Broadcom Told the Truth. The Market Hasn’t Heard the Rest of It Yet.

Key Takeaways

• AI revenue hit $10.8 B, up 143 % YoY
• Google’s deal hedges against sole‑source reliance on Broadcom
• AI‑ASIC mix is diluting Broadcom’s blended gross margin
• Marvell’s interconnect margins stay above 58 % while Broadcom compresses
• Sector sell‑off shows investors treat AI chips as a single category

Pulse Analysis

Broadcom’s Q2 2026 earnings painted a paradoxical picture: record AI semiconductor revenue and aggressive top‑line guidance collided with a market‑driven price decline. The company’s $10.8 billion AI sales, driven by custom ASICs for hyperscalers such as Google, Amazon and Meta, eclipsed its own forecasts, yet the after‑hours sell‑off reflected deeper concerns raised during the Q&A. Analysts focused on Google’s long‑term agreement, which explicitly acknowledges a diversification strategy, and on the CFO’s admission that the surge in AI‑specific silicon is eroding gross margins. This mix‑shift reality signals that the fastest‑growing revenue line is also the least profitable, a nuance many investors missed in the hype.

The contrast with Marvell’s performance sharpened the narrative. While Broadcom’s compute silicon margins slipped, Marvell’s optical interconnect and networking business posted a non‑GAAP margin of 58.9 % and accelerated revenue growth, underscoring that the data‑pipeline layer of AI infrastructure remains more resilient. The market’s indiscriminate sell‑off of every semiconductor name with an AI tag reveals a categorical mispricing: investors have been treating the AI boom as a monolithic chip story, ignoring the divergent economics of compute versus connectivity. This misalignment forces a reassessment of valuation models that previously assumed uniform margin expansion across the sector.

Beyond the balance sheet, the episode highlights a hidden supply‑chain risk: the human‑capital layer that translates design specifications into silicon. Indian and Israeli design‑service firms, such as Tata Consultancy Services, Wipro and Intel’s Haifa labs, execute the RTL, verification and physical implementation for custom ASICs. A disruption in this talent pool could delay hyperscaler roadmaps, throttling AI‑capex efficiency and undermining the very growth assumptions that drove the recent rally. As investors digest Broadcom’s margin confession and Marvell’s contrasting success, the focus will shift toward the broader AI‑infrastructure ecosystem—where chips, interconnects, power, cooling and engineering talent together determine the true profitability of the AI supercycle.