Built by Grok 4.6 High

AVGO | Earnings Review — FQ3 FY2026

Verdict: ACCELERATING — and narrowing to AI. FQ3 revenue $29,591 million (+85.5% YoY vs $15,952 million) and nGAAP EPS $3.32 (+96.4% vs $1.69) are the fifth consecutive top-line acceleration off the FQ2'25 +20.2% trough. AI semiconductors $16.7 billion (+221% YoY) are now 56% of sales and 84% of dollar growth. Non-AI is a flat ~$4.2 billion sleeve; software re-accelerated to +29% after the VMware lap.

Print (clean, modest double-beat that is 100% AI). vs FMP $29,241M / $3.22 = +$350M / +1.20% revenue and +$0.10 / +3.11% nGAAP EPS. vs own guide $29.4 billion: +$191M / +0.65%. Bridge of that overshoot: AI +$0.7B vs $16.0 billion; non-AI −$0.3B vs $4.5 billion; software −$148M vs $8.9 billion. Treating this as a broad-based beat is a category error. L12Q EPS 12/12, revenue 10/12; L4Q 4/4. Magnitude is stable, not deteriorating.

Quality of earnings. nGAAP GM 75.0% ($22,191 million / $29,591 million) −336 bps YoY on mix — and still beat the 74% guide. nGAAP OM 67.9% ($20,095 million) +237 bps YoY, +90 bps vs the 67% guide. Semiconductor OM 61% (+400 bps YoY); semi opex is 6% of segment sales. FCF $13,665 million (+94.5%, 46% of revenue). Use ~67% semiconductor GM, not the uncorrected 76% Amie walked back on the same call (Daloopa 76% is the slip).

Guidance. Q4 first issuance $34.8 billion (+93% YoY) sits on FMP $34.86B. AI $21.7 billion (+236%); software resets to $8.7 billion; GM 73% / OM 66%. The raise that matters is the AI stack: FY26 $56 billion → $58 billion; FY27 $100B+ → $115 billion; first FY28 $230 billion. Hock: $350B of AI semis over FY27–28, “we try to be conservative,” demand can ship “significantly more.” Amie: will not update $115/$230 quarterly. FMP FY28 sales $281.6B still sits below a $230B-AI stack.

Tone. More bullish on the level of AI (first $115B point, first $230B, first “exceed $30” FY28 EPS), louder on conservatism, more directive on what to model (OM not GM). Softer on software and non-AI. Classic Hock: raise the destination, haircut the near-term rest of P&L, kill the raise-every-quarter game.

Contradictions (5, two high). Same-call 900 bp semi-GM slip (76% vs 67%); Q1 “GM will not be affected” vs Q4 73% (−400 to −500 bps); $/GW “will increase” vs stable $20–30B; racks vs only-chips vs “in some cases racks”; “modest” RVG vs ~$29B max on the $35B first tranche.

Catalysts into 2026-12-09. Hit $34.8B and show the $21.7B AI mix (Ironwood, TPU v8i HV, Jalapeño continuing, Meta MTIA first production). A quiet reiteration of $115/$230 is the June-style tape. Street FY28 is the gap.

AI went from 33% of sales in FQ3'25 to 56% in FQ3'26 and is compounding at +221% while non-AI is a cyclically recovering ~$4.2B sleeve. That mix is why consolidated YoY jumped +20% → +86% after the VMware anniversary, and why nGAAP GM compressed 336 bps even as nGAAP OM expanded 237 bps. Software’s +29% is the second engine: it lapped VMware a year ago and is now growing on price/ARR, not deal math. ARR slowed 17% → 15%.

Segment drivers

*FQ1'26 AI YoY +105% is $8.4B / $4.1B. Daloopa’s AI YoY KPI prints 16% for that quarter and does not reconcile to the dollar series — discarded.

Driver read. Three things at once; only one is the print.

1. AI XPUs are the growth machine. $16.7 billion vs $5.2 billion. XPU shipments >3.5× YoY and 73% of AI revenue. Ironwood TPU v7 in high volume to Anthropic and Google; TPU v8i production begun; Jalapeño (OpenAI) shipped. AI networking >2.5× YoY.
2. Software re-accelerated after a dead FQ1, and still missed its own guide. $8,752 million +29% vs $8.9 billion. They rounded to $8.8B and did not acknowledge the miss. Software GM 94%; OM ~84% (transcript). Q4 “stabilize” at $8.7 billion is a labeled pause, not a VMware thesis break.
3. Non-AI is dead money. $4.2 billion +5% YoY, flat sequential, missed the $4.5 billion guide. Broadband and server storage up, wireless down. Q4 $4.3 billion — no snap-back.

Consolidated P&L

GAAP is not the operating EPS series. FQ3 GAAP $2.68 (calc. $13,088 million / $4,887 million sh) vs nGAAP $3.32 is a −$0.64 gap (SBC, amortization). FQ3'24 GAAP was a loss. Street underwrites nGAAP.

FQ3 Adj. EBITDA is calc. nGAAP OI $20,095 million + CF depreciation $171 million ≈ $20.3B (margin ~68.5%). Daloopa’s posted Adj. EBITDA recon is not yet loaded for this quarter.

Absolute revenue — 12 quarters ($B)

Revenue and nGAAP EPS YoY — 12 quarters

Annual arc (FY2021–FY2025) + FY2026 implied

FY2021–FY2023 EPS shown split-adjusted. FY2024 software/revenue step-up is VMware, not organic. FY2026E revenue = FQ1–FQ3 $71,089 million + Q4 guide $34.8 billion ≈ $105.9B, in-line with FMP $106.0B (29 estimates).

FQ3 was a modest double-beat vs Street that is 100% an AI-semiconductor overshoot — and a software miss vs Broadcom’s own guide. Revenue surprise is small and clean (+1.2% vs Street; +0.65% vs a point guide). EPS surprise (+3.1%) is the operating-leverage echo of the AI dollar overshoot plus +90 bps of OM vs the 67% guide — not a one-timer, not a tax item.

Pattern: consistent beater, magnitude stable. EPS has not missed Street in 12 quarters. Revenue missed twice, both FQ4s, both <130 bps, and has not missed in the last 7. Company revenue guides are structurally conservative (8/8 above). Mix inside the beat is narrowing to AI.

Heatmap — last 8 quarters

This quarter vs consensus and own guide

Mgmt variance story. Hock: “Q3 demand was simply hot, and we’re just getting started.” The FQ2 guide already baked in AI $16.0B; clearing it by $0.7B is a true operational overshoot, not a sandbag of the $29.4B total. Amie: GM 75% was “better than our guidance of 74%” as AI mix rose; “stop focusing on gross margin… look at where it matters, operating margin.” They did not attribute the EPS beat to tax, buybacks, or below-the-line items. Non-GAAP tax is guided ~16% — a headwind vs history, not a tailwind.

Broadcom issues a point next-quarter P&L (revenue, segments, GM, OM, tax, shares, capex) plus a multi-year AI-revenue trajectory. It does not issue quarterly EPS or a full-year P&L range. Q4 is a first-issuance print. The only official numbers that were revised are the multi-year AI stack.

FQ4 FY2026 — first issuance (applies to 2026-12-09 print)

EBITDA % was dropped. FQ2 issued 68% for Q3; FQ3 is the first in this 8-quarter set with no EBITDA guide. They are herding the street onto OM.

AI stack that was revised

Waterfall — prior guide → print → new guide → street

Guide-accuracy tracker (7 converted quarters)

Revenue 7/7 above the point (never a miss). AI 6/7 above, 1/7 in-line. This is informative conservatism — the same posture Hock just applied to $115B / $230B. The leaky lines are non-AI and software.

What to underwrite. FQ4 revenue is not the debate. FQ4 AI $21.7B is the first checkpoint for the $58 / $115 / $230 stack. They just beat a $16.0B AI guide by $0.7B; the analogous beat here is ~$22.5B+. OM 66% is the P&L they will defend. GM 73% will look ugly vs 78% a year ago — that is mix, and they said so twice.

Revenue YoY inflected from a +20.2% FQ2'25 trough to +85.5% in FQ3'26 — two consecutive violent accelerations (+1,841 then +3,763 bps) — and nGAAP EPS is running ahead of revenue at +96.4%. FQ4 guide of $34.8 billion implies +93.2% YoY, a third acceleration. Trajectory over absolutes: this is worse → better → best, not an 85% grower rolling over.

Three acts: (1) FQ4'24 last easy VMware comp at +51.2%; (2) anniversary air pocket FQ1'25 / FQ2'25 once VMware sat in both years; (3) AI re-acceleration FQ4'25 → FQ3'26. The last two steps are signal, not noise.

Inflection points

Segment decomposition of FQ3 dollar growth ($29,591M − $15,952M = +$13,639M):

What could break the second derivative. (1) Supply, not demand — LPS, HBM, substrates, lasers. If LPS slips, the YoY rate in FY27 still looks huge off a $58B FY26 AI base; the sequential ramp is what would slip. (2) Customer concentration — one semiconductor customer is 50% of direct sales YTD; six XPU customers, four huge. (3) GM dilution becoming OM dilution if memory content moves faster than opex leverage. (4) Software can go quiet (FQ1'26 was +1.4%).

The stock-moving debate into December is not whether FQ4 can print ~$34.8 billion — FMP has already parked at $34.86B. It is whether the FY27/FY28 AI path ($115 billion then $230 billion) is a supply-capped conservative outlook Hock will not update quarterly, or a land/power/shell haircut the Street still will not fully capitalize. FMP FY28 revenue of $281.6B sits below a mechanical $230B AI + software + non-AI stack (~$288B). That is the gap.

11 analysts, 13 scored items. 8 Well Answered / 5 Deflected. Street used the hour to pressure-test the $115B FY27 / $230B FY28 AI trajectory — supply vs demand, LPS, $/GW vs NVIDIA, and the Apollo/Blackstone XPV residual-value backstop. Zero questions on software/VMware, non-AI, buybacks, tax, or customer-mix %.

Hock quantified $350B of AI semis over FY27–28 and $20–30B content per GW. Amie refused to size XPV residual-value guarantees or say how much of 15 GW is Broadcom-financed.

Pattern. Hock will quantify product and GW math. Amie / Hock will not quantify Broadcom’s role as financier (RVG max, XPV share of 15 GW). Charlie will not give networking attach rates. Underwrite the semiconductor trajectory; treat XPV quantum, Ultra attach, and Google mix as unanswered until the 10-Q / next print.

Five genuine contradictions — two high, three medium. Labeled guidance raises ($56B → $58B; $100B+ → $115B; new $230B) are not contradictions.

Use 67% semiconductor GM, not 76%. Model GM dilution. Do not linearize disclosed GW × $20–30B. Rack vs chip is not settled. “Modest” RVG vs $29B needs an IR bridge before treating XPV as off-balance-sheet and low-risk.

This is a demand-and-ecosystem call, not a classic macro call. Zero comments on FOMC, CPI, China, or export controls. The industrial backdrop is land/power/shell + HBM + substrates + lasers, not the Fed.

One-line takeaway. Frontier labs (Anthropic then OpenAI) become AVGO’s largest XPU accounts through 2028 at claimed half-GPU cost; Tomahawk still rides NVIDIA GPU clusters; land/power/HBM/lasers — not orders — cap the doubling to $115 billion / $230 billion.

Wireless is the only soft print (non-AI +5% YoY, wireless down). Consumer silicon is not participating. Pillar Two is now an explicit 16% non-GAAP tax floor.

Customer mix is rotating inside a doubling pie. Google stays strategic; Anthropic is guided to be largest XPU customer in 2027–28; OpenAI #2; Meta a 3 GW sidecar. That is a lab-first, not hyperscaler-only, AI cycle.