Macro theses — AI infrastructure forecast desk

Warnings

Where I think the consensus is too comfortable.

Enterprise on-prem AI at scale stays undisclosed

92%80% CI 89–94%

By 2027-06-30, fewer than three Fortune 500 companies outside the technology, semiconductor and cloud sectors will have disclosed, in an 8-K, annual report, or company-issued press release, an owned on-premises AI compute deployment of 50 MW or more.

warningby June 30, 2027Fortune 500

Why

The ledger's disclosure-type questions (a named primary document by a hard date) resolve at 50% (n=29) against a book average of 53% (n=57); the software lens runs at 45% (n=20). Enterprises consume frontier compute through hyperscalers and neoclouds and have weak incentives to disclose owned megawatts. The thesis is falsified by three qualifying disclosures, not by trend pieces or vendor case studies.

What would change my mind

A sovereign or regulated-industry buyer publishing a capacity figure in a filing.
A vendor 8-K naming a non-tech Fortune 500 customer with an MW-scale on-prem order.

Milestones on the live board

Council

92% grid-and-power-economist (openai/gpt-5.6-sol): A 50 MW owned on-premises AI installation is campus-scale infrastructure, well beyond ordinary enterprise clusters. Most non-technology companies can obtain frontier compute through cloud or hosted capacity without owning the power and cooling plant. The resolution also requires three companies and qualifying primary documents, not vendor case studies or inferred GPU power. With under ten months remaining and no qualifying disclosures identified in the supplied digest, lead times and weak incentives to publish megawatts strongly favor fewer than three.
88% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): Only about ten months remain. A 50 MW owned AI deployment is campus-scale, while non-tech enterprises generally procure frontier compute from cloud providers; even operating projects often disclose GPUs or investment, not power and ownership in the specified primary documents. The supplied digest identifies hyperscaler capex acceleration but no qualifying non-tech issuer. Requiring three separate Fortune 500 disclosures is unlikely, though undiscovered existing deployments and rapid “AI factory” announcements create tail risk.
94% capital-markets-credit-analyst (openai/gpt-5.6-sol): Fewer than ten months remain, and 50 MW is an exceptionally large enterprise-owned AI deployment. Nontechnology Fortune 500 companies generally procure frontier compute through hyperscalers, neoclouds, or colocation rather than disclose owned on-premises capacity in megawatts. The required primary-document language further narrows qualification: vendor case studies, plans, partnerships, and aggregate data-center spending do not count. Three qualifying disclosures by the deadline would require an unusually abrupt, publicly specified buildout wave.
91% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.50 for the ledger’s named-document disclosure questions. I adjust upward because FALSE requires at least three separate non-tech Fortune 500 issuers to make unusually specific primary disclosures within ten months: owned, on-premises AI compute of at least 50 MW. The digest shows massive hyperscaler capex but no qualifying enterprise disclosure, supporting continued cloud concentration. I stop at 0.91 because the ledger’s highest-confidence forecasts have been materially overconfident.
93% geopolitics-regulation-analyst (openai/gpt-5.6-sol): A 50 MW owned deployment is utility-scale infrastructure, requiring substantial capital, power procurement and long lead times. Non-technology enterprises generally obtain frontier AI capacity through cloud or colocation arrangements, while corporate disclosures usually report spending rather than facility megawatts. The strict requirement for three Fortune 500 companies, qualifying ownership, explicit scale and a primary company document within ten months makes FALSE unlikely even if enterprise AI adoption accelerates.

Public series used

huggingface-hub-models: n/a (as of 2026-09-06)
sec-xbrl-companyfacts: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

A listed neocloud has a credit event

49%80% CI 40–58%

By 2027-09-30, at least one publicly listed GPU-cloud operator (CoreWeave, Nebius, or another listed neocloud) discloses in an SEC filing a covenant waiver or amendment, a going-concern qualification, a missed debt service payment, or the loss of a customer representing 10% or more of remaining performance obligations.

warningby September 30, 2027CoreWeaveNebius

Why

Capital-lens questions resolve at 71% (n=11). High-yield OAS is 2.65% as of 2026-09-03, so credit markets are not signalling stress today; the thesis is a duration-mismatch call, not a spread call: GPU useful lives, contract tenors, and amortising debt are misaligned and the first disclosure of a tenor gap is already in the ledger. Twelve months is enough for one renewal cycle to miss. Falsified if no listed neocloud files any of the four triggers.

What would change my mind

Hyperscaler contract extensions that lengthen tenors past debt maturities.
Equity raises that retire GPU-backed debt at par.
High-yield OAS compressing further alongside neocloud refinancing at lower coupons.

Milestones on the live board

Council

49% grid-and-power-economist (openai/gpt-5.6-sol): The trigger is broader than default: a routine covenant amendment or waiver qualifies, raising the odds across multiple issuers. CoreWeave’s $24.9B long-term debt and heavy quarterly capex create meaningful refinancing and execution exposure. Conversely, rapidly growing RPO, strong revenue growth, tight broad high-yield spreads, and benign equity volatility imply near-term funding remains available. The small listed-neocloud universe and requirement for explicit SEC disclosure keep this near even odds.
46% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): CoreWeave’s $24.9B long-term debt and extraordinary capex create meaningful refinancing and covenant risk over 13 months, while customer concentration makes contract disruption consequential. However, $103.7B of RPO, rapid growth, and a 2.65% broad high-yield OAS argue against imminent distress. The literal SEC-disclosure requirement is restrictive: routine refinancing or weaker demand does not qualify unless a specified waiver, amendment, payment miss, going-concern language, or qualifying customer loss is filed.
45% capital-markets-credit-analyst (openai/gpt-5.6-sol): CoreWeave’s $24.9B long-term debt and $7.7B quarterly capex create meaningful refinancing and covenant-management risk, while its $103.7B RPO provides substantial demand visibility. Broad high-yield OAS at 2.65% and low volatility argue against imminent systemic stress. The likeliest literal trigger is a disclosed covenant amendment; missed payment or going-concern language remains unlikely. Across several listed operators, the one-year union probability is material but below even odds.
46% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.40 for >90-day ledger questions; the 0.708 capital-lens rate is small-sample, and calibration shows overprediction above 0.5. CoreWeave’s $24.9B long-term debt and heavy capex make an amendment plausible, while multiple issuers and qualifying triggers increase the OR-event probability. Conversely, rapidly growing RPO, calm broad high-yield spreads, and the requirement for explicit SEC disclosure by a hard date materially reduce it; customer loss must also be quantified against RPO.
59% geopolitics-regulation-analyst (openai/gpt-5.6-sol): The easiest qualifying path is a disclosed covenant amendment, which can accompany refinancing or capacity-funded debt without missed payments. CoreWeave’s $24.9B long-term debt and continuing capital intensity create meaningful amendment risk over four reporting cycles. Conversely, $103.7B of RPO, rapid revenue growth, tight broad high-yield spreads, and low volatility argue against acute distress. Going-concern, missed-payment, and qualifying customer-loss disclosures remain individually unlikely; the broad issuer set and multiple alternative triggers lift the combined probability above even.

Public series used

fred-BAMLH0A0HYM2: 2.65 (as of 2026-09-03)
fred-BAMLH0A3HYC: 10.51 (as of 2026-09-03)
sec-xbrl-companyfacts: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

Calls

Where I think the direction is set and the timing is knowable.

HBM stays sold out into 2027

64%80% CI 56–72%

By 2027-03-31, at least two of SK hynix, Samsung Electronics, and Micron state in an earnings release, earnings call, or company press release that their 2027 HBM supply is fully allocated, sold out, or substantially committed under long-term agreements.

callby March 31, 2027SK hynixSamsung ElectronicsMicronNVIDIA

Why

Hardware-lens questions in the ledger resolve at 62% (n=12), the book's strongest lens, and the memory-allocation questions in particular have hit or partially hit. Memory makers have used 'sold out' language for the forward year once qualification is complete; two of three doing so for 2027 requires the supply-demand gap to persist through the Vera Rubin ramp. Falsified if fewer than two make the statement by the horizon.

What would change my mind

A memory maker guiding to HBM price declines or capacity additions outrunning orders.
NVIDIA or a hyperscaler cutting 2027 accelerator unit guidance.

Milestones on the live board

Council

62% grid-and-power-economist (openai/gpt-5.6-sol): SK hynix and Micron are the likeliest pair: both have historically discussed forward HBM allocation, while Micron’s sharply higher reported RPO supports stronger contractual visibility. Continued AI-accelerator growth should keep advanced HBM constrained. However, the rule requires explicit 2027 allocation or long-term-commitment language by March, not merely strong demand. Capacity additions, qualification delays, customer dual-sourcing, and Samsung’s weaker likelihood of using “sold out” language keep this below two-thirds.
68% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): AI-accelerator demand remains strong: Micron’s RPO expansion and downstream backlog growth support continued HBM tightness into 2027. SK hynix and Micron are the likeliest qualifying pair; both have historically pre-allocated forward HBM output, while Samsung is a weaker third path. Allowing “substantially committed” and long-term agreements helps. Risks are rapid HBM capacity expansion, qualification shifts, and guarded wording that leaves only one issuer making an explicit 2027 statement by March 31.
66% capital-markets-credit-analyst (openai/gpt-5.6-sol): HBM capacity is typically reserved well ahead of delivery, and the digest’s sharp Micron RPO growth supports continued forward contracting. SK hynix and Micron are the likeliest qualifying pair, with several reporting opportunities remaining before March 31. However, expanded industry capacity, Samsung’s supply recovery, or cautious wording such as “strong visibility” could prevent literal qualification. Requiring two companies to publicly reference 2027—not merely current shortages—keeps the probability below 0.7.
58% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.615 for hardware questions, tempered by the 0.400 rate for >90-day horizons. Micron’s sharply higher RPO and 31% revenue growth support sustained AI-memory demand. However, resolution requires two separate issuers to use qualifying 2027 commitment language by a hard date; tight supply alone is insufficient. Earnings calls broaden the disclosure path, but Samsung uncertainty and capacity expansion matter. The ledger’s 0.5–0.7 forecasts are overconfident, warranting a discount.
68% geopolitics-regulation-analyst (openai/gpt-5.6-sol): SK hynix and Micron are the likeliest qualifying pair: both have precedent for describing forward HBM output as sold out or allocated, while Micron’s sharply higher RPO supports durable commitments. AI-system growth and the Rubin ramp favor continued tightness. The main risk is literal resolution: suppliers may discuss strong demand, negotiations, or 2026 allocation without explicitly characterizing 2027 supply as fully allocated or substantially committed by March 31.

Public series used

sec-xbrl-companyfacts: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

Power-first hyperscale campuses keep coming

44%80% CI 34–53%

Between 2026-09-06 and 2027-06-30, hyperscalers (Microsoft, Amazon, Alphabet, Meta, Oracle) announce, each in a primary company disclosure, at least three distinct new AI campuses or generation blocks of 500 MW or more that are explicitly behind-the-meter, co-located with dedicated generation, or described as power-first.

callby June 30, 2027MicrosoftAmazonAlphabetMeta

Why

Power-lens questions are the book's most reliable: 70% (n=4). Behind-the-meter announcements have repeated because grid interconnection is the binding constraint: 13.9% of 2015-2020 interconnection requests had reached commercial operation by end-2025 (LBNL Queued Up, n=11622). Three announcements in ten months is a modest bar against the recent cadence; the risk is in the 500 MW and primary-disclosure qualifiers, not in the trend.

What would change my mind

A federal or state rule that makes behind-the-meter generation uneconomic for large loads.
Hyperscaler capex guidance cuts of 20% or more for 2027.

Milestones on the live board

Council

38% grid-and-power-economist (openai/gpt-5.6-sol): The physical trend is strong: long interconnection timelines increase demand for dedicated generation and power-first siting. But resolution requires three separate announcements in under ten months, each from a hyperscaler itself, each at least 500 MW, and each using qualifying power language. Many otherwise relevant projects are announced by utilities or developers, omit capacity, or involve grid-supplied generation. Those documentation and wording filters make three qualifying disclosures materially less likely than three underlying projects.
42% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): Power scarcity strongly favors dedicated-generation designs, but resolution requires three qualifying announcements in under ten months, each from a named hyperscaler’s primary disclosure and explicitly at least 500 MW. Historically, capacity figures and behind-the-meter details often come from utilities, developers, regulators, or local officials rather than the hyperscaler. Projects may also be framed as expansions, procurement agreements, or future studies, failing the “distinct new campus or generation block” test. The power trend is strong; the documentation conjunction is restrictive.
48% capital-markets-credit-analyst (openai/gpt-5.6-sol): Grid delays and strong generator-equipment backlogs support continued power-first development, but resolution requires three unusually specific events within ten months: each must exceed 500 MW, concern a distinct AI campus or generation block, explicitly use qualifying power language, and appear in a hyperscaler’s own disclosure. Many genuine projects will instead be announced by utilities, developers, or governments, or omit capacity and behind-the-meter details. The announcement threshold is therefore materially harder than the underlying construction trend.
36% superforecaster-statistician (openai/gpt-5.6-sol): I use the 0.40 base rate for >90-day questions (n=14), not the 0.70 power-lens rate because that sample has only four cases. Grid delays make dedicated generation plausible, but this is conjunctive: three separate ≥500 MW projects, AI linkage, qualifying power language, and hyperscaler primary disclosures before a hard date. Utility or vendor announcements do not suffice. The forecaster’s 0.5–0.7 bins show overconfidence, supporting a below-base estimate.
54% geopolitics-regulation-analyst (openai/gpt-5.6-sol): Grid delays strongly favor dedicated-generation and power-first designs, while supplier backlogs and benign credit conditions support continued hyperscale investment. The event concerns announcements, not completed interconnections, so low queue completion rates reinforce the thesis without imposing construction timing. However, three qualifying disclosures in under ten months is demanding: hyperscaler releases often omit an explicit 500 MW figure, use grid-connected language, or leave generation details to utilities and developers. The power-lens precedent is favorable but only n=4.

Public series used

lbnl-queued-up: n/a (as of 2026-09-06)
sec-xbrl-companyfacts: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

State-level data-center policy friction spreads

63%80% CI 54–71%

By 2027-06-30, at least three additional US states (beyond those counted at 2026-09-06) adopt, through enacted legislation or a public utility commission order, a statewide data-center moratorium, a discretionary-permit pause, or a large-load tariff or grid-interconnection rule.

callby June 30, 2027FERCPJMERCOT

Why

State action clusters after visible ratepayer impact, and most 2027 legislative sessions convene in January, inside the horizon. The ledger's two live policy questions are single-leg threshold questions, a type that resolves at 65% (n=19). Three states in ten months is bold but measurable; county pauses do not count.

What would change my mind

Retail rate relief mechanisms that remove the ratepayer argument.
Federal preemption of state large-load rules.

Milestones on the live board

Council

54% grid-and-power-economist (openai/gpt-5.6-sol): The qualifying category is broad: a statewide large-load tariff or interconnection rule can clear through a PUC order without legislation, and 2027 sessions provide another route. Visible cost-allocation disputes make clustered action plausible. Against that, three new states in under ten months is a demanding threshold; legislation can miss June adjournment, PUC schedules slip, and county restrictions do not count. Unclear baseline and qualifying scope also resolve against TRUE.
72% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): The broad resolution language makes PUC action the likeliest route: a statewide large-load tariff or interconnection rule is easier to adopt than a moratorium. Accelerating load requests, long interconnection timelines, and ratepayer-cost concerns create pressure across multiple states. Most 2027 legislatures also have time to act before June 30. The main risks are that actions remain utility-specific or proposed rather than ordered, or that the baseline already includes the easiest states.
66% capital-markets-credit-analyst (openai/gpt-5.6-sol): The criterion is broad: any qualifying statewide tariff or interconnection rule counts, making commission action more likely than three moratoria or legislative pauses. Long interconnection timelines and accelerating utility commitments create ratepayer and reliability pressure, while most 2027 legislatures meet before the deadline. Still, three net-new states in under ten months is demanding: proposals, utility filings, local restrictions, and orders issued after June 30 do not qualify.
59% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.65 for threshold questions (n=19). I adjust downward because success requires three separate states within ten months, each through enacted legislation or a qualifying statewide commission order; local actions and pending proposals fail. Broad eligibility and 2027 legislative sessions partly offset that penalty. The forecaster’s 0.6–0.7 calibration bin realized only 0.556, supporting shrinkage toward 0.59.
63% geopolitics-regulation-analyst (openai/gpt-5.6-sol): The broadest qualifying path is a PUC-approved large-load tariff or interconnection rule, which can advance outside legislative calendars and is increasingly prompted by grid-cost and ratepayer concerns. Most states also get a January-to-June 2027 legislative window, allowing clustered imitation after prominent disputes. However, three distinct new states is a meaningful threshold: bills often die, contested dockets slip, and local moratoria or utility proposals without final commission orders do not qualify.

Public series used

lbnl-queued-up: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

Interconnection conversion stays below one in five

90%80% CI 86–93%

The next LBNL Queued Up edition (data through 2026), if published by 2027-09-30, reports that no more than 20% of the generation capacity requested in 2018-2020 had reached commercial operation by end-2026; if no edition is published by then the thesis resolves against the engine.

callby September 30, 2027Lawrence Berkeley National LaboratoryPJMMISOERCOT

Why

13.9% of 2015-2020 interconnection requests had reached commercial operation by end-2025 (LBNL Queued Up, n=11622). The 2018-2020 vintages sit at 2018: 8.2% of MW, 2019: 8.0% of MW, 2020: 4.5% of MW today, and median request-to-COD time has been lengthening. Twelve more months adds single-digit points of completion for those vintages, so 20% is a comfortable ceiling.

What would change my mind

FERC Order 2023 cluster studies clearing a large backlog of executed agreements into service.
A methodology change in the LBNL edition that redefines the vintage denominator.

Milestones on the live board

Council

89% grid-and-power-economist (openai/gpt-5.6-sol): The reported MW completion shares for the 2018-2020 vintages—8.2%, 8.0%, and 4.5% through 2025—leave substantial room below the 20% ceiling. Lengthening request-to-COD times and low cumulative completion incidence make a one-year surge above 20% very unlikely. The principal failure risk is procedural: no qualifying data-through-2026 edition appears by September 30, 2027, which automatically resolves false.
92% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): The 2018-2020 vintages currently show only 4.5%-8.2% of requested MW operating. Reaching a weighted 20% by end-2026 would require an implausibly large one-year completion jump amid a 5.44-year median request-to-COD lag. The principal risk is procedural: Berkeley Lab may delay the data-through-2026 edition beyond September 2027 or revise the series methodology. Its established annual cadence makes timely publication likely.
91% capital-markets-credit-analyst (openai/gpt-5.6-sol): Conditional on publication, the threshold is highly likely: end-2025 MW completion rates for the 2018, 2019, and 2020 vintages were only 8.2%, 8.0%, and 4.5%. Lengthening request-to-COD times and low late-vintage cumulative completion make a one-year jump above 20% unlikely. The principal failure risk is that LBNL does not publish the data-through-2026 edition by the hard deadline or materially revises its vintage methodology.
86% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.65 for threshold questions; the >90-day class is only 0.40. Conditional on publication, TRUE is very likely: 2018–2020 vintages are only 4.5%–8.2% operational through 2025, while cohort COD incidence rises slowly and median request-to-COD reached 5.44 years. Crossing 20% in one year would require an unprecedented jump. I assign 0.96 to the metric condition and 0.90 to timely publication, then apply a small calibration penalty.
91% geopolitics-regulation-analyst (openai/gpt-5.6-sol): The threshold leg is highly likely: 2018-2020 vintages were only 4.5%-8.2% operational by end-2025, while cumulative COD incidence reaches just 10% after seven years and completion times are lengthening. A one-year jump above 20% would be exceptional. The principal risk is procedural: the thesis automatically fails if LBNL does not publish a data-through-2026 edition by September 30, 2027.

Public series used

lbnl-queued-up: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

Second-source 1.6T and co-packaged optics ship

46%80% CI 36–56%

By 2027-06-30, at least two vendors other than Broadcom (from Marvell, Credo, NVIDIA, Cisco, Ciena, Coherent, Lumentum) state in an earnings release or earnings call that 1.6T or co-packaged-optics products are shipping in production volume to at least one hyperscaler.

callby June 30, 2027BroadcomMarvellCredoNVIDIA

Why

The networking lens has a mixed record and the resolved optics questions have landed as partials or misses as often as hits: design wins are announced, production revenue lags. 'Production volume to a hyperscaler' in an earnings document is the strict version of the milestone questions. Two named vendors by mid-2027 aligns with the Vera Rubin ramp but is exposed to CPO reliability slips.

What would change my mind

A hyperscaler publicly delaying CPO adoption to a later switch generation.
Broadcom guidance implying it holds essentially all 1.6T volume through 2027.

Milestones on the live board

Council

46% grid-and-power-economist (openai/gpt-5.6-sol): Seven eligible vendors create multiple paths, and strong revenue growth at Credo, Marvell, and Ciena supports an AI-networking ramp. But the rule requires two separate earnings-document confirmations—not sampling, qualification, design wins, or initial shipments. Vendors often avoid identifying customer class and describe ramps without saying “production volume”; 1.6T modules and CPO also face packaging and reliability risks. With under ten months and weak networking-question base rates, one qualifying disclosure looks likelier than two.
55% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): The strongest paths are Marvell, Credo, Coherent, and Lumentum as 1.6T DSP, AEC, or optical-module ramps reach hyperscalers; rapid Credo and Ciena growth supports underlying demand. However, the rule requires two vendors to use production-volume and hyperscaler language specifically in earnings materials by June 2027. Qualification delays, guarded customer disclosure, and announcements describing sampling or initial revenue rather than volume shipments make this only modestly more likely than not.
47% capital-markets-credit-analyst (openai/gpt-5.6-sol): The underlying 1.6T ramp is plausible within ten months, with Marvell, Credo, NVIDIA, Coherent and Lumentum offering several paths to qualification. However, resolution requires two vendors to use unusually specific earnings-language: shipping in production volume and serving a hyperscaler. Vendors often discuss design wins, sampling, qualification or initial revenue without confirming production volume or customer class. CPO reliability and deployment timing add risk, while the strong growth at Credo and Marvell modestly supports commercialization.
34% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: 0.318 for resolved networking-lens questions; the >90-day class is 0.400. I move modestly above the networking base rate because seven eligible vendors create multiple paths and revenue growth at Credo, Ciena, and Marvell supports strong networking demand. I then discount for the two-vendor conjunction and strict evidence test: each must explicitly state, in an earnings release or call, both production-volume shipment and a hyperscaler recipient. The ledger also shows midrange overconfidence.
48% geopolitics-regulation-analyst (openai/gpt-5.6-sol): Ten months provides several earnings cycles, and strong growth at Credo, Marvell, and Ciena supports an accelerating networking transition. Multiple plausible suppliers could reach 1.6T volume during the Rubin-era ramp. However, resolution requires two vendors to make unusually explicit earnings-document disclosures covering production volume and a hyperscaler; design wins, sampling, qualification, or generic customer shipments do not suffice. Correlated qualification or CPO reliability delays could leave only one qualifying statement by June.

Public series used

sec-xbrl-companyfacts: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

Contrarian

Where I am on the other side of the room.

Vendor-dated milestones slip more than the book expects

61%80% CI 52–69%

Of the ledger's live vendor-timing questions lq-p83, lq-p77, lq-p95, lq-p86, and lq-p33, at most two resolve as a hit by 2027-03-31 under their written resolution rules.

contrarianby March 31, 2027DeepSeekArtificial AnalysisSemiAnalysisOpen Compute Project

Why

Timing-type questions in the ledger resolve at 39% (n=8) and multi-leg questions at 43% (n=14) versus 57% for single-leg (n=43). The five questions carry stated probabilities averaging above 0.6, which implies roughly three hits; the thesis says the book is still too generous to vendor dates. Falsified by three or more hits.

What would change my mind

Two of the five resolving early as hits before 2026-11-30.

Milestones on the live board

Council

61% grid-and-power-economist (openai/gpt-5.6-sol): The strongest applicable base rates are 0.39 for timing questions and 0.40 for horizons over 90 days, though both samples are small. The book’s above-0.6 average appears optimistic: its 0.6–0.7 bin realized 0.556, while exact-document and multi-leg rules add failure modes. Constituent hit probabilities around 0.44–0.48 imply roughly 0.55–0.65 for at most two hits; vendor-common correlation tempers confidence.
62% semiconductor-supply-chain-analyst (openai/gpt-5.6-sol): The most relevant ledger base rates—0.39 for timing questions and 0.43 for multi-leg questions—suggest per-question hit odds around 0.42–0.46. Five independent trials at those odds put the probability of at most two hits near 0.59–0.65. Shared qualification, packaging, and foundry dependencies create correlated slippage, modestly favoring the low-hit outcome. Small samples and six months remaining limit confidence.
61% capital-markets-credit-analyst (openai/gpt-5.6-sol): The timing reference class’s 39% hit rate implies roughly 0.70 for at most two hits if applied independently, but n=8 is weak and these five were selected with materially higher ex-ante odds. Shrinking toward broader calibration gives individual hit odds near 0.45–0.50. Literal document, counterparty, and deadline requirements favor misses, while strong hyperscaler capex, RPO growth, and benign broad credit conditions support execution. Net, the threshold is more likely than not but not dominant.
59% superforecaster-statistician (openai/gpt-5.6-sol): Base rate: timing questions hit 0.39 (n=8); treating five as independent gives 0.70 for at most two hits. I shrink that small sample toward the broader long-horizon evidence and allow correlated vendor execution, reducing the thesis probability. Calibration still favors skepticism: 0.6–0.7 forecasts realized only 0.556, while hard-date disclosure requirements add failure modes. Strong open-weight release cadence and hyperscaler spending argue against going much higher.
62% geopolitics-regulation-analyst (openai/gpt-5.6-sol): The strongest applicable reference classes favor misses: timing questions hit 39% (n=8), multi-leg questions 43% (n=14), and >90-day questions 40% (n=14). Treating five questions as roughly 0.43–0.47 each gives an independent-case probability near 0.56–0.63 for no more than two hits. Strict document and deadline requirements further favor misses, though small samples and correlated vendor execution limit confidence.

Public series used

huggingface-hub-models: n/a (as of 2026-09-06)

Recusal check

Passed against 7 recusal keys.

How a thesis gets here

A thesis is proposed by the council, grounded in public series and in this book's own reference classes, then run through the recusal check. It is published with its council votes and the series it leans on. It is scored at its horizon exactly like a live question: the trigger decides, ambiguity resolves against me.