Broadcom's $230 Billion AI Bet Gets Real

Broadcom expects AI semiconductor revenue to reach $115 billion in FY2027 and $230 billion in FY2028, underpinned by secured supply and planned deployments from six XPU customers totaling about 30 GW. Anthropic could contribute 5 GW of deployments in 2027 and 10 GW in 2028, while OpenAI targets more than 5 GW. The outlook indicates exceptionally strong AI infrastructure demand, with execution capacity rather than customer demand becoming the principal risk.

Analysis

The investable signal is not merely higher AI silicon demand; it is a potential step-change in custom-XPU adoption that shifts value from merchant accelerators toward ASIC design, high-radix networking and advanced packaging. AVGO should gain operating leverage disproportionately if incremental revenue is weighted to semi-custom compute and Ethernet switching, but the market will demand evidence that gross margin holds as customer-specific programs scale. NVDA is not necessarily a direct loser: custom silicon expands total cluster spending, though it raises the risk that the largest cloud buyers cap their marginal GPU mix over the next 12-24 months.

The binding constraint is now execution across power delivery, grid interconnection, data-center construction, HBM/advanced packaging and optical interconnects. This favors VRT, ETN, CEG, TSM and optical suppliers such as COHR and LITE, although the latter two carry materially greater qualification and pricing risk. A deployment slip of even one or two quarters would not necessarily impair long-run demand, but could produce sharp multiple compression in names priced for uninterrupted capex conversion.

Consensus may be underestimating the concentration risk embedded in AVGO's opportunity: a small number of buyers will determine both volume ramps and pricing, giving customers leverage once programs reach production. The near-term upside case requires quarterly proof in backlog conversion, not only design-win commentary; without disclosed customer commitments, the long-dated revenue targets should be treated as capacity planning assumptions rather than contracted earnings. Over 6-18 months, a successful ASIC ramp would justify a higher semiconductor mix and FCF trajectory, while a delayed ramp would likely hit AVGO harder than diversified infrastructure beneficiaries.

Key Decisions for Investors

• Accumulate AVGO on post-results volatility rather than chase guidance-driven strength; target a 6-12 month long, contingent on AI semiconductor growth translating into stable or expanding segment gross margin and no reduction in customer-program timing. Thesis is falsified by two consecutive quarters of AI revenue growth deceleration accompanied by margin erosion or weaker backlog conversion.
• Express the infrastructure bottleneck through a 6-12 month basket long VRT and ETN versus a short SOXX hedge, rather than a pure AVGO beta trade. This captures power-and-cooling content per deployed MW while reducing broad semiconductor multiple risk; exit if hyperscaler capex guidance falls materially or lead times normalize without corresponding revenue acceleration.
• Use a tactical AVGO/NVDA relative-value framework over the next 1-3 months: favor AVGO only if custom-program revenue visibility improves while NVDA's customer concentration or gross-margin commentary weakens. Do not initiate solely on the projected market size; the missing datapoints are program-specific volume commitments, wafer allocations and expected ASIC gross-margin profile.
• Place watch alerts on TSM and COHR/LITE for evidence of advanced-packaging and data-center optics tightness. A confirmed capacity squeeze would broaden the AI trade beyond compute, but absent order-book or utilization confirmation these remain monitoring candidates rather than recommended positions.