Broadcom

Who they are

Broadcom's corporate lineage is genuinely tangled: it traces back to a Hewlett-Packard semiconductor division from 1961, which became part of Agilent Technologies in 1999, was spun out as Avago Technologies in 2005, and finally acquired the original Broadcom Corporation in 2016 for $37 billion — keeping the more recognizable Broadcom name for the combined company. Hock Tan, who has run the business since the Avago days in 2006, is the person who turned that patchwork history into the semiconductor industry's most disciplined acquirer, and lately, into the quiet engine behind most of the industry's "we're building our own AI chip" headlines.

What they do

When Google says it designed its own TPU, or Meta says it built its own MTIA accelerator, what usually happened is that Broadcom did the actual chip design and engineering under contract, using the customer's specifications. This is Broadcom's XPU business, and as of Q1 FY2026 earnings the company had confirmed six major customers: Google (its longest-standing partner, seven TPU generations deep since 2014), Meta, OpenAI, Anthropic, and Apple, with two more reportedly unnamed. Anthropic alone is ramping from 1 gigawatt of TPU capacity in 2026 to 3-plus gigawatts in 2027. The technology underpinning all of it is Broadcom's 3.5D XDSiP platform, which combines face-to-face 3D chip stacking with the same CoWoS packaging TSMC uses for Nvidia's GPUs — meaning Broadcom's custom chips and Nvidia's merchant GPUs are quietly competing for the same packaging capacity.

How Broadcom makes money

Design fees and, more importantly, long-term chip supply contracts once a design ships into production. AI semiconductor revenue hit $8.4 billion in Q1 FY2026, up 106% year-over-year, with Q2 guided to $10.7 billion. CEO Hock Tan has told investors Broadcom has "line of sight" to more than $100 billion in annual AI chip revenue by 2027, backed by a disclosed $73 billion order backlog scheduled for delivery over the next 18 months. That backlog now includes a 10-gigawatt custom accelerator deal with OpenAI signed in October 2025, with first deployment targeted for the second half of 2026.

The bigger trend

Nvidia still holds an estimated 70% of the overall AI chip market, but custom ASICs are the fastest-growing slice of it: ASIC-based AI server shipments are projected to reach 27.8% of the market in 2026 — the highest share since 2023 — growing at nearly triple the rate of merchant GPUs. Broadcom's own market share of the custom AI accelerator design business is now estimated above 70%, up from the 60–80% range analysts cited earlier in 2026, meaning its customer wins are compounding rather than getting diluted by new entrants. The strategic logic customers cite is straightforward: a workload-specific XPU can offer 30–50% lower total cost of ownership than a general-purpose Nvidia GPU for the exact task it's built for — the tradeoff being less flexibility if the workload changes.

Related companies