factor stringclasses 7
values | metric stringlengths 20 82 | value stringlengths 3 806 | source stringlengths 3 99 | vintage stringlengths 4 35 |
|---|---|---|---|---|
1 Capex vs revenue + credit | Combined 2026 hyperscaler capex (Alphabet+Amazon+Meta+Microsoft) | ≈$720-745B (GOOGL 195-205 + AMZN ≈220 + META 130-145 + MSFT 175 stated); capex increases reflect component inflation as well as capacity | v6 BOUNDARY NOTE: the MSFT $175B is FY2027 capex AND finance leases, net of a finance-to-operating lease reclassification following the extension of DATACENTRE AND OFFICE BUILDING l... | company guidance | Jul-2026 |
1 Capex vs revenue + credit | Incremental annual AI revenue (industry) | ≈$20-30B (analyst estimate, register's weakest number) | analyst estimate | 2026 |
1 Capex vs revenue + credit | AI accelerator refresh / obsolescence clock | ≈2-3 years | industry | 2026 |
1 Capex vs revenue + credit | Data-centre lease commitments signed but not commenced (off balance sheet) | ≈$662B; larger than the same firms' combined on-BS debt | Moody's | early-2026 |
1 Capex vs revenue + credit | AI-related debt financing projected through 2030 | ≈$4.1T (within ≈$5.5T total AI capital spend); ≈15% of the corporate-bond universe; >$300B issued in 2026 YTD | JPMorgan | 2026 |
1 Capex vs revenue + credit | Fed policy rate (the rate-sensitivity lever) | held 3.50-3.75% for 5th consecutive meeting (9-3 vote); median end-2026 3.8% from Jun-2026 SEP | FOMC (Warsh) | 29-Jul-2026 |
1 Capex vs revenue + credit | Market-live tape and financing gradient INVERSION | Meta $589.85 (-8.4% since 21-Jul); Oracle $150.52 (+18.5% since 21-Jul, off BBB- trough ≈$121); Micron $971.66 (-22.6% from $1,255.50 peak, flat vs $970.82 21-Jul); financing gradient inverted on tape | Nasdaq historical | 14-Aug-2026 |
1 Capex vs revenue + credit | Single-counterparty concentration (the structural sub-factor) | Oracle cut to BBB- on customer concentration; OpenAI ≈half of a ≈$638B backlog; circular financing connects exposures | S&P / market | 9-Jul-2026 |
1 Capex vs revenue + credit | Memory-spend crowd-out landing in earnings | IBM -25.21%, worst single day on record, ≈-$68.8B; clients pulled memory/hardware spend forward, deferred software/mainframe/consulting | company pre-announcement | 14-Jul-2026 |
1 Capex vs revenue + credit | Capex aggregate - stated as a RANGE not a point | component guides sum to ≈$720-745B (GOOGL $195-205B; AMZN ≈$220B; META $130-145B; MSFT $175B stated post asset-life extension); reported aggregates track $670-710B | v6 BOUNDARY NOTE: the MSFT $175B is FY2027 capex AND finance leases, net of a finance-to-operating lease reclassification following the extension of DATAC... | company guidance / press | Jul-2026 |
1 Capex vs revenue + credit | ① TRIGGER RECALIBRATION - 4 quarters without a cut; focus shifts to Nvidia gap | Q2 prints passed with zero guidance cuts (GOOGL $195-205B, AMZN ≈$220B, META $130-145B, MSFT $175B stated). Trigger failed to fire while financing strain appeared in FCF and debt. The informative signal shifts to 26 Aug 2026: hyperscaler capex vs Nvidia take-rate gap via custom silicon (TPU, Trainium) and AMD | v6 BOUN... | company calls / analyst | Aug-2026 |
1 Capex vs revenue + credit | PART OF THE RAISE IS PRICE, NOT CAPACITY (Amazon and Microsoft decompositions) | Amazon raised capex to ≈$220B explicitly citing memory chip inflation (Jassy: 'The higher cost of memory is pushing this number up from our prior estimate of about $200 billion'); Microsoft attributes ≈$25B to component pricing; Meta cites memory pricing. A raise driven by factor 4 is cost pass-through, not added convi... | company guidance / calls | Jul-2026 |
1 Capex vs revenue + credit | DEFLATION - the $20-30B revenue figure is the register's weakest number | Microsoft AI business at a $37B annualised run rate (+123% YoY); Google Cloud +63% YoY to $20.02B in Q1-2026 (+82% Q2); Alphabet describes itself as compute constrained. Company-reported and defined differently from an industry increment, so not a replacement, but the capital-to-revenue gap is probably narrower than $7... | company disclosures | Jul-2026 |
1 Capex vs revenue + credit | Cash generation under the cycle | Amazon TTM free cash flow ≈$1.2B, about -95% YoY, at a firm whose demand is not in question | company disclosure | Q1-2026 |
1 Capex vs revenue + credit | Circular financing (why concentration is not one balance sheet) | labs buy cloud from hyperscalers; hyperscalers and chip vendors hold lab equity; chip vendors invest in the deployers who buy their chips; hyperscalers prepay memory makers years ahead. Oracle is where it surfaced first (named counterparty, public rating), not the only place it sits | author's reading | 21-Jul-2026 |
1 Capex vs revenue + credit | Credit conditions (the deflation) | IG OAS 80 bps (BAMLC0A0CM, up from 74 bps in Jun-2026); HY OAS 2.82% (BAMLH0A0HYM2, up from 2.75% in Jun-2026); stress remains issuer-specific | ICE BofA via FRED | 14-Aug-2026 |
1 Capex vs revenue + credit | Single-campus property reinsurance capacity & TIV limits | multi-gigawatt campuses represent $15-40B+ in Total Insured Value; catastrophe and fire treaty capacity limits act as a structural ceiling on bank debt syndication | Lloyd's / reinsurance circulars | 2026 |
2 Grid / power | PJM data-centre-attributed capacity cost (BRA 26/27 + 27/28) | ≈$13.8B ($7.3B + $6.5B) | PJM market monitor (Monitoring Analytics) | 2025-26 |
2 Grid / power | PJM capacity clearing price, 2026/27 vs 2024/25 | $329.17/MW-day vs $28.92 | PJM BRA | 2025-26 |
2 Grid / power | PJM 2028/29 BRA - the framing shift (price capped, constraint shows as volume) | cleared at the $325/MW-day cap (≈2.5% below prior cap); 6,831 MW SHORT; only 525 MW new generation; collar held it (would have cleared $554.72 / ≈$30B vs $16.4B) | PJM BRA | 14-Jul-2026 |
2 Grid / power | PJM post-auction regulatory expansion (large-load framework) | FERC filings 31 Jul ($20B procurement for generation serving datacentres) and 7 Aug (bring-your-own-power framework for >=50 MW loads post-2027); transforms grid bottleneck into active regulatory rationing | PJM / FERC filings | Aug-2026 |
2 Grid / power | Interconnection-queue dropout at first decision | 28-31% | PJM Reliability Resource Initiative disclosures | Q1/Q2-2026 |
2 Grid / power | Skilled high-voltage electrical labour (the non-tradable constraint) | certified electricians (SOC 47-2111) and line installers (SOC 49-9051) require 4-5 yr apprenticeships; non-tradable human pipeline; wage growth outrunning general ECI; utility commissioning delayed on crew availability | US BLS / DOL RAPIDS | 2026 |
2 Grid / power | Prime movers (the on-site generation queue) | heavy-duty gas turbine slots backlogged 3-5 yrs at GE Vernova, Siemens Energy, MHI; on-site bypass has its own manufacturing queue | GE Vernova / Siemens 10-Q | Jul/Aug-2026 |
2 Grid / power | Turbine hot-section superalloy castings (the prime-mover floor) | single-crystal turbine blades constrained at Howmet Aerospace and Precision Castparts; aviation and power compete for vacuum investment casting capacity | Howmet 10-Q / industry | 2026 |
2 Grid / power | Large power transformer (LPT) lead time | 128 weeks | Wood Mackenzie survey (corrob. DOE/trade press) | Q2-2025 |
2 Grid / power | Generator step-up (GSU) transformer lead time | 144 weeks | Wood Mackenzie survey (corrob. DOE/trade press) | Q2-2025 |
2 Grid / power | Extra-high-voltage transformer lead time | 36-48 months (to ≈60mo extreme) | Wood Mackenzie / US DOE | Q2-2025 |
2 Grid / power | Grid equipment beyond transformers (switchgear, breakers, HVDC) | HV gas-insulated switchgear, SF6-free circuit breakers, and HVDC cables run 3-5 yr lead times (Hitachi Energy, Prysmian, NKT) | OEM order books | 2026 |
2 Grid / power | US data-centre electricity demand | 176 TWh 2023 (≈4.4% of US elec) -> 325-580 TWh (6.7-12%) by 2028 | Lawrence Berkeley National Laboratory | 2024 (->2028) |
2 Grid / power | Global data-centre electricity demand | ≈416 TWh 2024 -> ≈946 TWh 2030 (Base); range 669 (Headwinds) to 1,264 (Lift-Off); PUE 1.41 -> ≈1.29 | IEA 'Energy and AI' data annex | 2025 (->2030/35) |
2 Grid / power | Projection-range deflation (the caveat on every figure above) | 2030 projections span ≈200 to ≈8,000 TWh across 100+ studies (≈40x); triangulated 2023 ≈300-380 TWh; widest figures from temporal extrapolation | IEA 4E EDNA (Kamiya & Coroama) | Mar-2025 |
2 Grid / power | AI rack power draw vs conventional | 30-100+ kW vs 7-10 kW | industry | 2026 |
2 Grid / power | Per-query energy - REVISED DOWN | ≈0.24 Wh (Google median Gemini text) / ≈0.34 Wh (OpenAI), vs the widely-cited ≈2.9 Wh; heavy reasoning/multimodal ≈18-40 Wh; agentic use inverts the per-TASK story | Google / OpenAI | 2025 (published 2026) |
2 Grid / power | CORRECTION - Lake Powell minimum power pool (3,490 ft) | breach DEFERRED past 2026: release cut 7.48 -> 6.00 maf; July-2026 Probable Minimum trace projects 3,503.65 ft at 31-Dec-2026 (≈13.7 ft above 3,490 ft power pool). Reclamation traces carry vintage mismatch (Min Jul-2026, Max May-2026; unregulated inflows 36% of average) | US Bureau of Reclamation 24-Month Study (Jul-2026 Min) | Jul-2026 |
2 Grid / power | Water siting as a physical constraint (not an externality) | municipal moratoria and basin allocations remove candidate sites outright; USBR Colorado River shortage guidelines and ADWR groundwater rules cap arid-basin deployment | USBR / ADWR / municipal dockets | May/Aug-2026 |
2 Grid / power | Nuclear baseload & SMR fuel chokepoint (HALEU) | NRC reactor restart licensing takes 3-5 yrs; commercial HALEU (5-20% U-235) enrichment pilot-stage in US (Centrus), historical supply Russia-dominated | US NRC / Centrus Energy 10-Q | 2026 |
2 Grid / power | On-site generation externality ('bring your own power') | Colossus 2, Memphis: 59 unpermitted gas turbines; federal air-permit loophole closed Jan-2026; substation lead times 3-5 yrs drive the bypass | press / EPA | 2026 |
2 Grid / power | Energy-shock feedback into factor 1 | Brent ≈$87 2026 average ($85 Q3, $78 Q4, $69 2027); recovery delayed to early 2027 on Hormuz transit constraints; energy cost moves from tail risk to 2026 baseline friction | EIA STEO | 11-Aug-2026 |
2 Grid / power | Hyperscaler renting OUT OF SHORTAGE (against the glut read) | Alphabet 22 Jul 2026: 'still in a supply-constrained environment', demand outpacing 3 years of rising investment; will expand THIRD-PARTY cloud capacity in Q3 as a bridge while it builds its own = a hyperscaler renting because it cannot build fast enough | company call / CNBC | 22-Jul-2026 |
3 Taiwan leading-edge | TSMC share of leading-edge logic | ≈90% | industry / TSMC | 2026 |
3 Taiwan leading-edge | Advanced packaging (CoWoS) concentration | effectively 100% Taiwan, no second source; leading edge Taiwan-bound through ≈2028-30 | industry | 2026 |
3 Taiwan leading-edge | Substrates (ABF) manufacturing concentration | Ajinomoto Build-up Film substrates for large CoWoS packages concentrated in Japan/Taiwan (Ibiden, Shinko, Unimicron); yield drops non-linearly with package size | Ibiden / Shinko filings | 2026 |
3 Taiwan leading-edge | Specialty EUV photoresists and polymer precursors | >85% of advanced EUV photoresists, PAGs, and fluorinated precursors controlled by Japanese chemical firms (Tokyo Ohka Kogyo, JSR, Shin-Etsu) under METI export regime | Japan METI / company reports | 2026 |
3 Taiwan leading-edge | High-purity quartz (HPQ) crucibles (grade leader; world share NOT published) | Two producers operate around Spruce Pine NC (Sibelco, The Quartz Corp); material exceeds 99.999% purity, the highest published grade. WORLD SHARE WITHDRAWN AT v6: v4/v5 claimed >80-90% citing USGS MCS, but the MCS 2026 Quartz chapter states "World Mine Production and Reserves: This information was not available" and no... | USGS Mineral Commodity Summaries 2026, Quartz (High-Purity and Industrial Cultured Crystal) chapter | 2026 (corrected 19-Aug-2026) |
3 Taiwan leading-edge | Lloyd's JWC listed areas (the trigger's live reading) | Taiwan Strait NOT listed as of JWLA-032/033; that amendment extended the Gulf instead. Baseline hull war ≈0.15-0.25% of hull value; Gulf now ≈1% | Lloyd's Market Association Joint War Committee | 3-Mar-2026 |
4 Chokepoints (made + licensed) | DRAM/HBM shortage horizon | to ≈2028; Micron/SK Hynix sold out through 2027-2028; SK Hynix CEO sees tightness extending beyond 2028 to 2030 | earnings / industry | Jul-2026 |
4 Chokepoints (made + licensed) | Memory incidence inversion (allocation effect holds; margin motive INVERTED at v6) | HBM takes ~22% of DRAM wafer INPUT in 2026 (18% 2025, 30% 2027) while supplying only ~9% of DRAM BITS; that wafer-to-bit gap is the crowding-out mechanism. MARGIN MOTIVE NO LONGER HOLDS: TrendForce June 2026 per-wafer analysis finds HBM wafer revenue was overtaken by DDR5 64GB RDIMM in Q1 2026, with HBM profitability b... | TrendForce (2 June 2026 per-wafer analysis) | Jun-2026 (corrected 19-Aug-2026) |
4 Chokepoints (made + licensed) | Automated Test Equipment (ATE) memory test scaling | HBM3e/HBM4 stacks require 4x to 6x the test duration of commodity DRAM across wafer probe and package test; Advantest and Teradyne test slot capacity is an inelastic throughput governor | Advantest / Teradyne 10-Q | 2026 |
4 Chokepoints (made + licensed) | Optical transceivers & InP CW laser diodes | scale-out 800G/1.6T networking requires Continuous-Wave InP laser diodes for silicon photonics, concentrated at Coherent, Lumentum, Broadcom | Coherent / Lumentum 10-Q | 2026 |
4 Chokepoints (made + licensed) | Strategic equity linkage (① and ④ coupling) | Anthropic Series H ($65B, $965B post-money, 28 May 2026) brought Micron, Samsung, SK Hynix as strategic partners alongside Amazon ($5B); chokepoints and frontier demand become one financing node | company filings / press | 28-May-2026 |
4 Chokepoints (made + licensed) | Nanya (DRAM) revenue | +583% YoY (Q1, off trough; price-driven, >70% DRAM price spike, not volume) | Nanya Q1-2026 earnings | Q1-2026 |
4 Chokepoints (made + licensed) | Memory de-cyclicalisation (may cut against the boom-bust read) | supply contracts lengthened to 3-5 years with prepayments up to 30% | SK Hynix / press | Jul-2026 |
4 Chokepoints (made + licensed) | SECOND packaging chokepoint (new in v3) | ASE Technology (largest OSAT) raised 2026 capex to $8.5B; critical node for co-packaged optics; distinct from TSMC CoWoS | company / press | 2026 |
4 Chokepoints (made + licensed) | EUV lithography optics | ASML / Zeiss sole-source | industry | 2026 |
4 Chokepoints (made + licensed) | China share of rare-earth refining/separation (historical framing) | China refines most REEs and dominates heavy-REE separation/magnets; earlier editions gave ≈91% refining / ≈60% mining (IEA / Rhodium Group) | IEA / Rhodium Group (historical) | 2024-26 |
4 Chokepoints (made + licensed) | China share of magnet-grade REE mining | processing not mining is the structural bottleneck | IEA | 2024-25 |
4 Chokepoints (made + licensed) | Western magnet alternatives at full production | not expected before 2027 (MP Materials, USA Rare Earth) | industry | 2026 |
4 Chokepoints (made + licensed) | China export-control ladder on materials | gallium/germanium 2023; US-specific Ga/Ge/antimony ban Dec-2024; Apr-2025 per-shipment licensing over 7 medium+heavy REEs and magnets | China MOFCOM | 2023-25 |
4 Chokepoints (made + licensed) | EU heavy-REE price (Dy/Tb) vs China, at peak | ≈6x | IEA (European Parliament corrob.) | 2025 |
4 Chokepoints (made + licensed) | Oct-2025 EXTRATERRITORIAL expansion + the dated trigger | any foreign product with >=0.1% Chinese-origin REE content would need a licence; suspended one year under truce; REVIEWED AROUND NOVEMBER 2026 (v3 states the month: the precise 27 Nov date carried in v2.x was not confirmed at source) | China MOFCOM | 2025-26 |
4 Chokepoints (made + licensed) | Helium (Ras Laffan strike) | RECLASSIFIED AT v6 TO THE NON-RELOCATABLE TAIL. Halt 2 Mar 2026; QatarEnergy force majeure 4 Mar; strikes 18-19 Mar damaged two LNG trains and a GTL facility, ~17% of Qatar LNG export capacity. Share of global helium affected: 30-38% as capacity share vs ~11% as annualised 2026 volumetric loss (AKAP Energy) - a boundar... | press / company statements | Mar-2026 (reclassified 19-Aug-2026) |
4 Chokepoints (made + licensed) | SEVERITY SPLIT within factor 4 (do not flatten these) | CORRECTED AT v6. v5 classified helium as an elastic BUFFER problem = months while the same register recorded repairs up to 5 yrs; those cannot both hold. Primaries support the multi-year reading, so helium now sits in the tail alongside heavy-REE separation. The buffer observation stands as a timing effect: a 3-6 month... | author's reading | 19-Aug-2026 |
5 Commoditisation from below | Open-weight frontier datapoint | Kimi K3 (Moonshot), 2.8T MoE, $3/$15 per M tokens; near the top closed tier on one aggregate index while trailing the leading closed model overall; reported to beat Fable 5 and GPT-5.6 Sol in blind front-end coding tests | Moonshot / Axios | 16-Jul-2026 |
5 Commoditisation from below | MEASUREMENT CAVEAT - least quantified of the six | the stated lens (yield on a deployed megawatt) is NOT directly observable: no firm reports revenue per MW of deployed capacity. Token price index is a stand-in, not a measure. Carried on a directional argument (token prices falling while capital cost per MW rises on components), not a measured ratio | author's reading | 21-Jul-2026 |
5 Commoditisation from below | Frontier token price index | ≈-20% from its May-2026 peak on substitution to cheaper models | Citadel (reported) | Jul-2026 |
5 Commoditisation from below | Industrial policy behind open weights | Xi framed open-weight models as a global public good (World AI Conference) | state media / press | 17-Jul-2026 |
5 Commoditisation from below | INVERSE-RISK ROW - the shortage and glut co-existence (Micron paradox) | Micron sold out through 2027 but stock closed $971.66 on 14-Aug (-22.6% from $1,255.50 peak); Samsung -33%; SK Hynix -15% broke ADR issue price; IDC projects 2026 smartphone shipments -14% (largest decline on record); Micron exits consumer memory. Shortage tail and glut tail co-exist in the same company | Nasdaq / IDC | 14-Aug-2026 |
5 Commoditisation from below | Inverse row - the FIRST PRICED market response (two-sided, not demand-side) | on the 1 Jul report Meta +≈9%; CoreWeave -≈14% in a session, Nebius -≈17%, both -≈35% over the following month. Meta had contracted ≈$48B from the two (≈$27B Nebius Mar-2026; $35.2B total CoreWeave after an Apr expansion), so a self-supplying Meta WITHDRAWS a buyer and ADDS a seller = reallocation of who sells compute,... | press | Jul-2026 |
5 Commoditisation from below | Inverse row - two constraints on the strong version | Meta's CoreWeave contract reportedly carries a NO-SUBLEASE provision barring resale of rented capacity; and several observers infer the opposite from the same fact, that surplus worth selling implies a market large enough to absorb it (growth, not glut) | press / analyst | Jul-2026 |
5 Commoditisation from below | Inverse row - why not base case | leasing is the ordinary cloud model; capital efficiency, burn mitigation, or demand STRENGTH (Anthropic, short of compute, reportedly proposed it) all fit the same facts; talks preliminary. Counter-indicators: memory sold out and contracted 3-5yr ahead with prepay; PJM cleared 6,831 MW short | author's reading | 21-Jul-2026 |
5 Commoditisation from below | Inverse row - UPGRADE TRIGGER (none present at 15 Aug 2026) | multiple hyperscalers turn net seller; published idle-fleet utilisation confirms sustained low use; HBM contracts renegotiated DOWN; or capex guidance CUT outright | author's trigger | 15-Aug-2026 |
6 Export-control two-stack | First export control applied to a MODEL, not a chip | BIS 'Is-Informed' letter (no rulemaking required): export licence needed for any foreign person to access Fable 5 / Mythos 5. Models went dark worldwide; trusted-partner carve-out 26 Jun; withdrawal 30 Jun; Fable restored 1 Jul | US Commerce / BIS | 12-Jun-2026 |
6 Export-control two-stack | Contagion beyond one firm | OpenAI limited the release of GPT-5.6 at the administration's request within days | press | Jun-2026 |
6 Export-control two-stack | The substitution it generates | Macron (G7): a switch thrown 'from one day to the next' is a reason not to buy US models; spur to sovereign open-weight procurement across the EU, Gulf, Japan | press | Jun-2026 |
6 Export-control two-stack | Hardware two-stack, blocked at both ends | H200 China sales approved Dec-2025 under a 25% export tax (codified Jan-2026); trade stalled by Mar-2026, US security screening one end, Chinese import restrictions the other | press / policy | 2025-26 |
6 Export-control two-stack | Direction of travel on Chinese models (reported, not yet a rule) | administration reviving restrictions post-Kimi K3: Entity List designations, federal procurement limits, security advisories, liability requirements; deliberately short of an outright ban. Contested internally (David Sacks). Deflation: open weights may make it hard to enforce | Axios (reported) | 20-Jul-2026 |
6 Export-control two-stack | TRIGGER (new: every factor now carries a falsifier) | a second Is-Informed letter or equivalent against any frontier model; OR Entity List designation reaching a Chinese AI lab; OR a sovereign procurement mandate (EU/Gulf/Japan) requiring locally-hosted weights. Not tripped at 15 Aug 2026 | author's trigger | 15-Aug-2026 |
6 Export-control two-stack | Governance bloc split | WAICO (China-led, 29 founding states, governance) vs Pax Silica (US-led, launched Dec-2025 with ≈2 dozen, reported 35 by its 2nd summit, supply-chain). Kazakhstan in both | press | Jul-2026 |
6 Export-control two-stack | State as prospective owner (conceptual) | OpenAI proposed donating 5% of equity (≈$42.6B on an $852B Mar mark) to a US public wealth fund; precedent = 10% US stake in Intel after $8.9B CHIPS. Anthropic states it is not in talks | Financial Times | 2-Jul-2026 |
6 Export-control two-stack | Standards as the real rate-limiter | EU AI Act GPAI enforcement from 2 Aug 2026 (to 3% of global turnover, or an order to restrict/withdraw a model); Annex III high-risk DECOUPLED from the calendar and gated on harmonised CEN-CENELEC standards (up to ≈16 months + 6-month transition) | EU Commission / EUR-Lex | 2026 |
Amplifier (not a factor) | US average effective tariff rate | ≈9.1% (highest since 1946); statutory basis in flux after the 20-Feb-2026 SCOTUS IEEPA ruling, duties replaced same-day under Section 122 | US trade data / SCOTUS | 2026 |
Amplifier (not a factor) | Cooling refrigerant & F-gas quotas (AIM Act / EU F-Gas III) | HFC phasedown quotas increase chiller replacement and chemical top-up expenditure without physically halting thermal management (pricing amplifier) | US EPA / EU Commission | 2026 |
Amplifier (not a factor) | Ratepayer tariff riders and large-load structures | utility large-load demand charges and minimum take-or-pay structures reallocate grid infrastructure costs to hyperscalers (pricing amplifier) | state PUC dockets | 2026 |
Amplifier (not a factor) | Datacentre land and dark-fibre routing premiums | real estate and metro dark-fibre price premiums in tier-1 hubs reallocate capital to secondary regions (pricing amplifier) | industry market reports | 2026 |
The AI Constraint Relay: A Supply and Demand Risk Assessment
A supply-side empirical assessment of the AI compute build-out, tracking moving bottlenecks across power, labour, equipment, upstream materials, and cost incidence.
AI disclosure: the research is the author's; this text was drafted with AI assistance (Anthropic Opus 4.8-5.0 and Google Gemini 3.7) and reviewed by the author. The models, and the conflicts they create, are named in the Conflict of interest and scope section.
- Author: NM AI Research (independent analyst)
- ORCID: 0009-0003-4213-7769
- DOI: https://doi.org/10.5281/zenodo.20586863
- Source and code: https://github.com/NMAIResearch/ai-constraint-relay
Version
These files are v6, byte-identical to Zenodo record 22015680 (published 19 August 2026). Cite the concept DOI above, which always resolves to the latest version.
v6 withdrew the Spruce Pine high-purity quartz world share, because the source it was attributed to does not publish one. It moved semiconductor-grade helium from an inventory-buffered problem to the non-relocatable tail, on a three to five year repair timeline. It excluded the Microsoft $175B from the factor 1 guidance reading, as an accounting boundary change rather than a cut.
Files
supply_constraints.csv: structured metrics with value, source, motive tier, and vintage.reproduce.py: standard-library reproduction script tabulating and citing every figure from the CSV. Python 3, no dependencies, no network.LICENSE: Creative Commons Attribution 4.0 International.
The paper itself, in Markdown and PDF, is on the Zenodo record and in the GitHub repository.
Reproduce
python3 reproduce.py
Every figure in the paper is tabulated from supply_constraints.csv and cited back to its source row. Edit the CSV and re-run to reproduce the register.
Scope
The register documents displacement, not collapse. It makes no L or S score assertion, and it is not a crash call. Non-relocatable tail risks are named as physical boundaries where re-routing is impossible, which is a different claim from predicting that they bind.
- Downloads last month
- 35