What We Revised, Refuted, and Retired
This is the full list of times the platform changed its mind — including the times it was right. Every entry carries a before and after confidence, the claim that was replaced, and what the evidence still cannot say. A product that only publishes its successes is a marketing product; this ledger is what makes the rest of the platform credible.
3
Revised4
Refuted1
Retired4
Called EarlyWe modelled the sequential accelerator revenue acceleration as demand arriving at a supply-constrained company, which is the conventional reading of a revenue step-up in this industry and the reason guidance gets treated as a demand signal.
The cluster documents two quarters bounded by advanced packaging allocation, and that constraint clearing. Data center revenue reached $54.2 billion with full-year guidance raised at both ends, from $295-305 billion to $310-315 billion, in the same quarter. The step-up was supply, and raising the floor of the range rather than the midpoint is the tell.
What we still cannot sayThe forward constraint moved to facility power and cooling, and the Rubin's roughly 3x rack draw has no published time-to-energization figure for any reference installation. A per-watt throughput gain does not shorten a grid interconnection queue, so the next binding variable is the one the corpus currently measures least well.
A 4 to 8 percent CET1 range was wide enough that no rational bank would change its distribution plan before the final rule, so the $50 billion authorization would proceed on its own timetable and the re-proposal was noise for capital return.
The earnings call contradicted that directly. The CFO said the buyback pace would be calibrated to the re-proposal, and management said it would wait for the final rule before adjusting capital return plans. A $50 billion authorization and a rule-conditioned pace are the same decision, not two independent ones.
What we still cannot sayWe still cannot say which end of the 4 to 8 percent range JPMorgan lands at, because the GSIB surcharge formula is not fixed. Nor can we size how much of the authorization actually gets used: an authorization is a permission, and the 90-day comment period plus the unfixed formula decide the pace. Two consecutive quarters of disclosed repurchase volume would settle it.
The live question going into the quarter was whether the guidance increase to roughly $94.5 billion would come from loan growth or from the catch-up in fixed-rate asset repricing. We had no view and modelled it as volume-driven, which is the conventional reading of a revenue guidance raise.
The release answered it against us. Average loans were flat at $1.35 trillion, deposits rose 2 percent, the margin gained nine basis points, and the CFO attributed the beat to repricing and disciplined deposit pricing rather than to lending, with roughly $40 billion of Treasuries maturing into a reinvestment premium.
What we still cannot sayThe size of the securities reinvestment benefit was called significant without a figure, so no quarterly disclosure currently separates the catch-up effect from the run-rate effect. Management also said the beat was structural but not repeatable in the same breath, which is why the confirmed read is about the mechanism and not about how much of it survives into 2027.
Before the award the government book read as a durable annuity with a contested edge. Maven was the largest program but a competitor protest seeking to void the expansion was live, so the correct model was steady government revenue with a probability-weighted timing haircut and the growth story sitting almost entirely in US commercial.
The award made Palantir the program of record across five combatant commands with deployments funded through fiscal 2029, the UK and Australia inside the delivery framework, and a $212 million option for 2028 disclosed before exercise. The Army's filing to deny the stay reframes the protest as a scheduling variable rather than a demand risk.
What we still cannot sayGAO data cited by the protester puts similar challenges at a 94-day average delay, and no ruling on the preliminary injunction in No. 26-1182C exists yet. We also cannot price the retention cost of the model: stock-based compensation sits at 24 percent of revenue with a 20 percent target stated twice and no repurchase authorization announced.
Azure growing 34 percent was read as new AI workload demand landing in the cloud, which is the reading that extrapolates beyond the installed base and the one we used.
Management attributed the acceleration to AI attach inside existing enterprise agreements rather than new migrations, with remaining performance obligations at $287 billion and Copilot at general availability for $30 per seat with 37 percent weekly penetration. Higher quality and much harder to extrapolate, because the revenue attaches to customers who were already spending.
What we still cannot sayAttach-rate growth shows existing customers spending more; it cannot show whether AI workloads are additive to cloud demand or cannibalizing migrations that were already planned. The company is the only party holding the denominator, having declined an AI run-rate breakout on cadence grounds and introduced a bookings metric with no prior-year comparison.
Silicon hiring, a shipped developer API and a monetized cloud fallback described a deliberately two-tier deployment that looked sequenced and largely on track, with the launch date treated as a marketing choice rather than an open question.
The fiscal Q4 call reframed the same evidence as a program running later than its architecture. The CFO attached $1.4 billion of incremental model training cost to the quarter and called it temporary without an end date, and the CEO declined a launch window three separate times while describing an orderly migration.
What we still cannot sayNo independent thermal or throughput measurement of the on-device model at the 2GB residency floor exists, so we cannot tell whether the refusals were discretion around a settled plan or management of a slip. The $1.4 billion is labelled temporary with no end date attached, and no filing quantifies yield learning at the new process nodes that would date the hardware transition.
This question opened on the assumption that Waymo's fleet growth would follow model performance as it did in earlier autonomy programmes, so scaling would be gated by how well the system drove rather than by what the company was permitted to do.
It resolved against itself. Nineteen of fifty-eight new deployment roles reference driverless operation or remote assistance policy, eleven at staff level or above, and the lead role owns the operating policy rather than the software. The safety-monitor ratio was published and the driverless ratio withheld, and the 2,000-vehicle target arrived without a capital figure.
What we still cannot sayThe premise being refuted does not mean the constraint is now known. The permit conditions attached to driverless operation in the two authorized markets have not been published, the fully driverless ratio behind the fleet target was withheld alongside a published safety-monitor ratio, and no capital figure has ever been attached to that target.
The read was that a reported revenue decline signalled a company still being dragged down by pandemic annualization, with oncology growing inside a base that could yet shrink as the COVID comparison rolled off.
Ex-COVID revenue grew 8 percent operationally with oncology up 16 percent, roughly $3.5 billion of COVID inventory was retired with no further write-downs expected, and a perioperative Padcev label added an estimated 18,000 eligible US patients annually. The denominator changed permanently and the fourth-quarter print is the first one with no comparison cushion.
What we still cannot sayThe perioperative label carries a 12-month treatment duration and shares its economics with a partner, so the 18,000-patient figure is an addressable population rather than a revenue figure. Nothing in the collection sizes the oncology franchise net of the shared economics, and the fourth quarter has not printed yet, so the confirmation is about the reporting frame rather than about the growth rate.
After the annual filing the working read was that export controls were the dominant structural variable. License denials had already stopped shipments to some customers, the revenue impact was left unquantified, and additional jurisdictions were signalling similar frameworks.
An eleven-year sovereign compute agreement carrying phased delivery beyond the current hardware generation inverted that frame. Demand that does not depend on the licensed destinations the filing worried about arrived with a decade of visibility, so compliance-by-design now reads as an enabler of diversified demand rather than only a cost of doing business.
What we still cannot sayThree numbers stay undisclosed: the revenue impact of license denials across four consecutive filings, the sovereign share of data center revenue which was declined twice on the record, and the interconnection schedules for the announced campuses. Direction is well supported and magnitude is not, so the confidence gain belongs to the thesis, not to its size.
The granted CDK4 dosing claims would hold against a first biosimilar challenge to the successor program, on the strength of a 2043 expiry, two pending continuations, and a franchise already defended once without a trial.
The question is retired rather than carried at low confidence. The decisive evidence lives in the FDA Orange Book and in ANDA paragraph certifications, and neither is reachable through the current scraper set, so continuing to collect Pfizer-originated signals cannot resolve it. Every additional signal would have been activity that looked like diligence.
What we still cannot sayWhether the granted claims survive their first validity challenge is unknown, and two filers have already certified against the 2031 and 2032 formulation patents while leaving the 2043 dosing claims alone. We cannot say whether that skip was a valuation, a route choice, or an oversight, and we cannot size the at-risk market because the notices do not disclose it.
The premise under collection was that vertical integration would supply the energy growth engine before the third-party fallback was needed, which is what the cost and margin story had been resting on.
The annual filing acknowledged that two in-house cell projects have experienced production ramp delays, described long-term contracted volumes as a partial mitigation, and prioritized LFP explicitly as a cost decision over energy density. The fallback they built is now the primary path, and the margin story that priced the integration premium has to be re-run without it.
What we still cannot sayNo supplier is disclosed above a stated threshold and the word partial carries no quantified remainder, so the size of the dependency is unknown. The concession is that this was not a collection failure but an evidentiary one: the filing is complete enough to raise the question and not complete enough to settle it, and only a subsequent filing stating those projects reached volume would confirm the refutation.
Management guided the embedded segment to bottom in the second half, so we carried the recovery as guided and treated channel data showing revenue down roughly 18 percent with fourteen weeks of industrial inventory as timing noise.
Two Japanese customers deferred 2027 orders, design wins moved to Arm competitors in edge vision and control applications, and test capacity was reallocated toward data center parts, extending industrial lead times. The recovery did not materialise on the guided schedule and the question is closed rather than carried into the next fiscal year.
What we still cannot sayChannel data is not a filing, and nothing in the quarterly segment disclosure breaks embedded revenue down by design win or order date. We know the recovery did not arrive on the guided schedule and we cannot say whether it was deferred, cancelled, or repriced, because the two deferred 2027 orders were never individually disclosed.
We carried the €38.7 billion backlog as evidence that the multi-year product cycle is intact, treating the China decline from 47 percent of system sales to 28 and then 18 as a known and quantified correction with a visible end.
The China correction is quantified and the backlog is not. ASML discloses aggregate bookings but never their geographic or product composition, so the same €5.2 billion can mean very different things depending on how much sits in licensed destinations and how much is High-NA. We have stopped implying a resolution in either direction.
What we still cannot sayWe cannot say whether the current licensing framework is the floor or an interim step before computational lithography, spare parts, and field-service rights come under restriction. The annual report describes only a portion of the €11.2 billion inventory balance as built for restricted destinations, and the restricted portion is never sized.
This ledger is the discipline that makes the rest of the product credible. Every entry here was forced by evidence the platform could not ignore, and every entry carries a mandatory field — what we still cannot say — that prevents the ledger from becoming a victory lap. A product that only publishes its successes is a marketing product, not an analyst product.