Broadcom (AVGO) might be the best AI infrastructure bet nobody's talking about enough

Been digging into AVGO and wanted to share the thesis because I think it's underrated relative to how much airtime NVDA gets.

TL;DR: Broadcom is two businesses duct-taped together — explosive custom AI silicon growth, hedged by a software cash cow (VMware) that prints predictable FCF. That combo is why the stock supports a premium multiple that would normally scare me off.

The numbers: FY2025 revenue +24% to $63.89B, adj. EBITDA $43B (+35%), FCF $26.9B. Q2 FY26 revenue hit $22.19B (+48% YoY), AI segment alone did $10.8B on 143% growth. Guidance for Q3 has AI revenue growth above 200%. This isn't a "story stock" — the growth is showing up in the actual financials quarter over quarter.

Why the model is different from NVDA: NVIDIA sells merchant GPUs under its own brand. Broadcom instead helps hyperscalers (Google, Meta, OpenAI) design their own custom ASICs, while the hyperscaler eats the capex and inventory risk at TSMC. Broadcom just collects IP licensing + design/integration fees. Result: 78.6% gross margin in this segment — actually higher than NVDA's ~73.5%. It's basically a royalty business wrapped in a semiconductor wrapper.

Why do hyperscalers bother? A custom ASIC tuned to one workload cuts cost-per-token by 50-67% vs. general-purpose GPUs. At billions of tokens/day that pays back the chip design spend inside a year. Broadcom + Marvell are basically the only two players (Broadcom ~70% share, Marvell 20-25%), and Broadcom's confirmed AI backlog is north of $73B.

Client list is the real moat:

• Google — TPU dev deal through 2031
• Anthropic — reserved 3.5GW of TPU capacity starting 2027 (Mizuho thinks this alone is worth $21B-$42B to AVGO by 2027)
• Meta — custom MTIA chip on 2nm, contract through 2029 (Hock Tan literally left Meta's board to advise Zuck directly on silicon)
• OpenAI — 6th ASIC customer, $10B project, >1GW deployment targeted 2027
• ByteDance + one undisclosed name

That's basically every major AI lab locked into multi-year hardware roadmaps with AVGO as the design partner. The $100B AI revenue target for 2027 isn't a slide-deck number, it's contracted backlog math.

The networking angle people miss: Everyone talks about chips, fewer people talk about the Ethernet vs. InfiniBand fight happening underneath. InfiniBand (NVDA/Mellanox) has been the standard for GPU-to-GPU interconnect, but at 10k+ GPU cluster scale the optics start failing constantly — 4,000 GPUs need 8,000+ transceivers, and at >5% annual failure rates you're looking at multiple hardware failures a week.

Broadcom backed the Ultra Ethernet Consortium to push Ethernet as the open alternative — comparable performance, way better fault tolerance (30x faster failover). Their Tomahawk 6 switch (102.4 Tbps, shipped 2025) has Scale-Up Ethernet, which lets you link up to 512 accelerators over standard Ethernet instead of NVDA's proprietary NVLink. NVIDIA's competing Spectrum-X1600 isn't expected until 2H26 at earliest — that's a real lead, not a rounding error in this industry.

VMware = the boring part that makes the whole thing work: Post-acquisition, Broadcom killed perpetual licenses, collapsed 70+ SKUs into 4 mega-bundles, forced a 16-core-per-socket minimum, and gutted the partner network from 4,500+ to ~300. Customers are seeing 200-500% cost increases on renewal. Brutal for the customer, but software gross margin is now 93% and segment operating margin is 78%. High switching costs mean enterprises grumble and pay rather than migrate off. This is the part of the business that funds AI R&D without touching debt.

Risks I'm tracking:

• Marvell is growing faster off a smaller base (~60% YoY vs AVGO's ~40%), positioned as the "second source" for AWS/Microsoft
• NVDA just bought Groq for $20B (LPU chips) — already drawing antitrust attention from Warren/Blumenthal
• Cerebras IPO'd at $56.4B, wafer-scale approach could disrupt the ASIC paradigm
• Apple is moving to in-house Wi-Fi/BT chips (Proxima), though AVGO keeps the RF filter business + new server AI chip deal
• $62.65B long-term debt post-VMware, though leverage dropped to 1.2x and S&P upgraded to A-

Valuation: Trailing P/E ~62x looks scary in isolation, but forward FY27 P/E is ~34x, and PEG sits at 0.83 (sub-1.0 = growth outpacing price). Median PT $420-507, bulls (Susquehanna, BofA) see $640.

Curious what others think — anyone modeling AVGO independent of the NVDA narrative, or is this still mostly "buy the AI basket" thinking? Also curious how people are thinking about the VMware churn risk longer-term once contracts start coming up for renewal in bulk.