Cantor Fitzgerald reiterates Broadcom stock Overweight on AI chip demand

Cantor Fitzgerald reiterated its Overweight rating on Broadcom, citing a planned 10GW TPU-chip supply arrangement for Anthropic in fiscal 2028, following 1GW in FY2026 and 5GW in FY2027. The implied Anthropic TPU capacity reaches roughly 15GW across FY2027-28, while separate neocloud agreements add 812MW of Nvidia-based compute capacity. Analyst views remain mixed after Broadcom's slightly above-consensus fiscal Q3 sales and uneven Q4 outlook, although KeyBanc highlighted increased FY2027 AI-revenue guidance of $115B.

Analysis

The important mechanism is not incremental accelerator demand, but a potential shift in AI compute mix from merchant GPUs toward customer-specific ASICs. AVGO captures high-value silicon design, networking attach and likely recurring generation-to-generation redesign work, while NVDA's exposure is concentrated in the portion of training/inference workloads that remain economically superior on a general-purpose CUDA stack. If the disclosed ramp is converted into binding wafer starts and funded data-center buildouts, AVGO's AI revenue durability could warrant a premium multiple rather than being treated as a cyclical semiconductor supplier.

The bottleneck is physical deployment, not chip design: multi-gigawatt plans require power interconnects, substations, cooling, networking, land and debt/equity financing. Over the next 1-3 months, the market will likely reward evidence of signed capacity, foundry allocation and customer prepayments; over 6-18 months, delays in utility energization could defer revenue recognition despite intact demand. The cleanest second-order beneficiaries are networking and optical vendors with higher port content in scale-out clusters, including ANET and CIEN, although these names need order-book confirmation before being underwritten as direct beneficiaries.

Consensus may overstate the near-term displacement risk to NVDA. A custom accelerator program does not automatically replace GPUs: customers commonly retain GPU fleets for frontier-model training, ecosystem flexibility and burst capacity, while ASICs are most compelling in stable, high-volume inference workloads. Conversely, AVGO's valuation sensitivity is asymmetric if future capacity statements are non-binding or if customer concentration becomes more visible; a guidance miss would compress both AI-growth assumptions and the conglomerate/software multiple.

Key Decisions for Investors

• Maintain or initiate a 3-6 month long AVGO / short NVDA pair, sized beta-neutral, only after confirming that AVGO's next revenue guide embeds material custom-silicon growth. The trade monetizes ASIC share gains while limiting broad AI-capex risk; exit if AVGO does not raise AI revenue expectations or if NVDA demonstrates sustained accelerator demand acceleration that exceeds current expectations.
• For directional AVGO exposure, prefer 6-9 month call spreads rather than outright stock after a post-earnings pullback: buy an at/near-the-money call and sell a 15-20% out-of-the-money call. This targets confirmation of customer conversion and limits downside if deployment schedules slip; avoid chasing a sharp pre-guidance rally because the cited capacity path is long-dated and execution-dependent.
• Put ANET and CIEN on a watch list rather than buying immediately. Upgrade to longs only if hyperscaler/neocloud order commentary shows accelerating AI back-end network and optical demand; falsification is flat sequential AI-related backlog or customer capex shifting from buildout to power-constrained deferrals.
• Use the next AVGO earnings release as the key catalyst: require disclosure of customer deposits, backlog quality, foundry supply visibility or a higher AI revenue outlook. Absent those markers, treat the announcement as narrative support rather than a basis for increasing risk.