Google's Custom Chips Are Cracking Nvidia's AI Monopoly

"If money were not a factor, I am 100% using a custom chip," says Gabriel Rasskin, a software engineer on Google's Gemini AI team. His reasoning is brutally simple: "Every second of compute matters."

That mindset is spreading fast among hyperscalers, and it's starting to crack Nvidia's seemingly unbreakable grip on AI computing. Google's tensor processor units (TPUs), co-designed with Broadcom, successfully trained Gemini 3—the November launch that put Alphabet back in the conversation as a serious challenger to OpenAI's ChatGPT.

When Specialized Beats General Purpose

Custom chips are built for one thing: doing specific high-volume tasks incredibly well. That's something Nvidia's general-purpose GPUs, despite their versatility, simply can't match in pure efficiency.

Jensen Huang recently dismissed this threat, telling Jim Cramer that "Nvidia can address markets that are much, much broader, not just chatbots." But the numbers tell a different story. Broadcom CEO Hock Tan announced that AI revenue jumped 65% year-over-year to $20 billion, driving total semiconductor revenue to a record $37 billion.

More telling? That mysterious fourth customer who placed a $10 billion order turned out to be Anthropic—the company behind Claude AI. The custom chip momentum isn't just Google anymore.

The Great Vendor Diversification

Amazon, Microsoft, and Meta are all developing their own custom chips while remaining massive Nvidia customers. The strategy isn't replacement—it's insurance against vendor lock-in.

"Nvidia's customers, especially the largest ones, don't want to be beholden to one vendor," explains Gil Luria, analyst at D.A. Davidson. "This is the free market, and profits attract competition."

Yet barriers remain sky-high. Custom chip development requires massive upfront investment and years of development time. Plus, there's limited manufacturing capacity at foundries like Taiwan Semiconductor (TSMC), where both Nvidia and Broadcom actually produce their designs.

An Apple source familiar with chip development points out that smaller companies simply can't afford the time and cost. Apple itself took over a decade to replace third-party silicon across its devices.

The Fragile Nature of Custom Chip Success

Broadcom is riding high, but its position is more precarious than Nvidia's. "Its biggest customer by far is Google," Luria notes. "If Google ever decided to go directly to TSMC the way Apple did, that would represent a much bigger risk for Broadcom than anything Nvidia is facing."

Still, Wolfe Research analysts struck a bullish tone on Broadcom this week, upgrading it to buy-equivalent with a $400 price target—implying 21% upside. They expect roughly 7 million TPU shipments by 2028 as "Google's willingness to make TPU available to third parties creates a true competitor to Nvidia."

Morgan Stanley maintained buy ratings on both companies but noted a "preference for Nvidia despite rising ASIC enthusiasm." They're betting on Nvidia's upcoming Vera Rubin platform launching in the second half of 2026.

Market Reality Check

Luria expects Nvidia to maintain over 50% market share for at least five years and over 70% for the next three. The company's CUDA software ecosystem remains deeply entrenched among developers worldwide.

Both stocks face near-term headwinds. Nvidia battles multiple compression as investors grow less willing to pay premium valuations, while ongoing U.S.-China tensions create uncertainty. Broadcom has dropped nearly 20% from its record close, despite strong fundamentals.

Jim Cramer called Wolfe's Broadcom upgrade "timely," noting he doesn't understand "the poor action in Broadcom." For Nvidia, he maintains his "own it, don't trade it" stance while waiting for Jensen's presentation of Vera Rubin chips at the annual GTC conference in mid-March.