Broadcom agrees to expanded chip deals with Google, Anthropic

Broadcom agreed to produce future AI chips for Google and signed an expanded deal giving Anthropic access to roughly 3.5 gigawatts of compute on Google's AI processors; Broadcom shares rose ~3% in extended trading while Alphabet was flat. CEO Hock Tan previously said Anthropic placed a $10 billion order; Broadcom reported 1 GW of TPU compute starting in 2026 with demand expected to surge to >3 GW in 2027. Broadcom is also working with OpenAI on custom silicon, while OpenAI has committed to 6 GW of AMD GPUs (first 1 GW arriving in H2).

Analysis

Broadcom’s move into bespoke AI silicon materially changes revenue quality: bespoke ASICs and associated software/ops contracts produce multi-year visibility and meaningfully higher incremental margins versus commodity GPUs. A rule-of-thumb to model: every $1B of high-margin custom silicon revenue can translate into roughly $200–300m of incremental EBIT (20–30% incremental margin) once software/recurring services are included, so model sensitivity to order cadence is large and non-linear over 12–36 months.

The immediate second-order constraint is capacity and components, not pure design: advanced-node foundry windows, high-bandwidth memory (HBM) supply and advanced packaging create lumpy delivery profiles that compress near-term gross margin if ramp timing slips. Expect 6–18 month lead-time risk clusters where bookings look great but cash realization and margin expansion lag, and where cloud providers push harder on pricing for long-term commitments.

Competitive dynamics will bifurcate the market: vertically integrated hyperscalers and diversified GPU suppliers can blunt ASIC pricing power by offering software+hardware bundles, while alternative GPU supply ramps reduce one vendor’s spot-market leverage. Over 2–4 years, vertical integration by the largest model owners and rapid architecture shifts in LLMs pose existential demand-risk for monolithic ASIC designs unless chip software is deeply reconfigurable.

Net stance: asymmetric upside for a silicon vendor that nails execution and foundry timing, but substantial execution and customer-concentration risk. Positioning should be tactical around execution milestones (foundry allocations, first production shipments, and cloud deployment dates) rather than headline bookings alone.

Key Decisions for Investors

• Long AVGO (core position): accumulate on >5% pullbacks and add around confirmed production-shipment milestones; prefer 18–30 month LEAPS calls to amplify upside (target 2.5x upside if ASPs and services ramps hit; downside 25–35% if ramps slip). Hedge with 6–9 month puts sized to limit one-way exposure to customer cancellations.
• Pair trade (capital-efficient hedge): long AVGO / short NVDA in a 1:0.5 dollar notional ratio for a 9–15 month horizon to play margin-rich custom silicon vs general-purpose GPU cyclicity. Rationale: capture bespoke-silicon margin expansion while partially hedging broader GPU demand pullbacks; risk if NVDA sustains platform moat or drives GPU price cuts.
• Selective AMD call spread (satellite): buy 9–15 month calls or call spread on AMD to capture cloud-provider diversification upside from non-NVDA GPU uptake. Keep position size <5% portfolio due to execution and backlog risk; expected payoff asymmetric if AMD ramps large cloud commitments.
• Event-driven GOOGL option play (speculative): tactical long-dated call exposure to a major cloud provider sized for 3–6% portfolio risk, timed to their next capital-allocation / cloud-VM announcement window. Rationale: cloud margin capture from bespoke silicon is optionality-rich but binary; trim or exit on signs of vertical integration or regulatory headwinds.