Broadcom is establishing itself as a credible Nvidia competitor in AI through custom silicon and networking solutions for hyperscalers and cloud providers.
Nvidia has established its position as the leading provider of hardware for AI training and is projected to achieve over $350 billion in revenue this year. However, fewer people are focusing on what comes after. The expense involved in training a model is primarily a one-time investment, but ongoing operation incurs continuous costs. Every query handled by ChatGPT, every AI-aided search result, and every action by automated agents represent inference, which runs ceaselessly across billions of users each day. At such volumes, the expense of each computation is not merely an engineering detail—it differentiates a lucrative AI product from one that fails to generate profits.
This is where Broadcom plays a crucial role. Instead of offering a general-purpose GPU to any potential client, Broadcom collaborates directly with major tech firms to develop chips specifically designed for their unique workloads. The client retains ownership of the architecture and intellectual property, while Broadcom translates those designs into manufacturable silicon, overseeing power requirements, thermal management, memory integration, and final preparations for production at a facility like TSMC. This process demands years of shared technical expertise and is not easily duplicated by competitors starting from the ground up. Once a hyperscaler integrates its software architecture and data center setup around a custom chip, substituting it becomes excessively costly.
The price advantage is substantial. The price for a top-tier Nvidia GPU is thought to be between $30,000 and $40,000 per unit, while Broadcom's custom TPUs are estimated to cost Google about $12,000 per unit in 2026. The latest generation of Google's chip is claimed to be 67% more efficient in energy consumption than equivalent GPUs for inference tasks. Over millions of continuously running servers, that efficiency gap significantly impacts operating costs.
Broadcom's financial performance reflects this opportunity. AI semiconductor revenue surged 106% year-over-year during Q1 2026 to $8.4 billion, with full-year FY2025 AI revenue estimated at around $20 billion. Semiconductor gross margins reached about 69% in the last quarter, with operating margins at 60%, slightly behind Nvidia's gross margins nearing 75%. Broadcom's AI silicon backlog surpasses $73 billion, making up nearly half of a total order book of $162 billion—an exceptional level of forward visibility rarely seen in hardware industries.
However, Broadcom's AI semiconductor segment carries multiple risks. Revenue concentration exists among a limited number of hyperscalers, and a strategic change at Alphabet or Meta could materially affect results. The design of custom chips also involves extended development cycles, meaning errors in any generation can be expensive and slow to rectify. Marvell is making headway by producing custom ASICs for Amazon and has strengthened its collaboration with Nvidia in networking silicon. Broadcom's $11 billion partnership with Anthropic and its multi-generational collaboration with Meta on MTIA somewhat alleviate concentration risks, but the business remains reliant on the capital allocation choices of a select few enormous entities.