Print-season model review - the numbers stepped up once, then stopped

A The scorecard - both graded calls were right, one for the wrong reason

Only two of this cycle’s four prints were covered by the prior edition, which was written on 23 August, before Marvell and Rubrik reported. Those two are the whole gradeable set.

Applied Materials - vindicated on both legs. The prior call was EPS up, multiple down. Consensus did step up, with the reverse-DCF’s consensus revenue CAGR moving 16.8% to 22.6% across the print [expectations:AMAT], and the multiple did compress against those rising numbers, hard: the stock ground from a $534.54 base to $458.39 by 31 August, down 14.2%, while sell-side houses raised estimates and cut applied multiples in the same notes [evidence:AMAT:amat-ev-0614] [evidence:AMAT:amat-ev-0615] [evidence:AMAT:amat-ev-0641] [evidence:AMAT:amat-ev-0642]. The honest qualification is that the EPS leg has already expired - consensus has been flat since 15 August - and that part of the “gap is narrowing” story is price, not consensus, because the reverse-DCF workbook is frozen at a 17 August build for every later row [expectations:AMAT].

Cerebras - vindicated, on a short. The prior call was EPS up, multiple down. Consensus reset up in a single hop on 13 August, with FY26 revenue $864.9mn to $887.3mn and FY27 EPS $0.965 to $1.297 [doc:Market Data\CBRS_estimates_daily.csv], and then stopped. Against that unmoved denominator, EV to forward sales compressed from 33.19x on 13 August to 24.74x on 28 August [doc:Market Data\CBRS_multiples_daily.csv]. A 25% de-rate with the numerator doing all the work is precisely the split the call named.

A correction that matters more than either grade. I was working from a reading that all six Applied Materials series were decelerating post-print, and that framing is wrong in a way that would have driven a bad call. The log shows one discrete step at the 15 August anchor - FY27 EPS slope 19.2 to 13.8 - held flat at 13.8 to 14.2 through 31 August [slope:AMAT:EPS FY27 (USD)]. The delta_pp column stays negative only because its comparison window still reaches back before the print. A slope that collapses within two days of a print and then sits still is the quarter’s actual re-basing inside a trailing window, not accumulated estimate cuts, and the levels remain solidly positive. The same artifact is visible on Marvell, where a single post-print reading flipped four series to decelerating [slope:MRVL:EPS FY27 (USD)], and in inverted form on Rubrik, where the FY27 EPS slope reads +106.7pp [slope:RBRK:EPS FY27 (USD)] on a consensus EPS base near zero. The real Rubrik move was 31 cents to 50 cents [doc:Market Data\RBRK_estimates_daily.csv]. Nineteen cents on a thirty-one cent base is a 59% level change, not a 189% one.

Rows proposed for the calibration ledger

Three candidates for the analyst to log. None is auto-logged.

• The revision-slope log’s delta_pp is structurally misleading for roughly two weeks after any print, because slope_now has re-based on the delivered quarter while slope_prior still spans the pre-print window. It reads as a revision-cycle roll-over when it is an arithmetic artifact. It appeared on three of the four names this cycle and would have produced a wrong EPS call on Applied Materials.
• A percentage slope on a near-zero denominator is reported without a level check. Rubrik’s +106.7pp encodes a nineteen-cent move.
• The landed roster returns after-close prints on the day itself, before they have printed. Broadcom appears on this cycle’s roster with a print scheduled for the evening of the run.

B The cycle read - four prints, four subsectors, no lane to read

This is the finding that shapes the section rather than something inside it. The four names map to four different subsectors - Applied Materials to semi capital equipment, Marvell to custom silicon and AI optics, Cerebras to AI accelerators, Rubrik to cyber security and data resilience [doc:Universe\universe.md]. No lane printed twice. There is no shared customer, no common shortage, no pull-forward visible across two reporters in the same business, because no two reporters were in the same business.

What does cut across all four is not a subsector thread but a tape behaviour, and it is worth separating from the fundamentals carefully. Each of the four beat, three raised a guide, and each was sold. Applied Materials surrendered a 6 to 7% beat-and-raise inside minutes [guidance:AMAT:amat-gd-0001] [guidance:AMAT:amat-gd-0002]. Cerebras beat its own guide by 8.2% on core revenue and raised all three full-year lines [guidance:CBRS:cbrs-gd-0010] [guidance:CBRS:cbrs-gd-0011] [guidance:CBRS:cbrs-gd-0012], and 89% of its drawdown landed before the call had run. Marvell cleared all three guided midpoints on the store’s own unpromoted rows [guidance:MRVL:mrvl-gd-0001] [guidance:MRVL:mrvl-gd-0003] [guidance:MRVL:mrvl-gd-0005] and faded through the call. Rubrik popped 8.3% and gave up twenty points in 54 minutes.

The competing read deserves stating rather than dismissing: all four are AI-capex-levered or AI-adjacent, three sit in semis, and August was a poor month for the semis tape - the SOXX fell from 559.12 on 17 August to 508.62 on 28 August. If the de-rate were a cohort event, that would be the explanation and there would be no company-level finding here at all. The benchmark decomposition is what rules it out. Applied Materials’ t+1 excess was -5.06pp of a -5.12% move, meaning the sector explained essentially nothing [doc:Market Data_event_path\AMAT.csv]. Cerebras’ was -12.60pp on a day the SOXX rose [doc:Market Data_event_path\CBRS.csv]. Marvell’s derived excess is -7.09pp and Rubrik’s -12.82pp. These names were not carried down; they were individually sold in a weak tape.

C The calls, name by name

Long book

Applied Materials - EPS hold, multiple down. The raise has been fully absorbed and there is nowhere left for the earnings chain to bend up. Revenue drivers are intact and largely un-lapped, with a Q4 guide 6 to 7% above pre-print consensus and China flipping to a guided growth contributor at 27.5% of revenue [evidence:AMAT:amat-ev-0320], but consensus already took that step and has not moved since 15 August [expectations:AMAT]. Margins are the only place the chain can move and they point the other way: Q4 gross margin is guided flat at roughly 50.4% into a 12% sequential revenue ramp, and management declined to raise long-term targets where a peer raised its own [evidence:AMAT:amat-ev-0650]. Everything else on margin confirms hard - a thirteenth consecutive quarter of year-on-year gross margin expansion and a record 34.0% non-GAAP operating margin [evidence:AMAT:amat-ev-0318] [evidence:AMAT:amat-ev-0599] - so this is a ceiling question, not a floor question. The downgrade from last cycle’s up-call is absorption, not breakage.

Management’s benign read on the flat margin is a ramp and service-hiring headwind, and it is credible and transient by construction. The competing read is leading-edge pricing erosion, sharpened by a roughly 150 basis point gross-margin gap to a US peer that was raised on the call and not answered. The 12-13 October investor event is the scheduled discriminator, because the wide-WFE-scenario margin targets were explicitly deferred to it.

The multiple call carries lower conviction than last cycle for an honest reason. Roughly 45% of the implied-versus-consensus revenue CAGR gap has already closed, from 32.3pp in mid-July to 17.6pp on 31 August [expectations:AMAT], and the portion after 17 August is price-only against a frozen workbook, so the true current gap is smaller than 17.6pp and unmeasured. What holds the call is that the selling did not stop at the print: the stock round-tripped above its base on 17 August and then fell in eight of the next ten sessions on falling highs. That is distribution, not a single-session expectations reset.

The claim that moved is the de-rating risk claim (AMAT thesis 4), and it was right for a reason it did not specify. Its falsification condition requires that “forward EPS estimates keep inflecting UP … AND the multiple holds or expands.” The first conjunct fired and the second did not, so the AND is broken and the risk stands. But the claim’s own script was de-rating on the first downward revision, and what actually happened was de-rating despite upward revisions. That deserves an adjudication note rather than a clean tick. Separately, the China cap claim (AMAT thesis 5) is moving toward its falsification rather than its confirmation - its premise is a step down to a 15 to 20% floor, and China stepped up to 27.5% with growth guided in both CY26 and CY27 [evidence:AMAT:amat-ev-0603]. Neither is falsified on one print; both need the analyst.

Marvell - review missing, run /earnings-review MRVL. Everything below is tape and store rows, explicitly not a print verdict.

The ladder is the one genuinely readable thing, and it has the classic shape: a 1.67% pop on the headline, reversed to -3.13% by the release settle, then -5.97% by the end of the call and -7.80% into the after-hours settle [doc:Market Data_event_path\MRVL.csv]. The erosion concentrates after the release settled, which is where guidance and Q and A sit. A pop the call erases indicts the guide, not the quarter. The store never wrote the t plus one or benchmark legs, so I derived them: the 28 August close of $216.62 against the $241.45 base is -10.29%, the SOXX fell 3.20% the same session, leaving roughly -7.09pp idiosyncratic [doc:Market Data\MRVL_ohlcv.csv] [doc:Market Data\SOXX_ohlcv.csv].

Against the store’s unpromoted guide rows the quarter cleared every midpoint - revenue $2,739mn against a $2,700mn midpoint, EPS $0.94 against $0.93, gross margin 58.9% against 58.75% [evidence:MRVL:mrvl-ev-0033] [evidence:MRVL:mrvl-ev-0053] [evidence:MRVL:mrvl-ev-0062] - and four broker houses raised FY27 and FY28 into and after the print [evidence:MRVL:mrvl-ev-0034] [evidence:MRVL:mrvl-ev-0045] [evidence:MRVL:mrvl-ev-0078]. Candidate mechanisms for the fade are visible but not adjudicable here: a Q3 gross margin guided to 58.0% against a 58.5% street number and attributed to custom mix [evidence:MRVL:mrvl-ev-0063], and diluted shares at 921.2mn against 915mn guided with stock compensation stepping to 11.9% of revenue from 8.6% [evidence:MRVL:mrvl-ev-0031] [evidence:MRVL:mrvl-ev-0077].

That last pair is why the missing review is expensive. The custom-XPU margin dilution claim (MRVL thesis 4) has a two-part falsifier requiring gross margin to stabilise at or above 58.5% for two consecutive quarters while data-centre mix rises, and the ledger records exactly that pattern - 58.9% held flat for a second quarter with mix up to roughly 79% [evidence:MRVL:mrvl-ev-0039] [evidence:MRVL:mrvl-ev-0047]. Whether both legs are now met is a judgment the review produces and this edition must not manufacture. No claim status should move on the strength of this block.

Rubrik - review missing, run /earnings-review RBRK. Tape and store rows only.

The ladder here inverts the canonical pattern and the inversion is the finding. The stock peaked 8.34% up at 16:06, then round-tripped the entire twenty points to -11.93% by 17:00, before the call began - and the call then recovered it 7.8% off that trough to close the session at -7.01% [doc:Market Data_event_path\RBRK.csv]. Whatever did the damage was in the press release, and management talking was net additive. That is the opposite of Marvell on the same evening.

The base deserves a caveat that changes the reading. Rubrik rose 11.33% on 27 August itself, into an after-close print, so every percentage above is measured off a mark that had already jumped double digits. The derived t plus one is -13.05% against a flat SPY, roughly -12.82pp of adverse excess. Measured instead from the 26 August close, the whole reaction is -3.2% [doc:Market Data\RBRK_ohlcv.csv] [doc:Market Data\SPY_ohlcv.csv]. Both framings are arithmetic. Which is the honest reaction depends on whether that run-up was pre-positioning, and that is a judgment the review has to make.

Consensus rose after the print, but read the mechanism before crediting it: FY27 revenue marks landed at $1,690.68mn, inside the raised guide range, which is the street snapping to a midpoint rather than running ahead of one. The post-print evidence is genuinely two-sided, 15 confirms against 11 challenges, including a Q3 EPS guide of $0.07 to $0.09 against the $0.20 just delivered [evidence:RBRK:rbrk-ev-0421], Q3 revenue growth of 23% against Q2’s 38% [evidence:RBRK:rbrk-ev-0245], free cash flow margin contracting to about 15% from 19% [evidence:RBRK:rbrk-ev-0425], and ARR accelerating to 33% [evidence:RBRK:rbrk-ev-0249].

This is the most costly missing artifact in the cycle, because the print was the pre-registered test. The ARR stabilisation claim (RBRK thesis 2) names as its falsifier “net new ARR flat-to-down YoY at the Q2 FY27 print despite the seasonal step-up” - this print, by name. The operating-leverage claim (RBRK thesis 1) carries an analyst note deferring a live bull-bear split to “the next 1-2 subscription-ARR prints”. The system designated this quarter as the discriminator for two claims and then did not adjudicate it.

Short book

Cerebras - EPS hold, multiple down. The uncomfortable half first, because on a short it is the half that matters. This was a beat-and-raise on all three full-year core lines, core gross margin cleared its guide, and the bull-case risk claim (CBRS thesis 6) confirmed at the print with RPO at $25.4bn, an AMD partnership for Q4 2026 production and AWS on Bedrock for Q1 2027 [evidence:CBRS:cbrs-ev-0214] [evidence:CBRS:cbrs-ev-0216]. All of that is bad for the short.

The earnings chain is a hold with the skew down, and the skew is specific rather than atmospheric. FY26 consensus of $887.3mn sits at the top of management’s own $880 to $890mn band, so there is no headroom without a further beat, and Q3 is guided to just 2.4% sequential growth. The unresolved line is FY27, where consensus of $2,951.6mn sits roughly 11% above the approximately $2.66bn floor implied by management’s own language, and the street revised FY27 up rather than converging toward it. That gap has to resolve; this print did not decide it.

The multiple call rests on supply and sentiment, and I am naming that limit rather than dressing it as valuation. The cycle’s single most decision-relevant fact is not in the evidence ledger at all: across 14 to 21 August, in the window around the 19 August lock-up tranche, the insider store records 138 open-market sale rows including CEO Andrew Feldman, CTO and co-founder Sean Lie, and directors Eric Vishria and Steven Vassallo, summing to roughly $245mn of proceeds, against on the order of 118,000 shares across the name’s entire post-IPO history before that window [doc:Market Data\CBRS_insiders.csv]. The tranche calendar continues on 2 and 16 September, 30 September, 14 and 28 October, and expires 10 November. That is mechanical supply meeting an already-broken tape, and it materially confirms the lock-up overhang claim (CBRS thesis 4).

The benign read is real and must be carried: the largest Feldman, Lie and Vishria tranches carry 10b5-1 flags, so they are pre-scheduled plan sales set at listing rather than fresh decisions, and several of the 18 August sequences share an identical price ladder across four insiders, which reads as one broker-executed settlement rather than four independent judgments. What discriminates is the September tranches - plan selling that continues at materially lower prices is mechanical, while fresh non-plan sales in the $180s would not be.

The claim that has met its condition on its face is the OpenAI dependence claim (CBRS thesis 5), and it is still unadjudicated. Its falsification names “the 10-Q customer table shows the largest customer below 50% of revenue while total revenue still tracks the street’s out-year slope.” Customer A printed at 34%, down from 70% a year ago, and FY27 and FY28 consensus were revised up [evidence:CBRS:cbrs-ev-0195]. The competing read, on its own terms: Customers A and D are disclosed related parties inside the same sovereign complex, so a 34% top name beside a 32% second name may be counterparty re-labelling rather than diversification, and EMEA remained 53% of revenue growing 31%. What discriminates is the undisclosed identity of Customer B at 32% and Customer C at 10%. If B is OpenAI and C another sovereign, the table is a duopoly; if either is a hyperscaler, the claim is genuinely falsified. That is an IR question, not an inference.

One further fact cuts against the short’s tape leg: short interest rose from 12.90mn shares on 31 July to 15.84mn on 14 August [doc:Market Data\CBRS_shortinterest.csv]. The squeeze fuel was loaded and did not ignite this cycle.

D Adjudication

The highest-probability read is that the previous edition’s regime call is holding, but its mechanism was mis-stated, and the correction is not cosmetic. Last cycle described a revision cycle that was intact while the market refused to pay for it. This cycle’s evidence says consensus is not still inflecting on these names: it took one step on each print and stopped, in every case within days. Applied Materials flat since 15 August, Cerebras frozen since 13 August, Rubrik snapping to a guide midpoint, Marvell showing a single post-print reading. The multiple, by contrast, kept compressing for two to three more weeks after each print. “Numbers up, multiple down” was the right observation with the wrong tense. The accurate version is that the numbers went up once and stopped, while the de-rating continued - which is a materially worse setup, because the earnings leg that was supposed to carry these names has already been paid out.

The strongest evidence for the regime reading is that it reproduced in a completely non-overlapping name set. Last edition’s ten names shared none with this edition’s four, and the four here sit in four unrelated subsectors, yet all four de-rated from their print base and three of four carry idiosyncratic excess between -5 and -13pp. That is not four company stories.

The counterweight, stated in the headline rather than a footnote. Half this edition has no adjudicated review. Marvell and Rubrik contribute tape, ladders and evidence counts, not analysis, and both abstentions are recorded as holds - which flatters the call distribution, because two of the four “holds” are silences rather than judgments. Any reading of “0 up, 0 down, 4 hold” as a balanced season is wrong. The real distribution is two down-calls on the two names we can actually see, and two blanks.

What consensus is currently mispricing, ranked by how cleanly the evidence supports it.

First, the Cerebras insider supply. Roughly $245mn of open-market selling by the CEO, the CTO and two directors, against a name whose entire prior post-IPO insider selling was on the order of 118,000 shares, into a tranche calendar running to 10 November - and none of it is in the evidence ledger, so no downstream reader of that store would see it. It is the clearest gap between what is knowable and what is priced in the whole cycle, and it sits on the short.

Second, the Rubrik print being the designated discriminator for two claims and passing un-adjudicated. This is not a mispricing the market holds; it is one the system holds. The raw material is already in the workspace - the transcript and 8-K are in the corpus and 34 evidence rows were extracted on 1 September - and only the adjudication is missing.

Third, the Applied Materials margin ceiling. Consensus is carrying a Q4 gross margin recovery that management guided flat into a 12% sequential ramp while declining to raise the long-term target a peer just raised. The October investor event is the scheduled resolver, and it is the only leg that can bend the earnings chain down on this name.

Fourth, and most contested: whether Marvell’s custom-XPU dilution claim has now met both legs of its falsifier. A second consecutive quarter of 58.9% gross margin on rising data-centre mix is on the record. If the second leg is met, a risk claim dies and the name’s margin debate closes in the bull’s favour; if the revisions leg fails, it stands. The evidence is in the store and the judgment is not.

E Data limitations and gaps

Ranked by what would most change a call if closed.

1. Two missing earnings reviews - Marvell and Rubrik, both printed 27 August. Every judgment on half this cycle is blocked on them, and Rubrik’s is the pre-registered test for two claims. The corpus already holds both transcripts and 8-Ks.
2. Cerebras insider selling is absent from the evidence ledger, which stops at 12 August. The most claim-relevant fact of the period is invisible to any store consumer.
3. Event-path benchmark and t plus one legs missing for both 27 August prints. Neither Marvell nor Rubrik has an excess-versus-benchmark row at any phase. I derived both by hand from the price store; the daily leg remains recoverable via the adapter’s backfill, and the method verifies exactly against Applied Materials’ stored decomposition.
4. The revision-slope log has no rows for Cerebras at all, so every statement about its revision cycle rests on raw estimate levels rather than a measured slope.
5. Guidance registries are behind the prints. Cerebras’ Q3 and FY26 rows remain staged rather than promoted and its actuals file carries no Q2 FY26 row; Marvell has seven staged rows and no promoted file, with the 27 August forward guides absent entirely; Rubrik’s registry stops at the June vintage. In each case the comparator a falsification condition names is not queryable.
6. Stance stores did not ingest any of these prints. Applied Materials holds two pre-print placeholder rows with no rating or PT, Cerebras ends 24 June, Marvell carries no ratings at all, and Rubrik has no stance file. There is no sell-side reaction read anywhere in this cycle.
7. The Applied Materials reverse-DCF workbook is frozen at a 17 August build, so all later gap-narrowing is price movement, not consensus movement.
8. Rubrik’s benchmark resolves to SPY because its theme string matches no token in the divergence config, and no security-software cohort tape is wired in.
9. Broadcom and Credo sit on the roster without prints. Broadcom’s Q3 FY26 is scheduled for the evening of this run and has not reported; Credo’s Q1 FY27 date passed on 1 September with no print materials in the corpus, which is either an ingestion lag or a slipped vendor date and cannot be distinguished from the workspace. Neither is counted as a printed name here.
10. Thesis adjudication backlog across all four names. Applied Materials, Cerebras and Marvell claims carry null analyst adjudication; Rubrik’s claims 5 through 7 are marked analyst-confirmed while their own notes still read “staged - review + promote or reject”. The 30-day auto-promotion clock runs out on 16 September.