Broadcom Pivots to "Chips-Only" Custom Silicon Strategy as AI Revenue Surges 143%

Broadcom Pivots to "Chips-Only" Custom Silicon Strategy as AI Revenue Surges 143%¹

Broadcom has firmly established itself as the dominant incumbent in the custom AI semiconductor market, with its fiscal Q2 2026 AI semiconductor revenue skyrocketing 143% year-over-year to $10.8 billion (on total revenue of $22.2 billion). Mizuho estimates Broadcom’s AI revenue will grow 103% YoY to $40.4 billion in fiscal 2026 and nearly double to $78 billion by fiscal 2028, positioning CEO Hock Tan to comfortably achieve his public target of $100 billion in AI revenue by 2027.

This explosive growth is driven by Broadcom's custom ASIC (application-specific integrated circuit) business with hyperscalers, particularly Google's seventh-generation TPU.

Google TPU v7 (Ironwood) Architecture and Economics

Google officially announced the general availability of its TPU v7 (Ironwood), designed specifically to drive down costs in the "age of inference." Google’s TPU program represents the primary driver of Broadcom's custom silicon revenue, with HSBC projecting Google TPUs to account for ~78% of Broadcom's ASIC revenue ($22.1 billion) in FY2026.

Key architectural and economic details of Ironwood include:

• Performance and Memory: Each Ironwood chip delivers 4.6 petaFLOPS of dense FP8 (surpassing Nvidia's B200 at 4.5 PFLOPS), packaged with 192 GB of HBM3E at 7.4 TB/s bandwidth via eight stacks.
• Extreme Scale-Up: Ironwood scales to 9,216-chip superpods linked with Inter-Chip Interconnect (ICI) networking at 9.6 Tb/s, sharing 1.77 Petabytes of memory.
• Optical Switching: Pods are interconnected via Google's Apollo Optical Circuit Switch (OCS) fabric (supplied by Lumentum), drawing only ~100W per switch compared to ~3,000W for traditional electronic switches.
• Cost Advantage: Individual Ironwood TPUs are estimated to cost around $13,000 per chip to manufacture, less than half the cost of Nvidia's Blackwell GPUs ($30,000 to $40,000 for a B200).
• Fabrication and Memory Supply: Fabricated on TSMC's N3P (3nm) node, with Samsung holding ~60% of the HBM3E supply to Google via Broadcom.

Landmark Anthropic and Hyperscaler Commitments

The TPU ecosystem received a massive commercial validation with Anthropic signing a landmark gigawatt-scale TPU deal with Google and Broadcom. Anthropic has committed to deploying up to 1 million TPUs (including Ironwood) online in 2026–2027, a deal worth tens of billions, as Claude's run-rate revenue surpasses $30 billion. Broadcom is also in active negotiations with Microsoft and OpenAI to co-develop custom silicon.

Google Diversifies: Active Negotiations with Marvell

To reduce pricing power and eliminate single-source capacity risk, Google is actively negotiating TPU development projects with Marvell Technology (MRVL). Marvell, whose custom silicon business runs at a $1.5 billion annual rate across 18 cloud design wins, is in talks with Google to co-develop:

1. A specialized memory processing unit to optimize HBM performance.
2. An inference-optimized TPU variant (Google's version of a Language Processing Unit, or LPU) to counter Nvidia's push into high-performance inference via Groq.

While Broadcom remains locked in as Google's primary training design partner through 2031, Google's multi-partner strategy (Broadcom for training, MediaTek and Marvell for inference/edge) introduces competitive pricing dynamics into the custom silicon supply chain.

Verbatim Quotes

• Broadcom CEO Hock Tan on the massive shift toward inference:

  "In fact, what we’ve seen recently is that they are doubling down on inference in order to monetize their platforms. And reflecting this, we may actually see an acceleration of XPU demand into the back half of 2026 to meet urgent demand for inference on top of the demand we have indicated from training."

• Brendan Burke (Futurum Research) on the Broadcom-Anthropic TPU deal:

  "Anthropic’s TPU expansion with Google and Broadcom highlights rising compute demand, long-term infrastructure commitments, and the growing importance of custom silicon in AI market competition... Broadcom’s CEO Hock Tan has publicly targeted more than $100 billion in AI revenue by 2027, and multi-year supply commitments of this nature provide the demand signal."

• Chris Zeoli (Data Gravity) on Google's multi-partner ASIC strategy:

  "Google is in active discussions with Marvell for two chips: a memory processing unit and an inference-optimized TPU variant... The multi-partner strategy—Broadcom for training, MediaTek for inference, Marvell for edge inference and memory processing—creates pricing competition and eliminates single-source capacity risk."