Broadcom reported a massive surge in its artificial intelligence chip business this week, signaling that major technology companies are moving aggressively to design and deploy their own custom processors rather than relying entirely on off-the-shelf hardware.
The semiconductor and software company said its AI semiconductor revenue more than tripled year-over-year to $16.7 billion in its fiscal third quarter. Total quarterly revenue rose 86% to $29.6 billion, easily clearing Wall Street expectations.
But it was the company’s long-term forecast that drew the most attention. In a rare move for a semiconductor supplier, Broadcom Chief Executive Officer Hock Tan laid out a multi-year financial roadmap based on guaranteed manufacturing supply and firm customer orders. The company raised its fiscal 2026 AI revenue target to $58 billion. For fiscal 2027, Broadcom expects AI chip revenue to double to roughly $115 billion, before doubling again to $230 billion in 2028.
“Demand was simply hot, and we’re just getting started,” Tan told analysts on the company’s earnings call Wednesday.
The results show how the AI infrastructure build-out is maturing. In the initial rush to develop large language models, technology giants bought standard graphics processing units (GPUs) as quickly as manufacturers could produce them. Now, major cloud providers are increasingly running their workloads on custom-designed processors—often co-developed with Broadcom—to lower operating costs and reduce power consumption.
Unprecedented Multi-Year Guidance
Technology hardware companies typically avoid forecasting specific revenue numbers more than one or two quarters in advance due to the cyclical nature of the semiconductor industry. By publishing firm targets through 2028, Broadcom is showing unusual confidence in its order book and supply chain.
Tan noted that the multi-year figures are based on secured supply agreements for wafer fabrication, advanced packaging substrates, and high-bandwidth memory (HBM). The projections also rely on conservative estimates of data center power availability over the next two years.
Broadcom expects its fourth-quarter AI revenue to reach $21.7 billion.
The Rise of Custom Silicon
Broadcom’s growth is heavily concentrated among a small group of large technology companies. Six major customers currently account for the bulk of the company’s AI chip revenue. Shipments to these companies increased more than 3.5 times year-over-year.
The most prominent of these partnerships is with Google. Broadcom confirmed it is currently shipping Google’s seventh-generation Tensor Processing Unit, known as Ironwood (TPU v7), in high volume.
At the same time, Broadcom has initiated production of Google’s next-generation chip, the TPU v8i. During the call, Tan stated that the new processor equals or surpasses the performance of upcoming merchant silicon, specifically referencing Nvidia’s planned Vera Rubin GPU. The two companies recently signed a new long-term agreement covering tens of billions of dollars of TPUs annually over the next several years.
Why Hyperscalers Want Custom Chips
To understand Broadcom’s momentum, it is helpful to look at how data centers are built. Standard GPUs are versatile and can handle a wide variety of computational tasks. However, this versatility comes with a trade-off in physical chip space and electrical efficiency.
When a technology giant knows exactly which software workloads it will run for the next three years, it can strip away the unnecessary components found in standard GPUs. By working with Broadcom, a company like Google can design a processor specifically optimized for its own proprietary network architecture. This results in an application-specific integrated circuit (ASIC), which Broadcom refers to as an XPU.
According to Tan, custom accelerators perform specialized workloads more efficiently and cost less than half the price of standard commercial equivalents. While the upfront engineering costs to develop custom chips are high, companies operating at a massive scale save billions of dollars over time through improved efficiency and lower hardware costs.
Financing the Next Generation of AI Labs
The massive scale of computing infrastructure has created new financial structures within the technology industry. Developing advanced models now requires data centers capable of drawing gigawatts of power, pushing construction costs into the tens of billions of dollars.
To facilitate this growth, Broadcom detailed its involvement in a specialized financing platform created in partnership with private equity firms Apollo and Blackstone. The initiative, known as the AI XPV platform, is designed to fund large-scale computing deployments for AI research laboratories.
Broadcom disclosed that the first $35 billion tranche of this funding closed in June. The capital is currently financing a one-gigawatt deployment for Anthropic, an AI research company heavily backed by Amazon and Google. The financing vehicle aims to support more than 20 gigawatts of compute capacity for companies like Anthropic and OpenAI by the end of 2028, allowing independent AI labs to build infrastructure at the same scale as major public cloud providers.
The Networking Backbone
While processors receive the most public attention, connecting thousands of chips within a data center requires highly specialized networking hardware. Broadcom holds a dominant market position in this category, and its networking division is expanding alongside its compute business.
The company reported that its AI networking revenue grew more than 2.5 times compared to the same period last year. Broadcom’s newest high-capacity data center switch, the Tomahawk 6, is now deployed across nearly all major AI hyperscalers.
Crucially, this networking equipment is purchased both by customers building custom chips with Broadcom and by companies using standard processors from competitors. Even when a data center relies entirely on Nvidia GPUs for computation, it often uses Broadcom Ethernet switches to manage the flow of data between the server racks.
By the Numbers
Broadcom’s focus on high-margin AI infrastructure has resulted in significant cash generation.
In the fiscal third quarter, the company’s operating income grew 92% to $20.1 billion, resulting in a record operating margin of 67.9%. Gross margins stood at 75% of revenue, down slightly from the previous quarter as the hardware-heavy semiconductor business represented a larger portion of the total sales mix.
The company generated $13.7 billion in free cash flow, accounting for 46% of total quarterly revenue. Broadcom used a portion of this cash to pay down $5.6 billion in long-term debt during the quarter. The debt reduction comes as the company continues to integrate software maker VMware following its acquisition.
Market Implications
The financial results validate a broader shift in the semiconductor industry. While Nvidia remains the undisputed leader in selling merchant graphics processors to the broader enterprise market, the largest cloud providers are successfully diversifying their hardware supply chains.
Broadcom’s strategy focuses exclusively on these massive buyers. The company is not attempting to sell custom chips to standard enterprise customers. Instead, it provides the engineering resources and intellectual property necessary for the world’s wealthiest technology companies to build exactly what they need.
“We will ship $350 billion of AI semiconductors to these customers in the next two years,” Tan told investors.
Despite the massive revenue increases, the cost to deploy data centers on a per-gigawatt basis is stabilizing. Tan estimated that the cost per gigawatt remains between $20 billion and $30 billion. Although individual chip prices are rising, the newer processors handle significantly more computation per watt of electricity, meaning companies need to buy fewer physical chips to achieve the same total computing power.
This dynamic suggests that while total capital expenditures by technology giants will continue to rise, the underlying efficiency of the hardware is improving. For Broadcom, the primary challenge over the next two years will not be finding buyers, but rather securing enough manufacturing capacity to meet its own ambitious production targets.
