Phase Space AI

Valuation

Broadcom [AVGO]

Broadcom Inc. [AVGO] — Valuation (v2, re-underwritten 2026-07-29)

Spot $380.46 (2026-07-28) · two outputs, both required · references/valuation.md · MEMO_BRIEF v2

Supersedes archive/AVGO_Valuation_2026-07-29_v1.5.1_superseded.md. What changed and why is §0. The prior conclusion is preserved unaltered in archive/ for comparison.

---

0. What changed from v1 — and what did not

The premise this re-run was commissioned on was half right. The Tier-1 screen did clamp AVGO's terminal margin to 15.6% from a measured 39.9%, and that cap is why the screen recorded "reverse DCF unsolvable in range." But v1 had already overridden it to 40.0% and said so explicitly. The live defect was smaller and subtler: 40.0% sits below Broadcom's trailing actual (43.66%) and 2pp below its latest quarter (49.0%), on a margin series that rose in eight of the last nine quarters. v1 committed the framework's canonical error in miniature — a conservatism haircut applied to a number that was already the conservative one. The correction is worth 4.2pp of required CAGR and it changes the verdict.

The second live defect: reverse_dcf.py was terminal-only, and the CLI still is

Every figure in this document that reads 28.3% in an earlier draft was produced by the terminal-only reverse DCF — it discounted the terminal value alone and omitted five years of interim free cash flow, which overstates required CAGR in proportion to cash generation. The module's solve() was fixed and now accepts fcf_margin, but the argparse CLI exposes no --fcf-margin flag, so any caller invoking it as a command still receives the terminal-only answer. A fix in the library that never reached the interface.

AVGO is the largest instance of this defect measured anywhere in this project, because AVGO has the highest FCF margin of any name run here: a demonstrated 43.41% (TTM OCF $33,622m less capex $860m on revenue $75,465m; fabless, so capex is 1.1% of revenue). At that margin the correction is worth −3.56pp:

Construction	Required 5y revenue CAGR	Margin vs trailing 26.4%
Terminal-only (defective; the earlier headline)	28.34%	−1.9pp
Interim-FCF-inclusive (fcf_margin=0.4341) — used throughout	24.78%	+1.6pp
Correction	−3.56pp	+3.5pp

Note the direction. The defect ran against the name, and correcting it removes the only negative valuation margin this document reported: the required path goes from marginally above the trailing CAGR to marginally below it, so the PASS no longer depends on preferring the forward comparator over the trailing one. The framework's recurring failure mode is manufacturing false FAILs, and this is that failure mode in the tooling rather than in the judgement. Every memo in this project run through the command line carries it, scaled by the name's FCF margin (+1.22pp at 5%, +3.57pp at 15%, +6.04pp at 26%).

Recommendation: add --fcf-margin to the reverse_dcf.py CLI. All figures below were produced by calling solve() directly with fcf_margin=0.4341.

---

1. Verified inputs, and the defects found

Defects found

1. AV balance-sheet commonStockSharesOutstanding = 4,876,000,000 is not shares outstanding. It is the Q2 FY26 diluted weighted average. EDGAR's cover page gives 4,757,580,198. Using AV's balance-sheet field overstates market capitalisation by $45.1bn (+2.49%). v1 used the correct figure. This is the CRWD-class failure: no within-vendor check catches it, because AV computes its own EPS off the same 4,876m and the two agree. The generalisable form, verified across all three names in this cluster: AV's OVERVIEW.SharesOutstanding is correct on all three (AVGO 4,757,580,000 ✓; AMD 1,630,601,000 vs EDGAR 1,630,600,639 ✓; MRVL 875,766,000 vs 874,800,000, +0.11%), while AV's BALANCE_SHEET.commonStockSharesOutstanding is the diluted weighted average and overstates by 1.19%–2.49%. Two AV endpoints disagree with each other and the OVERVIEW one is right.
2. AV ebit ≠ operating income on AVGO: $33,272m vs $32,945m, a $327m / 0.43pp gap. Small here, but see MRVL_Valuation.md §1 defect 2, where the same field is wrong by 22.09pp. ebit was used nowhere in this memo.
3. AV's ebitda field is unusable on AVGO, with a magnitude. For the quarter ended 2025-04-30, AV reports ebitda $10,194m against operating income $5,829m + cash-flow D&A $2,166m = $7,995m — a 1.27x overstatement, exactly as the brief predicted. The cause is visible: AV's income-statement depreciationAndAmortization for that quarter is $4,340m against the cash-flow statement's $2,166m, a 2.00x divergence. av_vs_edgar.py returned FAILED on AVGO for this flag alone; revenue disagreed on 0 of 5 quarters.
4. AV OVERVIEW.OperatingMarginTTM = 0.49 is not a TTM figure. It equals the latest quarter (Q2 FY26, 49.0%), overstating the true TTM 43.66% by 5.3pp. This resolves the apparent contradiction in the re-run mandate — "43.4% from filings and 49.0% on Alpha Vantage's TTM figure" — both numbers are right and the second is mislabelled. Anything that consumed OperatingMarginTTM as a trailing margin on a name with steeply rising margins inherited the peak.
5. AV's opex lines do not close the bridge. TTM gross 66.96% − R&D 15.89% − SG&A 5.64% = 45.44% against an actual 43.66%: an unallocated $1,345m residual = 1.78pp (acquisition-intangible amortisation and restructuring, which Broadcom reports on separate lines AV does not carry). SG&A did not come back zero, as it did on KLAC/LRCX, but the bridge cannot be closed from AV alone. The residual is carried explicitly as a line in §2 rather than absorbed.
6. Splits checked explicitly. No trap. AVGO's only split is 10:1 on 2024-07-15 — inside the multiple history window, but Alpaca bars are split-adjusted and the share count is current-basis, so the series is internally consistent. No split in 2026 on AVGO, MRVL or AMD (MRVL's are 2006 and 2004, AMD's 2000). The KLA failure mode — pre-split share count against post-split price, understating market cap 89% — does not apply to any name here. Sanity invariant holds: computed market cap $1,810bn against AV OVERVIEW $1,762bn, a 2.7% gap from a stale AV price, not a share-basis break.
7. longTermDebt did not return 0 in the newest quarter on any of the three (62,655 / 4,961.3 / 2,350), and all three tie to EDGAR exactly.
8. v1's terminal margin was 3.66pp below trailing and 9.0pp below the latest quarter. Magnitude: −4.2pp on the required CAGR, and a verdict change.
9. v1 valued a shifting segment mix on one multiple. Software goes from 32% of Q2 FY26 revenue to a guided 30% in Q3 and ~27% across NTM, at a 93% gross margin against semiconductors' 70%. Consolidated gross margin is guided down 3.5pp in one quarter on mix alone. Magnitude: a flat EV/Sales anchor overstates the 12-month target by ~7%.
10. Not a defect — a confirmation. Net financial debt and share count were both right in v1. The project's "net_cash wrong on ~20 of 22 names" pattern did not recur, because v1 read the 10-Q directly instead of taking the scanner's field. That is the fix working, and it is worth recording as such.

COMPANY STATE: A — mature and structurally stable

Evidence, and the objection answered. State A requires profitability, low operating-margin variance across ≥5 years, and no structural regime change. Broadcom's GAAP annual operating margin over six years is 16.8% / 31.0% / 42.8% / 45.3% / 26.1% / 39.9% — a 28.5pp range, which reads like State B or C.

It is not. The entire variance is purchase accounting, and it is identifiable. The two troughs are the two large acquisitions: FY2020 (Symantec/CA/Brocade amortisation at full weight, 16.8%) and FY2024 (VMware closed 2023-11-22, 26.1%). On the company's own non-GAAP basis, which excludes acquisition-intangible amortisation and SBC, the operating margin ran 65% (Q2 FY25) → 67% (Q2 FY26, a record) → 67% guided (Q3 FY26) — low variance, monotonically rising, and guided flat while revenue rises 32% sequentially. Gross margin is 67.0% GAAP / 77.5% non-GAAP on a fabless cost structure. This is a mature platform whose reported margin oscillates with deal amortisation, not a cyclical or a scaling business.

Consequence of State A: the terminal margin is built from the company's own normalised economics plus an explicit forward bridge; industry data is a sanity band and may not override — no clamping to a percentile. Instrument: fix the underwritten margin, solve for price-implied growth, show the neighbouring sensitivity surface. Evidence grade: A.

The VMware acquisition is a structural regime change to the revenue base — that is why the trailing CAGR is unusable and why §3 benchmarks against the semiconductor segment and against forward guidance instead. It is not a regime change to the economics, which is what State A tests.

---

2. Terminal EBIT margin: 47.0%, built not assumed

The segment economics, verified by arithmetic

Q2 FY26 (quarter ended 2026-04-30), from the CFO's prepared remarks, 2026-06-03. Non-GAAP — the only basis on which Broadcom reports segment margin.

The implied and reported figures reconcile to within rounding — 0.70×0.676 + 0.93×0.324 = 77.5% gross; 0.62×0.676 + 0.79×0.324 = 67.5% operating. That check is what makes the split usable rather than illustrative. Segment opex is separately disclosed for software ($1.0bn on $7.2bn).

Q3 FY26 guidance, same source: semiconductors $20.5bn (+124% YoY), software $8.9bn (+31% YoY), total $29.4bn (+84% YoY), consolidated gross margin down to ~74% on mix, consolidated operating margin flat at 67%. Verbatim: "This decline in gross margin does not represent a structural change in semiconductor margin. Rather, it reflects product mix between semiconductors and Infrastructure Software … Regardless of the impact to gross margin, we expect Q3 operating margin to be 67%, which is flat quarter on quarter, demonstrating our strong operating leverage."

That sentence is the single most important input here. Broadcom is telling the market that a 3.5pp adverse gross-margin mix shift, on 32% sequential revenue growth, nets to zero change in operating margin. Operating leverage exactly offsets mix. It is a statement about the terminal margin, made by the company, one quarter forward, and it is falsifiable next quarter.

The bridge, on GAAP — the basis the reverse DCF requires

Terminal mix: semiconductors 85% of revenue, software 15% — the direction guidance implies (semis go 67.6% → 70% in one quarter as AI compounds >200% while software grows 31%), taken to a plateau rather than to an extreme.

m_gross,T = 0.85 × 62.0%          +  0.15 × 90.0%         =  52.7% + 13.5%  =  63.0%
            ^semis, GAAP: 8pp          ^software, GAAP:
             below today's 70%          3pp below the 93%
             non-GAAP, for TPU/XPU      non-GAAP figure
             mix dilution

m_EBIT,T  = 63.0% − R&D 11.5% − SG&A 3.5% − residual amortisation/other 1.0%
          = 47.0%

Every line justified, and the residual is taken from the measured gap rather than assumed:

• Gross margin 63.0%. Below today's GAAP 66.96%. The semis leg is haircut 8pp from its current non-GAAP 70% for the GAAP/non-GAAP wedge and for management's repeated warning that "as the TPUs continue to accelerate, there will be pressure overall on gross margin." Hard constraint m_EBIT,T ≤ m_gross,T satisfied with 16.0pp of headroom.
• R&D 11.5%. TTM actual 15.89% ($11,991m). R&D dollars have been flat at $2.95–3.05bn/quarter for six quarters while revenue rose 48%. Holding R&D dollars growing at half the revenue rate for five years takes the ratio to ~11.5%. Operating leverage on a fixed-cost design house, not assumed cost cuts.
• SG&A 3.5%. TTM actual 5.64% ($4,253m). SG&A dollars flat at $1.02–1.11bn/quarter for nine quarters against revenue up 78%. Flat dollars on 2.5x revenue gives ~2.2%; 3.5% allows growth.
• Residual / other 1.0%. This line exists because AV's opex fields leave a measured $1,345m (1.78pp) unallocated (defect 5) — acquisition-intangible amortisation and restructuring. It is the load-bearing forward item: TTM cash-flow D&A is $8,753m against essentially no fab capex, i.e. overwhelmingly acquisition-intangible amortisation (VMware, Symantec, CA, Brocade). VMware closed 2023-11-22; the bulk of its intangible base amortises over five to eight years and is substantially exhausted inside the horizon. That roll-off alone is worth roughly 6–8pp of GAAP operating margin on today's revenue. Taking the residual from 1.78pp to 1.0% is therefore conservative — it assumes only ~44% of the wedge closes.

Why 47.0% is neither too high nor too low

The quarterly GAAP series: 23.7% → 29.0% → 32.9% → 42.0% → 38.8% → 36.9% → 41.7% → 45.0% → 49.0%. Eight increases in nine quarters, 25pp of expansion, with the company guiding the non-GAAP equivalent flat at a record while revenue rises another 32%. A terminal margin below the trailing actual on that series would require an argument that operating leverage reverses. No such argument exists in the filings: the only margin warning management gives is about gross margin, and it explicitly disclaims any effect on operating margin.

The honest statement of what 47.0% is: 20pp below the operating margin Broadcom reports and guides today, the entire gap being purchase-accounting amortisation and stock compensation. It sits inside the company's own evidenced range and above its trailing actual, which is what State A requires.

---

3. Output 2 — the implied-path test (the Valuation Criteria)

assets/reverse_dcf.py. Solving for the revenue CAGR today's price requires.

Result: required revenue CAGR 24.8% (v1: 32.5%; terminal-only construction: 28.3%)

Implied FY2031 revenue $228.3bn (from $75.5bn — 3.0x).

Both constructions are reported because the difference is a measured tool defect, not a modelling choice: 28.34% terminal-only, 24.78% with the five years of interim free cash flow the price also buys, at AVGO's demonstrated 43.41% FCF margin. The −3.56pp correction is the one used. See the boxed note in §0.

The benchmark that decides this name

v1 measured 32.5% against a trailing 3.5-year CAGR of 26.4% and called it −6.1pp. That is the wrong comparator for a business whose next six quarters are already published.

Only one comparator is now negative — the organic-semiconductor-only CAGR at −3.1pp, which excludes both the software segment and the AI ramp and is therefore the strictest possible read.

The decisive arithmetic. The required 24.8% compounds $75.5bn to $105.2bn at the 1.5-year mark (FY2027 year-end). Broadcom's own dated guidance for FY2027 is >$158bn: AI semiconductors in excess of $100bn, plus ~$20bn non-AI semiconductors, plus ~$38bn software at the guided Q3 run-rate. The required path is exceeded by ~50% at the eighteen-month checkpoint, using nothing but company guidance. Going from $158bn in FY2027 to the required $228.3bn in FY2031 needs 9.6% a year for four years — roughly a third of Broadcom's slowest recent growth, and slower than its non-AI semiconductor cyclical recovery alone.

That is what a PASS looks like. It is not a claim that Broadcom will earn $228bn; it is the observation that the price requires materially less than the company has already told the market, on the record, twice, under challenge.

Sensitivity over the exit multiple — and the identity constraint

Per valuation.md rule 4 the exit multiple is not free: EV_T/EBIT_T = (1−t)(1−g/ROIC)/(WACC−g).

The identity puts the warranted exit at 11.9x–21.2x. The Tier-1 anchor of 24.2x sits above all three. Reported rather than buried: the base case uses a generous exit multiple, and the sensitivity does the work.

All required CAGRs below are the interim-FCF-inclusive figures; the terminal-only value each one supersedes is shown in the last column, so the size of the tool defect is visible at every point on the surface rather than only at the base case.

Sensitivity over the terminal margin at the 24.2x anchor: 40% → 28.5%; 44% → 26.3%; 47% → 24.8%; 52% → 22.5%. (Terminal-only equivalents: 32.5% / 30.0% / 28.3% / 25.8%.)

Every cell clears the company's guided FY2027 growth rate. Even at the identity's lowest warranted multiple — 11.9x, a 5.6x EV/Sales exit, below Broadcom's own historical minimum of 7.84x — the required 41.5% is below the FY2027 guide. The verdict does not turn on the exit multiple. That is unusual and it is the strongest single fact in this valuation, because the exit multiple is the parameter the framework has found least reliable.

Terminal value is 100% of modelled EV under this construction, far above the 60% threshold, so the reverse DCF is mandatory as the primary long-horizon output and this table — not a probability-weighted scenario tree — is the analysis.

Verdict: PASS

Upgraded from v1's PASS WITH ARGUMENT. criteria.md reserves PASS WITH ARGUMENT for cases where the price requires more than demonstrated and a named mechanism bridges the gap. Here the price requires less than the company's own dated, twice-defended guidance for the next eighteen months, at a terminal margin below its latest reported quarter and at an exit multiple above what the warranted-multiple identity supports. The argument is no longer load-bearing; it is corroboration.

Ranking input for the strategy: margin = +1.6pp against the corrected trailing CAGR, +25.2pp against company-guided FY2027, +23.1pp against the latest quarter. The forward figures are the informative ones on a name whose growth rate quadrupled in five quarters; the trailing figure is reported because it is the conservative one.

What would break it. Not the multiple and not the margin — the guidance. If FY2027 AI semiconductor revenue lands at $70bn rather than >$100bn, FY2027 total is ~$128bn, the eighteen-month checkpoint is met with ~22% of headroom instead of ~50%, and the verdict weakens to PASS WITH ARGUMENT. Below ~$47bn it fails. That is a single, dated, falsifiable test and it is in the catalyst calendar.

---

4. Output 1 — the 12-month target

Own EV/Sales history, 2022-03-10 → 2026-07-28 (1,099 trading days)

Today's multiple sits at the 80.3rd percentile of Broadcom's own four-and-a-half-year history. The EV inputs are unchanged from v1 to within 0.001%, so the distribution and the percentile carry forward unmodified.

Construction: current shares and net debt held constant, price varied against point-in-time TTM revenue lagged to filing dates. Limitation, stated: share count rose 4.1bn → 4.76bn and net debt $28bn → $45bn across the window on the VMware financing, so the early series is not a like-for-like capital structure. That bias makes the historical multiple look higher than it was, so the 80.3rd-percentile reading is conservative. The 2024-07-15 10:1 split is inside the window; Alpaca bars are split-adjusted and shares are held at the current basis, so the series is consistent.

EV/EBIT is declared UNIDENTIFIED as an anchor. A continuous TTM EBIT series exists only back to the Q1 FY25 filing — 151 trading days. Per valuation.md a history that short cannot support a percentile and substituting a peer median is forbidden. For reference only: today's 56.3x sits below the ~63.6x median of that short window, because margins expanded faster than price.

NTM revenue — built from the disclosed segment guidance, by segment

Two checks. (a) The Q4 FY26 AI residual of $20.9bn is what the $56bn FY2026 guide requires given $19.1bn shipped in H1 and $16.0bn guided in Q3 — and it matches Hock Tan's prepared statement that "we expect our AI semiconductor revenue to double from what we shipped in the first half of the year" (H1 $19.1bn → H2 $36.9bn = 1.93x). (b) An analyst put the higher case to management: Harlan Sur (JPMorgan) argued the build implied "AI revenues over $60 billion with sequential growth in fiscal Q4, but you gave us the $56 billion number." Tan did not raise the guide. $56bn is therefore the conservative end of management's own range, tested in Q&A — which is the direction of bias this framework should want.

Committed-backlog cross-check: $164.6bn of RPO on contracts "where customers do not have termination rights" at 2026-05-03, ~30% (~$49.4bn) expected inside twelve months. The build requires $88.1bn of non-committed revenue against a Q3 guided run-rate annualising to $117.6bn. The build is below the current guided run-rate annualised; it is aggressive only relative to history.

The gross-margin mix haircut v1 omitted

A constant EV/Sales multiple on a mix shifting toward the lower-gross-margin segment overstates value. Quantified: NTM GAAP gross profit ≈ $137.5bn × 62% = $85.3bn against TTM $50.5bn. Gross profit grows +68.9% while revenue grows +82.2% — a 7.3% gap. Every EV/Sales-derived target below is haircut 7%, which holds EV/gross-profit constant instead. This is the CFO's "model the segments separately" instruction reduced to a single number.

Target

12-month target: $609, +60.1% to spot. Base is the own-history 75th percentile — a de-rating from today's 80.3rd percentile, so the return comes from revenue rather than from multiple expansion. Bear $297 (−22.0%) at p25. Bull $720 (+89.2%) at p90.

Read correctly: revenue +82%, gross profit +69%, multiple −6.5%, mix haircut −7%, netting +60.1%. The number is large because the guided revenue growth is large.

Sum-of-the-parts cross-check — and why it is not the primary output

The brief is right that one multiple on a two-margin business is an error, and the split is now in the memo. But a SOTP needs two exit multiples and neither leg has a valid growth-matched comparator set:

• Semiconductors grow +124% next quarter at a 62% segment operating margin. valuation.md rule 6 requires n ≥ 5 mature profitable firms with a matching operating model and dispersion in the dimension being matched. No set of five public companies growing >100% at >60% operating margins exists. The semis-leg multiple is UNIDENTIFIED.
• Infrastructure software at a 79% operating margin and 19% ARR growth has a plausible class. Pairing a defensible software multiple with an undefined semis multiple puts all the error in the leg that cannot be anchored.

Illustrative arithmetic, for transparency: NTM software $37.1bn × 79% = $29.3bn non-GAAP EBIT; NTM semis $100.4bn × 62% = $62.2bn non-GAAP EBIT. At 18x and 22x respectively the gross SOTP is ~$1,896bn, before capitalising the non-GAAP-to-GAAP wedge (~$24.7bn/yr, of which the recurring SBC element at 20x removes ~$180bn), giving ~$1,716bn EV — below the consolidated p75 target and close to today's EV of $1,855bn. That divergence is a real flag and is stated rather than suppressed: the consolidated own-history anchor is more generous than a multiple-based SOTP, and the reason is that the SOTP's semis multiple is a guess. The consolidated own-history percentile is retained as primary because it is the only anchor in either method drawn from an observed distribution rather than asserted — which is exactly what valuation.md "12-month target" rule 3 requires. A reader who prefers the SOTP should read the target as bounded below by ~$360 rather than by the p25 $297.

External sanity band. No professional target for AVGO is on file in this project, so no gap can be reported. Framework item B16 (16 of 16 house targets below spot, median 46.1% below Street) is the failure mode this output is checked against; a target 60% above spot on a name guiding +84% next quarter is the expected shape, not an anomaly.

---

5. Disclosed AI revenue — the numbers, the wording, and the evidence grade

Broadcom is the only name in this cluster that discloses an AI revenue line item. Prepared-vs-Q&A separation from the AV transcript title field.

Evidence grade on the >$100bn: high. It was volunteered in prepared remarks in March 2026 — the weaker form, per the brief — and then survived direct challenge from three analysts in June. Harlan Sur pressed that the FY2026 build implied ">$60bn" against the guided $56bn; Tan declined to raise it. Stacy Rasgon asked whether the 2027 gigawatt targets had changed: "we indicated about 10 gigawatts of shipments in 2027 … nothing has changed." Tan restated the figure twice unprompted: "$100 billion in 2027, which is what we indicated last quarter, and we continue to say that it will be over $100 billion in 2027 … and it is on the same trajectory as we are seeing in the back half of 2026."

Capacity-unit disclosure — new, and the harder evidence. Broadcom has begun quantifying in gigawatts rather than dollars: ~10 GW of shipments in 2027, ~20 GW expected in 2028, inside a 10 GW-by-2029 agreement with its lead TPU customer, plus a further named commitment "which includes XPUs and our networking, has been received and will start delivery in the second half of 2027." Three analysts independently pinned revenue content at $15–20bn per gigawatt (Rasgon, Arya, Buchalter) and Tan did not dispute the range, saying only that content per gigawatt "will trend up" generation to generation. 10 GW × $15–20bn = $150–200bn, which brackets and exceeds the >$100bn dollar guide — a second, independently-derived route to the same number. That is the cross-source verification the brief requires before adopting an unprecedented figure, and it holds. Per MEMO_BRIEF: this is unprecedented-and-verified, and it is adopted.

Also disclosed: ~$30bn of AI bookings in the quarter (raised by Joe Moore, Morgan Stanley, in Q&A) against $10.9bn shipped; RPO went $33.3bn → $45.0bn → $164.6bn across three reporting dates.

Customer concentration (see AVGO_Research.md §3 for the full table): largest customer — a distributor — is 42% of net revenue, up 21% → 28% → 32% → 42% in four years, and 44% of net accounts receivable, up from 18% a year earlier. Top five end customers ~45%. Six "core" AI customers are now named as a group in prepared remarks. The mitigant, and it is the one that separates AVGO from CRDO and MRVL: $164.6bn of firmly committed, non-cancellable RPO. Concentration with a non-cancellable order book is a materially different risk from concentration without one.

---

6. Mention frequency over time — transcripts, prepared vs Q&A

Required core metric. Source: AV EARNINGS_CALL_TRANSCRIPT, ten consecutive fiscal quarters, split at the first utterance whose title contains "Analyst". This supersedes v1's filing-word-count proxy, which was labelled a proxy because transcript-level data was not available then. Format: prepared / Q&A.

Three readings, and the Q&A split changes what each means.

1. "Gigawatt" goes 0 → 7 → 15 → 34 combined in four quarters, weighted 27/7 to Q&A. A new disclosure unit that management introduced and analysts then drove is materially stronger evidence than one management repeats to itself. Broadcom has changed the denominator it sells in — "we market our chips in gigawatts instead of dollars" — and the market has adopted it. Highest-quality signal in the table, and it did not exist a year ago.
2. VMware collapses from 44 combined mentions (FY24 Q1) to 4 (FY26 Q2), Q&A down to 1. The integration story is finished as a topic. The segment is still 32% of revenue and ~38% of operating profit, which is exactly why §2 splits it out: disclosure attention and economic weight have decoupled, and a memo following attention alone would drop a third of the earnings.
3. XPU / custom-accelerator mentions run persistently Q&A-heavy (12/34 in FY26 Q1, 6/24 in FY25 Q4). Analysts pull on this line harder than management pushes it. Combined with the gigawatt pattern the shape of the evidence is unusual and favourable: on the two items that matter most, management is being interrogated into disclosure rather than promoting it.

---

7. Cluster note

TTM YoY revenue growth: AVGO +32.3%, AMD +35.0%, MRVL +34.1% — a 2.7pp band across three fabless SIC-3674 names selling into the same hyperscaler capex budgets. A shared growth-matched EBIT exit anchor in the 20–24x range is legitimate on level.

A shared terminal margin is not, and that is where the screen failed all three. Corrected TTM operating margins, computed as operatingIncome / totalRevenue: AVGO 43.66%, MRVL 16.20%, AMD 11.65% — a 3.7x spread. The screen applied a single 15.6% "industry p75" cap to AVGO (from a measured 39.9%) and MRVL (from 16.1%) while leaving AMD at its own 10.7%. Margin is the one axis on which this cluster is not homogeneous at all, and it is the axis the screen flattened. v2's terminal margins are 47.0% / 26.0% / 24.0%, each built from that company's own bridge and its own formally-stated target model.

Trajectory, restated with the newest data — and v1 had MRVL's direction wrong. v1 recorded MRVL as decelerating (+63% → +28%) while AVGO and AMD accelerated. True of the trailing series, but it has since reversed: Marvell guided Q2 FY27 to +35%, FY2027 to ~+34% (~$11bn), and FY2028 data centre to +55%, accelerating, having raised its FY2027 outlook three times ($9.x bn → $10bn → ~$11bn). All three names are now accelerating. The separate low-end exit anchor v1 assigned MRVL for deceleration no longer has that justification; see MRVL_Valuation.md §0.

---

8. Sources

Alpha Vantage normalized INCOME_STATEMENT / BALANCE_SHEET / CASH_FLOW (78 / 73 / 75 quarterly periods), SPLITS, OVERVIEW, EARNINGS_CALL_TRANSCRIPT FY24 Q1 – FY26 Q2 (ten quarters, with speaker / title); assets/av_vs_edgar.py --symbols AVGO,MRVL,AMD --quarters 6, re-run after the 52/53-week fiscal-calendar fix (AVGO: 5 periods compared, 0 revenue disagreements, FAILED on the AV-ebitda flag only); SEC EDGAR companyconcept for dei:EntityCommonStockSharesOutstanding, LongTermDebt, LongTermDebtCurrent, LongTermDebtNoncurrent, CashAndCashEquivalentsAtCarryingValue (CIK 0001730168); Broadcom 10-Q 2026-05-03 (filed 2026-06-09), 10-Q 2026-02-01, 10-K 2025-11-02; earnings release 8-K EX-99.1 2026-06-03; Alpaca daily bars 2026-07-28; assets/reverse_dcf.py.