Financebotresearch desk研究台

AVGO › analyze

AVGO · Analyze from before

ACCUMULATE Semiconductors

A portfolio-specific passage was removed from the public build.

1. What This Updates

The baseline was written six days before the event it was built entirely around: Q3 FY2026 earnings, reported September 2, 2026. That print has now happened, the stock has moved through it, and the Street has repriced. This pass exists to test the baseline's own named break trigger (AI semi revenue <$10B) against real numbers for the first time, and to find out whether the two risk vectors it flagged — Google/Marvell dual-sourcing and the Anthropic SPV credit exposure — hardened into something worse or resolved.

Event list since August 27: 1. Q3 FY2026 results (Sep 2): revenue $29.59B (+86% YoY, beat), GAAP EPS $2.68 (net income $13.1B, more than tripled YoY), non-GAAP EPS $3.32 (beat the ~$3.24 Zacks consensus by ~2.5%), free cash flow $13.7B (46% of revenue, a record), capex just $532M, cash and equivalents up to $24.0B. AI semiconductor revenue $16.7B, +221% YoY, +54% QoQ. 2. Q4 FY2026 guidance: total revenue ~$34.8B vs $35.03B Street consensus (a ~0.7% miss on the total line) — but AI semiconductor revenue guided to $21.7B, +236% YoY, accelerating, not decelerating. The guidance miss sits entirely in the non-AI mix (Infrastructure Software $8.8B/+29% YoY and non-AI semiconductor), not in the segment the thesis is actually about. 3. Stock reaction: fell ~3.5-6% in the days after the print, intraday lows reported as low as the mid-$340s, before settling around $360.83 today — a guidance-driven pullback, not a fundamentals-driven one; Bernstein's Stacy Rasgon and most sell-side coverage explicitly framed the underlying quarter as strong. 4. Analyst reaction, post-print (12 actions found, Sep 3-10): 1 initiation (Piper Sandler, Overweight, $460), 4 raises (Citi $500→$515, Raymond James $450→$475, Morgan Stanley $502→$505, Rosenblatt $500→$600, BMO $455→$575, Cantor $525→$600), 3 modest trims while staying Buy/Outperform (Evercore $582→$578, TD Cowen $500→$475, Truist $550→$520), 1 upgrade (Macquarie Neutral→Outperform, $490 — the same firm the baseline noted had downgraded in June), 1 cut to Neutral unchanged rating with a lower target (DA Davidson $400→$350). Net effect: the aggregate mean target rose ($525.97 → $533.41) even as the stock price fell — the classic signature of the market repricing the multiple down while the Street prices the business up. 5. AI revenue targets extended and raised: FY2026 AI semiconductor revenue guide raised to ~$58B (from $56B at baseline). FY2027 target specified at $115B (was ">$100B", now a number). FY2028 target given for the first time: $230B — roughly 4x this year's figure. CEO Tan told investors actual customer demand already exceeds the $115B FY2027 number — the binding constraint is packaging/HBM supply, not orders. 6. OpenAI's Broadcom-built "Jalapeño" chip moved from rumor to public specification, with a stated target of >5GW of deployment by 2028 — the baseline's "4th hyperscaler substantively confirmed" claim is now fully public, and reporting now describes six named or credibly-rumored custom-silicon relationships (Google, Meta, OpenAI, Anthropic, plus Apple/ByteDance unconfirmed). 7. The Anthropic SPV credit exposure moved from rumored to company-disclosed. Broadcom itself now states a maximum credit exposure of ~$29B from backstopping Anthropic's lease obligations on the $35B TPU-financing SPV (Apollo/Blackstone) over 5-year terms. Broadcom holds no first lien on the racks; recovery on custom XPUs in a default is explicitly less certain than on standardized GPU clusters. The CFO declined on the call to generalize beyond this one, disclosed number — future tranches are "evaluated deal by deal." Separately, multiple outlets report Broadcom in talks for another $60-70B facility that could bring the total program to ~$100B; none of that incremental exposure has been quantified by the company. 8. New, single-source risk: OpenAI is reportedly working with Samsung on next-generation chips — role unspecified (memory, fab, design, or a future competing design win). Structurally this is the same shape as the Google/Marvell story six weeks ago: an early second-source signal at a customer the thesis leans on. 9. Data-quality finding, for the record: Yahoo's aggregate freeCashflow field (TTM) now reads $30.5B, while the bottom-up sum of the four most recent quarterly cash-flow statements (Q4 FY25 through Q3 FY26, all pulled from the same MCP) is $39.4B — a 22% understatement. The quarterly statements individually reconcile with the press release ($13.7B Q3 FCF checks out exactly). This looks like a vendor-aggregate lag or definitional mismatch, not a data-entry error on any single quarter. Flagged for a pitfall-* note; not written here per this pass's scope lock.

2. The Delta Ledger

Leading with what changed in kind, not just in number:

# Claim (baseline, 2026-08-27) Type Status Evidence
A AI semi revenue break trigger: <$10B for two consecutive quarters fires the bear case Judgment ✅ RESOLVED — NOT FIRED, thesis-confirming Q3 actual $16.7B, Q4 guide $21.7B — both roughly 2x the trigger threshold, accelerating QoQ (+54%) not decelerating. This was the single sharpest falsification test the baseline set, and it cleared decisively. Closed out; no longer a live trigger this cycle.
B SPV/Anthropic exposure: "flagged UNTESTED," baseline's own upgrade condition hoped for "credit-enhancement optics rather than a direct guarantee" State/Structural 🔁 REFRESHED, unfavorably Now company-disclosed: $29B maximum credit exposure, no first lien, uncertain recovery on custom silicon — a real backstop, not optics. The baseline's own upgrade condition for this row did not resolve favorably; it resolved as "yes, it's real," just quantified and (for now) bounded. A further $60-70B facility is reported but not yet company-confirmed or sized.
C TTM FCF ~$32.76B, "+21.7%" trajectory; FY2026E $42-46B (base case) Trend 🔁 REFRESHED, favorably Q3 FCF $13.7B (46% margin) replaces the weakest quarter in the old TTM window. New TTM FCF ≈ $39.4B (bottom-up quarterly sum; see the vendor-field discrepancy noted above — do not use Yahoo's aggregate field for this name right now). Nine months of FY2026 FCF already total $31.97B; at the current ~46% FCF margin applied to the Q4 revenue guide, full-year FY2026 FCF lands near $48B — above the baseline's own base case, near its bull case.
D Net Debt/EBITDA ~1.08x State 🔁 REFRESHED, favorably Debt fell to $59.42B (from ~$64.9B), cash rose to $23.98B (from $19.63B at Q2), TTM EBITDA rose to $52.06B on the print. Net Debt/EBITDA ≈ 0.68x — materially better than 1.08x, and the fastest single-quarter improvement in the series so far.
E Analyst consensus "Strong Buy," mean target $525.97 State 🔁 REFRESHED, favorably Mean target now $533.41, 46 analysts, "1.2 – Strong Buy." Net of the post-print actions (see event #4), targets moved up on balance despite the price falling — a re-rating-setup signature (see §3).
F Row 25 (baseline): 4th hyperscaler "substantively confirmed," reportedly OpenAI Judgment 🔁 REFRESHED, fully confirmed and extended OpenAI's Jalapeño chip is now public (unveiled June, detailed further at/after the print), with a stated >5GW 2028 deployment target. Reporting now counts six named/rumored custom-silicon relationships against the baseline's "four."
G Row 17 (baseline): XPU lock-in DRIFTED — Google dual-sourced attach silicon to Marvell Structural ✅ CARRIED at DRIFTED, unchanged The 8/19 SEC filing detail is now corroborated by trade press: Broadcom retains core ASIC design, advanced packaging, HBM integration, SerDes and switching; Marvell has the attach layer (inference accelerators, storage/NIC/memory controllers). No further erosion since baseline — Broadcom+Marvell together are described as ~95% of the custom-AI-ASIC co-design market, and the deal is read by trade press as evidence the TPU market is growing, not that Broadcom is being displaced. Held at DRIFTED rather than upgraded back to CARRIED because the dual-sourcing precedent itself doesn't reverse; held rather than downgraded further because nothing new eroded the core-ASIC role. Same break level as before: Marvell winning core TPU ASIC work, not attach silicon.
H AI revenue guidance ladder: FY2026 $56B, FY2027 ">$100B" State 🔁 REFRESHED, favorably FY2026 raised to ~$58B. FY2027 specified at $115B (a number, not a floor). FY2028 given for the first time: $230B. Management states actual demand already exceeds the FY2027 figure — packaging/HBM supply, not orders, is the binding constraint. This is new information with no baseline counterpart on the FY2028 point.
I Row 19 (baseline): VMware/Infrastructure Software, UNTESTED both prior passes Trend 🔁 REFRESHED (partial) Infrastructure Software segment (which includes VMware) reported $8.8B, +29% YoY in Q3 — real evidence of continued growth, sourced from the primary press release. The specific 77% op-margin figure was not independently re-verified this pass; that sub-claim stays UNTESTED. Two consecutive passes without the margin figure being re-checked — flag for the next pass rather than defer a third time.
J Row 5 (baseline): ROIC ex-goodwill ~47% (FY2025 basis) State ⏳ UNTESTED, second consecutive pass No new annual print (FY2026 year-end is still ~7 weeks out). This is the second straight differential pass carrying this forward unverified — noting it explicitly per protocol rather than letting it quietly persist a third time.
K NEW: OpenAI reportedly evaluating Samsung for next-generation chips Structural risk 🆕 NEW, embryonic, single-source Role unspecified (memory/fab/design/competing design). Structurally the same shape as the Google/Marvell story at an earlier stage — worth naming now so it isn't a surprise later, not yet weighted as a material risk.
L Analyst qualitative pushback (Moorhead/Moor Insights): bookings-to-revenue conversion, AI networking mix, first-gen XPU yields, credibility of the FY2027 revenue bridge Judgment 🆕 NEW, single-source, low-weight No hard numbers attached; recorded as a standing qualitative headwind, not a finding. Explicitly: "does not prove broad Nvidia displacement."
M Row 26 upgrade condition: "Marvell's Google win stays confined to attach silicon through at least two more quarters" Judgment ✅ CARRIED, condition still holding One quarter in (not two), and holding — see row G.
N Row 21 (baseline): sector-wide AI-capex repricing (SOX -28.6% peak-to-trough) as market backdrop Price/context ✅ CARRIED The post-print pullback (guide-miss-driven, not fundamentals-driven) is the same phenomenon continuing: the market is pricing AI-capex risk more aggressively than AVGO's own disclosed numbers currently justify.

Carried without material change (compact list): dividend CAGR trajectory (row 12, no new annual data); CEO Tan's routine 10b5-1 selling pattern and no fresh insider cluster activity since the July transactions the baseline already saw (row 14 — the insider-transaction feed shows nothing dated after 2026-07-10, i.e., nothing new to report either direction); institutional positioning stable-to-growing (row 16, not independently re-pulled this pass); CFO transition settled and uneventful (row 15); AI Networking Silicon/Tomahawk as a second lock-in layer, unchallenged (row 18); process leadership/TSMC relationship (row 20); Scenario D — Nvidia total-cost undercut still not a near-term threat (row 22); entry zone $360-420 and strong-buy-under-$360 (row 3 — market-tested a third time, including an intraday dip into the mid-$340s on the post-print selloff, which if anything strengthens the case that the zone is real and gets bought).

Ledger count: 14 explicit rows this pass (A-N) + 9 carried-unchanged items named above. Of the 14: 1 resolved/closed (A), 6 refreshed favorably (C, D, E, F, H, I), 1 refreshed unfavorably (B), 2 carried at prior drifted/held status (G, M), 2 new (K, L), 1 carried context (N), 1 untested a second time (J).

3. How the Close Calls Were Decided

Row B (SPV exposure) vs. the print's fundamentals — this is the central judgment of this pass. The baseline set an explicit upgrade condition: primary-source confirmation that the exposure is "credit-enhancement optics rather than a direct guarantee." That confirmation arrived, and it resolved in the unfavorable direction — Broadcom itself discloses a real $29B backstop with no first lien and uncertain recovery value on custom silicon. Weighed against that: the number is disclosed (not rumored), bounded (the company explicitly would not generalize it to future tranches), and small relative to what the business now throws off — $29B of contingent exposure against $48B of expected FY2026 FCF and $24B of cash on hand. One financial-press analysis characterized it as "primarily a valuation-multiple risk rather than an immediate earnings problem — unless future tranches require increasingly large Broadcom backstops." That conditional is exactly right and exactly why this is refreshed, not retracted or escalated to a break trigger: the risk is real and now quantified, but it has not yet grown to a size that threatens the balance sheet, and the next test is whether the reported $60-100B expansion actually happens and on what terms.

Row G (Google/Marvell) — holding at DRIFTED rather than moving either direction. New reporting corroborates the division of labor (Broadcom keeps the core ASIC, packaging, HBM, SerDes, switching; Marvell has the attach layer) but adds no new erosion. Per the framework's own instruction from the baseline, the level that would force a further downgrade — Marvell winning core TPU ASIC work, not attach silicon — has not happened. Moving this back to CARRIED would overstate the resolution; moving it further down would overstate a non-event. DRIFTED, unchanged, is the honest read.

Row C/D (FCF, leverage) vs. Row N (sector sentiment) — the re-rating-setup pattern. Fundamentals (Trend/State rows) improved by more than the price did; the price fell. This is not a contradiction to resolve — it is the textbook shape the framework calls out when it says a big price move against high thesis persistence is "the classic re-rating setup: the business held and the multiple moved." The analyst-target move (mean target up while price down, row E) independently corroborates that this is a sentiment/multiple event on the total-revenue guide miss, not a reassessment of the AI-semiconductor engine the thesis is actually about.

Row A (break trigger) — adjudicated cleanly, no close call. The baseline set a specific, falsifiable number six days before the data existed to test it. The number cleared by roughly 2x on both the reported quarter and the guide. This is exactly the kind of test this command exists to make honest: it doesn't get to quietly disappear now that it's favorable, any more than it would have been ignored if it had fired.

4. Thesis Persistence and Conviction Delta

Persistence: Of the rows that carry the verdict (A through N, treating the 9 carried-unchanged items as fully persistent), the count is: 1 resolved favorably, 6 refreshed favorably, 1 refreshed unfavorably, 2 held at a previously-downgraded status (not further eroded), 1 untested a second time, 2 new, 9 carried clean. Excluding Judgment rows (A, L, M) and the untested row (J) from the denominator, 15 of 19 Structural/Trend/State rows moved favorably or held; 1 moved unfavorably; 2 are new. This is higher clean persistence than the baseline's own June→August pass (~67-75%), and unusually one-directional for a differential update — most of what changed, changed in the thesis's favor, with the SPV exposure as the one real offsetting force.

Conviction: 8.0 → 8.3. Driven specifically by: - Up: Row A (the sharpest bear test on the table cleared by ~2x), Row C (FCF trajectory running ahead of the baseline's own base case, near its bull case, on a single quarter), Row D (leverage improved faster than the baseline's own dollar-pace projection implied), Row H (FY2026 guide raised, FY2027 specified, FY2028 given for the first time, demand explicitly exceeding the FY2027 target), Row F (the "rumored 4th hyperscaler" is now a named, public, specified relationship, and the count of relationships is now six, not four), Row E (the Street raised its aggregate target into a falling stock price — an external, independent signal pointing the same direction as the internal numbers). - Down (the offset that keeps this at 8.3 rather than higher): Row B (the SPV exposure is now confirmed real and quantified at $29B, not resolved away as hoped), and Row K (a new, early-stage OpenAI/Samsung second-source signal that has the same shape as a risk that already partially materialized once this year). - Unchanged weight: Row G (Google/Marvell) neither helped nor hurt this pass — it is exactly where the baseline left it.

No row was RETRACTED. Nothing in the baseline was wrong when written — the two things that moved unfavorably (Row B, and the new Row K) are genuinely new information arriving after the baseline's date, not corrections to it.

5. What Is Genuinely New

  • The $230B FY2028 AI revenue target (Row H) — a full additional year of management-disclosed visibility that did not exist at baseline.
  • The $29B disclosed maximum SPV credit exposure (Row B) — converts a rumor into a number, in the unfavorable direction relative to what the baseline hoped for.
  • OpenAI/Samsung as a second-source signal (Row K) — no baseline counterpart, structurally analogous to the Google/Marvell story.
  • The Yahoo freeCashflow vendor-field discrepancy (event #9) — a data-quality finding with no baseline counterpart; worth a pitfall-* note centrally.
  • The Beta-position share-count discrepancy (22sh vs. the baseline's stated 52sh) — flagged, not resolved, out of this file's scope.

6. Updated Verdict

ACCUMULATE at $360.83 — at the bottom edge of the validated $360-420 zone, three months and one full earnings cycle after the baseline's window opened. Conviction [8.3] (was [8.0]).

The print did what the baseline's break trigger was built to test, and the thesis passed: AI semiconductor revenue is accelerating (not decelerating), free cash flow beat the baseline's own base case, leverage improved faster than projected, and the Street raised its aggregate price target into a falling stock — the signature of a multiple compressing under a business that is still executing. The offsetting, genuinely new risk is that the Anthropic SPV credit exposure is now a real, company-disclosed $29B backstop rather than the unverified rumor the baseline flagged, and a second early-stage customer-diversification signal (OpenAI/Samsung) has appeared in the same shape as the Google/Marvell story from six weeks ago. Neither is large enough yet to threaten the thesis; both are the correct things to watch into the next print.

Valuation (re-derived from scratch, Price rows never carry forward)

Sanity check — FY2026E, now ~90% actual. Nine months of FY2026 FCF are booked ($31.97B); at the ~46% margin the last two quarters have held, the Q4 guide ($34.8B revenue) implies full-year FCF near $48B. Applying a 33-38x EV/FCF band (the baseline's own base-case multiple, unchanged) and netting today's actual net debt ($35.44B) against 4.758B shares outstanding gives an EV/FCF-implied equity value of roughly $330-400/share — a range spot ($360.83) already sits inside. This says the current fiscal year is fairly priced, not cheap and not expensive — consistent with a business trading in line with results that already happened.

Forward-looking, FY2027E (the more decision-relevant horizon, consistent with how the baseline set its trim):

Scenario FY2027E revenue FY2027E FCF EV/FCF multiple Implied equity value Implied Price/Share
Bear (Marvell/OpenAI-Samsung erosion, or an AI-capex pause) ~$140B ~$54-56B 22-25x ~$1.14-1.34T $247-282
Base (management's $115B AI target broadly met) ~$165-167B ~$73-76B 28-32x ~$2.05-2.35T $431-494
Bull (demand exceeds target, as management already says it does) ~$178-182B ~$84-87B 34-38x ~$2.87-3.21T $603-675

Assumptions stated plainly: non-AI revenue (Infrastructure Software + non-AI semiconductor, ~$48B in FY2026) grows modestly (~5-8%) across scenarios; FCF margin holds near the current ~44-46% in base/bull and compresses to ~38-40% in bear on rising packaging capex and SPV-related costs; multiples are held roughly in line with the baseline's own bands, narrowed slightly at the top to price in the now-confirmed SPV exposure. This is a scenario range, not a point estimate, and the AI-revenue inputs are management's own disclosed targets, not an independent forecast.

Fair value: $425-500 (base case), a modest raise from the baseline's $390-480 — justified by the FCF beat, the FY2028 target extending visibility a year further out, and the Street's own target increase, tempered from raising further by the now-quantified SPV exposure and the embryonic OpenAI/Samsung signal.

  • Graham's Number: Still not meaningful — GAAP book value per share is now positive ($20.92) and produces a nominal Graham figure of ~$60.70, but $97.8B of FY2025's $81.29B equity is goodwill from M&A (VMware, CA, Symantec), so the figure doesn't reflect real net-asset backing. Unchanged conclusion from both prior passes; the exact tangible-book-value dollar figure was not independently re-verified this pass (no new annual print).
  • Bogle: 0.71% yield (flat vs. 0.73% at baseline) + FY2026→FY2027 EPS growth (~66.5% on consensus, even more front-loaded than baseline's estimate) — still not a clean single-year input given the ramp's lumpiness; directional support only, unchanged conclusion.
  • DYT: 0.71% current yield vs. 1.76% 5yr average — still reads "expensive," barely moved from baseline (0.73% vs 1.80%). Same structural distortion (VMware-driven equity growth vs. price appreciation corrupts the comparison window) — yellow flag, unchanged, not a sell signal.

  • DDM: Still not applicable — growth assumptions exceed any reasonable discount rate.

  • Entry: $360-420 (carried, now market-tested a third time, including an intraday dip into the mid-$340s on the post-print selloff — the strong-buy-under-$360 sub-zone was briefly touched and bought). Spot $360.83 sits essentially at the floor of the zone.

  • Strong buy: <$360.
  • Trim: 25x fwd (carried, unchanged multiple) — epsForward (FY2027 consensus) is $19.38, barely moved from $19.50 at baseline, so the dollar level recomputes to ≈$484.5 today, down slightly on the EPS estimate alone. Held unchanged rather than raised despite the favorable fundamentals, because it already sits close to the top of this pass's own re-derived base case ($494) — raising it further would require the Street's FY2027 consensus EPS to actually move, which it has not done yet even after the beat-and-raise. Re-derive next pass once FY2027 consensus estimates have had a full quarter to react to the $115B/$230B framework.

Break Triggers (carried, one resolved, one added)

  • ~~AI semiconductor revenue below $10B for two consecutive quarters~~ — resolved, did not fire. Removed as a live trigger; the relevant forward-looking test now is whether Q4's $21.7B guide is met.
  • Google or Meta ends its XPU partnership outright — not fired; watch for Marvell winning core TPU ASIC work (not attach silicon) as the lower-bar warning sign, unchanged from baseline.
  • VMware/Infrastructure Software ARR declines YoY — not fired (+29% YoY this quarter), but the specific margin sub-claim remains unverified for a second pass.
  • Net Debt/EBITDA rises above 3x — currently ~0.68x, wide berth, improving.
  • FY2027 AI revenue guidance ($115B) formally withdrawn or cut.
  • New: the reported $60-70B additional SPV facility is confirmed and sized by the company at a level that materially exceeds the disclosed $29B current exposure, or Broadcom's own credit rating agencies flag the program explicitly (CDS has reportedly already widened; watch for a formal rating-agency comment as the harder confirmation).
  • New: OpenAI's Samsung relationship is clarified as a competing design win (not memory/fab supply) for a successor to Jalapeño.

Upgrade Conditions

  • Q4 FY26 print clears the $21.7B AI semiconductor guide with FY2027 $115B reaffirmed or raised.
  • FY2027 consensus EPS estimates move up meaningfully in response to the $115B/$230B framework (would justify re-deriving the trim multiple upward).
  • The additional SPV facility either doesn't materialize at the reported $60-70B scale, or materializes with Broadcom's exposure capped well below the current $29B/$35B-deal ratio.

7. What This Pass Did NOT Test

  • Row J (ROIC ex-goodwill, ~47% FY2025 basis) — untested for a second consecutive pass; no new annual print exists to test it against. Flagging this explicitly per protocol rather than letting it quietly persist a third time without comment.
  • VMware's specific 77% operating-margin figure — the segment's revenue growth (+29% YoY) was refreshed this pass, but the margin figure itself was not independently re-verified.
  • The reported additional $60-70B SPV facility — reported by multiple outlets but not confirmed or sized by the company; only the original $35B deal's $29B max exposure is company-disclosed.
  • The OpenAI/Samsung relationship's actual scope — single-source, no primary confirmation of what Samsung's role would be.
  • The exact new total-debt and cash balance-sheet detail for Q3 (beyond the $59.42B debt / $23.98B/$24.0B cash figures Yahoo's aggregate and the press release both independently support) — a full Q3 10-Q balance-sheet read (debt maturity schedule, finance-lease balances) was not performed this pass; flagged as the next natural extension given [[pitfall-lease-financed-capex-hides-from-the-fcf-line]] and the SPV-adjacent lease structures now confirmed to exist in this business.
  • The Beta-account share-count discrepancy noted in the header — read, not reconciled; outside this file's scope.

Sources