Skip to content

AI Market · Forecast desk · as of September 6, 2026

The public book.

Every AI-infrastructure call I have made in public since May 2026, scored against the trigger I wrote when I made it. 23 landed, 15 landed in part, 19 did not. The numbers are hashed into a chain so I cannot quietly edit them later.

What this is. A scored ledger, not a newsletter. Each row is a probability, a deadline, and a written resolution rule; ambiguity resolves against me. The book covers power, silicon, networking, capital, and software. It does not price anything on the recusal list. I am testing whether a transparent, public-data forecasting process can add signal on a bounded set of AI-infrastructure questions. Every forecast is timestamped and scored. The track record is still early, and I am not claiming a demonstrated forecasting edge.

Scoreboard

57 resolved of 104 issued. Hit rate 40% counting only clean hits, 54% with partial credit.

Resolved calls

57of 104

23 hit15 partial19 miss

Hit rate 54% with partial as half. 47 live, 0 overdue.

Brier score

0.20090% CI 0.160.24

Partial as miss: 0.315. Coin flip is 0.250, zero is perfect. Both policies published; the harsher one is the one to beat.

Skill vs coin flip

-8%partial as half

vs base rate -5% · partial as miss vs coin flip -26%. Positive means the stated numbers carry information beyond the baseline.

Overconfidence gap

+13 pts

Mean stated confidence minus hit rate. 90% CI +3 pts to +22 pts. This is the number I am working down.

Calibration error

0.14

Expected calibration error across 4 populated confidence bins. Lower is better; 0 is a forecaster whose 70% calls land 70% of the time.

Ledger integrity

verified291 events · sha256

head b374483ca574cfcf9ebed767f3854797da57430110221442cf1c643ce3008359

Every issued probability, repricing, and resolution is hashed into a chain. Edit one number after the fact and the head changes. Download the chain and recompute it yourself.

Calibration

Where the stated confidences land against reality, and which slices of the book carry the error.

Reliability diagram

0%0%25%25%50%50%75%75%100%100%n=6n=29n=17n=4stated confidencerealized hit rate

Each bubble is a confidence bin; size is the number of resolved calls in it. A perfectly calibrated forecaster sits on the dashed line. Bubbles below the line are overconfident. Almost everything I have called so far sits in the 55–80% band, so the diagram is honest about how little of the probability range has been tested.

By lens and horizon

Slicenhit ratebriergap
software2045%0.240+19 pts
hardware1263%0.124+4 pts
networking1030%0.256+36 pts
capital1173%0.143-4 pts
power475%0.243-4 pts
horizon <=30d1450%0.169+17 pts
horizon 31-90d2962%0.184+4 pts
horizon >90d1439%0.262+27 pts

Gap is mean stated confidence minus hit rate for the slice. Small n per slice; read direction, not decimals.

Called it

23 clean hits. These are the ones that were least obvious when called, sorted by how little confidence I gave them. Every card links to the public evidence.

hitcalled at 43%software2026-W34

Z.ai publishes GLM-5.3 weights to Hugging Face by September 15, 2026, closing the two-week window promised at the model's August 14 announcement.

Z.ai published the full 753B-parameter GLM-5.3 weights to Hugging Face at zai-org/GLM-5.3 on August 27–28, 2026 — in-window and inside the trigger's September 15 window, distinct from GLM-5.2 — after GLM-5.3-Flash MIT weights landed Aug 26. The material nuance is licensing, not availability: GLM-5.3 ships under a bespoke GLM-5.3 license rather than MIT, requiring Z.AI security review before commercial use by any Model-as-a-Service operator whose group revenue exceeds $10B over any 12 consecutive months.

deadline By September 15, 2026 · brier 0.325

hitcalled at 60%power2026-W22

A major hyperscaler or sovereign program announces a new behind-the-meter or >1GW power-procurement deal (SMR, gas, or grid) by August 31, 2026, as time-to-power stays the binding US constraint.

Amazon confirmed in August that it acquired the GW Ranch site in Pecos County, Texas, for an AI campus powered by Pacifico Energy's planned 7.65GW behind-the-meter gas-and-storage system — a named hyperscaler >1GW BTM generation deal inside the Aug 31 window.

resolved September 1, 2026 · brier 0.160evidence ↗

Where I was wrong

Ranked by Brier contribution: the high-confidence calls that did not land cost the most. Each one carries the error class it taught me.

misscalled at 71%networking2026-W29other

The largest colocation operators' Q2 prints (starting July 29) show interconnect/fabric revenue growth again outpacing overall revenue growth, sustaining the Metcalfe read for a second consecutive quarter.

Equinix Q2 FY26 — the largest operator's Jul 29 print — grew interconnection revenue 11.3% YoY versus total revenue up 16.4%; interconnect did not outpace overall growth, and no other major operator's Q2 filing showed fabric/interconnect beating total revenue either.

resolved September 1, 2026 · brier 0.504evidence ↗

misscalled at 70%capital2026-W17timing-slip

At least one frontier lab announces a customer-funded compute commitment greater than 2 GW.

No frontier lab announced a single customer-funded commitment >2 GW by June 30. The period's biggest customer-funded deal — Anthropic paying SpaceX/xAI up to $45B for Colossus access (disclosed May 20 in the S-1) — covered ~300 MW; Meta-Crusoe (June 18) topped out at 1.6 GW. Stargate additions remained structural build-out, not a single >2 GW customer-funded announcement.

resolved July 18, 2026 · brier 0.490evidence ↗

Error classes · 34 misses and partials

Conjunctive question (two legs had to land)9
Named primary disclosure by a date2
Timing slip13
Threshold not reached5
Other5

What changed because of this: 50% is banned as a stated probability, every live question now carries a written resolution rule and an ambiguity policy, conjunctive questions are split before they are issued, and the book is capped at 40 live questions so each one gets repriced.

Live board

40 frozen questions with written rules, sorted by deadline. The large number is the current probability; the small line is what I called when the question was issued. Repricings come from the offline engine and are chained.

software
lq-p2
At least one Fortune 500 enterprise discloses an on-prem AI workload greater than 50MW equivalent by September 30, 2026.1 caution2 points24%was 60%24d2026-09-30
capital
lq-p5
At least one neocloud loses an anchor tenant or sees backlog growth turn negative by September 30, 2026.capex durability2 points15%was 35%24d2026-09-30
hardware
lq-p9
Samsung HBM4 reaches greater than 25% share of NVIDIA Vera Rubin BOM by Q3 2026.1 caution2 points13%was 60%24d2026-09-30
software
lq-p11
By Q3 2026, at least two frontier labs publicly disclose a separately-priced gated cyber offering.council split1 caution2 points21%was 60%24d2026-09-30
hardware
lq-p15
A second public open-weights frontier-class model trained end-to-end on AMD silicon by September 30, 2026.2 cautions2 points36%was 55%24d2026-09-30
software
lq-p19
At least one Fortune 500 enterprise discloses a production specialist-multimodal deployment >$25M annualized by September 30, 2026.1 caution2 points25%was 60%24d2026-09-30
hardware
lq-p33
At least one major OEM announces customer shipment or formal order availability of Vera Rubin systems by September 30, 2026.1 caution2 points52%was 70%24d2026-09-30
power
lq-p36
A hyperscaler announces another >500MW power-first AI campus or behind-the-meter generation block by September 30, 2026.energization2 cautions2 points35%was 60%24d2026-09-30
networking
lq-p45
A second non-Broadcom vendor cites a 1.6T or co-packaged-optics hyperscaler design win by September 30, 2026.1 caution2 points34%was 55%24d2026-09-30
capital
lq-p50
Groq announces at least one named Fortune 500 or hyperscaler inference-cloud customer by September 30, 2026.1 caution2 points37%was 57%24d2026-09-30
networking
lq-p60
A second named vendor or operator announces a commercial cross-data-center scale-across product by September 30, 2026.1 caution2 points35%was 61%24d2026-09-30
software
lq-p71
An independent evaluator publishes completed-task cost showing GPT-5.6 Luna cheaper than the stated bar by September 30, 2026.1 caution2 points44%was 62%24d2026-09-30
software
lq-p83
DeepSeek publishes the V4-Pro-0813 build weights to Hugging Face by September 30, 2026.1 caution2 points46%was 64%24d2026-09-30
software
lq-p96
Highest single ISO week of OpenRouter aggregate token volume in September 2026 stays below the stated bar.council split1 caution2 points25%was 40%24d2026-09-30
power
lq-p65
At least one additional US state announces a moratorium or discretionary-permit pause on data centers by October 31, 2026.power policy1 caution2 points44%was 57%55d2026-10-31
software
lq-p72
An independent party reproduces at least a 15-point ARC-AGI-3 improvement from harness memory by October 31, 2026.1 caution2 points39%was 44%55d2026-10-31
hardware
lq-p73
SK hynix or Samsung states in a primary release that 2027 HBM capacity is substantially committed by October 31, 2026.2 cautions2 points66%was 81%55d2026-10-31
software
lq-p77
Artificial Analysis publishes Endpoint Accuracy Index results covering at least the stated expansion by October 31, 2026.1 caution2 points54%was 83%55d2026-10-31
software
lq-p91
Anthropic or a named enterprise customer publishes an independently attributed Skills case study by October 31, 2026.1 caution2 points44%was 58%55d2026-10-31
hardware
lq-p95
SemiAnalysis publishes AgentX v3 multi-turn results for OpenAI Jalapeño by October 31, 2026.1 caution2 points44%was 36%55d2026-10-31
networking
lq-p99
Meta contributes MetaRoCE specification through OCP at the October 2026 Global Summit.1 caution2 points41%was 44%55d2026-10-31
networking
lq-p74
No publicly listed global colocation operator reports Q3 2026 interconnection revenue declining sequentially.interconnect conversion2 cautions2 points46%as called70d2026-11-15
hardware
lq-p89
Marvell discloses the first attributable Custom Products revenue tranche under the Google relationship by November 30, 2026.1 caution2 points23%was 34%85d2026-11-30
hardware
lq-p98
NVIDIA Q3 FY2027 earnings disclosure states Vera Rubin contributed more than 25% of datacenter revenue.1 caution2 points44%was 74%85d2026-11-30
capital
lq-p75
A definitive agreement of at least $100B in vendor-guaranteed AI data-center financing is documented by December 31, 2026.capex durability1 caution2 points27%was 34%116d2026-12-31
power
lq-p76
A power service agreement for the Paducah AI campus is filed with the Kentucky PSC by December 31, 2026.energization1 caution2 points52%was 63%116d2026-12-31
capital
lq-p84
A second publicly traded neocloud discloses a GPU contract tenor gap in an SEC filing by December 31, 2026.residual duration1 caution2 points54%was 69%116d2026-12-31
networking
lq-p86
OCP Open Silicon Photonics for AI Systems submits its first specification by December 31, 2026.1 caution2 points45%was 61%116d2026-12-31
power
lq-p92
A second US state with major hyperscale exposure publishes a data-center grid rule by December 31, 2026.power policy2 cautions2 points49%was 39%116d2026-12-31
capital
lq-p102
A second AI infrastructure contract above $500M discloses both a future service date and a queue position by December 31, 2026.interconnect conversion2 cautions2 points46%was 68%116d2026-12-31
networking
lq-p103
Equinix publishes Fabric One beta documentation exposing approval, rollback, and schema controls by December 31, 2026.2 cautions2 points38%was 57%116d2026-12-31
power
lq-p69
A second US multi-gigawatt AI campus publicly commits to at least 250 MW of utility curtailment flexibility by January 31, 2027.energization1 caution2 points46%was 49%147d2027-01-31
networking
lq-p79
A major model-serving platform publishes per-endpoint accuracy, price, and latency by January 31, 2027.1 caution2 points43%was 31%147d2027-01-31
capital
lq-p81
At least two frontier labs publish network isolation or containment requirements for eval vendors by January 31, 2027.2 cautions2 points49%was 44%147d2027-01-31
software
lq-p87
A frontier lab publishes turn count or task-completion cost as a headline model-card metric by January 31, 2027.1 caution2 points40%was 26%147d2027-01-31
hardware
lq-p85
NVIDIA publicly confirms a Rubin Ultra memory configuration at or below 512GB, or explicitly walks it back, by March 31, 2027.1 caution2 points40%was 37%206d2027-03-31
networking
lq-p94
A second major networking or accelerator vendor ships co-packaged optics commercially by March 31, 2027.1 caution2 points40%was 47%206d2027-03-31
hardware
lq-p68
At least one named customer reports receiving a production AMD Helios rack by June 30, 2027.2 cautions2 points55%was 68%297d2027-06-30
hardware
lq-p80
AMD publicly names a Taalas-derived product or roadmap item tied to a specific manufacturing node by June 30, 2027.2 cautions2 points44%was 46%297d2027-06-30
power
lq-p101
Google accepts at least 500 MW of Fervo's conditional expansion and the parties disclose it by June 30, 2027.energization2 cautions2 points39%was 41%297d2027-06-30

Macro theses

Macro calls with a date and a measurable trigger. Each one rolls up milestone questions from the live board and passes the recusal check before it is published.

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

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.

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

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.

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

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.

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

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.

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

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.

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

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.

Engine

The offline engine reprices the live board from public series, reference classes, calibration models, and a five-persona council. It stays in shadow until it beats my stated numbers out of sample.

Shadow scoreboard · backtest on 57 resolved calls

No trainable fold yet. The strict scheme only lets a challenger learn from resolutions that were knowable before each fold's first issue date. The book is young enough that 0 of 5 folds qualify, so every challenger falls back to my stated number and the gate counts that as a loss. The diagnostic cuts below are the real comparison; they are not eligible for promotion.

Forecasterbrierpartial = misslog lossn
Stated (me)0.2000.3150.72857
Recalibrated +0.0000.2000.3150.72857
Feature model +0.0000.2000.3150.72857
Gradient boosting +0.0000.2000.3150.72857
Base rate +0.0000.2000.3150.72857
Coin flip 0.0150.1840.2500.69357
Platt scaling +0.0000.2000.3150.72857

Walk-forward by resolution date · diagnostic

Forecasterbrierpartial = misslog loss
Stated (me)0.2000.3150.728
Recalibrated +0.0150.2150.3110.757
Platt scaling +0.0250.2250.3210.777
Feature model +0.0440.2430.3350.823
Gradient boosting +0.0700.2700.3550.885
Base rate +0.0250.2240.3160.776
Coin flip 0.0150.1840.2500.693

Leave-one-out Brier (not time-ordered, optimistic): Recalibrated 0.195 · Gradient boosting 0.187 · Logistic (features) 0.182 · Stated (me) 0.200. Read this as an upper bound on what recalibration could add, not as evidence it has.

Backtest on resolved predictions, time-ordered folds by issue date; training uses only resolutions knowable before each fold's first issue date. Untrainable folds fall back to the stated probability and count as losses for the gate. 'walkForward' re-cuts the same rows by the date each resolution became known (diagnostic only).

Promotion gate

shadow mode0 of 5 folds won

calibrated or model must beat stated Brier on >= 4 of 5 folds and n >= 80

0.30f0f1f2f3f4

Until the gate passes, the engine's numbers are published next to mine, not instead of mine. Grey bars are my stated Brier per fold; accent bars are the challenger.

Data sources · run 2026-09-06T22-42-59

Live: fred-csv, sec-xbrl-companyfacts, huggingface-hub-models, lbnl-queued-up. Unavailable this run: magellan-mcp.

Full ledger

Every prediction ever issued, 104 rows, oldest deadline first. Click through for the rule, the evidence, and the chain events.

Method and firewall

Scoring follows the published weekly methodology: Brier with partial as half and as miss, bootstrap intervals, skill against coin flip and base rate, ten-bin reliability, and a SHA-256 chain over issued, repriced, and resolved events. Inputs are public URLs only. The private evidence engine I use at work never emits a public probability and never appears as an authority here.