Broadcom

Refresh delta (since 2026-07-07)

What changed (material moves, one line each):

1. Apple, through 2031 — the franchise broadened. 8-KForm 8-KA filing for something that happened between reports and matters enough to tell shareholders now. Item 8.01, filed 2026-07-06: Broadcom and Apple "expand their long-standing technology collaboration through 2031" via new multi-year agreements for "a range of custom ASIC silicon products for use in multiple generations of Apple products". Sized at >$30B through 2031 and covering custom RF, Wi-Fi/BT connectivity and networking silicon. AVGO +6.49% on 2026-07-07.

2. The stock re-rated +13.8% on no new quarter — and crossed above Nvidia on forward P/E. $373.90 → $425.41, cap $1.78T → $2.02T. Fwd P/E 27.15x AVGO vs 22.43x NVDA. This reverses the prior dossier's Lens 7 finding that "AVGO now arguably [screens] cheaper on forward P/E." It does not screen cheaper any more.

3. Hock Tan conceded Google multi-sourcing on the record — and the prior dossier missed it. Answering Jefferies on the FQ2 call: "we fully expect that there will be some diversity of sources for them… but our commitment from them is a very substantial dollar amount". Macquarie cut Outperform→Neutral, PT $513→$437 on Broadcom's share of Google TPU revenue falling ~95% (2026) → 80% (2027) → 65% (2028) and a 21% cut to 2028 EPS. Morgan Stanley (Joseph Moore, PT $502) counters that MediaTek "is real, but not disruptive" and Broadcom retains ~80%. Unresolved — surfaced, not adjudicated.

4. The "no material regulatory findings" clean bill is VOID. The prior dossier's Lens 10 said "No material regulatory or legal findings." That conclusion rested on regulatory/regulatory-findings.md, which checks only SEC EDGAR LR + AAER and whose own footer says non-SEC agencies "are NOT indexed in EDGAR — run the following web search in Lens 10." That search was not run. It has now been: (a) the European Commission has a formal investigation into VMware access/pricing/contract terms, with a CISPE complaint filed 2026-03-19 over the EU CSP-program closure, and Broadcom lost its EU General Court bid to block a US-documents request; (b) a US ITC investigation on Netlist memory patents names Broadcom alongside Google and Nvidia over AI-server memory. All — none is in the 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited. Note 10 contingencies the prior run relied on.

5. XPUs are the LOWER-margin part of the AI business — the opposite of what the prior SOTP assumed. CFO Kirsten Spears, FQ2 call: "our ASICs, TPUs, some of the wireless business have lower margins. So as the TPUs continue to accelerate, there will be pressure overall on margins… the AI networking side of the business has very rich margins, so it will offset it somewhat." Consolidated non-GAAP GM guided 77.1% → ~74% for Q3. The prior dossier's SOTP assumed AI was "the richest slice" and applied a ~55% AI EBITDA margin. That assumption is now contradicted by the CFO and is repaired in Lens 11 below.

6. Deleveraging started, and governance flashed amber. Broadcom completed an upsized $3.0B cash tender (from $2.5B) for six note series on 2026-06-17/18 — ~$5.5B tendered, ~$2.9B principal accepted. Separately, at the 2026-04-20 annual meeting say-on-pay passed with 1,232,879,962 against vs 2,433,503,375 for — ~33.6% opposition, and director Harry L. You drew ~954.7M against (~26%).

Also new to the shelf, pre-boundary but absent from the prior dossier (surfaced honestly as gaps the last run left, not as fresh events):

• Apr-6-2026 8-K (read direct from EDGAR this run): a Long Term Agreement with Google for custom TPUs across future generations, plus a Supply Assurance Agreement for networking/other components in Google's next-gen AI racks through up to 2031; and a three-way expansion under which Anthropic accesses ~3.5GW of next-gen TPU-based compute via Broadcom from 2027 — with the sentence that matters: "The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success."
• AI XPV/XPU Platform, announced 2026-06-09: Apollo-led $35B first tranche with Blackstone Credit & Insurance as anchor investors, targeting >20GW through 2028, first serving Anthropic's >1GW at Fluidstack sites from mid-2026. This is the commercial dress on the 10-Q Note 11 $29B lease backstop the prior dossier flagged as a subsequent event — the structure is larger and more institutional than "an investor partner."
• FQ2 AI bookings >$30B against $10.8B shipped (~3x book-to-billBook-to-billNew orders divided by orders filled. Above 1 means the backlog is growing; below 1 means the company is working through it faster than it is replacing it.); non-AI bookings >$6B; Tan: "Our visibility runs all the way to 2028 right now. Just three months ago… visibility ran pretty much to 2027".

What held. The structural thesis stands and is broader than it was. Five weeks ago the dossier's central fear was single-program reset risk — "the customer owns the design." Since then Broadcom has put Apple through 2031 and Google through up to 2031 on paper, on top of OpenAI (1.3GW contracted for 2027 inside the 10GW/2029 deal, silicon delivered, production late-2026), Meta (3GW MTIA-X through end-2028, initial 1GW ordered, ships H2-2027), Anthropic (~3.5GW from 2027) and two unnamed customers with $6B of POs received. Fabless economics, near-zero capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. ($231M in FQ2 on $22.2B revenue), 62% semi op margin and the software annuity all hold — and software is accelerating, not churning (Q3 guide +31% YoY, ARR +17%), which retires one leg of the prior bear case.

---

Phase A — Understand the Business

Company Overview — REFRESHED (contract structure changed materially)

Unchanged in shape: Semiconductor Solutions ($15,009M FQ2, +79%, 62% segment op margin) + Infrastructure Software ($7,178M, +9%, 79% segment op margin). What changed is the customer book, which is now a portfolio of dated, multi-generation agreements rather than a handful of programs:

Conflict surfaced, not resolved: Tan said on the call Anthropic gets "another 5 gigawatts"; the Apr-6 8-K says "approximately 3.5 gigawatts." The 8-K is the filed document and is the number to underwrite.

A prior-dossier correction. The 2026-07-07 Lens 2 said Broadcom "is creeping up the stack toward systems… supplies, in some cases, racks." Tan, asked directly on the FQ2 call: "No racks. It is all a chip business only. We only chips. Only chips.". The correct read is narrower and more interesting: Broadcom sells chips only, but has built a financing vehicle (the AI XPV Platform, Apollo/Blackstone, $35B first tranche) so that customers who cannot fund racks can still buy the chips. It is not moving up the stack — it is underwriting the demand for its own product, which is a different and more contingent kind of exposure.

Supply Chain — unchanged, see the previous dossier

One datapoint worth carrying forward from the newly-ingested transcript: asked by UBS whether he can get incremental wafers and HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips., Tan said "We are very comfortable that we have been able to secure supply… for our needs for 2026 and 2028. Working on 2028 and 2029 right now," and confirmed customers "have been coming to us incrementally over the last few months" and he can generally serve them. That softens the prior dossier's CoWoS-cap falsifier — management asserts supply is not the binding constraint through 2028. Management assertion, not verification; the falsifier stays live.

Competitive Advantages — REFRESHED (the share moat took a real hit)

The prior dossier's moat leg #1 was "~60-70% custom-AI-ASIC co-design share" with the RPORemaining performance obligationsRevenue a company has already signed contracts for but has not yet delivered or booked. as "the moat made legible." Two of those three supports moved:

1. Co-design lock-in — WEAKENED at the largest customer. Google is explicitly multi-sourcing, confirmed by Tan himself. The reported split assigns Broadcom "Sunfish" (training) and MediaTek "Zebrafish" (inference, at a claimed 20-30% lower cost), with Marvell in talks on memory processing. Broadcom's counter is the Apr-6 LTA's "very substantial dollar amount" commitment — i.e. the defence is now contractual minimums, not design lock-in. That is a materially weaker moat than the one the prior dossier described, and it is the single most important negative finding of this refresh.
2. Backlog — STRENGTHENED and broadened. >$30B AI bookings in one quarter against $10.8B shipped, visibility extended from 2027 to 2028, and the customer set widened to include Apple through 2031. Concentration risk at any one customer fell even as the share at the biggest one fell.
3. Networking — still the strongest leg, and management guided it DOWN as a mix. Tomahawk 6 (100T Ethernet, shipping >1yr) with a 200T switch taping out in FQ2; Jericho 3/4 fabric; de-facto standard in CPO/1.6T DSPs/lasers; 200G-400G SerDes for scale-up copper. But networking was ~40% of AI revenue in FQ2 and Tan said that is "as high as that percentage would go… closer to around 30%" going forward. Since networking is the rich-margin half and XPUs the thin half, the mix is moving the wrong way for margins even as revenue compounds.
4. Software switching costs — STRENGTHENED. VCF 9.1 shipped in the quarter with heterogeneous GPU/CPU support; Q3 software guided +31% YoY on strong on-prem server demand; ARR +17%. The "VMware churn" bear leg looks wrong on the numbers — though see Lens 10: the reason it is growing (tenfold price increases, forced bundling, program closure) is exactly what the EU is investigating.

Bargaining-power read, updated: the prior dossier's asymmetry call — strong over suppliers, weak over customers — is now evidenced rather than inferred. Google exercised it.

Segments — REFRESHED

Two segment-level moves the prior dossier did not have:

• Non-AI semis have turned. $4.2B revenue (+6%) but bookings >$6B — a book-to-bill well above 1 in the cyclical half of the business. Q3 guided +12%. The cyclical recovery is real and is a source of FY27 revenue the AI-only models ignore.
• Software re-accelerated hard. +9% → guided +31%. The mechanism Tan gave: high CPU core counts shipping alongside GPUs drag VCF licences with them — i.e. the software annuity is now levered to the same AI build rather than being the diversifier it was described as. That is good for revenue and bad for the "annuity smooths the cyclicality" argument in the prior dossier's Lens 4.

Phase B — Measure Performance

Latest Earnings, Guidance, Balance Sheet, Market Reaction — REFRESHED

No new print in the window. The latest remains FQ2 FY26 (period ended 2026-05-03, reported 2026-06-03): revenue $22,187M +48%, GAAP net income $9,310M, GAAP diluted EPS $1.91, non-GAAP EPS ~$2.44 vs ~$2.32 est. What the newly-ingested transcript adds to the prior dossier's account:

• FCF $10.3B = 46% of revenue (the prior dossier's ~$10.0B was a back-computed estimate; $10.3B is the CFO's stated figure). Capex $231M in the quarter. Cash $19.6B, up from $14.2B.
• Inventory days 86, up from 68 — the CFO's own framing, "in anticipation of accelerating AI semiconductor growth in the second half." The prior dossier flagged the $4.3B inventory build; the days figure quantifies it. An 18-day build is consistent with a ramp, but it is the number to check against the Q3 print on 2026-09-02.
• Non-GAAP tax rate ~16% for Q3/FY26 on global-minimum-tax and geographic mix; Q3 non-GAAP diluted share count ~4.94B. Both are needed to model FY27 and neither was in the prior dossier.
• Guidance now captured properly on the shelf — five Q3 lines (semi $20.5B, software $8.9B, GM ~74%, op margin ~67%, plus AI $16.0B) and the FY27 10GW shipment plan were appended to our figures this run from the primary transcript.

Balance-sheet flags. Unchanged from the 10-Q and still the right watch list: total debt $64,907M, goodwill+intangibles $126.1B = 70% of a $179.2B balance sheet, quarterly acquisition amortization $1,967M, SBC $2,092M/qtr. New: the June tender retired ~$2.9B of principal — real, but ~4.5% of the stack. Net debt improves to roughly $42B.

Market reaction — the delta that matters. The prior dossier's entry was $373.90 the day after a −12.6% beat-and-fall. Since then: +6.49% on the Apple 8-K (2026-07-07), −3% to $381 on 2026-07-16 in a sector-wide AI-capex-slowdown scare (Marvell −8%, AMD −5%, Intel −5%), +6.1% on 2026-08-04 on Palantir's print as an AI-demand proxy. Net +13.8% to $425.41, ~14% below the $495.00 52-week high set 2026-06-03. The market has re-priced the thesis without a single new financial disclosure.

Date conflict, flagged: a webscrape (moomoo) puts the Q3 print at 2026-08-29 with consensus revenue $29.44B / EPS $2.55. IR's own closing statement on the FQ2 call says "Wednesday, Sept. 2, 2026," and stockanalysis agrees. Use 2026-09-02. The $2.55 EPS consensus also fails an arithmetic check against Broadcom's own guide (see Lens 11) and should be treated as stale or differently-based.

Earnings-Call Sentiment — REFRESHED (now on primary text for the first time)

The prior dossier characterised Tan's tone from press coverage because the transcript could not be scraped. It can now, and the primary text supports the prior read while adding two tells that press coverage lost.

The trend across the last several calls is escalating confidence with a widening time horizon: the headline metric moved from an 18-month backlog figure (>$73B at FQ4-FY25) to a bookings-flow figure (>$30B in one quarter) to an explicit horizon statement — "Our visibility runs all the way to 2028 right now. Just three months ago, I can tell you visibility ran pretty much to 2027." The "insatiable" vocabulary is verbatim and unchanged ("Demand for XPUs and networking is simply insatiable").

Tell #1 — the discipline is real and it is deliberate. Pressed by JPMorgan on whether the 18-month backlog now sits at "$200 billion or better," Tan refused the number: "we are not trying to guide you every quarter what 2027 would be like… it might be, based on what we are doing, very much on track, if not stronger." He reiterated >$100B for FY27 rather than raising it — which is precisely what triggered the −12.6% on June 4. A CEO who declines to raise into a hot tape while conceding a share loss at his largest customer is a CEO whose numbers you can weight more heavily, not less.

Tell #2 — the hedge is now audible. Two concessions in one call that the prior dossier's press-based read did not have: the Google multi-sourcing admission, and the CFO's XPU-margin statement. Both were volunteered under analyst pressure rather than in prepared remarks. What is still conspicuously not volunteered: the RPO split by customer, XPU gross margin as a number, and — most tellingly — the reconciliation between "10GW shipping in 2027" and "$15-20B per gigawatt," which would arithmetically imply $150-200B of FY27 AI revenue against a >$100B guide. TD Cowen asked this directly. Tan answered about content-per-gigawatt trajectory and did not reconcile the two figures. That gap is either the largest unpriced upside in the model or evidence that revenue-per-gigawatt is falling well below the old framing. It is the highest-value question outstanding.

Comps — REFRESHED (the screen inverted)

Two inversions since the last run, and they point opposite ways.

• Against Nvidia, Broadcom got more expensive. The prior dossier concluded AVGO "now arguably [screens] cheaper on forward P/E" (23.8x vs 31.7x). Five weeks later it is 27.15x vs 22.43x — Broadcom re-rated up, Nvidia de-rated down, and the premium is back on both multiples. The prior dossier's relative-value comfort is gone. This is the strongest single support for the active call's "wrong price" leg.
• Against Marvell, Broadcom decoupled violently. MRVL fell 37% in July — cap $272B → $188B — on hyperscaler capex-forecast revisions and "intensifying competition in the custom ASIC market," while AVGO rose. Read carefully: the market did not de-rate the ASIC theme. It de-rated the sub-scale ASIC player and paid up for the scaled one. For a thesis premised on owning the workflow rather than the theme, that is confirmation — and it is also the mechanism by which AVGO's relative-value cushion disappeared. On the pure-ASIC comparison Broadcom now looks cheap (27x vs 47x forward) because Marvell broke, not because Broadcom is cheap.

Catalysts and ±5% Moves — REFRESHED (window: 2026-07-07 → 2026-08-10)

Pattern update. The prior dossier's rule was "up-moves from AI disclosure, down-moves from AI expectations." The window confirms the first half — the only two >5% up-days were a signed 2031 contract and an AI-demand datapoint — and adds a nuance: the down-move on 2026-07-16 was a sector beta event Broadcom absorbed at −3% while Marvell took −8%. Broadcom is now trading as the quality name inside the ASIC complex, which cuts both ways: it dampens drawdowns and it removes the discount.

Forward catalysts. FQ3 FY26 print, 2026-09-02 (AI $16.0B guide, consolidated $29.4B, GM ~74%) — the first read on whether the 86-day inventory build converts and whether FY27 gets raised. Then: OpenAI production start (late-2026), Meta 1GW delivery (H2-2027), the EU VMware decision (no date), and any second tranche of the AI XPV Platform.

Phase C — Judge People & Books

Management — REFRESHED

The CFO transition completed. Kirsten Spears retired 2026-06-12 after 12 years; Amie Thuener — Alphabet's VP/Corporate Controller and Chief Accounting Officer since 2018, ex-PwC, CPA — took the seat, having joined 2026-05-04. Package: $700k base, 100% target bonus, $1M sign-on, 50,000 RSUs + 50,000 PSUs (TSR-based, max 200%); Spears consults to 2027-03-15 with continued vesting capped at target. She signed the June and July 8-Ks — she is operating. Read: hiring Alphabet's chief accountant is a substantive signal in two directions at once — a genuinely strong technical-accounting appointment for a company whose earnings quality is its weakest point, and a hire from the customer whose multi-sourcing is the largest open risk.

Governance amber — ~33.6% against say-on-pay. At the 2026-04-20 annual meeting, executive compensation passed 2.433B for / 1.233B against / 17.5M abstain, and director Harry L. You drew 954.7M against (~26%). A third of the voted shares objecting to pay is a real institutional protest, not noise, and it is new to this dossier.

Capital allocation — one genuine improvement, one escalation.

• Improvement: the upsized $3.0B tender (June 17-18) is the first visible move against the $64.9B debt stack. ~$5.5B was tendered against a $3.0B cap — holders wanted to sell more than Broadcom wanted to buy, which is a mild negative signal on where the notes trade, and the accepted ~$2.9B is only ~4.5% of total debt. Dividend $3.1B in FQ2 at $0.65/share.
• Escalation: the AI XPV Platform — Apollo-led $35B first tranche with Blackstone, >20GW through 2028 — is the prior dossier's $29B Note 11 backstop grown into an institutional programme. Tan's own framing: "creating, in partnership with guys with the best balance sheets around, a vehicle to basically have these chips funded for these LLM players, who otherwise might have difficulty getting access to our technology". That is an explicit statement that a portion of the backlog belongs to customers who cannot pay for it unaided. He also pushed back when Barclays called the Anthropic deal chip-backstopped — "We were the ones providing the compute capacity to Anthropic" — which is a correction that makes it worse, not better: it describes Broadcom supplying compute capacity, not merely selling chips, to a customer whose consumption is contractually conditioned on its own commercial success.

Net read on Tan: unchanged as an allocator of the first rank, now running two levers the prior dossier under-weighted — vendor-adjacent customer financing at $35B scale, and a debt paydown that has started but is small relative to the stack.

Forensic Red Flags — REFRESHED (one flag reversed, one new, one repaired)

Carried from 2026-07-07 and unchanged by anything this run: goodwill+intangibles $126.1B (70% of assets); $1,967M/qtr acquisition-amortization add-back; total debt $64,907M less the ~$2.9B tender; SBC $2,092M/qtr with $20,106M unrecognized; one distributor at 42% of net revenue and top-5 end customers at ~45%; "AI semiconductor revenue" as a management-defined non-GAAP metric [all research-layer: 10-q-2026-q2].

REVERSED — the regulatory clean bill. The prior dossier concluded "No material regulatory or legal findings," citing regulatory/regulatory-findings.md and 10-Q Note 10. That file searches SEC EDGAR LR + AAER only and its own footer instructs the analyst to run a non-SEC web search in Lens 10; the prior run did not. Running it surfaces two live matters:

• EU — formal Commission investigation into VMware access, pricing and contract terms for EU infrastructure customers, following the January-2026 termination of the VMware Cloud Service Provider programme in Europe (effective 2026-03-31) and a CISPE complaint filed 2026-03-19 seeking interim measures; allegations include ~tenfold price increases since the 2023 close, forced bundling and multi-year commitments. Broadcom lost its EU General Court bid to suspend an EU request for related US legal documents.
• US — ITC investigation on Netlist memory patents naming Broadcom together with Google and Nvidia, over memory used in AI servers.

Neither is accrued in Note 10, and both are -sourced only — the research-layer regulatory file has not been re-run and still reports zero findings. That is itself the finding: the tool's scope was mistaken for the answer. The exposure is concentrated in exactly the segment now guided to +31% growth — the EU is investigating the pricing behaviour that produces the software acceleration in Lens 4.

NEW — conditional demand inside the "firmly committed" backlog. The Apr-6 8-K states Anthropic's expanded ~3.5GW consumption "is dependent on Anthropic's continued commercial success". The prior dossier's central hardening claim was that the ~$164.6B RPO represents "firmly-committed, non-cancellable" contracts with "the committed, no-termination-right" character being "the whole ballgame." That claim needs a carve-out: at least one named multi-gigawatt tranche carries an explicit consumption condition tied to the counterparty's commercial performance. The RPO figure itself is a 2026-05-03 GAAP disclosure and stands; the interpretation of it as uniformly unconditional does not.

REPAIRED — the AI-margin assumption. Not a red flag so much as a modelling error the prior dossier committed and the transcript corrects: XPUs/TPUs carry lower gross margins than networking, per the CFO. Any SOTP that applies a premium margin to the AI franchise is wrong in the direction of over-valuation. Repaired in Lens 11.

Phase D — Project & Stress-Test

Forward Projection — REFRESHED, and the prior run's base was arithmetically wrong

Model-workbook status, stated honestly: there is no model.xlsx for Broadcom. our model was not run because no new financials landed this window and, more fundamentally, our figures holds two rows — one quarter (2026-Q2) and one full year (FY2025) — against an eight-quarter 10-Q shelf. A workbook built on that would produce an opening balance sheet with nothing to check it against. Every number below is with arithmetic shown. Filling our figures from the eight 10-Qs already sitting on the shelf is the highest-value shelf task for the next refresh.

First, a correction. The 2026-07-07 dossier put FY26 revenue at "~$66B" and called it "the reconciled base." That is impossible on its own inputs: FQ1 $19.3B + FQ2 $22.187B + FQ3 guide $29.4B = $70.9B in three quarters. The FY26 base is roughly $106B:

H1 actual 19.3 + 22.187 = $41.5B · Q3 guide $29.4B · Q4: FY26 AI $56B − H1 AI ~$19B − Q3 AI $16B = $21B Q4 AI; + non-AI semi ~$4.6B + software ~$9.3B = ~$34.9B → FY26 ≈ $105.8B, +66% on FY25's $63.887B. Cross-check: stockanalysis reports TTM revenue $75.47B as of 2026-08-10; FY25 H2 (~$34.0B) + FY26 H1 ($41.5B) = $75.5B. ✓ The FQ1 $19.3B input is sound.

Non-GAAP EPS bridge, calibrated on the actual quarter: revenue × 67% non-GAAP op margin − ~$0.65B/qtr interest, × 0.84 (16% tax), ÷ diluted shares. Check on FQ2: 22.187 × 0.67 = $14.87B op (transcript says "$14.9 billion" ✓); − 0.65 = 14.22; × 0.84 = $11.94B; ÷ 4.88B = $2.45 vs the reported $2.44. The bridge holds.

At $425.41: base FY27 P/E ~24.9x · bull ~22.4x · bear ~30.4x. The GAAP haircut is severe and must be carried — FQ2 GAAP EPS $1.91 vs non-GAAP $2.44 = GAAP ≈ 78% of non-GAAP, so base FY27 GAAP EPS ≈ $13.35, i.e. ~31.9x GAAP.

Conflict flagged: a webscrape puts Q3 consensus non-GAAP EPS at $2.55. Broadcom's own guide ($29.4B at 67% op margin) arithmetically produces ~$3.24. A 27% gap between a company's guide and a quoted consensus is a data error, not an analytical disagreement — do not underwrite the $2.55.

SOTP, repaired for the XPU-margin correction:

• AI franchise (FY27): ~$100B revenue. The prior run applied ~55% EBITDA margin on the belief AI was the richest slice; the CFO says the opposite for XPUs, and networking drops from 40% → ~30% of the AI mix. Use ~50% → ~$50B AI EBITDA. At a de-rated 18-22x → ~$0.90-1.10T.
• Core semi (non-AI) + software (FY27): ~$19B non-AI semi at ~30% op + ~$35-36B software at 79% op → ~$33B. At 16-18x → ~$0.53-0.60T.
• SOTP EV ≈ $1.43-1.70T vs current EV ≈ $2.06T (cap $2.02T + net debt ~$42B).

Read. The gap between the conservative SOTP and the market widened from roughly 11-26% at the last run to 17-31% now — entirely from price, plus a small amount from the margin repair. The FY28 bridge still exists (base FY28 EPS ~$20.50 at a 22-25x forward multiple ≈ $450-515/sh) but it is a 2027-visibility trade now, not a 2026 one. Jefferies' FY28 EPS of "$30-$40" sits ~50-95% above this bridge's base and cannot be reconciled without either much higher revenue than the 10GW framing implies or a large buyback — surfaced, not adopted.

(Per the unattended-run rules, no our model create was run. The prior run's registered forecast — AVGO FY27 AI ≥ $100B at p=0.62 — is unchanged by this window: the RPO and bookings make delivery more likely, the Google share path makes it less so.)

Bull vs Bear — REFRESHED

Bull. The franchise stopped being a bet on a handful of programmes and became a portfolio of dated contracts: Google through up to 2031, Apple through 2031, Meta 3GW to end-2028, OpenAI 1.3GW contracted for 2027 inside a 10GW/2029 deal, Anthropic ~3.5GW from 2027, two more customers with $6B of POs in hand. Around it: >$30B of AI bookings against $10.8B shipped in a single quarter, visibility pushed from 2027 to 2028, non-AI semis turning (bookings >$6B on $4.2B revenue), software accelerating to +31%, 46% FCF margin, $231M of quarterly capex, and management asserting wafer/HBM supply is secured through 2028. And the market has now shown its hand on how it discriminates: it took 37% out of Marvell in July and left Broadcom up — it is paying for scale and share inside the ASIC theme, which is exactly what this name has.

Bear (impairment vectors, re-weighted).

1. Price is now the dominant risk, not a secondary one. +13.8% in five weeks on zero new financial disclosure; forward P/E above Nvidia's; ~17-31% above a conservative SOTP; the prior seed's whole base-case band consumed.
2. The moat lost its best evidence. Tan conceded Google multi-sourcing on the record. Broadcom's largest customer is deliberately building a second and third source, at a claimed 20-30% cost advantage for the inference part. The defence is contractual minimums — which is a floor, not a moat.
3. The margin mix runs the wrong way. XPUs are the thin-margin half and are the half that scales; networking, the rich half, is guided down from 40% to ~30% of AI revenue. Consolidated GM 77.1% → ~74% in one quarter.
4. Earnings quality is unchanged and the regulatory floor moved. $1.97B/qtr amortization add-back, $64.9B debt (−$2.9B), 42% single-distributor concentration — plus a live EU investigation aimed squarely at the pricing that generates the software growth, and an ITC patent matter.
5. The backlog contains its own financing. $35B Apollo/Blackstone first tranche, >20GW target, plus the $29B Note 11 lease backstop and Anthropic's explicit consumption condition. A meaningful slice of "committed" demand is committed by counterparties whose ability to consume is being underwritten by the seller's ecosystem.

Pre-mortem (18 months out, the thesis broke). It is 2028. FY27 AI landed at ~$88B, not the ">$100B" reiterated all year — not because demand failed but because MediaTek took Google's inference generation and Broadcom's Google share slid toward Macquarie's 65%. The 10GW/2027 plan shipped, but revenue-per-gigawatt came in far below the old $15-20B framing, and the question TD Cowen asked and Tan never answered turned out to have a bad answer. Gross margin ground from 74% toward 70% as XPUs became the majority of AI revenue. The EU landed a VMware remedy that capped the re-pricing, and software growth fell from +31% back toward single digits. Anthropic's consumption condition was invoked. At a forward multiple above Nvidia's going in, a merely-good outcome produced a 30-40% de-rate, and $62B of remaining debt turned a growth story into a deleveraging one.

Contrarian view. The prior dossier's contrarian line was that Broadcom is the riskier of the two AI names, not the safer. This window sharpened it into something more specific: the market has now proved it agrees with the thesis and disagrees only about who gets paid for it. It de-rated the theme's pure-play by 37% and re-rated its incumbent by 14% in the same month. That is not a market that doubts custom silicon — it is a market that has already concentrated the trade into the one name everyone can defend, and priced it accordingly. The uncomfortable implication for a LONG: you are no longer early to a contested idea, you are late to a consensus one, and the entry premium is the price of that consensus.

Devil's Advocate — REFRESHED

Take the two sentences management actually said this quarter and read them together. "We fully expect that there will be some diversity of sources for them." "Our ASICs, TPUs… have lower margins." Now say the bull case out loud: Broadcom will grow AI revenue from $56B to $100B+ by selling more of its lowest-margin product to customers who are actively qualifying its replacements — and the stock trades at a premium to Nvidia to reflect this.

The $164.6B RPO is the bull's shield, so examine it. It is a 2026-05-03 disclosure the prior dossier called "firmly committed, non-cancellable" — and yet the Apr-6 8-K says one named tranche is "dependent on Anthropic's continued commercial success," and Broadcom simultaneously stood up a $35B Apollo/Blackstone vehicle whose stated purpose is funding customers "who otherwise might have difficulty getting access to our technology." A backlog you have to arrange the financing for is a pipeline, not a receivable. Note the direction of travel: last quarter's disclosure was a $29B lease backstop; this quarter it is a >20GW, $35B-first-tranche institutional platform. That line grows faster than the revenue does.

Then the unanswered arithmetic. Tan says 10GW ship in 2027. Tan has said content runs $15-20B per gigawatt. Those two numbers multiply to $150-200B against a >$100B guide. TD Cowen asked; the answer was about content trajectory, not reconciliation. Either the guide is 50-100% conservative — in which case why refuse to raise it into a tape that punished you 12.6% for not raising — or revenue per gigawatt is collapsing, which is exactly what you would expect when your largest customer has just introduced a competitor bidding 20-30% below you. The bulls have chosen the first reading without evidence.

And the software annuity, the thing that supposedly de-risks the whole structure: it is growing 31% because Broadcom raised VMware prices roughly tenfold, forced bundling, and shut hundreds of European service providers out of the programme — conduct now under formal European Commission investigation, with Broadcom having already lost a court fight to withhold documents. The prior dossier recorded "no material regulatory findings." It had only checked SEC enforcement.

The most dangerous competitor is still the customer. But there is now a second: the customer's other supplier, whom the customer introduced on purpose, and whom Broadcom's own CEO has told you to expect.

Management Questions — unchanged, see the previous dossier

The fifteen questions stand. Three are now materially more urgent and should lead the 2026-09-02 call: (Q2) XPU gross/operating margin as a number, now that the CFO has confirmed it is dilutiveDilutionIssuing new shares, so each existing share owns a smaller slice of the same company.; (Q3) the Google renewal-reset question, now that multi-sourcing is conceded rather than hypothesised; (Q4) the recurrence and terms of customer financing, now that it is a $35B institutional platform rather than a single backstop. Add one new: what reconciles 10GW of 2027 shipments with $15-20B of content per gigawatt against a >$100B revenue guide?

Review finding — the ACTIVE MarketCall

Call: MEDIUM · "Broadcom: the right custom-silicon thesis at the wrong price." Finding: the new facts SUPPORT the call, and sharpen both halves of it.

• "The right custom-silicon thesis" — supported and broadened. Two 2031-dated agreements (Apple, Google), 3x book-to-bill on AI, visibility extended to 2028, and a market that de-rated the pure-play ASIC comp 37% while re-rating Broadcom. The thesis is more right than it was.
• "At the wrong price" — supported and strengthened. +13.8% with no new quarter; forward P/E crossed above Nvidia's; the SOTP gap widened from ~11-26% to ~17-31%; the prior dossier's entire base-case target band was consumed in five weeks.
• One qualifier the call should absorb: the reason the price is wrong has partly changed. It was valuation-versus-earnings-quality. It is now valuation-versus-earnings-quality plus a conceded share loss at the largest customer and a margin mix that runs the wrong way. That is a small step toward the falsifier, not away from it — falsifier #3 ("a hyperscaler re-bids… an XPU program") is now partially triggered: Google has introduced competing suppliers, though the LTA's contractual minimums mean no revenue line has reset.

No MarketCall row was read, edited or created. Conviction changes route to our position log, human-gated.