If You Only Own Nvidia for AI Exposure, You're Missing Half the Trade. Here's the Other Half.

Broadcom's fiscal 2025 AI-chip sales rose 65% to $20 billion, representing 31% of revenue, and the company projects nearly sixfold growth to $115 billion by fiscal 2027. The article argues Broadcom's custom ASIC accelerators are increasingly critical for AI inference, offering faster and more cost-efficient processing at scale than stand-alone GPUs. Analysts forecast fiscal 2025-2028 revenue and EPS CAGRs of 62% and 77%, respectively, while Broadcom trades at 22x next-year earnings versus Nvidia at 23x despite Nvidia's expected 59% revenue and EPS CAGRs through fiscal 2029.

Analysis

The investable issue is not whether inference grows, but who retains the economic surplus as workloads shift from generalized training clusters to repeatable, high-volume serving. Custom silicon can lower hyperscaler cost per token, but it also concentrates AVGO's upside in a small set of buyers with exceptional negotiating power and long design/qualification cycles. That makes reported AI revenue less durable than a GPU/software ecosystem revenue stream: a delayed tape-out, customer insourcing decision, or workload change can create an abrupt growth reset rather than a gradual demand slowdown.

NVDA's principal risk is mix and multiple, not an immediate volume collapse. If inference becomes a larger share of accelerator spend and customers increasingly optimize for specific models, NVDA may need to concede price/performance through lower-priced inference products, bundles, or more aggressive networking/software economics; that would pressure gross-margin expectations over the next 6-18 months. Conversely, ASIC adoption is not automatically bearish NVDA: rapid model turnover, multimodal workloads, and utilization uncertainty favor programmable GPUs, while custom chips work best after demand patterns stabilize.

The cleaner second-order beneficiaries are TSMC (TSM) and advanced-packaging equipment/supply-chain exposures, since both GPU and ASIC architectures raise leading-edge wafer and packaging intensity. Marvell (MRVL) is the more direct listed alternative ASIC optionality, but AVGO's customer concentration and execution history warrant a quality premium; the article's forward estimates should be treated as consensus assumptions, not evidence of contracted demand. Near-term, this is likely a narrative-driven relative-value setup rather than a fresh directional catalyst absent hyperscaler capex guidance, disclosed ASIC volumes, or evidence that cost-per-token is falling faster on custom silicon than on Nvidia platforms.

Key Decisions for Investors

• Maintain core long NVDA, but hedge inference-mix risk with a 3-6 month long AVGO / short NVDA dollar-neutral pair only after AVGO outperforms NVDA by 8-10% on relative strength. Thesis: ASIC design-win momentum can re-rate AVGO while NVDA faces inference-margin debate; exit if NVDA raises data-center gross-margin guidance or AVGO discloses a customer program delay.
• Add TSM as the lower-idiosyncratic picks-and-shovels expression over a 6-18 month horizon; it captures aggregate leading-edge accelerator demand irrespective of GPU-versus-ASIC share. Primary risk is Taiwan geopolitical premium or a broad hyperscaler capex cut, so size below single-name accelerator exposures.
• Put MRVL on an earnings-watch list rather than buying preemptively. Upgrade to a long only if management identifies incremental custom-silicon revenue, backlog conversion, or a credible ramp timeline; without those disclosures, it remains a lower-visibility substitute for AVGO rather than a confirmed beneficiary.
• For AVGO, avoid chasing a valuation-only thesis until the next customer capex updates from GOOG and META clarify whether custom-accelerator spend is incremental or displacing GPU purchases. Falsification is a reduction in AI semiconductor guidance, evidence of customer dual-sourcing, or weaker networking attach rates; any of these would challenge the expected operating-leverage case within 1-3 months.