NVDA + AVGO: Why I Track Them as an AI Infrastructure Pair
Contents
TL;DR
- I'm Jim Liu β I've tracked NVDA and AVGO since 2024-08 as a paired AI infrastructure position, not two separate bets
- NVDA covers the GPU compute monopoly; AVGO covers custom AI accelerators (Google XPU, Meta MTIA) plus the 800G network switching layer β together they cover different failure modes of the same infrastructure buildout
- The 30% withholding asymmetry is real: NVDA pays $0.16/yr per share (negligible), AVGO pays ~$2.36/yr per share β for a $10K position in AVGO, that's roughly $35/yr lost to withholding
- I currently run a ~60/40 NVDA/AVGO split and rebalance when one drifts more than 30% relative to the other
- This isn't investment advice β I get things wrong. But the pairing logic is more honest about uncertainty than picking one winner
Who I Am, and Why I Think in Infrastructure Pairs
I'm Jim Liu, a developer based in Sydney. I run LRTS (lowrisktradesmart.org), covering US and HK investment mechanics for retail investors in Hong Kong, Taiwan, and mainland China.
I've held US tech positions through an IBKR HK account since mid-2024. I'm not a professional manager β I track maybe 8 stocks and 3 ETFs. More than that, and analysis turns into noise.
The pairing idea came from a mistake. In early 2024, I sold most of my NVDA at around $680 (pre-split, roughly $68 adjusted) thinking the valuation was stretched. It kept going. I added AVGO instead at ~$140 post-split. Both ran. But what mattered more than which one I held was the ratio between them β that ratio is what I've spent the past year actually managing.
This article is the pairing logic I wish I'd had before I started.
The AI Infrastructure Duopoly: What NVDA and AVGO Actually Do
Most coverage frames them as competitors. They're not, at least not cleanly.
NVIDIA (NVDA) dominates the GPU training layer. H100, H200, Blackwell β plus the CUDA software ecosystem that maybe 90% of AI researchers use. The moat isn't just hardware; retraining teams away from CUDA takes years. When AWS, Azure, or Google expand AI capacity, NVDA is usually the default.
Broadcom (AVGO) plays three roles at once:
- Custom AI accelerators (XPUs): Google's TPU successors, Meta's MTIA, Apple's custom Neural Engine derivatives β AVGO designs these. These are nominally NVDA competitors, but AVGO gets the design revenue regardless of which architecture wins.
- Network switching: Tomahawk 5 and Trident series chips carry data between GPUs in large clusters. Multi-thousand-GPU training runs need Ethernet at 800G+. AVGO has roughly 60-70% share of AI cluster networking.
- VMware infrastructure: The 2023 acquisition added $4B+ in recurring software revenue β traditional IT rather than AI, but a revenue floor.
Together: NVDA owns the compute stack. AVGO owns the networking connective tissue and captures design revenue even from XPUs built to displace NVDA.
This isn't a hedge in the strict sense. An AI capex winter hits both. But they have different failure modes, which is the useful part.
The Asymmetry: Four Numbers, Two for Each Stock
I don't track them identically. Each stock has two numbers I watch.
For NVDA:
- Data center revenue (quarterly, absolute $B): The entire thesis lives or dies here. If it plateaus or misses consensus by more than 5%, I reassess. FY2025 Q4 was running ~$35B/quarter.
- Gross margin (%): NVDA's pricing power shows up here. They've held 73-75% GM through recent quarters. If it drops below 70% consistently, commoditization has started.
For AVGO:
- AI revenue run rate (quarterly β growth rate matters more than level): AVGO began disclosing AI-specific revenue in late 2023. As of FY2026 Q1, it was tracking ~$4.1B/quarter from AI networking plus custom chips. The year-on-year growth rate is what I watch.
- Dividend growth rate (%): AVGO has raised its dividend for 13 consecutive years. The rate of increase tells me more about management confidence than any forward guidance does.
I don't watch PE ratios on either. NVDA's forward PE usually runs 40-60x β absurd until you remember earnings grew roughly 10x in two years. AVGO's is lower at 30-40x but harder to interpret during the VMware integration. Neither number tells me much about timing.
Why a Pair: Allocation Math for HK/TW Investors
Here's how I sized it. This is a worked example, not a recommendation.
In late 2024, I had about HK$150,000 in my US-stock allocation. I put roughly 25% (~HK$37,500, ~US$4,800) into the NVDA+AVGO pair:
- NVDA: US$2,880 (~28 shares at ~$104)
- AVGO: US$1,920 (~10 shares at ~$192)
Why 60/40 and not 50/50? NVDA has the higher earnings growth rate β it gets the larger weight. AVGO gets 40% because of the meaningful dividend income component at this allocation size, and because the custom silicon exposure hedges part of the NVDA displacement risk.
Rebalancing rule: when one drifts more than 30% relative to the other β say NVDA swells to 75% of the pair β I trim the winner and add to the laggard. I've done this once, trimming NVDA in Q1 2025 and adding AVGO.
For HK investors on IBKR: US stocks settle T+2 in USD, no stamp duty, minimum commission US$1 per trade. Two rebalancing trades = $2 total. Compared to custodian brokerage (εΎ©ε§θ¨) at $18-20 minimum, rebalancing a paired position with IBKR is ~10x cheaper.
For Taiwan investors: Firstrade or IBKR Taiwan work similarly.
The 30% Withholding Reality: NVDA vs AVGO Side-by-Side
HK and Taiwan retail investors have no double-tax treaty with the US, so dividends from US stocks are withheld at the full 30% rate.
NVDA: $0.04/quarter = $0.16/yr per share. At ~$100/share, that's a 0.16% yield. 30% of $0.16 = $0.048/share/yr in withholding. For a $10K NVDA position (~100 shares): $4.80/yr. Essentially irrelevant to the investment thesis.
AVGO: ~$0.59/quarter = ~$2.36/yr per share. At ~$200/share, that's a ~1.2% yield. 30% of $2.36 = $0.71/share/yr in withholding. For a $10K AVGO position (~50 shares): ~$35.50/yr.
Not catastrophic, but it factors in if you're income-oriented. The pairing implication: the pair generates modest income via AVGO (after withholding ~0.84% net yield), while NVDA's dividend is negligible. If you're purely growth-oriented, the withholding tax on AVGO is a small drag on an otherwise growth-focused position.
For mainland China investors: the US-PRC tax treaty applies at corporate/fund level, but individual investors holding US stocks directly typically face the full 30% through most offshore brokers. Check your broker's 1042-S documentation for the actual rate applied.
What Could Break the Pair (and What Can't)
This matters more than the bull case.
What breaks NVDA but not AVGO: Google, Meta, and Apple are all building in-house AI chips β TPU v5, MTIA 2, Apple silicon derivatives. If XPUs reach performance parity with NVDA GPUs at lower power draw, hyperscaler GPU purchases slow. AVGO designs those same XPUs, so a bad quarter for NVDA is often a good quarter for AVGO's custom silicon revenue.
What breaks AVGO but not NVDA: NVDA's Mellanox acquisition gave them InfiniBand β a competing networking standard that some high-performance NVDA GPU clusters already use. If AI cluster networking shifts heavily to InfiniBand, Broadcom's Ethernet switch business faces slower growth. NVDA's InfiniBand revenue benefits directly.
What breaks both: An AI capex winter. If hyperscalers pull back datacenter spending β rate environment, regulatory pressure, or a genuine AI revenue shortfall β both stocks fall. The pair doesn't protect you here. This is the correlated tail risk you take on.
What doesn't break either easily: The fundamental AI scaling logic: larger models need more compute. NVDA and AVGO's core revenues don't depend on AI applications being commercially profitable. They depend on the capital commitment to build AI infrastructure. Even unprofitable AI products run expensive training jobs.
FAQ
Should I buy NVDA or AVGO if I can only pick one? If you believe in GPU dominance long-term: NVDA. If you want dividend income plus exposure to the custom AI chip trend: AVGO. I find the either/or framing slightly wrong β the pair is more honest about the uncertainty.
How do I buy NVDA and AVGO from Hong Kong? IBKR HK is the most straightforward. US stock accounts open in 1-2 days, US stocks trade at US$1 minimum commission, no stamp duty. moomoo HK works too at slightly higher minimums. Both are SFC-regulated.
Does the 30% withholding apply to capital gains? No. US withholding only applies to dividends. Capital gains from US stock sales are not taxed at source for non-US investors. HK has no capital gains tax either, so the growth component of both stocks is essentially untaxed for HK investors.
What's Broadcom's actual AI revenue today? As of FY2026 Q1 results, AVGO reported ~$4.1B in AI-related quarterly revenue β XPU design fees plus Tomahawk/Trident AI networking chips. Management guided AI quarterly revenue toward $4.7B+ by end of CY2026.
Is this pair risky for HK/TW retail investors? Yes. Both are USD-denominated, high-valuation growth stocks. FX risk (HKD/TWD vs USD), company-specific risk, and sector concentration all apply. Treat this as a speculative growth allocation, not a stable income vehicle.
How I Track Both Stocks (Methodology)
I use a single spreadsheet updated quarterly:
- NVDA: data center revenue + gross margin from NVDA Investor Relations, updated within 24h of earnings
- AVGO: AI revenue figure + dividend declared from AVGO Investor Relations + earnings call transcript
- Pair weight (NVDA% vs AVGO% of total paired position): updated after each trade
- Running 30% withholding estimate: dividends received Γ 0.30, reconciled against 1042-S forms in February
No Bloomberg terminal, no paid data service. All inputs are freely available on NVDA's and AVGO's investor relations pages.
Affiliate Disclosure: LRTS earns commission from IBKR and moomoo referrals. This article doesn't constitute investment advice. US stocks carry currency risk, company-specific risk, and market risk. Consult a licensed financial advisor before investing.