How Broadcom Ends Nvidia’s GPU Monopoly

How Broadcom Ends Nvidia's GPU Monopoly

Nvidia (NVDA) built the dominant hardware for AI training and is expected to generate over $350 billion in revenue this year. Most investors know that story.

Fewer are paying attention to what comes next.

Training a model is largely a one-time cost. Running it is not.

Every query processed by ChatGPT, every AI-assisted search result, every automated agent action - that is inference, and it runs continuously across billions of users every day. At that scale, the cost of each computation is not an engineering detail. It is the difference between a profitable AI product and an unprofitable one.

That is where Broadcom (AVGO) comes in.

What Does Broadcom's Chip Business Do?

Rather than selling a general-purpose GPU to anyone who needs one, Broadcom works directly with large technology companies to design chips tailored to their specific workloads. The customer owns the architecture and intellectual property. Broadcom converts those blueprints into manufacturable silicon, managing power, heat, memory integration, and the final preparation for production at a foundry like TSMC. The work requires years of accumulated technical collaboration and is not easily replicated by a competitor starting from scratch.

While AI chips are the company's biggest growth engine, its software business serves as the high-yield annuity powering the company's AI investments. See the three vectors powering AVGO stock.

How Much Can Hyperscalers Save With Custom Chips?

A flagship Nvidia GPU is estimated at $30,000 to $40,000 per unit. Broadcom's custom TPUs are estimated to cost Google approximately $12,000 per unit in 2026. Google's latest generation chip is reported to be 67% more energy-efficient than comparable GPUs for inference. Across millions of servers running continuously, that efficiency gap is material to operating economics.

What Are The Financials Like?

Once a hyperscaler builds its software stack and data center infrastructure around a custom chip, replacing it becomes prohibitively expensive.

This produces a revenue profile that compounds over successive hardware generations rather than resetting each cycle.

Broadcom's AI semiconductor revenue grew 106% year-over-year in Q1 2026 to $8.4 billion, with full-year FY2025 AI revenue at approximately $20 billion. Semiconductor gross margins stood at approximately 69% last quarter, with operating margins at 60%. This is just a notch below Nvidia, which posted gross margins of about 75%. (See AVGO margins vs peers) Broadcom's AI silicon backlog exceeds $73 billion, nearly half of a $162 billion total order book - a level of forward visibility unusual in hardware businesses.

What Are The Risks?

Still, Broadcom's AI semiconductor business carries several risks worth considering.

Revenue is concentrated among a small number of hyperscalers - a strategic shift at Alphabet (GOOG) or Meta (META) would have a material effect on results. Custom chip design also involves long development cycles, meaning missteps in any generation are costly and slow to correct.

On the competitive side, Marvell (MRVL) is gaining ground making custom ASICs for Amazon (AMZN) and has deepened its partnership with Nvidia on networking silicon. Marvell is also emerging as a formidable player in the AI interconnects space.

Broadcom's $11 billion Anthropic engagement and multi-generational Meta MTIA partnership address concentration to a degree, but the business remains exposed to the capital allocation decisions of a handful of very large companies.

Your Next Move

For investors seeking exposure to Broadcom's AI thesis without the concentration risk tied to customers, a natural question is how to capture the upside with less volatility than a single stock.

The Trefis HQ strategy offers one such approach, having outperformed its blended benchmark (S&P 500, S&P MidCap, and Russell 2000) with over 105% returns since inception. Why did HQ outperform? See HQ performance metrics with five reasons why - for the full story.