Broadcom (AVGO) Teardown — Q3 FY26 refresh, 8 September 2026

Deep dive · refreshed 8 September 2026 · Q3 FY26 reported 2 Sep

AVGO .

Broadcom Inc. — NASDAQ · Semiconductors & Infrastructure Software · FY ends early November · CEO Hock E. Tan

$369.43

▲ 3.22% · prev close 357.90 · vol 17.9M

Mkt cap $1.76T · EV $1.74T

The setup

KPIs & YoY

Balance sheet

Cost of debt

Projections

Technicals

Ownership

The street

Retail

Competitors

Jalapeño

VMware

SWOT

Scorecard

Sources

◐ Light

52-week range 290–495 39% band

From 52w high −25.4%

Q3 print reaction −2.7% low −6.8%

P/E fy26e · fy27e 31.7× · 19.1×

Consensus PT $533 +44%

FY27 AI rev guide $115B +98%

FCF ttm $30.5B

RSI(14) 38.0 weak

Inst · insider 89% · 1.95%

The setup, six days after the print

Q3 FY26 release & call, 2 Sep 2026 Stocklake · Alpha Vantage · Stocktwits · sell-side notes

Broadcom beat, raised every AI number it has, put a $230 billion figure on fiscal 2028 — and the stock fell. Then, over three sessions, it clawed most of it back. The report below is the 25 August teardown re-cut around what the print actually settled and what it left open.

The quarter itself was not close. Revenue of $29.59B (+86%) against a ~$29.4B guide; AI semiconductor revenue $16.7B (+221%) against a $16.0B guide; non-GAAP EPS $3.32 against $3.24 consensus; non-GAAP operating margin a record 67.9% ; free cash flow $13.7B, 46% of revenue . Broadcom repaid $5.6B of debt in the quarter and ended with $24.0B of cash.

The guidance did the thing the June print refused to do. FY2026 AI revenue was raised from $56B to $58B . The FY2027 "in excess of $100 billion" became a specific ~$115 billion . And Hock Tan added a year: "line of sight" to $230 billion in FY2028 — "$350 billion of AI semiconductors to these customers in the next two years," with supply "secured" for both. Anthropic alone is now 1 GW in 2026, 5 GW in 2027 and 10 GW in 2028; OpenAI is 1.3 GW in 2027 and "over 5 GW" in 2028; Meta 3 GW through 2028.

Why it sold off anyway. Three things. The Q4 revenue guide of ~$34.8B sat a hair under the $35.0B consensus and the $21.7B AI guide "ticked up but came in slightly below" the street. Infrastructure software printed +29% against its own +31% guide — a ~$150M miss with no explanation offered, and a Q4 guide decelerating to +25%. And when Melius asked for the maximum off-balance-sheet exposure to the XPV financing vehicles, the CFO's answer was "I cannot give you an overarching look at what is the max." That is the sentence the credit market heard.

Then it recovered. After a $342 intraday low on 3 September the stock closed −2.7%, went flat on the 4th, and is +3.2% today at $369 — back on top of the 200-day. Nine houses raised targets, Macquarie upgraded, and three cut (Truist, TD Cowen, DA Davidson to a street-low $350). Consensus FY27 EPS moved down 1% on the print, not up.

What is left is a cleaner argument than the one in August. The demand and the supply are no longer the question — management put numbers on both. The question is now entirely about who pays for the 20 gigawatts and what Broadcom is on the hook for if the two labs cannot.

What the print settled

FY27 AI is ~$115B , not ">$100B" — the number the street was already at. FY28 $230B is new. FY26 AI raised to $58B. Q3 AI beat by $0.7B. Gross margin held at 75.0% vs a ~74% guide. Interest-bearing debt is down to $59.4B with a disclosed 4.0% weighted coupon.

What the print left open

Software missed its own guide and decelerates again in Q4 (+25%). No backlog, bookings or book-to-bill figure was disclosed for a second straight quarter. No maximum XPV exposure was given — only "modest residual value guarantees… contingent liabilities we view as low risk." The reported >$60B second SPV was neither confirmed nor denied.

The number the market is now arguing about

Bernstein's Rasgon, on the call, worked Tan's gigawatts back to ~$11–12B per GW. Tan's answer: "$20 billion–$30 billion… just accept the fact that there is going to be a lot more gigawatts." Whether FY28 is $230B or something closer to half that depends on which of those two numbers is right.

KPIs and three years of YoY

Alpha Vantage income & cash-flow statements SEC-normalised GAAP · Broadcom IR releases

Fiscal years end the Sunday closest to 31 October. FY2024 is the distorted year — VMware closed 22 November 2023, dragging in $21.5B of software revenue plus purchase accounting, amortisation and a one-time $4.5B discrete tax item that pushed GAAP net income down 58% on a 44% revenue gain.

Revenue TTM $89.1B Q3 FY26 $29.6B, +86% YoY

Gross margin Q3 75.0% Non-GAAP · GAAP 69% · Q4 guide ~73%

Op margin Q3 67.9% Non-GAAP, +240bps YoY — a record

FCF TTM $30.5B Q3 $13.7B — 46% of revenue

Non-GAAP EPS Q3 $3.32 +96% YoY · vs $3.24 consensus

ROE TTM 44.2% Net margin 42.9% TTM

Net debt $35.4B $59.4B debt vs $24.0B cash · 1.2× TTM FCF

Intangibles / assets ~66% Tangible book/share ≈ −$5.0 (VMware)

Source: Alpha Vantage normalised income statement and cash-flow statement (GAAP, as reported to SEC). FCF = operating cash flow less capex. FY2024 net income reflects VMware purchase accounting and a discrete tax charge.

Fiscal year Revenue $M YoY Gross % Op inc $M Op %

Net inc $M YoY R&D $M R&D % EBITDA $M FCF $M FCF %

FY2021 27,450 — 61.4 8,519 31.0 6,736 — 4,854 17.7 14,691 13,321 48.5

FY2022 33,203 +21.0% 66.5 14,225 42.8 11,495 +70.7% 4,919 14.8 19,155 16,312 49.1

FY2023 35,819 +7.9% 68.9 16,207 45.2 14,082 +22.5% 5,253 14.7 20,554 17,633 49.2

FY2024 VMware 51,574 +44.0% 63.0 13,463 26.1 5,895 −58.1% 9,310 18.1 23,879 19,414 37.6

FY2025 63,887 +23.9% 67.8 25,484 39.9 23,126 +292.3% 10,977 17.2 34,714 26,914 42.1

FY2026e consensus 106,040 +66.0% ~74 — ~67 n-GAAP — — — — — — —

Revenue and the shape of the acceleration

Fiscal year revenue $B · YoY % direct-labelled

Two step-functions, not one trend: FY2024 is the VMware acquisition, FY2026 is the AI silicon ramp. The FY2026 bar is consensus, not reported.

Margin structure through the two shocks

GAAP gross, operating and FCF margin · % of revenue

VMware compressed reported margins for one year, then expanded them permanently — infrastructure software runs at 93% gross and ~79% operating margin. The forward risk is the opposite direction: XPU mix.

Quarterly detail — where the inflection actually happens

Quarterly revenue by segment

$B · Q1 FY24 → Q4 FY26 · pale bar is company guidance

Semiconductor Solutions has run $9.2B → $11.1B → $12.5B → $15.0B → $20.8B over five quarters and is guided to ~$26.1B. Software moved $6.79B → $8.75B and is guided flat at $8.7B. Its share of revenue has gone 45% → 30% → 25% guided — diluted, and now also decelerating.

AI semiconductor revenue ramp

$B per quarter · YoY % direct-labelled · Q4 is guidance

Q3 landed at $16.7B against a $16.0B guide — 73% XPU (~$12.2B, shipments up 3.5× YoY) and 27% networking (~$4.5B, up 2.5×). Q4 is guided to $21.7B, which sums the year to $57.6B against the raised $58B target. The step from here to ~$115B in FY27 requires the quarterly run-rate to reach ~$29B by late 2027.

AI semiconductor revenue and infrastructure software revenue are company-disclosed non-GAAP operating metrics from quarterly releases and earnings calls; Broadcom does not break out VMware separately within Infrastructure Software. Q3 FY26 is the reported quarter (2 Sep 2026); Q4 FY26 is management guidance given on that call. Q3 FY26 vs its own 3-June guide: revenue $29.59B vs ~$29.4B, AI $16.7B vs $16.0B, software $8.75B vs ~$8.9B.

Quarter Total rev $M YoY QoQ AI semis $B AI YoY Infra software $M SW YoY SW % of rev

Q1 FY25 14,916 +24.7% +6.1% 4.1 — 6,703 — 44.9%

Q2 FY25 15,004 +20.2% +0.6% 4.4 — 6,596 — 44.0%

Q3 FY25 15,952 +22.0% +6.3% 5.2 — 6,786 — 42.5%

Q4 FY25 18,015 +28.2% +12.9% 6.5 — 6,943 — 38.5%

Q1 FY26 19,311 +29.5% +7.2% 8.4 +105% 6,796 +1.4% 35.2%

Q2 FY26 22,187 +47.9% +14.9% 10.8 +143% 7,178 +8.8% 32.4%

Q3 FY26 actual 29,591 +85.5% +33.4% 16.7 +221% 8,752 +29.0% 29.6%

Q4 FY26 guide 34,800 +93.2% +17.6% 21.7 +236% 8,700 +25.3% 25.0%

The software line, checked

In August this table said the +31% Q3 software guide was "the second number to check on 2 September." It came in at +29.0% ($8.75B) — a ~$150M miss against Broadcom's own guide, still a $1.57B sequential step-up and the biggest software quarter on record, but the company offered no explanation for the shortfall on the call. Q4 is guided to $8.7B, +25% , sequentially flat-to-down. ARR growth was disclosed at 15% , versus 17% a quarter ago. The renewal-cohort question this raised in August is now sharper, not resolved.

Capital returns and the balance sheet

Dividends FY25 $11.1B $2.60/sh, 15th straight raise; 0.72% yield

Operating CF FY25 $27.5B vs $20.0B FY24, +38%

Capex FY25 $0.62B 1.0% of revenue — fabless, asset-light

Payout ratio 40.1% Of TTM earnings; 41% of FCF

Capex at 1–2% of revenue is the quiet strength of the model — Broadcom captures AI infrastructure economics without funding fabs, though Q4 capex is guided to $1.4B and a Singapore substrate fab starts in FY27, so that ratio is drifting up. The offset is that roughly two-thirds of the asset base is intangible and tangible book value per share is negative, so the balance sheet gives no downside support. Net debt of $35.4B against $30.5B of trailing FCF is comfortable today ; the XPV vehicles described in the Threats section are what make that number worth re-checking each quarter.

Balance sheet, three years of change

Alpha Vantage / SEC-normalised balance sheet FY22–FY25 Q3 FY26 from the 2 Sep release; 10-Q not yet filed

Total assets went from $73B to $166B in a single year and have barely moved since. Everything interesting on this balance sheet is either the VMware crater or the working-capital build under the AI ramp — and the Q3 print materially eased the second worry.

Total assets $188.1B Q3 FY26 · +10.0% vs FY25 close

Shareholders' equity $99.7B 53.0% of assets, up from 31% in FY22

Net debt $35.4B $59.4B debt less $24.0B cash · −$9.8B in a quarter

Net debt / TTM FCF 1.16× Debt / equity 0.60× — was 1.74× in FY22

Current ratio 2.50× FY24 trough 1.17× · Q2 FY26 2.24×

Goodwill + intangibles ~66% ~$124B of $188B · intangibles est. from Q2 less amortisation

Tangible book value ≈ −$24B ≈ −$5.0 per share (est.)

Days sales outstanding 42.3 ▼ from 68.9 in Q2 — see note

SEC-normalised balance sheet via Alpha Vantage, $M, through FY2025. Q3 FY2026 (2 Aug 2026) is from the condensed balance sheet in the 2 Sep 2026 earnings release; the 10-Q had not been filed at time of writing, so accounts payable, retained earnings and the intangibles/goodwill split are not yet available — intangibles are estimated as Q2's $28,333M less the quarter's $2,006M amortisation. Net debt / EBITDA is not shown for Q3 because Broadcom did not disclose adjusted EBITDA this quarter. "vs FY25" compares to the FY2025 year-end, not to a year-ago quarter.

$M FY2022 FY2023 YoY FY2024 YoY FY2025 YoY Q3 FY26 vs FY25

Total assets 73,249 72,861 −0.5% 165,645 +127.3% 171,092 +3.3% 188,148 +10.0%

Cash & equivalents 12,416 14,189 +14.3% 9,348 −34.1% 16,178 +73.1% 23,975 +48.2%

Receivables, net 2,958 3,154 +6.6% 6,332 +100.8% 12,150 +91.9% 13,707 +12.8%

Inventory 1,925 1,898 −1.4% 1,760 −7.3% 2,270 +29.0% 4,523 +99.3%

Total current assets 18,504 20,847 +12.7% 19,595 −6.0% 31,573 +61.1% 52,173 +65.2%

Goodwill 43,614 43,653 +0.1% 97,873 +124.2% 97,801 −0.1% 97,801 0.0%

Intangibles ex-goodwill 7,111 3,867 −45.6% 40,583 +949.5% 32,273 −20.5% ~26,300 ~−18%

Total non-current assets 54,745 52,014 −5.0% 146,050 +180.8% 139,519 −4.5% 135,975 −2.5%

Total liabilities 50,540 48,873 −3.3% 97,967 +100.5% 89,800 −8.3% 88,458 −1.5%

Accounts payable 998 1,210 +21.2% 1,662 +37.4% 1,560 −6.1% n/d —

Short-term debt 440 1,608 +265.5% 1,271 −21.0% 3,152 +148.0% 2,252 −28.6%

Total debt (carrying) 39,515 39,229 −0.7% 67,566 +72.2% 65,136 −3.6% 59,419 −8.8%

Total current liabilities 7,052 7,405 +5.0% 16,697 +125.5% 18,514 +10.9% 20,838 +12.6%

Shareholders' equity 22,709 23,988 +5.6% 67,678 +182.1% 81,292 +20.1% 99,690 +22.6%

Retained earnings 1,604 2,682 +67.2% 0 −100% 9,761 n/m n/d —

Shares outstanding (M) 4,230 4,270 +0.9% 4,778 +11.9% 4,853 +1.6% 4,887 +0.7%

Net debt 27,099 25,040 −7.6% 58,218 +132.5% 48,958 −15.9% 35,444 −27.6%

Net debt / EBITDA 1.41× 1.22× — 2.44× — 1.41× — n/d —

Tangible book value −28,016 −23,532 — −70,778 — −48,782 — ≈ −24,400 —

Equity as % of assets 31.0% 32.9% — 40.9% — 47.5% — 53.0% —

What the balance sheet is made of

Asset composition, $B · goodwill and intangibles vs everything else

Goodwill

Intangibles ex-GW

Cash

Receivables + inventory

Other assets

VMware added $54.2B of goodwill and $36.7B of intangibles in one year. Intangibles ex-goodwill have amortised from $40.6B to roughly $26B — non-cash charges that depress GAAP earnings while cash flow is unaffected. Q3's $9B asset jump is almost entirely cash (+$4.3B) and "other" current assets; goodwill is unchanged and unimpaired. Q3 intangibles are estimated pending the 10-Q.

Deleveraging after the acquisition

Total debt and shareholders' equity, $B

Through Q2 the deleveraging was all equity accretion. Q3 changed that: Broadcom repaid $5.6B of debt in the quarter and another $1.5B at maturity after quarter-end, taking carrying debt from $64.9B to $59.4B while equity rose $12.0B. Debt-to-equity has gone 1.74× → 1.64× → 1.00× → 0.80× → 0.74× → 0.60× . No buybacks.

Working capital: the stretch reversed in Q3

Days sales outstanding and days inventory · quarter revenue/COGS annualised

DSO went 30 → 31 → 41 → 62 → 69 days from FY22 through Q2 FY26 — and then dropped to 42 days in Q3 as receivables fell $3.0B while quarterly revenue rose $7.4B. Inventory days eased from 54 to 45 even as inventory dollars rose. Both computed on annualised quarterly revenue and COGS, the fairer method for a business growing this fast.

The line this report said to watch — and what it did

In August this note flagged receivables at $16.7B and DSO at 69 days , more than double FY22, as the number that would decide whether the working-capital build was a growth build or a collection problem. Q3 answered it: receivables fell to $13.7B on a quarter where revenue rose 33% sequentially, taking DSO to 42 days , and operating cash flow of $14.2B converted 48% of revenue. That is the growth-build reading, confirmed with cash. Inventory is the one line still climbing ($4.5B, +99% vs FY25 year-end), which is what a supply-secured ramp into $21.7B of Q4 AI shipments should look like. Keep watching it; stop worrying about receivables.

What the deleveraging does and doesn't tell you

Net debt of $35.4B against $30.5B of trailing free cash flow is comfortable by any measure, and Q3's $5.6B paydown makes it more so. But it measures on-balance-sheet debt only. The Apollo/Blackstone XPV platform sits outside it, and on the Q3 call the CFO confirmed Broadcom "may provide modest residual value guarantees, which are contingent liabilities we view as low risk" — while declining to size the maximum. S&P treats that support as a contingent, debt-like obligation added to adjusted debt . The reported leverage ratio and the rating-agency leverage ratio are still diverging; the Q3 10-Q's guarantee footnote is the next disclosure that narrows it.

Cost of debt: what Broadcom pays to borrow

SEC FWP pricing term sheets · 8-K · 10-K debt footnotes Moody's · S&P · Fitch · press reporting for the 2026 SPVs

Twenty-two tranches priced across six offerings since July 2024. The coupon has bounced between 4.15% and 5.70% — but almost all of that is the Treasury curve. Broadcom's own credit spread has compressed by roughly a third, and it tracks the ratings ladder almost exactly. On 2 September the CFO put a number on the blended book for the first time: 4.0% on $59.6B of fixed-rate principal, 7.4 years to maturity — and Broadcom retired $5.6B of it in the quarter.

Total debt $59.4B Q3 FY26 carrying · −$5.6B repaid in Q3, −$1.5B since

Wtd-avg coupon 4.0% CFO-disclosed 2 Sep on $59.6B principal · 7.4 yrs to maturity

Effective cost of debt ~4.85% Interest expense ÷ avg debt, FY26 H1 ann.

Interest expense $1.58B FY26 H1 · −4.0% vs $1.64B H1 FY25

5yr new-issue spread +60bp Jan 2026 · was +85bp Jul 2024

Ratings A3/A− Moody's / S&P · Fitch BBB+ · all Positive

5yr CDS ~122bp Record, 20 Aug 26 · +28bp in August

Undisclosed SPV debt $60–100B Sought, not priced · off balance sheet

Every coupon Broadcom has printed since 2022

Coupon % by pricing date · colour = tranche tenor at issue

3yr

5yr

7yr

10yr+

Twenty-two tranches. The January 2026 30-year at 5.700% is the highest coupon Broadcom has ever printed — and it is a deliberate duration extension, the first long bond in years, not a sign of deteriorating access.

Separating the Treasury from the credit

The 5-year tranche of each deal: coupon vs benchmark Treasury vs spread

Coupon %

Benchmark UST %

Credit spread (bp, right-hand labels)

Same issuer, same tenor, 85bp of coupon difference between Jan 2025 (5.050%) and Sep 2025 (4.200%) — of which only about 10bp is credit. The spread went +85 → +83 → +63 → +68 → +53 → +60bp while the rating went BBB → BBB+ → A−. Coupon is a rates story; spread is the Broadcom story.

Every offering, tranche by tranche

Compiled from Broadcom's SEC free-writing-prospectus pricing term sheets, the November 2023 8-K, and 10-K debt footnotes. Benchmark Treasury yields are derived as (reoffer yield − spread) from each FWP. The April 2022 deal was 144A with no FWP filed, so spreads are unavailable. There were no Broadcom senior notes issued in calendar 2023 — the VMware cash consideration was funded entirely by term loans, and VMware's notes were assumed rather than exchanged.

Priced Tranche Size $M Coupon Yield Spread Bmk UST

31 Mar 2022 144A Apr 2029 750 4.000% n/d n/d —

31 Mar 2022 Apr 2032 1,200 4.150% n/d n/d —

Apr 2022 exchange May 2037 2,500 4.926% n/d +215 —

22 Nov 2023 — VMware close. Drew the full $28,390,625,000 term loan: A-2 $9,195M @ SOFR+1.125% (7.11% eff.), A-3 $11,195M @ SOFR+1.250% (7.02%), A-5 $8,000M @ SOFR+1.625% (7.22%). Assumed $8,250M of VMware notes with coupons of 1.000%–4.700%.

8 Jul 2024 Jul 2027 1,250 5.050% 5.056% +65 4.406%

8 Jul 2024 Jul 2029 2,250 5.050% 5.080% +85 4.230%

8 Jul 2024 Nov 2031 1,500 5.150% 5.178% +95 4.228%

26 Sep 2024 Feb 2028 875 4.150% 4.198% +65 3.548%

26 Sep 2024 Feb 2030 1,500 4.350% 4.388% +83 3.558%

26 Sep 2024 Feb 2032 875 4.550% 4.593% +93 3.663%

26 Sep 2024 Oct 2034 1,750 4.800% 4.826% +103 3.796%

6 Jan 2025 Apr 2028 1,100 4.800% 4.847% +53 4.317%

6 Jan 2025 Apr 2030 800 5.050% 5.052% +63 4.422%

6 Jan 2025 Apr 2032 1,100 5.200% 5.221% +70 4.521%

7 Jul 2025 Jul 2030 1,750 4.600% 4.647% +68 3.967%

7 Jul 2025 Jul 2032 1,750 4.900% 4.946% +78 4.166%

7 Jul 2025 Jul 2035 2,500 5.200% 5.247% +85 4.397%

22 Sep 2025 Oct 2030 1,000 4.200% 4.231% +53 3.701%

22 Sep 2025 Feb 2036 2,250 4.800% 4.827% +68 4.147%

22 Sep 2025 Feb 2038 1,750 4.900% 4.927% +78 4.147%

6 Jan 2026 Jan 2031 750 4.300% 4.322% +60 3.722%

6 Jan 2026 Jan 2033 1,250 4.600% 4.641% +70 3.941%

6 Jan 2026 Jan 2036 1,250 4.950% 4.961% +78 4.181%

6 Jan 2026 Jan 2056 30yr 1,250 5.700% 5.747% +88 4.867%

Ratings history. Broadcom went from BBB− at the VMware announcement to A−/A3 in under three years — an unusually fast two-notch climb for a company that added $28B of acquisition debt en route.

Date Agency Action Outlook

Nov 2023 S&P BBB− → BBB on VMware close

15 Jan 2025 S&P BBB → BBB+ Positive

17 Sep 2025 S&P BBB+ → A− Positive

22 Sep 2025 Moody's Baa1 → A3 Positive

22 Sep 2025 Fitch BBB → BBB+ Positive

6 Jan 2026 S&P / Fitch New-issue A− / BBB+ "leverage neutral"

11 Jun 2026 S&P XPV tranche "credit negative" rating unchanged

Why the agencies moved

Moody's cited debt/EBITDA falling below 1.5× from ~3× at FY2024, FCF going $17B → $29B, and $10.7B cash plus a $7.5B revolver. S&P cited AI at 57% of semiconductor sales in Q3 FY25 versus 41% in FY24, leverage 2.0× → 1.2× → ~0.8×, and said plainly that it had been "too conservative over the past two years."

Interest expense is falling while debt is flat

Interest expense by fiscal year, $B · FY2026 is H1 annualised

Peak quarterly interest expense was $1,047M in Q2 FY2024 ; the run-rate is now ~$780–800M. That ~25% reduction is the term-loan paydown plus the refinancing of floating SOFR+ debt at ~7% effective into 4.2–5.2% fixed notes — the single clearest financial benefit of the 2024–26 issuance programme.

Maturity wall, termed out

Principal maturities at FY2024 close vs FY2025 close, $B

The FY2027 bucket fell from $11.7B to $2.5B in a single year while "thereafter" grew from $36.8B to $43.3B. A further ~$2.9B of the highest-coupon paper (the 4.926% 2037s, 4.900% 2038s and 5.050%–5.200% 2030–32s) was retired in the June 2026 tender.

Where the cost-of-debt story stops being reassuring

Everything above describes an issuer whose access is improving — and Q3 improved it further, with $5.6B repaid and no new issuance. The two things that still do not appear anywhere in it are the Apollo/Blackstone XPV platform ($35B first tranche closed in June for Anthropic's 1 GW; >20 GW targeted through 2028) and the >$60B second SPV reported in August. The 2 September call was asked directly. What it gave: the platform lets "sophisticated third-party financial partners… independently underwrite and capitalize the assets rather than providing the direct financing ourselves," and "where necessary, we may provide modest residual value guarantees , which are contingent liabilities we view as low risk." What it withheld, when Melius asked for the maximum: "I cannot give you an overarching look at what is the max… we are going to evaluate any strategic financing really on a deal-by-deal basis." Asked whether the 10 GW Anthropic and 5 GW OpenAI 2028 builds would all be XPV-financed, Tan: "Anthropic is well on its way to an IPO… its investment credit will change. OpenAI… it's different." BofA's credit desk had already moved Broadcom's bonds to Marketweight on 11 August, citing a $29B maximum loss exposure on the first tranche and up to $370B of XPV senior debt by mid-2029. Neither vehicle will ever produce a Broadcom pricing term sheet — the FWP history stops at 6 January 2026 — so the Q3 10-Q's guarantee footnote is the only forthcoming primary disclosure. Reported cost of debt: best since before VMware. Market-implied cost of credit: still the worst of the cycle, and the call did not close the gap.

Earnings and revenue projections

Alpha Vantage consensus (43–48 analysts) Broadcom management guidance · Bigdata.com estimates

Consensus has Broadcom nearly tripling revenue in two fiscal years — $63.9B in FY25 to $173.4B in FY27 — and management has now put its own numbers under that: ~$115B of FY27 AI revenue and "line of sight" to $230B in FY28. On consensus the stock trades at 19.1× FY27 earnings. On Cantor's FY28 EPS of $35, it trades at 10.6×.

Alpha Vantage consensus, retrieved 8 Sep 2026 (post-print). EPS is non-GAAP diluted. FY2025 actual non-GAAP EPS $6.82. Q3 FY26 row shows the pre-print consensus against the reported actual. Company guidance for Q4 FY26: revenue ~$34.8B, non-GAAP GM ~73%, op margin ~66%, ~4.94B shares.

Period Rev est YoY EPS est YoY Low–High EPS #

Q3 FY26 actual $29.59B vs 29.43 +86% $3.32 vs 3.24 +96% beat +2.5% 38

Q4 FY26 guide $34.8B $34.88B +94% $3.827 +96% 3.08 – 3.91 38

Q1 FY27 $37.16B +92% $4.135 +84% 3.61 – 4.55 34

FY2026 $105.97B +65.9% $11.638 +70.6% 11.35 – 11.73 43

FY2027 $173.43B +63.7% $19.372 +66.5% 17.47 – 22.44 45

FY26e P/E 31.7× On $11.64 consensus EPS

FY27e P/E 19.1× On $19.37 — below the S&P forward

FY26e EV/Sales 16.4× Trailing 19.5×

FY27e EV/Sales 10.0× PEG on FY27 growth: 0.48

Estimate revisions through the print

FY2027 consensus EPS went $19.57 → $19.37 (−1.0%) in the seven days across the report, even as six analysts revised up and one down in the trailing 30. FY2026 was unchanged at $11.64. Read that carefully: the company raised every AI number it has and the street's FY27 EPS went down , because the Q4 revenue guide was ~$230M light of consensus and the ~73% gross-margin guide is below where most models sat. Growth is not the debate; mix and margin are.

Management's own ladder — and the gap to the street

AI semiconductor revenue: guidance versus consensus

$B per fiscal year · company figures from the 2 Sep 2026 call · pre-print street range on FY27

On 2 September the FY27 figure went from "in excess of $100 billion" to "approximately $115 billion" — roughly where the street already was — and FY28 appeared for the first time at $230 billion , with "secured the supply" attached to both. That is a doubling, then a doubling. Bernstein's back-of-envelope on the call: ~10 GW in 2027, ~20 GW in 2028.

Guidance on the record — updated 2 Sep 2026

FY2025 actual: ~$20.2B AI revenue, +65% YoY.

FY2026: raised to $58B (+186%) from $56B. Q4 guided to $21.7B (+236%).

FY2027: "approximately $115 billion" — "we have secured the supply to again double AI revenue." Up from "in excess of $100 billion."

FY2028: "line of sight… to again double to $230 billion. Here again, we have secured the supply." To Rasgon: "$350 billion of AI semiconductors to these customers in the next two years."

Demand vs outlook: "Our demand actually exceeds this outlook, and we will work to improve supply in 2028." Moore (MS): "we try to be conservative… we can ship significantly more."

Backlog: for a second consecutive quarter, no backlog, bookings or book-to-bill figure was given. The last hard number remains $73B AI backlog at Q4 FY25.

Gross margin: Q3 printed 75.0% vs ~74% guide; Q4 guided to ~73% . Tan: "Stop focusing on gross margin… Look at operating margin… we expect to be able to sustain operating margin even as mix of products dilutes the gross margin." Q4 op margin guided ~66%.

$ per gigawatt: Tan — "will remain in the $20 billion–$30 billion level… content in dollars is relatively stable."

The six XPU/AI customers and disclosed deployment scale as of the 2 Sep 2026 call. Gigawatt figures are Hock Tan's on-call statements. No seventh customer was announced. Tan on the mix: "The other four customers are financially secure… With these two guys, Anthropic and OpenAI… every gigawatt of compute they deploy, they could achieve $30 billion of ARR."

Customer 2026 2027 2028

Google TPU v7 Ironwood high volume; v8i in production "multi-tens of billions annually" successive generations; no GW given

Anthropic 1 GW Ironwood 5 GW TPU v8i +10 GW "just for Anthropic"

OpenAI Jalapeño shipped, small volume 1.3 GW "on track" >5 GW

Meta MTIA, three generations through end-2027 3 GW cumulative

Two undisclosed ships from late 2026 — —

Technical picture

Stocklake daily OHLCV to 4 Sep 2026, 365 bars Indicators computed from close · spot 8 Sep intraday

The print produced a textbook capitulation candle — a $342 intraday low on 3 September on 2.4× average volume, closing well off it — and three sessions later price is back on top of the 200-day at $369. Momentum is still negative and the 20- and 50-day sit overhead at $378 and $384; the April golden cross remains intact.

AVGO daily close with 50- and 200-day moving averages

8 Sep 2025 – 4 Sep 2026 · 251 sessions · hover for values

Close

SMA 50

SMA 200

Event

Marked events, left to right: A 5 Sep 25, 4th XPU customer >$10B (+9.4%) · B 13 Oct 25, OpenAI 10 GW pact (+9.9%) · C 12 Dec 25, Q4 beat but GM warning (−11.4%) · D 6 Apr 26, Google TPU-through-2031 8-K (+6.2%) · E 2 Jun 26, all-time closing high $480.81 · F 4 Jun 26, Q2 print, Q3 AI guide light (−12.6%) · G 10 Jun 26, Apollo/Blackstone $35B XPV vehicle (−5.1%) · H 14 Aug 26, BofA models $370B XPV debt (−5.9%) · I 19 Aug 26, Marvell wins Google custom-chip warrant (−4.6%) · J 3 Sep 26, Q3 FY26 print — beat, FY27 $115B, FY28 $230B; −2.7% close, −6.8% intraday low.

RSI(14) and MACD histogram

Momentum · 12 months

RSI has lifted from 35 to 38 through the print without ever reaching the sub-30 washout that marked the September 2025 and March 2026 lows. The MACD histogram has narrowed from −6.2 in late August to −1.4 — momentum is no longer deteriorating, but it has not turned positive either.

Computed from Stocklake daily closes to 4 Sep 2026 and cross-checked against Stocklake's indicator block on 8 Sep (RSI 38.0, SMA50 383.66, SMA200 369.03). Spot is the 8 Sep intraday print at $369.43. Annualised volatility from log returns.

Indicator Value Reading

Last close (4 Sep) $357.89 Spot $369.43 (+3.2%)

SMA 20 $378.32 Spot −2.4%

SMA 50 $383.66 Spot −3.7%

SMA 200 $369.03 Spot +0.1% — reclaimed today

50/200 cross 17 Apr 26 Golden, still intact

RSI(14) 38.0 Weak, lifting from 35

MACD − signal −1.39 Narrowing from −6.2

Bollinger (20,2) 332.55 / 424.09 Mid-band $378

ATR(20) $12.53 3.4% of price / day

Realised vol 12m 47.5% 30-day 40.1%

Print-day range $342.33 – $359.40 Vol 2.4× average

Beta 1.46 vs S&P 500

Levels that matter

$342 — the 3 September print-day low, on the heaviest volume of the year. This is now the level. A close below it invalidates the "capitulation" read.

$369 — the 200-day, reclaimed intraday today after eleven sessions below it. Holding it on a close is the bulls' first job.

$378 / $384 — the 20- and 50-day, stacked overhead. Clearing both on volume turns the structure; a rally into them without volume is a lower high.

$401 — the late-August pre-selloff high; also the 10 July insider-sale print.

$480.81 / $495.00 — the all-time closing high (2 Jun) and intraday 52-week high.

$289.96 — the 52-week low, 30 Mar 2026. The September 2025 low has rolled out of the window.

The honest technical read

The event this section was waiting for has happened and the chart did something specific: it sold the news hard intraday, closed the gap, and reclaimed the 200-day within three sessions. That is the behaviour of a stock whose sellers have exhausted themselves on a known catalyst, not one starting a new leg down. It is not yet a reversal — that needs a close above $384 on volume — but the risk/reward at $369 against a defined $342 stop is a different proposition from the open-ended drawdown of late August. Realised vol is falling (40% 30-day vs 48% a month ago), which is the tape telling you the same thing.

Ownership: who is selling, who is buying

SEC Form 4 via Alpha Vantage, 12 months to 25 Aug 2026 — no new filings since SEC Form 13F, 5,185 filers, latest 30 Jun 2026 · next 13F window mid-Nov

Insiders sold $1.20 billion of stock in the last twelve months and bought $1.9 million. Institutions added on net into the drawdown. Those two facts point in opposite directions and both are worth understanding precisely.

Insider activity — Form 4

Open-market sells $1.20B 339 transactions, 3.44M shares, avg $348.67

Open-market buys $1.93M 5 transactions, 5,550 shares — one director

Sell : buy ratio 622:1 By dollar value

Insider ownership 1.95% ~92.7M shares of 4.76B outstanding

SEC Form 4 open-market dispositions with a stated price, 25 Aug 2025 – 25 Aug 2026. Excludes 926,724 shares of zero-price disposals (tax withholding, gifts, trust transfers) and 571,430 shares of zero-price acquisitions (RSU vests and grants). Most of these sales run through Rule 10b5-1 plans.

Insider Role Shares Value Avg px

Samueli, Henry Director / co-founder 2,125,321 $752.9M $354.26

Tan, Hock E. President & CEO 696,668 $235.1M $337.46

Brazeal, Mark D. Chief Legal & Corp Affairs 312,767 $109.5M $349.98

Spears, Kirsten M. CFO (to 2026) 108,566 $36.3M $333.99

Velaga, S. Ram President, ISG 102,594 $34.4M $335.51

Kawwas, Charlie B. President, Semi Solutions 84,394 $27.7M $328.56

Page, Justine Director 6,820 $2.4M $356.86

Delly, Gayla J. Director 2,890 $1.1M $376.01

Low, Check Kian Director 117 $0.05M $399.63

You, Harry L. Director — buyer +5,550 $1.9M $347.13

Insider open-market selling by month

$M of proceeds · Sep 2025 – Jul 2026

Clustered in four windows — September, December, March and June — i.e. the post-earnings open-window pattern of scheduled 10b5-1 plans, not opportunistic selling. Note nonetheless that the two heaviest months, March ($338M) and June ($263M) , both sold into strength within weeks of the June all-time high.

How to read $1.2 billion of insider selling

Two thirds of it is Henry Samueli , an 80-year-old co-founder with a decades-old diversification programme — this is estate and philanthropy mechanics, not a signal about the quarter. The two entries that carry information are: Hock Tan sold $235M and stopped in January , with no open-market sale in the eight months since; and Harry You is the only director buying , in size-consistent increments including 1,000 shares at $373.57 on 11 June, a week after the Q2 selloff. No Form 4 has been filed since 10 July — the trading window reopened after the 2 September print, so the next fortnight of filings is the first read on whether the June pattern of selling into strength repeats at $370 rather than $400.

Institutional ownership — Form 13F

Institutional % 89% 4.23B shares across 5,185 filers

Holders adding 2,724 +173.7M shares

Holders trimming 2,127 −128.8M shares

Net institutional flow +44.8M shares, ~$16B at spot — net accumulation

Top 12 institutional holders and quarterly change

% of shares outstanding · change in shares vs prior filing

The top ten hold 44.2% of shares outstanding and half of all institutional stock. Six of the top ten are index or quasi-index vehicles, so most of that number is mechanical.

SEC Form 13F holdings via Alpha Vantage. 4,924 of 5,185 filings are as of 30 Jun 2026; Vanguard Group's headline line is stale at 31 Dec 2025 (its other entities report 30 Jun 2026). "n/a" change % = new position.

Notable moves Change % Now

FMR (Fidelity) +20.2M +16.3% 144.4M

Jupiter Topco new +19.6M new 19.6M

BlackRock +12.3M +3.2% 398.2M

Capital World +5.7M +4.9% 121.9M

State Street +5.5M +2.9% 196.9M

Sanders Capital new +5.2M new 5.2M

MFS +3.1M +34.6% 12.1M

CPP Investment Board +2.7M +26.6% 13.0M

Arrowstreet Capital −10.1M −54.1% 8.6M

SG Americas −7.6M −35.6% 13.7M

Norges Bank −4.1M −5.9% 66.0M

Franklin Resources −3.9M −13.7% 24.7M

Jane Street −3.9M −82.9% 0.8M

Cantillon Capital −3.9M −95.4% 0.19M

D. E. Shaw −3.8M −58.3% 2.7M

Separating passive accumulation from active conviction

Strip out the index complex and the picture sharpens. Genuine active adds: Fidelity +20.2M (+16.3%) is the single largest conviction move in the file and it is a stock-picking house, not an indexer; MFS +34.6% and CPP +26.6% are meaningful percentage adds; two new positions of size (Jupiter Topco 19.6M, Sanders Capital 5.2M). Genuine active exits: Cantillon liquidated 95% of a long-standing position and D. E. Shaw cut 58% — but Arrowstreet (−54%), SG Americas (−36%) and Jane Street (−83%) are quant and market-making books whose 13F lines turn over routinely and carry little signal. The net read is fundamental money added into the drawdown while quant and multi-strat books de-grossed.

What the street says

Post-print actions 3–4 Sep 2026 from published notes Consensus: Stocklake (46) · MarketBeat (34) · Alpha Vantage

The print split the sell side for the first time this year. Nine houses raised, Macquarie upgraded, and three cut — Truist, TD Cowen and DA Davidson, whose $350 is now the first named target below spot. The mean is $505–533 depending on the panel, still 37–44% above the price.

Mean target $533 46-analyst panel, +44% · MarketBeat $505 (34)

Rating split 30/4/0 Buy / hold / sell (MarketBeat) · score 1.27 "strong buy"

Post-print raises 9 +1 upgrade (Macquarie) · Cantor & Rosenblatt to $600

Post-print cuts 3 Truist $520 · TD Cowen $475 · DA Davidson $350

Price target distribution versus spot

Named firm targets after the 2 Sep print · spot $369 marked

The centre of the distribution moved up (two houses now at $600, the $575 cluster grew) and the tail moved down: DA Davidson's $350 is the first named target below spot in this cycle. The distribution is wider than it was in August, which is what a genuinely contested print looks like.

Consensus target versus share price

Approximate consensus mean, Aug 2025 – Sep 2026 · $ · Sep = 46-analyst panel

The gap that opened in June is still ~$165. The print produced the first target cuts of the year and a modest price recovery — both ends of the gap moving, which is how it usually closes. Note the mean moved up on net despite three cuts.

Most recent published action per firm; rows shaded teal acted after the 2 Sep print, red rows cut. Compiled from Benzinga, MarketBeat, Investing.com, TheStreet and firm notes. Houses that asked questions on the call but for whom no post-print target could be located (JPMorgan, BofA, Goldman, Jefferies) are shown at their last confirmed level. Melius, Arete, New Street, Summit, Loop, Wedbush, Stifel and HSBC are omitted rather than carried stale.

Firm Analyst Rating Target Date Move

Cantor Fitzgerald C.J. Muse Overweight $600 3 Sep 26 ↑ from $525 · FY28 EPS $35

Rosenblatt Sajal Dogra Buy $600 3 Sep 26 ↑ from $500

JPMorgan Harlan Sur Overweight $580 4 Jun 26 "strong and conservative"; no new PT found

Evercore ISI Mark Lipacis Outperform $578 3 Sep 26 ↓ from $582 (trim)

KeyBanc John Vinh Overweight $575 3 Sep 26 reiterated

Bernstein Stacy Rasgon Outperform $575 3 Sep 26 ↑ from $550

BMO Capital Harsh Kumar Outperform $575 3 Sep 26 ↑ from $455 · "second only to Nvidia"

Jefferies Blayne Curtis Buy $550 4 Jun 26 no new PT found

Wells Fargo Aaron Rakers Overweight $545 14 May 26 —

StoneX Cody Acree Buy $545 3 Sep 26 reiterated · "not quite enough"

Oppenheimer Rick Schafer Outperform $535 4 Jun 26 —

Bank of America Vivek Arya Buy $530 4 Jun 26 no new PT found · credit desk at Marketweight

Mizuho Vijay Rakesh Outperform $530 21 Aug 26 —

Goldman Sachs James Schneider Buy · Conviction $525 3 Aug 26 no new PT found

Truist William Stein Buy $520 3 Sep 26 ↓ from $550 · "neither stellar nor clean"

Citigroup Atif Malik Buy $515 4 Sep 26 ↑ from $500

Deutsche Bank Ross Seymore Buy $515 4 Jun 26 —

Morgan Stanley Joseph Moore Overweight $505 3 Sep 26 ↑ from $502 · bull $640 / bear $300

Barclays Tom O'Malley Overweight $500 12 Dec 25 —

Macquarie Arthur Lai Outperform $490 3 Sep 26 upgraded from Neutral · from $437

Susquehanna Chris Rolland Positive $490 28 May 26 —

UBS Timothy Arcuri Buy $485 24 Aug 26 a 3 Sep "downgrade to Hold" appears on MarketBeat only — unverified

Raymond James Srini Pajjuri Outperform $475 3 Sep 26 ↑ from $450

TD Cowen Joshua Buchalter Buy $475 3 Sep 26 ↓ from $500 · "circular financing" skepticism

Piper Sandler Harsh Kumar Overweight $430 12 Dec 25 —

Baird — — $420 12 Dec 25 —

Wolfe Research Chris Caso Outperform $400 30 Jan 26 —

RBC Capital Srini Pajjuri Sector Perform $400 3 Sep 26 reiterated · supply concerns

DA Davidson Gil Luria Neutral $350 4 Sep 26 ↓ from $400 · street low, below spot

Lake Street — Buy — 3 Sep 26 upgraded

Consensus dispersion — read the fine print

Providers still disagree on both the mean and the panel: Stocklake/Yahoo $533.41 (46 analysts), MarketBeat $504.93 (34, range $350–600), Benzinga $505. The credible working range is now $350–$600 — wider at both ends than August's $400–$582. Consensus FY26 EPS remains disputed across providers ($11.64 vs ~$10.24) because they adjust off different FY25 bases ($6.82 non-GAAP vs $5.63). Do not mix the two.

What retail thinks

Stocktwits sentiment & message-volume API, 1-year, to 8 Sep Live symbol stream · 74,098 watchers

Retail read the print as a buying opportunity and said so loudly: the sentiment score jumped from 60 to 84 — "Extremely Bullish" — through a week in which the stock fell, and message volume is the highest of the year. Today's stream is price targets and dip-buying.

Stocktwits sentiment 84 /100 Label: Extremely Bullish · was 60 on 25 Aug

Bull / bear tags 80.6% vs 19.4% bearish · −2.9 delta on the day

Message volume EXTREME 78/100 now · 1-year percentile 91

Watchers 74,098 +2,013 since 25 Aug

Retail sentiment and chatter over twelve months

Stocktwits normalised sentiment (0–100, 50 = neutral) and message volume · weekly · last point is the print week

Sentiment score

Message volume

Neutral (50)

Retail sentiment is a coincident indicator here, not a leading one — and the print week made the point again: the score hit 84 in the same week the stock fell 2.7%, exactly as the 95 reading on 7 Jun 2026 sat on top of the 12.6% post-earnings collapse. Bulls buy the dip loudly. The pattern that has actually paid is the trough: readings in the high 20s (4 Jan, 29 Mar, 5 Jul) each preceded a recovery within four weeks.

The live stream, 8 September 2026

@nwelyf31 · 18:13 UTC · Bullish · 3 likes "$AVGO $550 soon"

@AVGOmysugardaddy · 18:25 UTC · Bullish "$AVGO $BA $INTC $META AVGO Cantor Fitzgerald New PT $600."

@LetsGetOurDailyBread · 18:19 UTC "$AVGO main position as of now. Cash set aside for potential dip buying opportunities."

@mramanathan · 18:26 UTC · Bullish "$AVGO innocent retail for peanut trade vs institutional shark attack 😝"

@mramanathan · 18:09 UTC · Bullish "$AVGO Institutions don't encourage for day trade and sudden spike in pre/post market hours coming soon...😀"

@Blackdyna · 18:08 UTC "$AVGO I miss the days when logic existed."

The retail signal that is actually informative

Not the 80.6% bull tag — that figure has sat above 80% through every drawdown this year and is close to useless as a directional input. What changed is the content . On 25 August the stream was arguing about Jalapeño benchmarks; today it is quoting sell-side price targets and declaring dip-buying cash "set aside." Retail has stopped debating whether the product is good and started debating whether $369 is the entry — the same shift the sell side made when it split into raisers and cutters on the same print. When retail and institutions are arguing about the same thing, the thing is usually the right question. Here it is price, not product.

Beyond Stocktwits

Bigdata.com news sentiment

Score 0.052 (slightly positive) against a 0.095 baseline, momentum −0.043 , 1-month z-score −1.2 — while media attention momentum is +14.9% . Deteriorating sentiment on rising attention. Bigdata flags its own evidence base as thin and source-concentrated, and several documents in the pool belong to Gen Digital rather than Broadcom, so treat as directional only.

Positioning chatter

Hedge-fund rotation out of AVGO toward Amazon, TSMC and Lam Research has been reported through August — characterised in coverage as "a vote on risk profile more than on AI demand." The 13F data corroborates this: quant and multi-strat books de-grossed while long-only added.

The retail-visible bear

The WSJ's 17 August piece on ~$3 trillion of off-balance-sheet AI commitments across nine large tech companies has circulated widely in retail forums and is the most-cited bear artifact on the name — more than VMware or Marvell. Retail is watching the $380 level as the reclaim.

The competitive set

Stocklake batch fundamentals, 25 Aug 2026 Company filings · Bigdata.com

Broadcom competes in three distinct arenas with three different competitor sets — and only one of them is Marvell. Conflating them is the most common analytical error on this name.

Custom AI silicon (XPU/ASIC). Direct rivals: Marvell , Alchip, GUC, MediaTek, and increasingly Astera Labs at the connectivity layer. The customer set is six hyperscalers and frontier labs. Broadcom is the incumbent; the 19 August Marvell–Google warrant is the first crack.

AI networking (Ethernet switching, SerDes, optics, retimers). Rivals: Arista (systems), Nvidia (InfiniBand/Spectrum-X), Cisco, Credo and Astera Labs at the component layer. Tomahawk 6 at 102.4 Tb/s is the anchor product and was ~40% of Q2 AI revenue.

Merchant GPUs. Broadcom does not compete here. Nvidia and AMD are the substitutes for Broadcom's customers' custom chips, not for Broadcom's products — the competitive question is whether hyperscalers keep building their own, not whether AVGO out-designs Nvidia.

Infrastructure software. Rivals: Nutanix , Red Hat OpenShift Virtualization, Microsoft Azure Local, Proxmox, and the public clouds. See the VMware section.

Stocklake batch fundamentals retrieved 25 Aug 2026 (prices intraday). Revenue growth is the latest reported YoY. Fwd P/E on each provider's own forward-EPS convention — comparable in direction, not to the decimal.

Ticker Company Arena Mkt cap Rev TTM Rev growth Op margin Fwd P/E PT upside RSI

AVGO Broadcom All four $1.71T $75.5B +47.9% 49.0% 18.5× +46% 35.4

NVDA Nvidia Merchant GPU + networking $5.14T $253.5B +85.2% 65.6% 16.3× +44% 44.7

TSM TSMC Foundry (supplier) $2.16T — +36.0% 60.3% 19.1× +33% 45.8

AMD AMD Merchant GPU/CPU $780B $41.3B +50.1% 17.3% 30.9× +28% 42.5

MRVL Marvell Custom ASIC + optics $218B $8.7B +27.6% 14.5% 38.6× +10% 52.9

ANET Arista AI networking systems $239B $10.5B +37.7% 45.4% 36.8× +28% 51.4

TXN Texas Instruments Analog (non-AI comp) $236B $19.5B +22.8% 42.6% 24.3× +25% 36.2

QCOM Qualcomm Wireless (non-AI comp) $168B $44.1B −4.0% 18.5% 15.7× +20% 41.6

CRDO Credo Connectivity / AEC $42B $1.3B +157% 35.7% 24.7× +26% 45.6

ALAB Astera Labs PCIe/CXL connectivity $49B $1.2B +105% 22.7% 44.4× +38% 39.1

MCHP Microchip Embedded (cycle read) $40B $5.1B +38.0% 24.0% 16.1× +47% 41.1

NTNX Nutanix VMware alternative $18B $2.7B +10.0% 10.0% 30.3× −6% 67.6

Growth versus forward multiple

Latest reported revenue growth % (x) vs forward P/E (y) · bubble = market cap

Broadcom sits in the lower-right quadrant — high growth, low forward multiple. Only Nvidia is cheaper on forward earnings, and Nvidia carries no VMware-style software drag or SPV credit exposure.

Marvell: the comparison everyone asks for

On 19 August, Google granted Marvell a warrant tied to roughly $120 billion of qualifying purchases through FY2033 for custom AI inference chips. AVGO fell 4.6% and shed ~$87B of market cap on the day. On 25 August MRVL rose 5.7% into its 27 August earnings.

Why it matters: Google was Broadcom's founding XPU customer and the TPU franchise is the single largest contributor to the FY27 guide. Coverage also points to MediaTek as another Google partner and rumours of AMD involvement in future TPU generations. Google is deliberately dual-sourcing.

Why it may matter less than the tape implied: JPMorgan pushed back on 20 August, noting Google's 3nm TPU v8i has started ramping with Broadcom and 2nm TPU v9 is expected early 2028 — the Marvell content is a later-generation, incremental award, not a displacement. And the scale gap is enormous: Marvell's entire TTM revenue is $8.7B against Broadcom's $89.1B , at 14.5% operating margin versus 54%.

What the Q3 call added: Marvell was not mentioned once. The Google second source Broadcom chose to name was MediaTek — Tan: "Broadcom is now shipping the TPU v8i ahead of the MediaTek v8t, which in fact was initiated earlier." That is the competitive claim management wants on the record: the incumbent is out-executing the challenger on Google's own roadmap. Arya's direct question — why isn't Google taking more? — was not directly answered.

The competitive question that actually decides the thesis

Not "can Marvell take share." It is whether the custom-silicon share of AI compute keeps rising against merchant GPUs at all. If it does, Broadcom's scale, packaging relationships and Ethernet franchise make it the structural winner regardless of a Google second source. If hyperscalers conclude that Nvidia's software stack is worth the margin, Broadcom's six-customer concentration becomes the problem rather than the moat.

Jalapeño — the OpenAI program, and what SemiAnalysis actually argued

SemiAnalysis 25 Aug 2026 (free portion; §8 paywalled) OpenAI & Broadcom releases · Q3 FY26 call, 2 Sep · The Information · CNBC

SemiAnalysis published "OpenAI' Jalapeño: Better Than Nvidia Blackwell" on 25 August; a week later Broadcom said on its Q3 call that the chip had shipped. The headline is the least interesting part of the SemiAnalysis piece. The argued claim — and the caveats they put on their own work — matter far more to the Broadcom thesis.

What Jalapeño is

OpenAI's first custom inference ASIC, co-developed with Broadcom, unveiled 24 June 2026, benchmarked publicly for the first time on 25 August, and confirmed as shipped on the 2 September call. Silicon specs come almost entirely from SemiAnalysis — OpenAI and Broadcom disclosed essentially none of them.

A reticle-sized ~840 mm² compute die on TSMC N3P with an N3E I/O chiplet and six HBM4 stacks at 15.4 TB/s, CoWoS-packaged, taped out November 2025. A0 silicon does 10.7 PFLOPS MXFP4 at 700 W ; B0, in fab, does 13.4 PFLOPS with ~25% better perf-per-watt. For reference, Nvidia's Vera Rubin is 17.5 PFLOPS dense NVFP4 at 900–1,150 W.

Broadcom's dollar content is the network, and it is Ethernet. 128 XPUs per rack meshed all-to-all through six Tomahawk 6 ASICs at 102.4 Tb/s each , giving 4.8 Tb/s per XPU locally; up to 2,048 XPUs in a single scale-up domain over an 8-rail architecture with optical circuit switches and 1.6T transceivers. The October 2025 release said the racks would be "scaled entirely with Ethernet and other connectivity solutions from Broadcom." Celestica does board and rack integration; Microsoft was named a deployment partner in June.

Software is entirely OpenAI's: kernels in "Gluon" (built on Triton), a serving engine called Teacup, with Codex used internally for compiler work.

OpenAI's published InferenceX results, 25 Aug 2026. Comparison systems' power ratings imply GB200 (1,200 W) / GB300 (1,400 W). Workload is 8k input / 1k output — which SemiAnalysis itself calls "a much easier workload to tune for."

Model Throughput/kW Latency tok/s/user

GPT-OSS 120B 1.9× 1.7× (1.03s vs 1.80s) 1,459 vs 535

DeepSeek R1 670B 1.7× 3.6× (1.65s vs 5.99s) 700 vs 169

Kimi K2.5 1T 1.5× 3.4× (1.56s vs 5.31s) 694 vs 182

Program timeline

Design start mid-2024 → CoWoS tape-out Nov 2025 → unveiled 24 Jun 2026 ("fastest ASIC development cycle ever achieved," per the parties) → initial deployment end-2026 in "very small volumes" → reliability data with datacenter partners through Jan 2027 → gradual 2027 ramp with most output scheduled for Q4 2027 → 10 GW complete end-2029.

SemiAnalysis's argument, in their words

The contrarian claims

Against the press narrative that this is a narrow OpenAI-model-specific part — which they blame on "throwaway comments from OpenAI" — they argue "OpenAI made a generalized chip for AI inference… capable of running all sorts of models, and all sorts of workloads."

"In general first generation chips are not competitive" — yet they assess Jalapeño as "industry leading and beating every Nvidia, AMD, and Google chip."

OpenAI rejected prefill-decode disaggregation , against current conventional wisdom, in favour of a unified "fungible fleet" because "workload mix changes over time."

On the gap being software, not silicon: "We don't think that Nvidia hardware is inferior, but more so that Jalapeño's software bring-up has progressed more quickly." Attributed to hardware/software co-design — "a cracked frontier lab ASIC team can excel over more established merchant silicon players."

The line that will get quoted back at Nvidia: "OpenAI models like GPT 5.6 Sol…have been used to design a chip that poses a real threat to the CUDA moat."

The caveats they put on themselves

"All numbers are provided to us by OpenAI. We verified the InferenceX runs in person in the lab, but we did not run the full suite… nor have we seen AgentX results." InferenceX is SemiAnalysis's own benchmark and OpenAI invited them.

"Comparison to Blackwell is somewhat incomplete and unfair. Jalapeño is really competing against chips like Rubin that also use HBM4."

On TCO against the right comparator: "Vera Rubin and Jalapeño are head-to-head, producing almost the same number of output tokens per $" — Jalapeño without speculative decoding, so with it implemented "will serve much lower cost tokens." Read carefully, that is a tie today with optionality , not a rout.

They refuse the lazy margin-arbitrage explanation: part of the TCO gain "comes from trading Nvidia's high margins for Broadcom's lower (though still high) margins" — but "not all of it."

Vera Rubin systems "are starting to ship to customers right now, while it will still be some time before OpenAI has anything beyond engineering samples."

Software gaps persist: "Surprisingly, OpenAI had no internal implementation of MLA kernels until they benchmarked DeepSeek with InferenceX."

The single most deflating number in the piece

SemiAnalysis's stated next milestone for the program is "The next goal is 100MW." Against a 10 GW headline that is 1% . The hurdles are, in their words, "mostly hardware: how much can they produce, how well can they deploy and operate datacenters" — because "software is already proven." Whatever else the article says, that sentence is the honest scale of where the OpenAI program sits in August 2026.

What it is worth to Broadcom — and what it is not

Get the customer attribution right first. A widely repeated conflation: the ">$10 billion of orders" from a mystery fourth XPU customer disclosed 5 September 2025 was not OpenAI. Broadcom's semiconductor president said so explicitly on 13 October 2025, and CNBC revealed on 11 December that it was Anthropic . That relationship has since been reported at ~$21B.

The math on OpenAI specifically. Bernstein's Rasgon uses ~$20B of Broadcom revenue per gigawatt and Hock Tan has said the real number is "not far from" that. Tan guided 1.3 GW of OpenAI shipments in 2027 — so roughly $26B , about a quarter of the FY27 ">$100B" guide. The full 10 GW through 2029 is on the order of $200B lifetime. FY2026 OpenAI contribution is essentially zero. Google, Anthropic and Meta carry more of FY27 than OpenAI does.

The structural risk is the financing, not the chip. The Information reported in May 2026 that the ~$18B "Nexus" production project hit a snag: Broadcom would only fund phase one if Microsoft committed to buying ~40% of the chips and leasing them back to OpenAI — and OpenAI's compute head Sachin Katti called that "financially unattractive" and "likely unworkable." Microsoft did appear as a deployment partner in June, which suggests something was reached, but the final terms have never been reported . Separately, on 21 August Broadcom was reported to be seeking $70–80 billion of debt through an SPV to finance compute leasing for AI companies, with Anthropic explicitly named. That converts a supplier-margin story into a credit story.

Status, 2 Sep call Shipped "We also shipped Jalapeño, OpenAI's first generation custom accelerator"

OpenAI 2027 1.3 GW "On track" · ~$26–39B at Tan's $20–30B/GW

OpenAI 2028 >5 GW New on the Q3 call — "line of sight"

FY26 OpenAI rev ~$0 Engineering-sample volumes only

Commentary — the trade this news sets up

Jalapeño is a genuine engineering achievement and a real validation of the co-design model Broadcom sells — and on 2 September it moved from "unveiled" to "shipped," with 1.3 GW in 2027 called "on track" and a new "over 5 GW" line for 2028. It remains, for FY2026, a zero-revenue event , and the 2028 number is the one that carries the FY28 $230B guide. If you are long AVGO on the Jalapeño narrative, you are underwriting OpenAI's balance sheet — and the call was candid about that: Tan separated "the other four customers [who] are financially secure" from "these two guys, Anthropic and OpenAI," whose builds may run through the XPV vehicles "in some cases." Whether Broadcom stands behind that paper is the question the CFO declined to size.

VMware — terms, product, competition, growth

Broadcom 8-K & PRs · EU Commission M.10806 · UK CMA · SAMR Quarterly segment disclosure · CISPE · Gartner / Forrester / CloudBolt

Financially the acquisition is a clear success. Operationally it is a clear controversy. Q3 gave one data point to each side: a record $8.75B software quarter at 94% gross margin — that missed its own guide and is guided to decelerate again.

Acquisition conditions

Consideration terms verified against Broadcom's merger announcement (26 May 2022) and closing 8-K (22 Nov 2023). The three figures in circulation are all correct for different dates: $61B equity at announcement, ~$69B including net debt, $84.2B actual consideration transferred at close after AVGO's share price roughly doubled.

Term Detail

Announced / closed 26 May 2022 / 22 Nov 2023

Election $142.50 cash or 0.25200 AVGO shares

Proration Subject to 50% cash / 50% stock

Blended at announcement $138.23/sh — a 44% premium

Equity value ~$61B (May-2022 AVGO price)

Net debt assumed $8B

Value at close $84.2B per Broadcom's own release

Financing $32B committed at signing; drew the full $28.39B term loan at close; termed out via bonds

Total debt FY24 year-end $67.6B → ~$64.9B today vs $19.6B cash

Regulatory conditions — narrower than commonly assumed

European Commission (M.10806), cleared 12 Jul 2023 with commitments. The Commission narrowed to a single theory of harm: degradation of interoperability between VMware and Marvell's Fibre Channel HBAs , Broadcom's only FC HBA rival. Three limbs — (1) interoperability access to VMware APIs, dev kits, pre-release materials and certification on terms equal to Broadcom's own FC HBA division; (2) a source-code grant of Broadcom FC HBA drivers to Marvell to reuse and modify; (3) organisational firewall between the FC HBA team and VMware certification. Duration: 10 years to ~Nov 2033, with a monitoring trustee and fast-track dispute resolution. The Commission explicitly rejected concerns on storage adapters, NICs, SmartNICs and bundling.

UK CMA — cleared unconditionally, 21 Aug 2023 , after a full Phase 2. Found no substantial lessening of competition. No remedies at all — the notable divergence from the EU and China.

China SAMR — conditional approval 21 Nov 2023 after a record 231-day pre-review. Six conditions imposed, four published and two kept confidential. The published four: (1) no tying or bundling of Broadcom FC adapters, storage adapters and Ethernet NICs with VMware in the Chinese market; (2) continued interoperability with third-party hardware; (3) continued driver development and certification for third-party virtualisation software; (4) confidential-information protection for third-party hardware makers. 10 years, auto-lapsing. China is the only jurisdiction that imposed an explicit anti-tying obligation. Confidence note: the four conditions are corroborated across Chinese media and an ABA antitrust analysis; the 10-year duration and "six, two secret" structure rest on Chinese-language reporting, not a retrieved English primary text.

The gap the remedies leave open

None of the EU, CMA or SAMR conditions constrain the thing customers actually object to — software pricing, SKU bundling, or channel access . That is precisely the gap CISPE's Article 102 complaint (19 Mar 2026, with a request for interim measures) is trying to close. CISPE has also sued in the EU General Court (Jul 2025) to annul the clearance decision itself . In May 2026 the General Court rejected Broadcom's attempt to suspend a Commission RFI on US in-house-privilege grounds. No formal Article 101/102 proceedings have been opened — treat "formal EU investigation" claims as imprecise.

What VMware is now

VMware Cloud Foundation (VCF) is effectively the only full-stack SKU. VCF 9.0 shipped 17 Jun 2025; 9.1 on 5 May 2026 . It bundles vSphere (compute), vSAN (storage, 1 TiB/core), NSX (networking and security), vSphere Kubernetes Service, VCF Operations and Automation (the former Aria), HCX, and VCF Private AI services. An "Advanced Services" add-on layer — cyber compliance, security, Avi load balancing, data services, network observability — is the current up-sell vector and a live customer grievance.

VMware Private AI Foundation with NVIDIA is the answer to the obvious bear case that AI workloads bypass the hypervisor. The VCF-side package is included in VCF; NVIDIA AI Enterprise licences are bought separately from NVIDIA. 9.0/9.1 added air-gapped deployment, multi-tenancy, GPU/vGPU monitoring, FIPS/STIG containers for government, and ConnectX-7 / BlueField-3 / Blackwell support. Hock Tan on whether genAI is displacing VMware: "We are not seeing it. If anything else… the high volume of core count of CPUs selling together with GPUs, it is driving some accelerated growth."

The commercial restructuring is what generated both the margins and the lawsuits: ~8,000 SKUs collapsed to four offers (Dec 2023); all perpetual licensing ended 11 Dec 2023; 16-core-per-CPU minimum, then a 72-core-per-transaction minimum from 10 Apr 2025 ; a 20% surcharge on late renewals; ~2,000 strategic accounts moved direct; and the cloud-provider program closed entirely — all remaining VCSP contracts non-renewed effective 26 Jan 2026, with subscription sales restricted to Pinnacle-tier partners. Reportedly ~19 authorised providers remain in the US and 9 in the UK.

Software gross margin 94% Q3 FY26 · op margin 84%, +650bps YoY

Q3 vs own guide −$148M $8,752M, +29% vs ~$8.9B, +31%

Q4 guide $8.7B +25% YoY · flat-to-down sequentially

ARR growth +15% Q3 FY26 · was 17% in Q2 · absolute undisclosed

On the "$8.5B → $20B run-rate" framing

Worth correcting: Broadcom's $8.5B figure is a commitment to incremental adjusted EBITDA within three years of close, not a revenue run-rate — and Tan said in Dec 2024 it was already being exceeded "in a tighter time frame." The revenue facts are that Infrastructure Software ran $27.0B in FY2025 and is guided to low-double-digit FY2026 growth (~$30B), against a pre-VMware Broadcom software base of roughly $7.6B in FY2023.

Competitive pressure and the churn evidence

Third-party survey and forecast data. Note the consistent structure across independent sources: a very high share of customers reducing footprint against a very low share who have actually completed a migration.

Source Finding

Gartner (2025) 35% of VMware workloads migrate away by 2028; share 70% (2024) → 40% (2029)

CloudBolt (Feb 26, n=302) 86% actively reducing footprint · only 4% fully migrated

Forrester (2026) 87% reducing · 4% complete — "maintain, migrate and modernise simultaneously"

Rimini Street (n=111) 98% using/considering alternatives · 36% already switched

CloudBolt on price 73% expected costs to >2× in 2024; only 14% saw >100% — modal outcome 25–49%

Nutanix Q3 FY26 ARR $2.435B, +15% · NDR 106% · 700+ new logos in the quarter

Red Hat (2025) VMs on OpenShift Virtualization +417% ; accounts running VMs +70% — off a small base

Litigation and regulatory sequel

AT&T v. Broadcom (NY, filed 29 Aug 2024) — ~75,000 VMs, 22,000 supporting public-safety agencies; AT&T alleged a "+1,050% in one year" proposed increase. TRO granted; settled in principle 13 Dec 2024 , terms undisclosed.

Broadcom/VMware v. Siemens AG (D. Del., 21 Mar 2025) — Broadcom suing the customer over ~23,000 alleged unlicensed US deployments. Ongoing, counterclaims pending.

Rijkswaterstaat (Dutch Ministry of Infrastructure) — Broadcom lost. 1 Jul 2025, The Hague found it "acts in breach of its duty of care," ordered up to two years of exit support at €250k/day penalty capped at €25M. Trigger was an 85% increase on systems running tunnels, locks and bridges.

Tesco plc (UK High Court, Sep 2025) — ≥£100M claim , ~40,000 workloads. Alleges 175% VMware and 350% mainframe increases. Trial window Nov 2027 – Feb 2028; Tesco migrating off by end-2027.

Frequently mis-cited: the UnitedHealth suit (Apr 2025) is about CA mainframe software , not VMware.

The structural read

The gap between "86–87% reducing" and "4% fully migrated" is the entire VMware story. Broadcom is monetising an installed base that is shrinking in seat count while paying far more per seat. Q3 showed both halves at once: record revenue and 94% gross margin, alongside a guide miss, decelerating ARR, and a Q4 that goes sideways. Whether revenue keeps growing depends on price/mix uplift outrunning core-count attrition — and that race resolves in the FY2027 renewal cohort , when the three-year subscriptions signed in the late-2023/2024 conversion wave come up from November–December 2026 onward. Management's own product answer is "VMware Private AI Cloud" and public-cloud repatriation; it gave no adoption metric for either.

Growth trajectory

Infrastructure Software revenue and YoY growth

$B per quarter, Q1 FY24 – Q4 FY26 · YoY % direct-labelled · Q4 FY26 is guidance

The +1% Q1 FY26 print was the first hard evidence that the price-for-volume trade has a ceiling. Q3 then delivered the largest software quarter on record — $8.75B, +29%, a $1.57B sequential step — and still fell ~$150M short of the +31% guide, with no explanation on the call. Q4 is guided to $8.7B, +25% : growth decelerating for a second consecutive quarter off the VCF 9.1 server-refresh bump. ARR growth slowed from 17% to 15%. The renewal-cohort question is now live rather than theoretical.

SWOT

Analysis · grounded in the data above

Ordered within each quadrant by how much each item actually moves the FY2027 outcome, not by conventional weight.

Strengths

The only credible full-stack custom-AI vendor. Compute ASIC design, Ethernet switching at 102.4 Tb/s, SerDes, optics, PCIe and rack-scale deployment under one roof. Competitors have pieces; Broadcom has the whole path from die to rack.

Margin structure that funds everything. 67.9% non-GAAP operating margin in Q3 FY26 — a record, up 240bps YoY — with 94%-gross-margin software subsidising the semiconductor build-out. Gross margin held at 75% against a 74% guide. Capex under 2% of revenue.

Cash generation. $30.5B TTM free cash flow; Q3's $13.7B was 46% of revenue. $5.6B of debt repaid in a quarter with no buybacks. Fifteen consecutive dividend increases.

Improving cost of capital. BBB− at the VMware announcement to A−/A3 by September 2025; the 5-year new-issue spread compressed +85bp → +53bp; interest expense is down 4% year over year on flat debt as ~7% floating term loans were refinanced into 4.2–5.2% fixed notes.

Visibility, now with numbers. FY27 AI "approximately $115B" and FY28 "line of sight" to $230B, both with "secured the supply" attached — $350B over two years. Management says demand exceeds even that outlook.

Customer lock-in through co-design. Multi-generation roadmaps with Google (through 2031), Meta (through 2028) and OpenAI (through 2029) are architectural commitments, not purchase orders.

Execution track record under Tan. VMware opex halved, the $8.5B incremental-EBITDA commitment beaten ahead of a three-year schedule.

Weaknesses

Six customers, two of them the growth. The FY27 ~$115B guide averages ~$19B per customer, and management itself separated "the other four [who] are financially secure" from Anthropic and OpenAI — whose 2028 builds (10 GW and >5 GW) carry most of the step to $230B and may be XPV-financed "in some cases."

Structural gross-margin dilution. XPU carries ~70% semiconductor gross margin against 94% software; consolidated non-GAAP GM went 77% → 75% and is guided to ~73% in Q4. Tan's answer — "stop focusing on gross margin" — is an argument, not a fix.

Infrastructure software deceleration. +1%, +9%, then +29% that missed its own +31% guide, now guided to +25% with ARR growth slowing from 17% to 15%. The ballast has become the swing factor.

No balance-sheet support. Roughly two-thirds of assets are intangible; tangible book value per share is about −$5 . $59.4B of debt against $24.0B cash before any XPV exposure.

Inventory is building. $4.5B, +99% since the FY25 year-end — appropriate for a supply-secured ramp into a $21.7B AI quarter, but the one working-capital line still rising after Q3 collapsed DSO from 69 to 42 days.

Customer relations as a liability. Four significant VMware suits, one already lost in the Netherlands, plus an active EU complaint. Reputational cost is now a pricing constraint in Europe.

Disclosure quality slipping. For a second straight quarter Broadcom gave no backlog, bookings or book-to-bill figure, and did not disclose adjusted EBITDA at all. It has also declined, when asked directly, to size its maximum XPV exposure.

Opportunities

Custom silicon taking share from merchant GPUs. The core structural bet. If ASIC share of AI compute keeps rising, Broadcom is the default beneficiary regardless of which lab wins.

Ethernet displacing InfiniBand for scale-up. Jalapeño's 2,048-XPU domain built entirely on Tomahawk 6 is the proof point Broadcom has wanted for three years. Networking was ~40% of Q2 AI revenue.

FY2028 is now a number. $230B "line of sight," roughly 20 GW on Bernstein's arithmetic, with Anthropic 10 GW, OpenAI >5 GW and Meta 3 GW disclosed individually. Cantor models FY28 EPS of $35 against a $27 consensus — the stock at 10.6× that.

Private AI inference on-prem. VCF Private AI with NVIDIA converts the hypervisor base into an AI attach point; Broadcom claims public-cloud inference usage fell 15 points YoY. If even directionally right, this reframes VMware from melting ice cube to distribution channel.

Two undisclosed XPU customers shipping from late 2026 (~$6B in orders) plus whatever the ~$21B Anthropic relationship becomes.

Multiple re-rating. At 19.1× FY27 consensus the stock is cheaper than the market on forward earnings, and the Q3 10-Q's guarantee footnote is the next chance for the SPV to prove benign. Macquarie's upgrade thesis: "the cleanest listed way to own the build" ahead of an Anthropic IPO.

Threats

The financing vehicle is the dominant risk — and the call did not size it. The $35B first XPV tranche is closed; a >$60B second raise is reported. Broadcom "may provide modest residual value guarantees," and when asked for the maximum: "I cannot give you an overarching look at what is the max." BofA's credit desk models up to $370B of senior debt by mid-2029 with $10.5–42B of Broadcom loss exposure and has the bonds at Marketweight. TD Cowen cut on exactly this: "circular financing."

Loss of Google exclusivity. The 19 Aug Marvell warrant covering ~$120B of qualifying purchases through FY2033 breaks the de facto monopoly on TPU. MediaTek is reportedly a second partner and AMD is rumoured for later generations.

Customer capex pause. A spending slowdown at any one of OpenAI, Anthropic or Meta materially breaks the FY27 guide. OpenAI's 10 GW commitment is not underwritten by its own cash flows.

The FY2027 VMware renewal cliff. Three-year contracts from the 2023–24 conversion wave renew from Nov–Dec 2026 into a market with mature alternatives. Even 10–15% churn among the top ~2,000 accounts weakens the recurring-revenue narrative.

EU regulatory action. CISPE's Article 102 complaint with interim-measures request could force unbundling or restored legacy licensing. Broadcom has already lost a privilege fight in the General Court.

Supply chain. Reticle-sized N3P dies plus HBM4 plus CoWoS put Broadcom in direct competition with Nvidia, Google and AMD for the same scarce TSMC advanced-packaging and HBM4 capacity.

The $/GW arithmetic. Rasgon's reconciliation of the disclosed gigawatts implies ~$11–12B per GW; Tan insists on $20–30B. If the lower figure is right, FY28 is closer to $115B than $230B — and the FY28 number is now the multiple's anchor.

Scorecard and what to watch next

Q3 FY26 print vs the August watch-list · company IR Bigdata.com events calendar · court schedules

The 25 August version of this section listed six things to check on 2 September. Here is how each resolved, and what replaces them.

The six things to check — checked

FY2027 AI guidance. Resolved, favourably. ">$100B" became "approximately $115B" — the street's number — plus a new FY28 $230B. The June set-up did not repeat.

Software vs the +31% guide. Missed. $8.75B, +29%, no explanation; Q4 guided to +25%. The renewal-cliff narrative hardened.

Gross margin vs ~74%. Beat. 75.0%. But Q4 is guided to ~73% and management is now openly steering attention to operating margin instead.

Cumulative AI backlog. Absent again. No backlog, bookings or book-to-bill figure for a second quarter.

SPV / XPV commentary and the backstop. Asked directly, not sized. "Modest residual value guarantees… low risk"; "I cannot give you… the max." The Q3 10-Q footnote is now the disclosure to read.

Google TPU / Marvell. Addressed, indirectly. "Never been stronger," TPU v8i in production "ahead of the MediaTek v8t." Marvell unmentioned. Arya's "why not more?" unanswered.

Dated catalysts from 8 September 2026. The 10-Q is the near-term one that matters: it is the first primary-source disclosure of the XPV guarantee terms since the Q2 filing's "backstop lease payments for five years" language.

Date Event Why it matters

~9–10 Sep 26 Q3 FY26 10-Q Guarantee / contingent-obligation footnote; AP, intangibles, retained earnings

Sep 26 Post-print Form 4 window Does the June pattern of selling into strength repeat at $370?

21 Sep 26 Ex-dividend $0.65 Paid 30 Sep

TBD >$60B second XPV / SPV pricing Tranche sizes, Broadcom guarantee, coupon — still unconfirmed by the company

TBD 2026–27 Anthropic IPO Tan: "well on its way… its investment credit will change" — de-risks the largest XPV counterparty

31 Oct 26 Microsoft stops license-included AVS Forces BYOL VCF — software tailwind into FY27

Nov–Dec 26 VMware renewal cohort opens Three-year 2023 conversions come due into a +25% growth quarter

mid-Nov 26 13F filings (30 Sep) Did fundamental money keep adding through the print?

~10 Dec 26 Q4 FY2026 print $34.8B / $21.7B AI / ~73% GM guide; FY26 AI $58B close-out; first FY27 quarterly guide

Ongoing EU Commission RFI / CISPE Art.102 No formal proceedings opened

Q4 2027 Bulk of Jalapeño output 1.3 GW; the OpenAI revenue actually lands

FY2027 Singapore substrate fab Broadcom building its own substrate capacity — a supply-chain tell

Where this leaves the name

In August the argument was that the operating business and the capital structure were being valued by two different sets of people. The print did not change that; it clarified it. Management answered the growth question about as completely as a company can — a specific FY27 number, a first FY28 number, gigawatts by customer, supply "secured" — and declined to answer the financing question at all. The stock's response was to sell the ambiguity intraday and then buy the certainty back over three sessions, which is roughly the right weighting. At $369, 19× FY27 consensus, the market is paying for the $115B and giving the $230B away for free — and charging a discount for not knowing what "modest residual value guarantees" adds up to. The 10-Q this week is the first document that can put a number on that. Anyone underwriting the $533 consensus is underwriting the FY28 figure; anyone underwriting DA Davidson's $350 is underwriting the guarantee. Both now have a specific thing to be wrong about.

Sources and data caveats

Provenance for every figure above

Known gaps and corrections — read before using any number

MSCI is connected but returns no usable data for this task — and the server says why. get_at_server_info reports "index": {"active": false, "registered": false} — the MSCI Index API module is switched off on this account , so every index and security call fails with HTTP 403 "Data access could not be verified" regardless of how it is constructed. Two further paths were tested and also denied: ESG Ratings returns access_denied: … not part of your current subscription , and Portfolio AI Insights returns ENTITLEMENT_DENIED: User has no entitled PAI workspaces . The only MSCI surface that responds is the sustainability/climate taxonomy search, which returns factor definitions and methodology text but no issuer values. Consequently Broadcom's MSCI USA / World / ACWI weights, the Information Technology sector weight, and MSCI sector performance and valuation are absent from this report and not estimated . To unlock them, MSCI needs to enable the Index API module (and, for the ESG angle, an ESG Ratings subscription) on this account.

No "$60B AI backlog" exists. The $60B figure in circulation is the 20 Aug 2026 Bloomberg report of Broadcom seeking $60B+ of debt. The real backlog disclosures are $110B total (Q3 FY25) and $73B AI of $162B total (Q4 FY25) — and nothing since; the Q3 FY26 call gave no backlog figure.

Q3 FY26 balance-sheet lines are from the earnings release, not the 10-Q , which had not been filed at time of writing. Intangibles, tangible book value and the intangibles/assets ratio for Q3 are estimates (Q2 intangibles less the quarter's $2,006M amortisation). Adjusted EBITDA was not disclosed for Q3.

A UBS "downgrade to Hold" dated 3 Sep appears on MarketBeat only and could not be confirmed from any primary or press source; UBS is shown at its last confirmed Buy/$485. Post-print targets for JPMorgan, BofA, Goldman and Jefferies could not be located and are shown at pre-print levels.

Post-print CDS and bond-spread moves were not found in any source; the credit data in the cost-of-debt section is as of late August.

Q4 FY25 AI revenue was corrected from $5.9B to $6.5B, the figure implied by Broadcom's own "+236% YoY" on the $21.7B Q4 FY26 guide and consistent with its ~$20B FY25 total.

The ">$10B mystery customer" was Anthropic, not OpenAI — confirmed by CNBC 11 Dec 2025.

"Nutanix poached 30,000 VMware customers" conflates Nutanix's total installed base with defectors. Annual net adds are ~2,700.

UnitedHealth v. Broadcom concerns CA mainframe software , not VMware.

The "$8.5B → $20B+ software run-rate" framing conflates an incremental-EBITDA commitment with a revenue run-rate. See the VMware section.

Consensus varies by provider ($492–$526 mean, 27–49 analysts) and FY26 EPS is disputed ($11.63 vs $10.24) because of different adjustment bases. Do not mix them.

Alpha Vantage institutional data is 4,924 of 5,185 filings as of 30 Jun 2026; Vanguard Group's headline line is stale at 31 Dec 2025.

SemiAnalysis §8 ("Direct Implications for NVIDIA, AMD and Cerebras") is paywalled and was not read. All Jalapeño silicon specs are single-source to SemiAnalysis.

Bigdata.com's sentiment score (0.052) is contaminated by Gen Digital documents in its evidence pool — Bigdata flags this itself. Directional only.

Company primary

Q3 FY2026 results, 2 Sep 2026

Q3 FY2026 8-K

Q3 FY26 call transcript

Q2 FY2026 results

Q1 FY2026 results

Q4 / FY2025 results

Q2 FY26 call transcript

OpenAI 10 GW collaboration

Jalapeño unveiling

VMware acquisition close

Data connectors

Alpha Vantage — income statement, cash flow, earnings estimates, Form 4 insider transactions, Form 13F institutional holdings, company overview

Stocklake — 365-day OHLCV, indicators, batch peer fundamentals

Stocktwits — sentiment & message-volume history, live symbol stream

Bigdata.com — company tearsheet, sentiment tearsheet, events calendar, news search

MSCI — connected, but Index API module inactive ; ESG Ratings unsubscribed; no PAI workspaces. Taxonomy search only. See caveats.

SEC EDGAR — FWP pricing term sheets, 8-K, 10-K debt footnotes (cost-of-debt section)

Jalapeño / OpenAI

SemiAnalysis, 25 Aug 2026

OpenAI — Jalapeño first results

CNBC — mystery customer is Anthropic

DCD — $18B financing snag

CNBC — $70B+ debt deal

VMware

EU Commission M.10806 decision

CMA unconditional clearance

SAMR conditional approval

VCF 9.1 FAQ

CISPE Article 102 complaint

Dutch court ruling

Nutanix Q3 FY26 transcript

Debt & ratings

Jan 2026 pricing term sheet

Sep 2025 pricing term sheet

Jul 2025 pricing term sheet

Jan 2025 pricing term sheet

Sep 2024 pricing term sheet

Jul 2024 pricing term sheet

FY2025 10-K (debt footnote)

Jun 2026 tender offer 8-K

Apollo $35B XPV release

Market events

Google–Marvell $120B warrant

Apollo/Blackstone XPV platform

CNBC — Q4 FY25 selloff

Benzinga post-print PT roundup

Motley Fool — why AVGO dropped 3 Sep

BofA credit — Marketweight on XPV

MarketBeat price targets

StockAnalysis forecast

Broadcom Teardown · compiled 25 August 2026 · refreshed 8 September 2026 for the Q3 FY26 print · prices intraday

Prepared for research purposes. This is not investment advice and I am not a financial adviser — it is a compilation of public filings, company disclosures, third-party data feeds and press reporting, plus my own reading of them, and it may contain errors. Prices and estimates move. Figures labelled "guide," "consensus," "implied" or "est." are not reported results. Independently verify anything you act on, and consider your own circumstances and a licensed professional before making any investment decision.