Broadcom Inks Major AI Chip Deals With Google and Anthropic

Broadcom signed a new five-year deal to co-design and build Google's Tensor Processing Units and will supply networking/rack components to Google through 2031, while Anthropic committed to 3.5 gigawatts of TPU-based compute starting in 2027. Broadcom reported fiscal Q1 2026 revenue of $19.3 billion (+29% YoY) and adjusted EPS of $2.05 (+28%), with AI semiconductor revenue of $8.4 billion and CEO Hock Tan projecting >$100 billion in AI chip revenue in 2027.

Analysis

The structural shift toward co-designed, workload-specific silicon creates a durable margin wedge for firms that can own system-level integration (chip + NIC + firmware + rack design). That wedge increases bargaining power with hyperscalers and lengthens contract durations, turning what used to be spot GPU demand into longer, more predictable multi-year procurement — a secular revenue smoothing that favors integrated solution providers and advanced packaging suppliers.

A meaningful second-order effect is a reorientation of data-center supply chains: demand will tilt from raw GPU die to advanced packaging, custom interposers and high-performance NICs, concentrating growth in a narrower set of suppliers and raising switch-in costs for new entrants. At the same time, incumbent GPU vendors retain a defensive moat via software, model-optimized kernels and developer ecosystems, so displacement is likely to be demand-share erosion over 2–5 years rather than a sudden market flip.

Key tail risks that could reverse the current trajectory include execution problems on high-density packaging yields, a hyperscaler decision to standardize on an alternative architecture, or a slowdown in enterprise model deployment that defers hardware refresh cycles. Near-term catalysts to watch are cloud capex guidance, unit-level ASPs for accelerated compute, and any public latency/efficiency benchmarks from large model providers.

Consensus currently prizes long-term capture of inference workloads but underweights switching friction from software and integration testing; that implies upside remains but should be accessed tactically with hedges. The most asymmetric opportunities are calendar spreads and pairs that exploit the mismatch between contract-driven revenue visibility and volatile market sentiment around general-purpose GPU demand.

Key Decisions for Investors

• Long AVGO via a directional, capital-efficient structure: buy 12–18 month LEAP calls (target delta ~0.55–0.65) and sell near-term (3–6 month) calls to finance ~40–60% of premium. Timeframe: 9–18 months. Risk/reward: limited downside = net premium; upside = >2x if ASIC adoption accelerates and multiple expansion follows stronger recurring revenue visibility.
• Pair trade (sector re-allocation): long GOOGL equity (or 12-month call LEAP) / short NVDA (buy 6–9 month puts) in equal-dollar notional for 6–12 months. Thesis: cloud margin and contracted co-design capture vs GPU cyclicality. Risk/reward: protects against GPU-led drawdowns while keeping upside to cloud margin improvement; cap losses if NVDA’s software moat drives further outperformance.
• Event hedge: buy 9–18 month OTM NVDA puts sized to cover 5–10% portfolio exposure as insurance against rapid GPU demand collapse or a negative benchmark release. Timeframe: maintain through next two cloud-capex seasons. Risk/reward: premium drag vs potential 5–8x payoff on a 30–50% NVDA sell-off.
• Income-backed buy-write on GOOGL: initiate a 6–12 month buy-write (long shares / sell 3–6 month calls) to capture premium while retaining upside to cloud cost-of-inference improvements. Timeframe: roll every quarter on improving guidance. Risk/reward: reduces cost basis and provides downside cushion equal to premiums received; caps extreme upside if multiple expansion accelerates.