Skip to content
Sign in
WATCHActiveMed1Y$AVGO$321.68Fri 23 Oct

Broadcom: the right custom-silicon thesis at the wrong price

By MenFem Editorial·Semiconductors·9 June 2026·Methodology·
ai-infrastructuresemiconductorscompute
Share
Broadcom: the right custom-silicon thesis at the wrong price

The argument, taken apart

Turn it, hover a ring, a moon or a weight
One call, taken apart
Rings are the falsifier’s legs. Moons are the catalysts it named. Weights are its risks, longer for higher severity.
The claim

The right custom-silicon thesis at the wrong price — watching the most expensive, lowest-quality way to play it.

What would make it wrong · 2 legs
  1. Leg 1 · Not fired
    The premium never compresses while the thesis compounds
    “Neither limb has fired.”
    read 9 Oct 2026
  2. Leg 2 · Not fired
    A hyperscaler re-bids or in-sources an XPU programme
    “Neither limb has fired.”
    read 9 Oct 2026
The window
9 Jun 2026 to 9 Jun 2027
38% gone · 226 days left
next print 10 Dec 2026
The reviews
Reviewed on 3 days · last 9 Oct 2026
3 Aug · 6 Sept · 9 Oct
Column = the claim, taller for more conviction · rings = falsifier legs, lit when firing · moons = catalysts · weights = risks, longer for higher severity
  • Leg 1, not fired: The premium never compresses while the thesis compounds, read 9 Oct 2026: “Neither limb has fired.”
  • Leg 2, not fired: A hyperscaler re-bids or in-sources an XPU programme, read 9 Oct 2026: “Neither limb has fired.”
  • Q3 FY26 earnings + FY27 AI update
  • Hyperscaler custom-chip deployments
  • high severity: Premium-multiple de-rating
  • medium severity: Single-customer program reset

Key Points

  • The named winner of 'custom silicon eats commodity inference' — AI revenue +143%, FY27 AI guided >$100B.
  • Trades at a premium to NVIDIA on every earnings multiple while earning a third of the ROE.
  • Lower earnings quality: acquisition-amortization add-backs, ~$64B debt, the same hyperscaler customers as NVIDIA.
Price at Call
$392.16
Broadcom

Broadcom is the most dangerous competitor the NVIDIA bulls underestimate — the named beneficiary of the one thing NVIDIA's own disclosures concede it loses, commodity inference to custom silicon, growing AI revenue +143% off a multi-year backlog of hyperscaler design wins. It is also the most expensive and lowest-quality way to own that thesis. Broadcom trades at a premium to NVIDIA on every earnings multiple while earning a third of the return on equity; its non-GAAP earnings are flattered by billions in acquisition amortization; it carries roughly $64B of debt; and it is exposed to the very same hyperscaler capex cycle as NVIDIA, in the same customers. The −13% beat-and-fall in June is the market reminding everyone the premium leaves no room for a merely-excellent quarter. We are WATCHING, not buying: a great business and the right structural thesis, but we would rather own the chokepoints than the levered, premium-priced challenger until the multiple resets or the FY27 AI number de-risks.

Research Log

Reviewed 9 Oct 2026 against the falsifier on this page. Neither limb has fired. AI revenue arrived ahead of guidance (FQ3 AI semiconductor revenue $16.7B, +221%; FY27 AI outlook raised to about $115B, Broadcom release 2 Sep), yet the stock sits 6.3% below the call price ($367.45 on 9 Oct against $392.16), so the waiting has not cost a move. No hyperscaler has re-bid or taken in-house a Broadcom XPU programme. Of the two things this call said it was waiting for, the FY27 AI number has now de-risked. The watch now waits on the other one alone: Broadcom's earnings multiple ceasing to trade at a premium to NVIDIA's while it earns a fraction of NVIDIA's return on equity. The FQ4 print in December is the next check. Insider filings over the last 30 days show one director sale under a 10b5-1 plan adopted in December 2025, and no buys. KEEP as WATCHING, conviction MEDIUM.

Source

Reviewed 2026-09-06. Tape: 392.16 → 357.90 (−8.74%), peak-to-trough −16.33%, vs QQQ −10.71pp. The waiting was right — and its reason has largely dissolved: FQ3 revenue $29.6B (+86%), AI semis $16.7B (+221%), FY27 AI raised to ~$115B (the prior dossier's bull case), FY28 introduced at ~$230B; forward P/E ~20.7× vs NVDA ~19.1×, an ~8% premium where it was at-or-above. The prior dossier's question was answered bearishly on content (~$11–12B revenue per GW, undisputed) and Marvell won a Google custom-chip programme. KEEP as WATCHING; the next review decides whether "the wrong price" still holds.

The most resilient position in the book — which is what a WATCHING call should look like.

Roughly −1.6% since the call was opened, against a book averaging materially worse. This call said the custom-silicon thesis was right and the price was wrong, and took no position accordingly.

Nothing in the Q2 earnings wave changed the custom-silicon argument in either direction. The entry-price question that kept this on the watch list is still open, and the sector de-rating since mid-July has moved it closer to interesting without a company-specific catalyst to force the decision.

Since this call

  1. Broadcom · Exchange offer for up to $654,004,000 of 4.926% Senior Notes due 2037 (issued 18 Apr 2022); expires 5:00pm NYC time on 19 Oct 2026; Broadcom receives no proceeds. · sec.gov
  2. Broadcom · Before (deep-dive-2026-07-07.md): The custom-silicon workflow winner, now with the thesis HARDENED by primary data — a firmly-committed ~$164.6B RPO (Note 2, incl. the OpenAI 10GW contract), ~70% co-design share, 62% semi-segment op margin, and near-zero capital intensity (TSMC carries it). LONG the structural ASIC winner as Nvidia's inference rent-share peaks. The bear case is entirely PRICE + earnings-quality (premium EV/EBITDA ~41x, $64.9B debt, $1.97B/qtr amortization add-back, 42% single-distributor concentration), not the thesis — which is exactly why the seed's stop is a valuation/backlog-slip stop, not a thesis-break stop. After (deep-dive-2026-08-10.md): The thesis got WIDER and the price got WORSE. Apple through 2031 and a Google long-term agreement through 2031 broaden the franchise beyond the single-program risk the last run feared — but the stock ran +13.8% on no new quarter, forward P/E crossed ABOVE Nvidia's, the SOTP gap widened to ~17-31%, and three items the prior dossier got wrong surfaced: Hock Tan conceded Google multi-sourcing on the record, the CFO said XPUs carry LOWER margins than networking, and the "no material regulatory findings" line is void (EU VMware probe + ITC patent investigation). Still LONG the structure. But this is now a fuller-priced LONG than it was five weeks ago, and the prior seed's entire base-case target was consumed in those five weeks. · broadcom
  3. Broadcom · Before (deep-dive-2026-06-09.md): The most dangerous competitor bulls underestimate — the named beneficiary of NVIDIA's own inference-share vulnerability, growing AI +143% — but priced at a PREMIUM to NVIDIA (45× vs 30× EV/EBITDA) on one-third the ROE, with lower earnings quality (acquisition-amortization add-backs, $64B debt) and the same hyperscaler concentration. The most expensive, lowest-quality way to play the custom-silicon thesis. NEUTRAL / WATCHING / MEDIUM. After (deep-dive-2026-07-07.md): The custom-silicon workflow winner, now with the thesis HARDENED by primary data — a firmly-committed ~$164.6B RPO (Note 2, incl. the OpenAI 10GW contract), ~70% co-design share, 62% semi-segment op margin, and near-zero capital intensity (TSMC carries it). LONG the structural ASIC winner as Nvidia's inference rent-share peaks. The bear case is entirely PRICE + earnings-quality (premium EV/EBITDA ~41x, $64.9B debt, $1.97B/qtr amortization add-back, 42% single-distributor concentration), not the thesis — which is exactly why the seed's stop is a valuation/backlog-slip stop, not a thesis-break stop. · broadcom

Bull Case

The named beneficiary of NVIDIA's ceded inference share, growing AI +143% with a $30B+ backlog and a 79%-margin VMware annuity funding it.

Bear Case

A premium-to-NVIDIA multiple on one-third the ROE, with amortization-flattered earnings, ~$64B of debt, and the same hyperscaler concentration — doubly levered to one capex cycle.

What would prove this wrong

A watching call is a claim about the WAITING, so it fails in two directions. The reason for holding it stops being true if the premium never compresses while the thesis compounds anyway — AI revenue arriving as guided with the multiple intact would mean the discipline cost a move that was there to take. It stops being true in the other direction if the structural case breaks: a hyperscaler re-bidding or in-sourcing an XPU programme would end the 'named winner of custom silicon' framing, and there would be nothing left to wait for.

Catalysts

Q3 FY26 earnings + FY27 AI updateEarnings

Whether the FY27 >$100B AI number gets raised, not just reiterated.

Hyperscaler custom-chip deploymentsProduct Launch

OpenAI and Anthropic custom silicon deploying through 2026-2027.

Risk factors

Premium-multiple de-ratingHigh

The −13% June beat-and-fall shows the premium leaves no room for a merely-excellent print.

Single-customer program resetMedium

A hyperscaler re-bidding or in-sourcing one XPU program resets a whole revenue line — the customer owns the stack.

Related Calls

LONG$TSMActive

TSMC: own the bottleneck, not the brand

Own the bottleneck, not the brand: the chokepoint every AI chip depends on, at a discount to all of them.

Fabricates and CoWoS-packages every leading-edge AI accelerator — NVIDIA, AMD, Broadcom, Apple alike.

~22x forward earnings: a discount to NVIDIA, AMD and Broadcom, the names that cannot ship without it.

ai-infrastructuresemiconductorsfoundry
TSMC
High3YSemiconductorsUpdated 3d ago9 Jun
LONG$000660.KSActive

SK Hynix: the largest HBM rent, valued like a Korean cyclical

The largest slice of the HBM scarcity rent, valued like a Korean cyclical — with a Nasdaq listing as the re-rating catalyst.

~62% HBM share and ~two-thirds of NVIDIA's HBM4 — the largest rent-capture in the cluster.

72% operating margin and ~61% ROE, yet a single-digit forward multiple — rent in the earnings, not the multiple.

ai-infrastructuresemiconductorsmemory
SK Hynix
Med1YSemiconductorsUpdated 3d ago9 Jun
LONG$MUActive

Micron: priced like a cyclical at the top, contracted like a utility

Priced like a peak-cycle memory-cyclical, contracted like a utility — a re-rating bet, not an earnings bet.

Entire CY2026 HBM output sold under multi-year fixed price-and-volume contracts.

Overtook Samsung into the #2 HBM slot; power-efficiency edge keeps it in every NVIDIA generation.

ai-infrastructuresemiconductorsmemory
Micron
Med1YSemiconductorsUpdated 3d ago9 Jun

Conviction

Conviction History

Med→Med

3 Aug 2026

Reviewed against the Q2 2026 earnings wave (SK Hynix 29 Jul, Samsung 30 Jul, and the late-July hyperscaler prints). Most resilient position in the book at roughly -1.6%; nothing in the wave changed the custom-silicon argument. Still WATCHING on price — no conviction change.

Weekly closes

entry 392.16

Sources

See all Semiconductors calls