The market delivered its verdict on Google’s chip strategy in two trading sessions: about $87 billion, subtracted from Broadcom’s market value.
Broadcom shares fell for two consecutive sessions on Aug. 19 and 20, wiping out roughly $87 billion in market capitalization, as investors reassessed the company’s role in Google’s AI supply chain. The trigger was Google’s expanded custom-chip agreement with Marvell, announced on Aug. 19, which analysts said signaled that Google intends to spread its custom silicon orders across more suppliers. Broadcom, which has long been one of Google’s principal partners for custom accelerators, saw its expected share of that business marked down.
The selloff reflects a specific change in expectations. Google’s TPU program, the custom chips that power much of its AI computing, has relied on a small group of partners, with Broadcom the most important. The Marvell agreement, covering accelerators, networking, storage and memory controllers, suggested to investors that Google is building a multi-supplier strategy. Each new partner dilutes the others’ share of a pie that, while growing, is no longer expected to flow through one company.
The warrant made the point vivid. Marvell disclosed that Google received a warrant to buy up to 58.97 million Marvell shares at $206.58 each, a structure that ties Google’s fortunes to Marvell’s success. The message investors took from the deal was not that Broadcom had done anything wrong, but that its monopoly-like position in Google’s custom chip plans had been broken. In a market where AI revenue is priced on expectations of exclusivity, the loss of that exclusivity is expensive.
Broadcom’s AI story has been one of the market’s most powerful narratives. The company’s custom chip business, which designs accelerators for hyperscale customers, has grown rapidly and commanded premium valuations as investors bet that AI would drive sustained demand. Broadcom’s own guidance for AI-related revenue has repeatedly exceeded expectations, and its stock has climbed accordingly. The Google news punctured the assumption that the growth would be concentrated.
Analysts were divided on the severity of the reaction. Some said the selloff was overdone, noting that Google remains a large Broadcom customer and that the overall demand for custom AI chips is growing fast enough to support multiple suppliers. Others said the market was right to reprice the stock, arguing that the economics of custom silicon favor competition and that Broadcom’s margins will face pressure as customers diversify. The two sessions of selling suggest the latter view carried the day.
The Google-Marvell deal also shifted the balance among the chip companies chasing hyperscaler business. Marvell, which had been viewed as the clear number two to Broadcom in custom silicon, now has two anchor customers, Google and Amazon, both with multi-year roadmaps. The warrant gives Google a direct financial stake in Marvell’s performance, a commitment analysts read as deeper than a typical supply agreement. The contrast with Broadcom’s situation was immediate and painful.
The episode illustrates how quickly the AI chip market’s structure is changing. For most of the past two years, investors treated custom AI silicon as a winner-take-most market, with Broadcom and a handful of others dividing the spoils. The hyperscalers, however, have consistently moved toward diversification, signing multiple suppliers, designing more in-house and negotiating harder. Each new agreement chips away at the exclusivity that supported the sector’s valuations.
The consequences extend beyond the two companies. The selloff rippled through other AI infrastructure stocks as investors reconsidered how much of the market’s growth is locked into single-supplier relationships. If Google is willing to spread its custom chip orders, the logic goes, so are Amazon, Microsoft and Meta, and every supplier’s expected share of the AI pie becomes less certain. The repricing is not specific to Broadcom; it is a broad reassessment of how the custom chip market will be divided.
Broadcom’s management has not changed its guidance, and the company’s executives have repeatedly said demand for AI custom chips exceeds supply. The question the market is asking is different: not whether the market grows, but whether any single supplier can hold the kind of share that the old assumptions priced in. The $87 billion drop was the market’s first, and largest, answer to that question, and it will not be the last.
The episode also illustrates how fragile the market’s assumptions about AI winners have become. Investors have swung between euphoria and panic over AI infrastructure stocks, pricing in both the industry’s enormous growth and the constant risk that the structure of the market changes overnight. The Google-Marvell deal was not a failure of Broadcom’s technology or its execution; it was a reminder that the hyperscalers hold the negotiating power and will use it. For Broadcom’s management, the selloff is a test of how it communicates its pipeline: the company has long argued that its custom chip revenue is understated by analysts, and the market’s reaction suggests those assurances have a limit. The coming quarters, when the first revenue from the Google-Marvell program appears, will show whether the diversification of Google’s supply chain shrinks Broadcom’s opportunity or merely slows its growth.
