The Custom Silicon Crown: Broadcom vs. Marvell — Who Wins?
The Battle for Custom AI Chips: Broadcom vs. Marvell Explained Simply
TL;DR: Two companies are building the custom "brains" inside the world’s biggest AI data centers. Broadcom (AVGO) is the dominant king with ~70% market share and $100B+ revenue targets. Marvell (MRVL) is the fast-growing challenger at ~20% share, betting on cloud giants wanting a "plan B" supplier.
Quick Read: The Big Picture
- Broadcom commands ~70% of the custom AI chip co-design market with $10.8 billion in quarterly AI revenue
- Marvell holds ~20% as the strategic "dual-source" alternative with accelerating growth
- Hock Tan (Broadcom CEO) projects FY2027 AI revenue above $100 billion across 10 gigawatts of compute for Google, Meta, OpenAI, and Anthropic
- Matt Murphy (Marvell CEO) targets ~$11.5B revenue by FY2027 and ~$16.5B by FY2028, with a $10B custom silicon run rate by FY2029
The Tale of Two Strategies
Broadcom: The "Whole Rack" King
Think of Broadcom as the general contractor who builds the entire building — not just one room.
| What They Sell | Why It Matters |
|---|---|
| Custom XPUs (AI accelerator chips) | The "brains" designed specifically for each customer |
| Tomahawk 6 Ethernet Switches | The "highways" moving data between chips |
| Jericho Fabrics | The "intersections" managing traffic flow |
| 1.6T Co-Packaged Optics | Light-speed connections using lasers instead of copper wires |
| Networking = ~40% of AI Revenue | A moat competitors haven’t matched |
The Master Plan: Broadcom’s AI XPV Platform (with investors Apollo and Blackstone) aims to fund 20+ gigawatts of deployed compute through 2028. That’s enough power for millions of homes — all running AI.
KEY INSIGHT: Once a cloud giant like Google builds their data center around Broadcom’s full stack, switching is incredibly painful. This creates massive customer lock-in.
Marvell: The "Pressure Valve" Challenger
Marvell is the specialist — and that’s exactly why cloud giants need them.
"Cloud giants want a second source." — Matt Murphy, Marvell CEO
If you’re Google or Microsoft, you cannot rely on a single supplier. Marvell is the insurance policy.
| Marvell’s Growth Engines | Target |
|---|---|
| Interconnect Growth | >70% in FY2027 |
| DCI Modules (Data Center Interconnect) | $1B annualized by FY2028 |
| Tier 1 XPU Program | Firm requirements locked in with a top-tier customer |
| Celestial AI & XConn Deals | Photonics & switching for scale-up optics |
| Custom Silicon Run Rate | $10B by FY2029 |
The Bet: If hyperscalers truly dual-source (buy from both), Marvell’s optics and interconnect roadmap offers steeper upside.
Head-to-Head Scorecard
| Business Driver | Broadcom (AVGO) | Marvell (MRVL) |
|---|---|---|
| Quarterly AI Revenue | $10.8B | Embedded in $1.83B Data Center |
| Marquee Customers | Google TPU, Meta MTIA, OpenAI, Anthropic | AWS Trainium, Microsoft Maia |
| Co-Design Market Share | ~70% | ~20% |
| FY2027 Revenue Target | >$100B (AI only) | ~$11.5B (total company) |
| Forward P/E Multiple | 21x | 55x |
| YTD Stock Performance | Steady compounder | +161.64% rally |
| Beta (Volatility) | Lower | 2.246 (high) |
What Smart Investors Are Watching Next
For Broadcom
- Networking Mix: Will AI networking stay near 40% of AI revenue or drift toward Tan’s 30% expectation as XPU volumes ramp?
- Execution at Scale: Can they deliver 10 gigawatts across 4+ massive customers simultaneously?
For Marvell
- Celestial AI Photonics: Can they execute on light-based chip connections without delays?
- Flagship XPU Program: Will the new Tier 1 customer chip ship on time?
How the Investment Setups Compare
| Broadcom: The Incumbent Play | Marvell: The High-Beta Challenger | |
|---|---|---|
| Thesis | Durable cash flow fortress | Steep upside if dual-sourcing accelerates |
| Strengths | 67% operating margin, 6 locked-in GW customers, VMware 93% software gross margin | Optics/interconnect leadership, Celestial AI optionality |
| Risks | Customer concentration, regulatory scrutiny | Valuation (55x), 2.246 beta = violent drawdowns |
| Entry Strategy | Buy on AI capex jitters, not fundamental cracks | Same — but size position for higher volatility |
IMPORTANT: Both stocks can drop sharply on AI spending fears (capex jitters) even if fundamentals are fine. These pullbacks historically offer better entry points than chasing all-time highs.
Summary
- Broadcom is the default choice for hyperscalers building massive AI factories. They sell the entire infrastructure stack — chips, networking, optics — creating a nearly unbreakable moat. $100B+ AI revenue by 2027 is the target.
- Marvell is the essential alternative. Cloud giants must have a second supplier. Marvell’s photonics, interconnects, and new XPU wins position them to capture meaningful share if dual-sourcing becomes standard.
- Valuation Gap: Broadcom at 21x forward earnings vs. Marvell at 55x. You pay a huge premium for Marvell’s growth slope.
- Bottom Line: Broadcom = lower risk, steady compounding. Marvell = higher risk, higher potential reward. Many investors own both for different roles in a portfolio.
FAQ
1. What is "custom AI silicon" and why does it matter?
Custom AI silicon (also called ASICs or XPUs) are chips designed specifically for one company’s AI workloads — like Google’s TPU or Meta’s MTIA. Unlike NVIDIA’s general-purpose GPUs, these chips do one thing extremely well, saving massive amounts of power and money at scale. Hyperscalers (Google, Meta, Microsoft, Amazon, OpenAI) spend billions building their own because it’s cheaper than buying off-the-shelf.
2. What does "co-design" mean?
Co-design means the chip maker (Broadcom/Marvell) works side-by-side with the customer (Google/Meta) from day one — defining architecture, optimizing software, validating silicon. It’s a multi-year marriage, not a one-night stand. This creates deep switching costs.
3. Why is networking 40% of Broadcom’s AI revenue?
AI isn’t just about fast chips — it’s about moving data between thousands of chips without bottlenecks. Broadcom’s Tomahawk/Jericho switches and co-packaged optics are the nervous system of AI clusters. Competitors have chips; few have the full networking stack.
4. What is "photonics" and why is Celestial AI a big deal for Marvell?
Photonics = using light (lasers) instead of electricity (copper wires) to move data between chips. It’s faster, uses less power, and generates less heat. Celestial AI brings photonics expertise that could let Marvell leapfrog in scale-up networking — connecting thousands of chips as if they were one giant chip.
5. Should I buy AVGO, MRVL, or both?
- AVGO if you want: Proven execution, reasonable valuation, durable moat, dividend growth
- MRVL if you want: Maximum exposure to AI networking/photonics, can tolerate 30-50% drawdowns, believe dual-sourcing accelerates
- Both if you want: Full coverage of the custom silicon duopoly with different risk/return profiles
This article is for educational purposes only and does not constitute investment advice. Always do your own research or consult a financial advisor.
Sponsored Note: SoFi Active Invest is offering new users up to $3,000 in complimentary stock with a $50 deposit. Terms apply. See their site for details.