Verdict Cards

All 82 published cards are listed here: settled calls by settlement date, newest first, then the rest by their scheduled check date. Results stay public whether we were right or wrong.

This industry predicts every day and rarely goes back to check. Verdict cards do the opposite: every forward call we publish is a public record — the call, what it rests on, what would overturn it, and when it is scheduled to be checked. At that date it is checked against the rule written in advance; a card that has no result yet says what it is waiting for. Each record can be checked card by card before it is cited. Methodology

Badges: Not yet due; Awaiting result (past its scheduled check, no result yet); Hit; Miss; Partly right; Can't be determined. Verdicts on outside claims are marked Supported, Contested or Refuted.

Settled

• Claim verdict: Contested Outcome: NoWill any frontier lab publish its flagship's absolute layer count?Settled 2026-11-01
• Claim verdict: Contested Outcome: YesWill Google sign a verifiable long-term TPU deal with an outside customer?Settled 2026-10-31
• MissWhether labs pull compute back from outside customers shows in prices and quotasSettled 2026-10-31
• Can't be determinedTSMC price hikes stay in a 5% to 15% band; memory carries this cycle's pricingSettled 2026-10-31
• Claim verdict: Contested Outcome: UnresolvedAnthropic's first profit: structural margin or compute-shortage rent?Settled 2026-10-31
• Can't be determinedThe US-China frontier AI gap will still measure between 4 and 14 months next timeSettled 2026-10-31
• Can't be determinedIn the next enterprise adoption report, Anthropic's share will drop 5 points or moreSettled 2026-10-30
• Can't be determinedWe bet OpenAI’s next system card again vouches for the wrong reasoning quantitySettled 2026-10-27
• Claim verdict: Contested Outcome: NoWill OpenAI explain the German website incident and when it knew?Settled 2026-10-18
• MissKimi K3 will cut its price to DeepSeek's level by mid-October (we give it 10% odds)Settled 2026-10-15
• Can't be determinedAugust's AI incident wave was mostly backlog; counts track disclosure, not dangerSettled 2026-10-04
• Claim verdict: Contested Outcome: HitWill analysts press CoreWeave on NVIDIA's backstop? Gurley's first testSettled 2026-10-03
• Claim verdict: Contested Outcome: MissAnthropic will end Sonnet 5's promo on schedule and charge full priceSettled 2026-09-30
• Claim verdict: Contested Outcome: Can't be determinedGurley says OpenAI credits count as Azure revenue; Nadella says noSettled 2026-09-30
• MissClaude Sonnet 5's promo price will revert to the standard list price after August 31Settled 2026-09-30

Not yet at the scheduled check

• Not yet dueThe Kimi K3 license guards Moonshot's own API price floor, not rent from cloudsDue 2026-11-05
• ContestedTPU up to 50% better per dollar: will an independent rerun confirm it?Due 2026-11-07
• Not yet dueModels tied to their coding tools: the format layer is fixable with strict modeDue 2026-11-12
• Not yet dueThe length of task an AI can finish on its own will pass 8 hoursDue 2026-11-15
• Not yet dueGPU price rises live in new short contracts; CoreWeave’s fleet average still lags in Q3Due 2026-11-16
• ContestedTo check Anthropic’s risk claim, publish incident counts and denominatorsDue 2026-11-22
• ContestedIs Anthropic’s risk line obviously false? We await a named real-world breachDue 2026-11-22
• Not yet dueOpenAI's Astra release names one monitor; the system card vouches with anotherDue 2026-12-04
• ContestedNo signatory officially names both an outside auditor and a board committeeDue 2026-12-06
• Not yet dueRetention of AI subscriptions over $250 a month will drop below the warning lineDue 2026-12-15
• ContestedDo smarter models get harder to watch? Three labs' next reports will tellDue 2026-12-20
• Not yet due“Backstop deals sink stocks” is wrong: the guarantor falls, the guaranteed side risesDue 2026-12-31
• Not yet dueThe cost of insuring NVIDIA's debt against default will stay high through year-endDue 2026-12-31
• Not yet dueAMD's 2027 MI500 optics and the co-packaged optics delay are two different thingsDue 2026-12-31
• Not yet dueVetted-list cyber models matter only until open-weight models catch upDue 2026-12-31
• ContestedNeither frontier lab clearly made tasks cheaper than its own last modelDue 2026-12-31
• Not yet dueEmployee-built AI agents caught on, but suite vendors are taking the moneyDue 2026-12-31
• Not yet dueTraining stays on GPUs; cloud giants move their own inference to in-house chipsDue 2027-01-31
• Not yet dueCompute glut or not: unresolved for now, with a window set to checkDue 2027-01-31
• Not yet dueMeta’s year-end cash and securities will drop below its long-term debtDue 2027-02-05
• ContestedOutsiders can’t confirm a frontier AI lab has slowed; inside auditors canDue 2027-03-15
• ContestedWhite House AI accord: the gap is who sets audit scope and sees resultsDue 2027-03-31
• ContestedThe “subsidy” in AI coding subscriptions is a compute bill, not a loss billDue 2027-03-31
• ContestedLong-doc AI gap is about reading, not code; subagents close most of itDue 2027-03-31
• Not yet dueKV compression makes HBM faster, not bigger: we bet Rubin Ultra stays at 288 GBDue 2027-03-31
• Not yet dueWe bet Kuaishou’s Kling earns at least RMB 850 million in the fourth quarterDue 2027-03-31
• Not yet dueBorrowing to build AI data centers will spread, but the only case so far is OracleDue 2027-04-30
• Not yet dueCopper will keep its near-term lead over optical links inside AI racksDue 2027-06-30
• Not yet dueData-center pauses moved to governor races, still by executive orderDue 2027-06-30
• Not yet dueMoratoria delay only 2.3GW, but miss withdrawals: projects are moving, not stoppingDue 2027-06-30
• Not yet dueAI agents' CPU work is splitting from GPUs; AMD should report CPU-only rack deliveriesDue 2027-06-30
• Not yet dueComputer-use AI will pass 60% on the long-task OSWorld 2.0 test by mid-2027Due 2027-06-30
• Not yet dueFor AI that must read a lot, long context will beat retrieval and per-task fine-tuningDue 2027-07-14
• Not yet dueAI memory: putting it back in the prompt is beating writing it into the modelDue 2027-07-14
• Not yet dueAI coding assistants and coding agents will coexist; neither will eat the otherDue 2027-07-14
• Not yet dueAgent harnesses: model makers' own stacks lead; no shared standard within a yearDue 2027-07-14
• Not yet dueCoding-AI harnesses: commoditizing for consumers, still locked in for enterprisesDue 2027-07-14
• Not yet dueFrontier models still run on attention; state-space layers join only as hybridsDue 2027-07-14
• Not yet dueSynthetic data’s ceiling splits: broken in math and code, intact elsewhereDue 2027-07-14
• Not yet dueEnterprise AI splits the work: knowledge in the prompt, behavior in the modelDue 2027-07-14
• Not yet dueBuy thinking time or distill a small model? Our call: both, split by type of workDue 2027-07-14
• Not yet dueRobotics is unsettled: world models are taking ground from VLADue 2027-07-14
• Not yet dueMultimodal AI leans native, but shipping products are still mostly assembledDue 2027-07-14
• Not yet dueDiffusion has replaced GANs in image generation, and we admit this call can barely loseDue 2027-07-14
• Not yet dueAt least 4 of 5 coding agents will still rely on long-context memory in July 2027Due 2027-07-14
• Not yet dueGitHub Copilot’s revenue and paid users will show no annual decline by July 2027Due 2027-07-14
• Not yet dueBy mid-2027, fewer than 2 firms will publicly confirm fully dropping copilot-style toolsDue 2027-07-14
• Not yet dueCan independent coding tools clear 3 adoption and model-switching tests by July 2027?Due 2027-07-14
• Not yet dueOpen-ended tasks: under 2 teams will replicate synthetic-data students beating teachersDue 2027-07-14
• Not yet due60%+ of surveyed firms will feed AI knowledge via retrieval and promptsDue 2027-07-14
• Not yet dueCompanies mostly fine-tune AI for tone and format; under 30% use it to add knowledgeDue 2027-07-14
• Not yet dueOur bet: in 2027 at least two of three big AI labs still charge extra for thinkingDue 2027-07-14
• Not yet dueNarrow distilled small models will reach at least 10 named company deploymentsDue 2027-07-14
• Not yet dueOnly a 30% chance a cross-vendor agent layer becomes the enterprise standardDue 2027-07-14
• Not yet dueAI coding: hard tasks still need the native model-wrapper pair, easy work swapsDue 2027-07-16
• Not yet dueSecondSource puts 35% on a firm showing swapped AI coding parts lose under 3 pointsDue 2027-07-16
• Not yet dueAI levels tasks that have a template, not judgment; the leveled work gets cheaperDue 2027-07-17
• Not yet dueAI labs profit on tokens now, but fat margins hold only in the top tierDue 2027-07-20
• Not yet dueUS chip controls widened the gap but pushed China to build its own supply chainDue 2027-07-21
• Not yet dueAI’s long solo tasks keep growing, but they are not yet a separate axis or a businessDue 2027-07-23
• Not yet dueThe AI capex debate has moved from whose math is right to a race between two clocksDue 2027-07-24
• Not yet due“LLMs will become swappable commodities”: our call—only below the top tier, so farDue 2027-07-25
• Not yet dueWill Anthropic’s first disclosed gross margin come in at 55% or below? We put it at 30%Due 2027-07-25
• Not yet dueAI labs aren't just selling models: they are becoming do-it-all integrated makersDue 2027-07-27
• Not yet dueNeutral AI model routers will hold up: OpenRouter's weekly volume won't collapseDue 2027-07-27
• Not yet due“Open weights” no longer means one thing: three business models share the labelDue 2027-07-28
• Not yet dueSoftware margins split on usage caps, downshifting and copyability, not pass-throughDue 2027-08-02