open MEDIUM
I-01 · due through 2028The binding constraint on AI capacity has moved from silicon to interconnection, and we do not expect it to move back before 2028. The scarce asset is no longer a chip. It is an energised building attached to a substation with capacity to spare.
Wrong if: Any of: (a) a hyperscaler states on the record in an earnings call or filing that it is holding energised data center capacity idle for lack of demand rather than holding accelerators idle for lack of power; (b) CBRE's North America Data Center Trends report shows primary-market vacancy above 5% in any half-year through 2028; (c) no RTO or ISO reports a large-load interconnection backlog exceeding 10 GW in its 2028 filings.
Full entry, evidence and who disagrees →open MEDIUM
I-02 · due observed, 12 monthsThe most under-reported number in this sector is that industrial electricity prices in the four PJM states we track rose by double digits over the past year while Texas industrial prices fell. We think this is mostly a market-design story rather than a data-center story — but that is an interpretation of a one-year divergence, not a demonstrated causal finding.
Wrong if: Two tests, because the correlational one alone does not test the causal claim. Correlational: the PJM-minus-Texas industrial price gap narrows or reverses in the May 2027 EIA data. Mechanistic: a PJM state's industrial prices fall or flatten through a period in which its capacity costs rise — which would show the capacity-market channel is not doing the work we attribute to it. Either result should move this inference to LOW.
Full entry, evidence and who disagrees →open MEDIUM
I-03 · due observedDemand forecasts for this sector are not converging as more data arrives. They are diverging. The spread between credible forecasts is now larger than the entire quantity being forecast five years ago, and that spread is the most honest single fact in the literature.
Wrong if: The next full round of 2030 US data center forecasts from EPRI, BloombergNEF, LBNL, S&P Global and the IEA lands with midpoints inside a band of plus or minus 20% around their collective median. Scored on the first publication from each after 1 January 2027.
Full entry, evidence and who disagrees →open HIGH
I-04 · due through 2027The interconnection queue is not a measurement of demand. It is a measurement of the cost of holding an option, and everyone inside the industry knows it — which is why queue totals keep setting records that nobody in the industry actually believes.
Wrong if: By 31 December 2027, ERCOT's audit concludes that more than half of its large-load queue is backed by executed contracts, financial security and site control. Or: any RTO publishes a realisation rate above 40% for large-load requests filed three or more years earlier.
Full entry, evidence and who disagrees →open MEDIUM
I-05 · due 2026–2029This is an installation-phase frenzy using Carlota Perez's installation/deployment distinction, which means two things at once that people keep insisting are contradictory: there will be a capital event that looks like a bust, and the physical infrastructure will mostly get used anyway. Bet against the financing structure, not against the concrete.
Wrong if: Two halves, tested separately, because the first draft's falsifier only tested the half we already said would not happen. On the capital half: no US-listed data center developer, neocloud or hyperscale financing vehicle experiences a distressed refinancing, a covenant breach, or a greater-than-40% peak-to-trough equity drawdown attributed by its own filings to data center exposure, by 31 December 2027. On the physical half: vacancy across CBRE's primary US markets rises above 8% for two consecutive reporting periods while under-construction volume stays flat.
Full entry, evidence and who disagrees →open MEDIUM
I-06 · due by 2029Electricity is being repriced from a variable cost into a strategic asset, and the largest buyers have already stopped behaving like ratepayers. Expect the majority of new US hyperscale capacity above 500 MW to arrive with contracted dedicated generation attached, not merely a utility service agreement.
Wrong if: Through 2028, fewer than a third of announced US campuses above 500 MW disclose dedicated or contracted generation, and standard utility service remains the norm at that scale.
Full entry, evidence and who disagrees →open MEDIUM
I-07 · due by 2028Flexibility is the cheapest gigawatt available in the United States, and it is about to stop being voluntary. Curtailability will become a standard, priced term in large-load tariffs across most RTOs before new baseload generation arrives to make it unnecessary.
Wrong if: By end-2028, fewer than half of US RTOs have a curtailment or flexible-load provision in an approved large-load tariff, and the DOE §202(c) precedent is not repeated.
Full entry, evidence and who disagrees →open MEDIUM
I-08 · due 2026–2032Nuclear is a 2032 story being marketed as a 2027 story. It is real, it is directionally correct, and it will not arrive in time to matter for this buildout. Gas turbines and fuel cells carry the load through 2029, and the industry's carbon accounting is about to get considerably more awkward.
Wrong if: Any purpose-built SMR reaches commercial operation serving a US data center before end-2029. Or: announced hyperscaler nuclear capacity exceeds 30 GW with executed, financed contracts rather than letters of intent.
Full entry, evidence and who disagrees →open MEDIUM
I-09 · due through mid-2027The political constraint will bind before the physical one. Public opposition moved faster in the last twelve months than any load forecast did, and siting — not generation — becomes the gating factor. Expect at least three more states to impose a moratorium or a mandatory large-load tariff by mid-2027.
Wrong if: Fewer than three additional US states adopt either a moratorium on new hyperscale data centers or a mandatory large-load rate class by 30 June 2027. Counted from the EEI large-load tariff tracker and state commission dockets.
Full entry, evidence and who disagrees →open MEDIUM
I-10 · due 2026–2028The 'data centers are raising your electricity bill' story is directionally right in capacity markets, unproven or reversed elsewhere, and almost universally reported without that distinction. Anyone publishing on this should state both sides or they are doing advocacy.
Wrong if: A well-identified study finds a positive, significant data-center effect on residential retail prices in vertically integrated states. Or: the EPRI result fails to survive peer review.
Full entry, evidence and who disagrees →open MEDIUM
I-11 · due through 2030Efficiency will not save the grid. Energy per token will keep falling fast, facility efficiency will keep not improving, rack power will keep climbing, and total consumption will rise through all of it. Anyone offering efficiency as the answer is describing the accelerant.
Wrong if: Global or US data center electricity consumption falls year over year while AI workloads grow. Or: industry average PUE drops below 1.4 within three years.
Full entry, evidence and who disagrees →open LOW
I-12 · due by 2027Cheapest-kilowatt-hour is already the wrong objective function for siting, and the industry's own metric will change to reflect it. Time-to-power replaces dollars-per-kilowatt-hour as the headline number in siting decisions and RFPs.
Wrong if: Of the ten largest US data center leasing or development transactions publicly reported in 2027, fewer than three have coverage or company statements that name energization timeline or time-to-power as a primary siting criterion, while $/kWh or colocation rate is named in the majority. Scored on CBRE, JLL and Data Center Frontier reporting.
Full entry, evidence and who disagrees →open SPECULATIVE
I-13 · due by 2028A standard 'large load compact' emerges — a template combining curtailability, full cost responsibility, a community benefit payment and mandatory water and emissions disclosure — and signing it becomes the price of entry in most states. The industry will end up asking for it, because a predictable national template beats fifty bespoke fights.
Wrong if: By 31 December 2028, no three US states have adopted large-load rules sharing at least three of the four components named here (curtailability, majority cost responsibility, community benefit payment, mandatory water or emissions disclosure), and no industry association has published a model template containing them.
Full entry, evidence and who disagrees →open SPECULATIVE
I-14 · due August 2027A dated, falsifiable house forecast, published so it can be scored against us. Define the PJM-state industrial price as the simple unweighted mean of the four PJM-state industrial prices in this index — Virginia 10.53, Illinois 10.20, Ohio 9.87, New Jersey 14.57 — which is 11.29 cents per kilowatt-hour in May 2026, against Texas at 6.33. The gap is 4.96 cents. Our forecast: in the May 2027 edition of EIA Table 5.6.A, that same four-state mean minus the Texas figure will be WIDER than 4.96 cents. We will publish the result next August whether we were right or wrong.
Wrong if: The four-state mean minus Texas, in EIA Table 5.6.A for May 2027, is 4.96 cents or less. We will report it in the same place with the same prominence.
Full entry, evidence and who disagrees →