Google Locks In Broadcom for Custom AI Chips as the Silicon Arms Race Deepens

Broadcom will design and supply Google’s custom AI chips through the end of the decade, under a long-term agreement announced Monday that signals just how committed big tech remains to building its own silicon rather than feeding Nvidia’s margins forever.

The deal covers future generations of Google’s tensor processing units — the homegrown chips the search giant uses to train and run AI models — as well as other components for Google’s next-generation AI server racks. It runs through 2031. Financial terms were not disclosed.

Broadcom shares rose roughly 3% in after-hours trading.

The Anthropic Connection

Alongside the Google agreement, Broadcom announced an expanded deal with Anthropic that will give the AI startup access to approximately 3.5 gigawatts of computing capacity powered by Google’s TPUs, starting in 2027. That is a staggering amount of infrastructure — for context, 3.5 gigawatts is roughly equivalent to the peak demand of a mid-sized US state.

Anthropic’s chief financial officer, Krishna Rao, called the partnership “a continuation of our disciplined approach to scaling infrastructure” in a blog post. The company said most of the new compute will be sited in the United States, building on a November 2025 commitment to invest $50 billion in domestic computing infrastructure.

Amazon Web Services remains Anthropic’s primary cloud provider and training partner, the startup was careful to note — a reminder that these deals are additive, not exclusive. Anthropic runs its Claude models across AWS Trainium chips, Google TPUs, and Nvidia GPUs.

The Numbers Behind the Demand

The infrastructure spending is being pulled by explosive revenue growth. Anthropic said its run-rate revenue has surpassed $30 billion, tripling from approximately $9 billion at the end of 2025. Over 1,000 business customers are now each spending more than $1 million annually on Anthropic products — a figure that has doubled in less than two months, according to the company.

Analysts at Mizuho, led by Vijay Rakesh, estimated after Broadcom’s last earnings call that the company could generate $21 billion in AI revenue from Anthropic alone in 2026, and $42 billion in 2027. Those are analyst projections, not confirmed figures — Broadcom’s Monday filing did not include dollar amounts.

What This Means for Nvidia

The deals underscore a structural shift in how the biggest buyers of AI compute are thinking about hardware. Google’s push to make TPUs a viable alternative to Nvidia’s market-leading GPUs has been underway for years. Reuters reported in December that TPU sales had become a crucial growth engine for Google’s cloud revenue, as the company works to demonstrate to investors that its AI capital expenditure is generating returns.

Broadcom is now the common thread connecting multiple challengers to Nvidia’s position. In addition to the Google and Anthropic deals, the chip designer is collaborating with OpenAI on custom silicon, according to CNBC. OpenAI has also committed to drawing on six gigawatts of AMD’s GPUs — a separate effort to diversify away from Nvidia dependence.

None of this means Nvidia is in trouble. The company’s GPUs still power the vast majority of AI training and inference worldwide. But the direction of travel is clear: the largest consumers of AI compute are working systematically to reduce their reliance on a single supplier, and they are willing to sign decade-long contracts to do it.

Broadcom’s Emerging Monopoly on Custom Silicon Design

What makes Broadcom’s position unusual is that the company isn’t trying to compete with Nvidia directly. It doesn’t sell general-purpose GPUs. Instead, it designs custom chips for specific clients — and in doing so, it has quietly assembled a portfolio of relationships with nearly every major AI player that wants to break free of Nvidia’s pricing power.

Google, Anthropic, and OpenAI are all working with Broadcom. The pattern is consistent: big tech builds its own silicon, and Broadcom is the design shop they keep turning to.

The Spending Question Nobody Is Asking

The implicit bet across all of these deals is that AI demand will continue to grow fast enough to absorb the enormous infrastructure being built. Anthropic’s revenue numbers suggest that, at least for now, demand is outpacing even aggressive projections. But 3.5 gigawatts of capacity coming online in 2027 is a forward commitment based on current growth rates — and growth rates in technology have a habit of flattening when the bill arrives.

For the moment, though, the money is real and the contracts are signed. Broadcom just locked in a decade of work with one of the world’s largest buyers of chips, and the market for AI compute that doesn’t bear Nvidia’s logo keeps getting bigger.

Sources

Discussion (12)

Linda M. Rojas

I think it's worth clarifying that Broadcom isn't actually manufacturing these chips—they're designing them. The manufacturing still goes to TSMC, which is a whole separate conversation about supply chain concentration. Everyone talks about Broadcom like they're competing with Nvidia on the manufacturing side, but that's not really what's happening here. They design the IP, the clients own it, TSMC prints it. Still significant, but the dynamics are different than a straight-up chip war.

8 ↑

To clarify my earlier point: I'm not saying this deal isn't significant. It absolutely is. But the reporting in a lot of outlets conflates design capability with manufacturing capability, and those are really different things when you're assessing competitive risk. Broadcom's stock jumped on the news, but TSMC's stock would also jump if there were concerns about capacity. Different risk profiles.

11 ↑

Fair point but the ecosystem argument still stands. CUDA has 15 years of tooling and optimization. TPUs have... what, Google's internal stack and a handful of researchers who learned to work around its limitations? Anthropic runs on AWS Trainium too, which means they're maintaining multiple codebases. That complexity tax adds up fast.

4 ↑

derek_v

This.

2 ↑

Okay but, has anyone considered that Google locking themselves into Broadcom through 2031 might be a mistake? What if custom silicon hits a wall? They'd be stuck with architecture they can't pivot from. Nvidia's ecosystem is enormous and has decades of software tooling behind it. Custom TPU love is cute but Google's been here before and the ecosystem never quite materialized the way they hoped.

5 ↑

marcus_tech

Finally someone is actually fighting back against Nvidia's pricing. I've been saying for years these cloud companies were just going to roll their own. TPU v5 incoming 2028 probably GOOG to the moon 🚀

1 ↑

jenny_adventures

When I was in Singapore last month visiting friends, we spent a whole afternoon talking about how Southeast Asia is basically becoming this massive test bed for AI infrastructure decisions. The energy costs alone are a huge factor—3.5 gigawatts sounds abstract but when you think about it in terms of what that actually means for data center siting, it's kind of insane. I drove past a Google facility near Jurong and my friend who works in semiconductor logistics said they've been expanding constantly. This deal makes so much more sense when you've seen the physical scale of what's being built.

techtruth42

The fact that nobody is talking about the power infrastructure required for this is wild to me. 3.5 gigawatts is not a number you just throw around. Who's paying for the grid upgrades? Spoiler: ratepayers and taxpayers. These companies get the compute, we get the bill for the power lines. Every time.

12 ↑

skeptical_sam

Anthropic claiming $30B run rate revenue tripling from $9B in like 15 months is insane. Either their numbers are real and they're printing money, or someone's stretching what "run rate" means. Run rate accounting is basically make-believe math.

6 ↑

carl_from_sales

Bro what is even happening anymore. 3.5 gigawatts. That's a number. A big number. I don't know what it means but it's a lot. The future is insane.

3 ↑

definitely_not_a_bot

This article is literally AI generated and posted on an AI news site. The irony of an AI writing about AI chip infrastructure is lost on no one except apparently the 47 people who liked this article. Wake up sheeples

7 ↑

chiphead99

The article kind of buries the lead here. Broadcom's actual moat isn't just custom design—it's that they have the physical verification infrastructure to do this at scale. You can't just hire 200 engineers and replicate what they do. That's why OpenAI, Google, and Anthropic all ended up at the same shop. The real story is the concentration of custom silicon expertise.

18 ↑