performance comparison that submits to review. | Vendors choose which configurations to submit; absence of a submission is itself information. Benchmark workloads lag production workloads. |
| SemiAnalysis | T3 paid | Continuous | Supply chain reporting on packaging capacity, cluster builds, networking architecture and cost-per-token modelling at a depth no free source matches. | Opinionated, occasionally wrong, and directionally influential enough to move markets — which is itself a reason for care. Corroborate the load-bearing claims. |
Metrics to maintain: data-centre revenue and gross margin by vendor · purchase commitments and prepayments · HBM contract price and allocation · CoWoS-class packaging capacity (estimate, flagged) · Korea/Taiwan export series · deployed H100-equivalents and IT power (Epoch).
L4Data centres and cloud
Moat durability: weak, capital-driven
The layer where announcement inflation is most severe and where the accounting assumptions matter more than the engineering. Depreciation schedule disclosures deserve standing attention: in the 1 GW model, the useful-life assumption moved project economics by roughly 8.7 times more than the choice of energy source.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| Hyperscaler 10-K / 10-QMicrosoft, Alphabet, Amazon, Meta, Oracle | T1 free | Quarterly | Capex, finance lease obligations, purchase commitments, remaining performance obligations, and — critically — the property and equipment useful-life policy note, where server life assumptions are stated and changed. | Capex excludes capacity taken through leases and third-party contracts, so headline capex understates capacity added. Useful-life changes are disclosed quietly and have a large, immediate earnings effect. Read the accounting policy note every quarter, not the press release. |
| Epoch AI — AI Data Centers hub | T2 free | Continuous | Project-level power, compute and capital cost estimates built from satellite imagery and permits, with build-stage tracking. The best available answer to "announced versus actually under construction". | Estimates with wide intervals; largest sites only; US-weighted. |
| Local permitting and utility filingsCounty planning portals, state air permits for backup generation, utility large-load service agreements | T1 free | Continuous | The ground truth on any specific project: site area, generator count and MW, water allocation, construction start. A filed air permit is far stronger evidence than a press release. | Labour-intensive and jurisdiction-specific. No aggregation layer exists; this is manual work per project. |
| CBRE and JLL data centre reports; datacenterHawk | T3 free | Semi-annual / quarterly | Vacancy, absorption, pre-leasing rates and asking rents by metro. Pre-leasing rates are the best available signal of genuine contracted demand versus speculative build. | Brokerages have a transactional interest in reporting tight markets. Definitions of "under construction" vary between houses. |
| Uptime Institute Global Data Center Survey | T2 free | Annual | PUE, cooling adoption, rack density and outage causes from an operator survey with a long consistent history. | Self-selected respondent base skews to enterprise operators; hyperscale AI facilities are under-represented. |
| Synergy Research; Dell'Oro; Omdia | T3 paid | Quarterly | Cloud market share, capex aggregation, data-centre physical infrastructure and networking equipment share. Omdia's accelerator shipment estimates feed the IEA projections. | Proprietary methods, frequent revisions. Note that using Omdia and IEA together is not two independent sources. |
| Data Center Dynamics; Data Center Frontier | T4 free | Daily | Project announcement detection and the reference that leads to the permit or filing. | Announcements are reported as fact with no build-stage qualification. Use to find, never to count. |
Metrics to maintain: hyperscaler capex plus finance leases · server useful-life assumption by company with change history · sixteen largest gigawatt-scale projects tracked at announced / permitted / under construction / energised · pre-leasing rates by metro · PUE trend.
L5Models
Moat durability: weak — no material chokepoint
Almost no audited financial data exists for the most important companies at this layer, because the largest are private. Revenue figures circulating in the press are unaudited, frequently annualised from a single month, and should never be treated as comparable to a reported figure.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| Epoch AI — Models, Capabilities, Chip Sales | T2 free | Continuous | Training compute, parameter counts and release history for 3,600+ models; benchmark performance tracking; documented methodology and downloadable data. Also publishes hyperscaler capex analysis and compute-cost-share estimates. | Training compute for closed models is estimated from disclosed and inferred parameters. Treat as estimate throughout. |
| Stanford HAI AI Index | T2 free | Annual · April | The standard annual compendium: capability, investment, adoption, policy and robotics chapters, with a published data appendix and citable underlying series. | Annual cadence means it is a lagging record, not a monitoring tool. Some series are re-published from IFR and other bodies rather than independently collected. |
| Lab primary documentsModel cards, system cards, technical reports, pricing pages | T1 free | Per release | Context window, modality, pricing per token, stated evaluation results and safety methodology. Pricing pages are the cleanest read on the direction of inference cost. | Self-reported benchmarks, self-selected comparisons. Pricing does not reveal gross margin. |
| Artificial Analysis; LMArena | T2 free | Continuous | Independent measured price, latency, throughput and quality across hosted models — the practical basis for judging whether model capability is commoditising. | Measured through APIs, so results reflect serving configuration as much as model. Arena rankings measure human preference, not task capability. |
| Private-company revenue reportingThe Information, Reuters, FT | T4 | Ad hoc | Directional signal on private lab revenue, funding and compute commitments. | Unaudited, single-sourced, usually annualised run-rate. Label as reported-not-verified wherever used. Do not construct valuation arithmetic on these figures. |
Metrics to maintain: frontier training compute trend · price per million tokens at constant capability · measured throughput and latency · disclosed compute commitments · capability-per-dollar trend.
L6Software, agents and applications
Moat durability: contested
Well covered by ordinary equity disclosure, and therefore the layer where the temptation is to over-source. The analytically useful numbers are few: remaining performance obligations, net revenue retention, and gross margin trajectory as inference costs are absorbed.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| SEC filings via EDGAR full-text search | T1 free | Quarterly / continuous | RPO and its current portion, net revenue retention, deferred revenue, stock-based compensation, and gross margin by segment. RPO is the disclosure that distinguishes contracted demand from pipeline narrative. | NRR definitions are company-defined and change without prominence. AI-attributed revenue claims in press releases are rarely reconcilable to segment disclosure. |
| Earnings call transcriptsvia Quartr / S&P connectors | T1 | Quarterly | Management's stated unit economics of AI features, pricing model changes, and — most usefully — what they decline to quantify when asked directly. | Prepared remarks are marketing. The Q&A is the evidence. |
| Ramp AI Index; Stack Overflow and JetBrains developer surveys | T3 free | Monthly / annual | Corporate AI spend penetration from card data; developer tool adoption. Useful early adoption reads not visible in financials. | Ramp's panel is skewed to US venture-backed firms. Developer surveys are self-selected. |
| Similarweb; Sensor Tower | T3 paid | Monthly | Consumer-facing usage and app-download trends where the company discloses no user numbers. | Panel-derived and known to diverge substantially from disclosed figures where both exist. |
Metrics to maintain: RPO and current RPO growth · net revenue retention · gross margin trend against inference cost · disclosed AI-attributed ARR with reconciliation status.
L7Embodiment: robotics, actuation, autonomy
Moat durability: moderate
Industrial robotics has excellent statistics; humanoid robotics has almost none. The most important variable in the humanoid thesis — global precision reducer manufacturing capacity — has no authoritative public source, and the register should say so plainly rather than borrow a number.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| IFR World Roboticsifr.org — plus January trends release and preliminary data at Automate | T2 paid (releases free) | Full report Sep · trends Jan · preliminary ~Jun | The authoritative industrial and service robot installation and operational stock dataset, collected from manufacturers. Preliminary 2025 data reported global installations up 15% to a record 621,000 units, with Asia at 79% of installations and US installations up 11% to 38,000; the market value of installations reached an all-time high of $16.7bn. | Full dataset is expensive; press releases carry headline figures only. Humanoids are not yet a separate reported category — IFR discusses them qualitatively in the trends release. Prior-year figures are revised. |
| A3 (Association for Advancing Automation) | T2 free | Quarterly | North American robot orders and shipments by industry, released faster than IFR — the better read on turning points in the NA market. | North America only; member-reported. |
| Component supplier filingsHarmonic Drive Systems, Nabtesco, THK, Yaskawa, Fanuc | T1 free | Quarterly | Order intake, capacity utilisation and capex for precision reducers and actuator components. These filings are the best available public proxy for the reducer capacity constraint — specifically, the capex line and any announced plant expansion. | Japanese primary filings; English summaries abbreviated. Unit capacity is not disclosed, only revenue and capex. Any global reducer capacity figure is a derivation, not a fact. |
| Chinese suppliersLeaderdrive, Zhejiang Shuanghuan, Sanhua, Tuopu — Shenzhen/Shanghai filings | T1 | Quarterly / annual | The clearest evidence on whether the Japanese reducer oligopoly is being displaced, and on Chinese humanoid component capacity build-out. | Disclosure standards, auditor quality and related-party practice differ materially from developed markets. Capacity announcements from Chinese local governments are promotional. Corroborate against customer disclosures. |
| NVIDIA robotics platform disclosures; GTC materials | T4 free | 2x per year | Platform architecture, edge compute specifications, partner announcements. | Keynote content is marketing with a product roadmap attached. Partner logos are not deployments. |
| Humanoid developer claimsTesla, Figure, Agility, Unitree, UBTech | T4 | Ad hoc | Stated production targets, pilot deployments and unit costs. | Treat every figure as marketing until corroborated by a customer disclosure, a component order, or a filed accounts entry. Production "capacity" claims in this sector have historically exceeded actual output by an order of magnitude. UBTech's listed status makes it the only one with audited numbers. |
Metrics to maintain: IFR installations and operational stock · A3 NA orders · reducer supplier order intake and capex · Chinese reducer share trend · humanoid units actually shipped, from audited or customer sources only.
04 — Cross-cutting
Sources spanning the whole stack
Company-level, market-level and macro sources that apply regardless of layer, plus the connected data pipeline already available in this workspace.
| Source | Tier | Cadence | Take from it | Caveat |
|---|---|---|---|---|
| SEC EDGAR full-text search | T1 free | Continuous | The primary document for every US-listed name: 10-K, 10-Q, 8-K, S-1, DEF 14A, Form 4. Full-text search across filings finds contract counterparties, warrant terms and risk-factor language changes that no summary captures. | Non-US issuers file 20-F annually with far less frequent disclosure; European and Japanese holdings need their home-market regulator instead. |
| Connected data pipelineQuartr · S&P Kensho · Morningstar · Interactive Brokers · Google Drive | T1/T3 | Continuous | Quartr for transcripts and filings; S&P for standardised financials and estimates; Morningstar for fair value, moat and uncertainty ratings as an external cross-check; IBKR for live positions, account metrics and prices. This is the workspace's own primary retrieval layer and should be used before web search for anything company-specific. | Standardised financials involve vendor adjustments that differ from as-reported figures; state which basis is used. Morningstar ratings are opinions and are treated as a comparison point, not an input. |
| Company IR primaryInvestor decks, 8-Ks, prospectuses, shareholder letters | T1 free | Continuous | Lock-up terms, share class structure, warrant agreements, PPA announcements. Two open items sit here: SpaceX lock-up terms from the prospectus, and whether the Roche line held is the bearer share or the participation certificate. | Investor decks are the most curated document a company produces. Prefer the filed version of the same information. |
| FRED (St. Louis Fed) | T1 free | Continuous | Rates, credit spreads, industrial production, PPI for electrical equipment — the macro series that drive discount rates and input costs. | Series revisions and definitional breaks; check the notes on any series used in a model. |
| SNB, ECB, Federal Reserve H.15 | T1 free | Daily / monthly | Official CHF reference rates, policy rates and yield curves — the basis for currency translation and for the cost of the USD margin loan against a CHF base. | Reference rates are not execution rates; broker financing spreads must come from the IBKR statement. |
| Federal Register; EUR-Lex; Swiss Federal Gazette | T1 free | Continuous | The actual text of export controls, tariffs, subsidy programmes and permitting rules across all layers. | Effective dates, grandfathering and licence policy are where the economic effect lives, and they are in the body of the rule. |
| 13F and Form 4 filings | T1 free | Quarterly / 2 days | Insider transactions (Form 4) are timely and meaningful, particularly sales outside a 10b5-1 plan. | 13F data is 45 days stale, long-only, and excludes derivatives and shorts. It is close to useless for inference about current positioning and should generally be omitted. |
05 — Release calendar
What arrives when
The monthly newsletter cycle should be built around this calendar rather than around news flow. Dates are typical publication windows and should be confirmed against each publisher; several shift year to year.
Monthly, recurring
- TSMC revenue (~10th)
- SEMI & SEAJ billings
- WSTS/SIA billings
- Korea 20-day exports (~21st)
- China customs export data
- EIA Short-Term Energy Outlook
- ERCOT queue update
- TrendForce contract prices
January
- IFR Top 5 robotics trends
- PJM load forecast
- CES product disclosures
February
- USGS Mineral Commodity Summaries
- IEA Electricity annual report
- Q4 / full-year filings begin
March–April
- Stanford HAI AI Index (Apr)
- LBNL "Queued Up" (~Apr)
- 10-K season completes
- NVIDIA GTC
May–June
- IEA Critical Minerals Outlook
- IFR preliminary installation data
- Automate / trade shows
- PJM capacity auction results
September
- IFR World Robotics full report
- Uptime Institute survey
- Hyperscaler fiscal year-ends begin
December
- NERC Long-Term Reliability Assessment
- Year-end capex guidance updates
Scheduled risk dates
- November 2026 — Chinese gallium/germanium and rare earth export control suspension expiry. Verify status directly against MOFCOM.
06 — Exclusions
Sources that should not enter the report
Not because they are always wrong, but because they cannot be audited and their errors are systematic rather than random.
| Category | Why excluded | What to use instead |
|---|---|---|
| Syndicated "market to reach $X by 2032" reportsGrand View, MarketsandMarkets, Precedence and similar | Sold on volume, methodologies unpublished, forecasts derived from press releases. The same market is frequently sized differently by a factor of five across vendors. | Trade body statistics (SEMI, WSTS, IFR), or bottom-up construction from company disclosure. |
| SEO content aggregators and "AI statistics 2026" listicles | Recycle figures without provenance, frequently mis-transcribe units, and create the illusion of corroboration by repetition. | The named primary source, located directly. |
| Press release capacity and production announcements | Announced capacity in this stack has diverged from delivered capacity by very large margins, particularly in Chinese robotics and in data centre siting. | Permits, filed accounts, customer disclosures, satellite-derived build tracking. |
| Sell-side price targets | Not evidence about a business. Useful only as a record of consensus positioning, and then only in aggregate. | Consensus estimate distribution and revision direction, treated as a sentiment indicator. |
| Social media supply chain "leaks" | Unverifiable, occasionally deliberate, and disproportionately influential on short-term pricing. | Wait for the filing or the customs data. |
| Any figure re-cited from a prior report in this project | Several figures in the existing project documents are explicitly flagged as the author's own derivations. Re-citing them as facts would launder an estimate into evidence. | Return to the original source, or carry the ESTIMATE label forward intact. |
07 — Known gaps
Where no reliable source exists
Honest accounting of the variables that matter to the thesis and cannot currently be sourced to the required standard. Each should appear in the report with the phrase "Data unavailable from accessible sources" rather than an estimate presented as fact.
| Variable | Why it matters | Best available proxy |
|---|---|---|
| Global precision reducer manufacturing capacity | Identified as the binding constraint on 10M+ unit humanoid deployment. The headline conclusion of the humanoid analysis rests on it. | Derived from industrial robot installation volumes — a derivation, not a source. Cross-check against Harmonic Drive and Nabtesco capex and any announced plant expansion. Flag as the figure most in need of independent verification. |
| High-purity quartz production volumes (Spruce Pine) | The most concentrated single-point dependency in the stack, upstream of every crucible and therefore every wafer. | Sibelco and Quartz Corp are private; the industry is described by researchers as deliberately opaque. Third-party volume estimates vary widely and none should be presented as precise. |
| Transformer and HV cable lead times | A schedule constraint on Layer 1 and Layer 4 that translates directly into project slippage. | Equipment maker backlog and book-to-bill, read alongside utility procurement commentary. No published lead-time series exists. |
| CoWoS-class advanced packaging capacity in wafers per month | The binding constraint on accelerator supply and therefore on the whole compute layer. | TSMC's qualitative commentary plus paid supply-chain research. Every specific number in public circulation is a third-party estimate; none is disclosed. |
| HBM allocation by customer | Determines which accelerator vendors can actually ship volume. | TrendForce estimates and memory maker commentary. Estimates only, and historically revised substantially. |
| Private lab audited financials | Layer 5 economics, and the durability of the demand driving the entire stack. | None. Reported run-rate figures are unaudited. This is a structural gap in any AI infrastructure thesis and should be stated as such rather than papered over. |
| Contracted GPU-hour pricing | The revenue input that determined project viability in the 1 GW data centre model. | Neocloud published rates and disclosed contract values, which reflect spot and short-term pricing rather than the long-term contracts that underwrite construction. |
Verification status of this document. The IEA data-centre demand figures, USGS Mineral Commodity Summaries 2026, Epoch AI data hub coverage statistics, and IFR 2025 preliminary installation figures cited above were checked against the publishers' own material on 4 September 2026. Remaining entries describe sources and their characteristics rather than asserting specific figures, and their publication schedules should be confirmed at point of use. Where a URL is given it is a publisher root or section path; deep links change and should be located by search rather than assumed.