For years, Broadcom was the chip company that didn’t need the AI hype machine. It had its own gravitational pull — steady cash flows from enterprise software, a dominant position in networking semiconductors, and a CEO in Hock Tan who ran the company with the discipline of a private equity shop. But something shifted. The company’s partnership with Google, long a known but underappreciated piece of Broadcom’s AI story, is now emerging as a signal that the custom silicon market could be far bigger than most analysts originally modeled.
That’s the argument gaining traction on Wall Street after a string of analyst upgrades and new data points suggesting Broadcom’s role as a designer of custom AI accelerators — known as TPUs, or Tensor Processing Units — for Alphabet’s Google is deepening. According to Yahoo Finance, the Broadcom-Google relationship is being read as a proxy for the strength of the broader custom chip market, one that could rival or even outpace the merchant GPU business dominated by Nvidia.
The timing matters. Google recently disclosed plans to spend $75 billion on capital expenditures in 2025, a staggering figure that even by hyperscaler standards raised eyebrows. A meaningful portion of that spending is directed at AI infrastructure, including the custom TPU chips that Broadcom helps design and manufacture. That’s not a rounding error. It’s a structural commitment.
And Broadcom isn’t just riding Google’s coattails. The company has been quietly building a custom silicon practice that now counts multiple hyperscale cloud providers as clients, though Google remains the most prominent and publicly acknowledged partner. During its most recent earnings call, Broadcom management projected that its AI-related revenue could reach $12 billion in fiscal 2024, a number that caught many off guard when first disclosed and has only grown more credible since.
What makes this story compelling for industry insiders is the competitive dynamic it sets up against Nvidia. Jensen Huang’s company has dominated the AI training market with its H100 and now H200 GPUs, commanding extraordinary margins and a near-monopoly on high-end AI training hardware. But the hyperscalers — Google, Amazon, Meta, Microsoft — have every incentive to develop their own chips. Custom silicon offers them better performance per watt for specific workloads, lower total cost of ownership, and critically, less dependence on a single supplier.
Broadcom sits at the center of that diversification effort. The company doesn’t compete with Nvidia directly in the merchant chip market. Instead, it acts as a design partner, working with cloud giants to architect custom ASICs tailored to their specific AI workloads. It’s a fundamentally different business model. Less flashy. But potentially just as lucrative over time.
Wall Street is starting to price that in. Shares of Broadcom have surged more than 60% over the past year, and several analysts have recently raised their price targets. The logic is straightforward: if the total addressable market for AI accelerators is $100 billion or more by the end of the decade — a figure multiple research firms have floated — then the custom chip segment could represent 30% to 40% of that total. Broadcom, as the leading ASIC design partner, would capture a disproportionate share.
Not everyone is convinced. Skeptics point out that custom chip programs are expensive, time-consuming, and carry execution risk. Designing a new generation of TPU can take two to three years from concept to volume production, and each generation requires re-engagement with Broadcom’s engineering teams. That’s a fundamentally different cadence than buying off-the-shelf GPUs from Nvidia, which can be deployed in months. There’s also the question of software. Nvidia’s CUDA platform has created deep lock-in across the AI developer community, and custom chips require custom software stacks — an ongoing investment that not every company is willing to make.
But Google is willing. More than willing. The company has been building TPUs since 2015 and is now on its sixth generation, called Trillium. Each successive generation has closed the performance gap with Nvidia’s best GPUs for certain workloads, particularly inference — the process of running trained AI models in production. Inference is where the real volume will be as AI applications scale. Training gets the headlines. Inference gets the revenue.
Broadcom’s role in this is more than just chip design. The company also supplies the networking components — switches, routers, optical interconnects — that tie thousands of accelerators together in massive AI clusters. This is a point that often gets lost in the narrative. Building an AI supercomputer isn’t just about the chips. It’s about moving data between them fast enough that the system doesn’t bottleneck. Broadcom’s Tomahawk and Jericho switch families are already deployed at scale in hyperscaler data centers, and the company’s recent push into 800-gigabit Ethernet and co-packaged optics positions it to capture networking revenue alongside its ASIC business.
That dual revenue stream — custom silicon plus networking — is what gives Broadcom a structural advantage that pure-play chip companies can’t easily replicate. It’s also why Hock Tan has been so aggressive about framing Broadcom as an AI infrastructure company rather than just a semiconductor firm. The $69 billion acquisition of VMware, completed in late 2023, added another layer: enterprise software that runs in the same data centers where Broadcom’s chips and networking gear operate. Whether that integration delivers the synergies Tan has promised remains to be seen, but the strategic logic is clear.
Recent market activity reinforces the bull case. According to reporting from Yahoo Finance, the strengthening Broadcom-Google relationship has been interpreted by multiple sell-side firms as evidence that custom AI chip demand is accelerating faster than consensus expectations. Some analysts now model Broadcom’s AI revenue reaching $15 billion or more by fiscal 2025, which would represent a tripling from just two years earlier.
The broader context here is a capital expenditure arms race among the hyperscalers that shows no sign of slowing. Amazon’s AWS has committed to spending tens of billions on AI infrastructure. Microsoft, through its partnership with OpenAI, is building out GPU clusters at unprecedented scale. Meta is doing the same for its Llama models and recommendation systems. Every one of these companies is evaluating — or already deploying — custom silicon alongside Nvidia GPUs. And for many of them, Broadcom is the partner of choice for ASIC design.
So what are the risks? Concentration, for one. Google is believed to represent a significant majority of Broadcom’s custom AI chip revenue today. If Google were to slow its TPU program, bring design capabilities fully in-house, or shift spending toward Nvidia GPUs, Broadcom’s AI growth story would take a hit. The company has been working to diversify its customer base, and there are credible reports of engagements with other major cloud providers, but the Google dependency is real and shouldn’t be dismissed.
There’s also the question of valuation. Broadcom trades at roughly 30 times forward earnings, a premium to its historical average and to the broader semiconductor sector. That premium is justified only if the AI revenue growth materializes as projected. Any stumble — a delayed product cycle, a shift in hyperscaler spending priorities, a recession that crimps cloud budgets — could compress the multiple quickly.
And then there’s competition. Marvell Technology, another ASIC design house, has been aggressively pursuing its own hyperscaler relationships and recently disclosed a growing AI chip pipeline. While Marvell is smaller and less diversified than Broadcom, it’s a credible competitor in the custom silicon space and has been gaining share with certain cloud providers. The market is big enough for both, but Broadcom’s dominance isn’t guaranteed.
Still, the weight of evidence points in one direction. The hyperscalers are spending at record levels. Custom AI chips are a growing share of that spend. And Broadcom is the company best positioned to capture it, thanks to its engineering depth, its existing relationships, and its ability to bundle chip design with networking infrastructure. The Google partnership isn’t just a contract. It’s a proof point — evidence that the custom silicon model works at scale and can compete with the best merchant GPUs in the world.
For investors and industry watchers, the key question isn’t whether custom AI chips will matter. They already do. The question is how big this market gets and how fast. Broadcom is betting its next chapter on the answer being: very big, and very fast. Based on what Google is spending, that bet looks increasingly well placed.
WebProNews is an iEntry Publication