SAN JOSE — In a session that wiped out hundreds of billions of dollars from chip stocks, Broadcom shares held their ground. The company’s AI-related order backlog now stands at $73 billion, according to figures presented to investors, a number large enough to absorb the latest round of speculation that Google is moving chip design in-house.
The backlog has become the market’s clearest window into real AI infrastructure spending, analysts said, because it is composed of binding orders rather than forecasts. Broadcom designs custom accelerators and networking silicon for the largest cloud operators, and each quarter’s backlog number tells investors how much computing capacity those customers actually intend to build.
The $73 billion figure also carried the company through a burst of news that would normally dent its shares. Reports circulated that Google, one of Broadcom’s largest customers, is expanding its own chip design efforts, reviving the question of whether the search giant will eventually replace Broadcom’s custom silicon with fully internal designs. JPMorgan analysts pushed back, arguing in a note that Google’s collaboration with Broadcom covers engineering capacity and networking expertise that an in-house team could not quickly replicate.
The same reports, analysts said, have surfaced in various forms for years. Google has designed its own tensor processing units since 2015, and it has worked with Broadcom on those designs for much of that time. The relationship has survived each wave of self-sufficiency talk, and the backlog figure suggests cloud customers see the division of labor as durable.
Broadcom’s finances have benefited accordingly. The company raised its dividend to a record level in the latest announcement, and its free cash flow has climbed as AI customers commit to multi-year supply agreements. Chief Executive Hock Tan, who built Broadcom through a string of acquisitions, has redirected the company’s focus from commodity semiconductors to the custom chips and networking gear that move data inside AI data centers.
The strategy carries concentration risk. A small number of customers account for a large share of the AI backlog, and the loss of any one of them would be visible in the numbers. The Google question is the one investors ask most often, and Broadcom’s answer, repeated in earnings calls, is that the switching costs are high: custom silicon developed with a partner is not easily transferred to a new team, and the networking products that tie clusters together are Broadcom’s to sell regardless of who designs the processors.
The company’s networking business has become as important as its chip designs. AI training clusters require switches that can move data between tens of thousands of accelerators without stalling, and Broadcom’s Tomahawk and Jericho switch families are standard equipment in the largest deployments. Analysts estimate that networking now accounts for a growing share of the AI backlog, giving the company a second source of demand that is largely insulated from the custom-silicon debate.
Tuesday’s session tested the thesis. While Nvidia fell as much as 15% and memory makers in Seoul dropped by double digits, Broadcom’s stock declined modestly and recovered most of the loss by the close, according to market data. Traders said the backlog number, published days earlier, gave investors a concrete anchor in a session driven by fear.
The longer question is whether the $73 billion backlog can keep growing. Hyperscale customers have signaled continued increases in capital spending, and the pace of AI model training shows no sign of slowing, but the history of technology cycles suggests that order books peak before demand does. Broadcom’s management has said it expects the AI business to keep compounding, and the dividend increase was framed as a sign of confidence rather than a payout of a maturing business.
The custom-chip market Broadcom serves is growing faster than the merchant chip market, and the company’s position in it is the result of a decade of investments. Broadcom’s engineers work with customers from the architecture stage, designing accelerators that are tailored to the way each cloud operator runs its AI workloads, and the resulting chips cannot be bought off the shelf from any competitor. The company’s relationships extend beyond Google: Meta and ByteDance have both been reported as customers, and analysts say the pipeline of potential custom-chip deals includes most of the largest AI spenders in the world. Each deal adds to the backlog and to the company’s knowledge of what the industry’s biggest buyers actually need, a moat that is difficult for a new entrant to cross.
The VMware acquisition, completed in 2023, has become a quieter contributor to the AI story. VMware’s software manages the servers that run enterprise workloads, and Broadcom has been integrating AI capabilities into that stack, selling customers a path from traditional data centers to AI-ready infrastructure. The combination gives Broadcom exposure to AI at two levels: the chips and networking gear that go into hyperscale data centers, and the software that runs AI in corporate data centers that are just beginning to adopt it. Analysts said the second level is the part of the backlog that is hardest to quantify and potentially the most durable, because enterprise AI adoption is expected to grow for years regardless of the quarterly swings in hyperscale spending.
Analysts said the company’s position is stronger than most in the AI trade because its revenue is booked rather than projected. The order book does not depend on Nvidia’s next product cycle or on a single model’s success; it reflects capacity that cloud operators have already committed to build. That visibility is why the stock held up on a day when the rest of the sector cracked, and it is why investors will keep reading the backlog number as the closest thing to ground truth on AI spending.
