Skip to content
Bullishmedium conviction · 1 yearActive · reviewed 1 day ago

Custom chips take share and NVIDIA keeps growing: the pie is bigger than the fight

By MenFem Editorial·NVIDIA·Microsoft·Broadcom·Alphabet·Amazon (AWS)·Meta Platforms·Marvell Technology·Semiconductors — Custom silicon·10 November 2026·Methodology·
ai-infrastructuresemiconductorscustom-siliconthematic
The next review, the day it is written.
Share

What would prove this wrong · verbatim, as published

The reason for holding this stops being true if custom chips stop gaining: Broadcom's AI semiconductor revenue growing more slowly, year on year, than NVIDIA's data-centre revenue for two consecutive quarters. It also stops being true if custom dents NVIDIA: NVIDIA's GAAP gross margin below 70% for two consecutive quarters, or its data-centre revenue falling quarter on quarter for two consecutive quarters.
  • leg 1 · Custom stops gaining: Broadcom AI semis growing slower y/y than NVIDIA Data Center for two consecutive quarters · not read yet
  • leg 2 · Custom dents NVIDIA: NVIDIA GAAP gross margin below 70%, or Data Center revenue down q/q, for two consecutive quarters · not read yet

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

Hyperscalers' own chips are growing about twice as fast as NVIDIA's data-centre business, from a fifth of its size. Bullish on the thesis: custom silicon gains share without denting NVIDIA's growth or margin, for now.

What would make it wrong · 2 legs
  1. Leg 1 · Not read yet
    Custom stops gaining: Broadcom AI semis growing slower y/y than NVIDIA Data Center for two consecutive quarters
    Not read at a review yet.
  2. Leg 2 · Not read yet
    Custom dents NVIDIA: NVIDIA GAAP gross margin below 70%, or Data Center revenue down q/q, for two consecutive quarters
    Not read at a review yet.
The window
10 Nov 2026 to 10 Nov 2027
0% gone · 364 days left
The reviews
Not reviewed since it was published
  • Leg 1, not read yet: Custom stops gaining: Broadcom AI semis growing slower y/y than NVIDIA Data Center for two consecutive quarters
  • Leg 2, not read yet: Custom dents NVIDIA: NVIDIA GAAP gross margin below 70%, or Data Center revenue down q/q, for two consecutive quarters
  • NVIDIA Q3 FY27 results (17 Nov 2026)
  • Broadcom Q4 FY26 results (Dec 2026)
  • AWS re:Invent (30 Nov–4 Dec 2026)
  • high severity: A demand pause
  • medium severity: No direct share data
  • medium severity: Custom stalls

The timeline · 1 entry · newest firstreviews · notes · moves · the book

  1. 10 Nov 2026
    Opened

    Custom chips take share and NVIDIA keeps growing: the pie is bigger than the fight

Key Points

  • Broadcom AI semiconductor revenue $16.7B (+221%); next-quarter guide $21.7B; backlog $179.2B.
  • NVIDIA data-centre revenue $89.0B (+117%); gross margin 75.0% (from 72.4%); hyperscale customers +102%.
  • Google sells TPU systems externally, a meaningful contributor to Cloud's 103% growth; no TPU revenue disclosed.
  • Proxy: Broadcom's AI revenue growth against NVIDIA's data-centre growth (Broadcom's figure includes networking).

The usual question about the AI chips Google, Amazon, Meta, Microsoft and OpenAI design for themselves is whether they will take NVIDIA's business. The 2026 filings say both are growing, and that the custom side is growing faster. Broadcom, which co-designs most of these chips, reported AI semiconductor revenue of $16.7B in its quarter to 2 August, up 221%, and guided the next quarter to $21.7B; its backlog of contracted revenue is $179.2B. Google now sells its TPU systems to other companies and says they were a meaningful contributor to Google Cloud's 103% growth. NVIDIA, over almost the same quarter, reported data-centre revenue of $89.0B, up 117%, and its gross margin rose to 75.0% from 72.4%. Its hyperscale customers, the ones with chips of their own, grew 102%, and Google is among the first buyers of its new Vera Rubin systems. The claim is that this holds for at least the next year: custom chips keep taking a rising share of AI chip spending, while the market grows fast enough that NVIDIA's revenue keeps rising and its margin holds. Conviction is MEDIUM because no company reports the share directly. Google does not disclose TPU revenue, Amazon's and Microsoft's chip volumes are not in their filings, and Broadcom's AI figure includes networking chips, so this call reads Broadcom's AI revenue growth against NVIDIA's data-centre growth as the best public proxy. NVIDIA reports on 17 November and Broadcom in December. It fits MenFem's calls as they stand: bullish on NVIDIA, Alphabet and Amazon, watching Broadcom and Marvell on price. BULLISH on the thesis, conviction MEDIUM, one year. This view was formed by MenFem's research process at Connor's request and adopted by him.

Bull Case

Every large buyer runs both: custom chips for its own high-volume models, NVIDIA for everything else and for customers. Broadcom's and Google's chip businesses keep compounding while NVIDIA's lead in systems and software keeps its margin near 75%.

Bear Case

What I'd have to believe for the other side: this is a split that cannot last. If custom chips are cheaper per token, every hyperscaler moves volume to them as fast as it can, and NVIDIA's growth and margin both fall once the build slows. Or the reverse: custom programmes slip and NVIDIA's lead widens.

Catalysts

NVIDIA Q3 FY27 results (17 Nov 2026)Earnings

Data-centre and hyperscale revenue, gross margin against the 74.0% guide.

Broadcom Q4 FY26 results (Dec 2026)Earnings

AI semiconductor revenue against the $21.7B guide; backlog.

AWS re:Invent (30 Nov–4 Dec 2026)Other

Trainium volumes and any external chip sales.

Risk factors

No direct share dataMedium

The call reads a proxy; Broadcom's AI figure includes networking.

Custom stallsMedium

A delayed XPU programme or an in-sourced design would slow Broadcom while NVIDIA keeps growing.

A demand pauseHigh

A capex slowdown would hit both sides and change the question.

Related Calls

LONGActive

Optics takes the new links, copper keeps the short ones: both grow through 2027

Co-packaged optics is arriving on schedule while copper cables inside the rack keep growing. Bullish on the thesis at medium conviction: the rack is big enough for both through 2027, and neither side's story of taking the other's place holds yet.

Coherent PhotonLink (CPO/NPO): >$100M in the March 2027 quarter, $250M a quarter before end-2027 (4 Nov slides).

Lumentum: >$5B of CPO/NPO commitments; Marvell to buy Lumentum (all-stock, ~$115B, 2 Nov), citing copper's limits.

ai-infrastructuresemiconductorsoptics
NVIDIABroadcomAstera LabsCoherent CorpCredo TechnologyLumentumMarvell TechnologyArista Networks
LONGActive

Memory is a rent for now: the contracts with price floors outlast this cycle's peak

Micron's revenue is up nearly fivefold in a year at an 87% gross margin, and its contracts carry price floors it says beat any past peak. Bullish on the thesis at low conviction: the floors hold margins above the old record through 2027.

Micron FQ4 FY26: revenue $54.2B (from $11.3B a year earlier); GAAP gross margin 86.8% (from 44.7%).

FQ1 FY27 guide: $61.5B at ~86% gross margin.

ai-infrastructuresemiconductorsmemory
NVIDIAMicronSK HynixLam ResearchSamsung Electronics
LONGActive

The old GPUs get cheap first: H100 rents fall as the new chips arrive, the newest hold

Every new NVIDIA generation pushes the last one down the price curve. Bullish on the thesis at low conviction: H100 rents fall over the next year while the newest chips hold, which hurts renters with spot exposure and spares those on contract.

MenFem vast.ai offers, four-week averages to 6 Nov: H100 SXM $2.38/hr (from $2.97 on 25 Sep); B200 $9.69/hr (from $9.38).

Azure list: H100 SXM spot $2.27 vs on-demand $12.29; unchanged for six weeks.

ai-infrastructurecompute-pricinggpu
NVIDIAOracleCoreWeaveNebius GroupCore ScientificIREN