Broadcom's fiscal third-quarter 2026 earnings report highlighted explosive growth in its custom AI chip (XPU) segment, which is delivering a compelling value proposition against traditional GPUs. Company executives provided an exceptionally strong long-term forecast, projecting that AI revenue will double annually over the next two years, a target that significantly exceeded market expectations.

For the quarter ending August 2, 2026, the company's net revenue surged 86% year-over-year to $29.59 billion, setting another record and slightly beating the anticipated $29.45 billion. This performance marked the highest single-quarter growth rate in over nine years, building on the previous quarter's momentum.

The primary engine behind these remarkable results is the AI semiconductor business, which saw revenue skyrocket 221% year-over-year and 54% sequentially to $16.7 billion. Looking ahead, Broadcom forecasts total fourth-quarter revenue of $34.8 billion, including a projected $21.7 billion from AI, a year-over-year increase of 236%. Consequently, the company has raised its full-year fiscal 2026 AI revenue guidance from $56 billion to $58 billion.

During the earnings call, CEO Hock Tan described the demand as intense, stating, "We are just getting started." He outlined an ambitious roadmap, projecting that AI chip revenue will double to approximately $115 billion in fiscal 2027 and further escalate to $230 billion in the following fiscal year. The company also anticipates earnings per share exceeding $30 by fiscal 2028, surpassing Wall Street's current forecasts.

Tan further detailed expectations for massive deployments of TPU accelerators over the coming years. He noted that Anthropic is slated to deploy 5 GW of TPU capacity in 2027, positioning it to become Broadcom's largest chip customer that year. OpenAI is also expected to deploy 5 GW of chips by 2028, becoming the company's second-largest customer.

Custom Chips Deliver Superior Performance at Half the Cost

The market's focus is keenly on Broadcom's future growth prospects, particularly the deployment plans of its six major XPU (custom accelerator) customers, which include tech giants like Google, OpenAI, Anthropic, and Meta. Tan emphasized that leading AI model developers are rapidly adopting XPUs to achieve superior performance, cost efficiency, and lower power consumption.

Broadcom disclosed remarkably strong order visibility across its customer base. For Google, the company is in high-volume production of the Ironwood (TPU v7) chip and has begun producing the next-generation TPU v8i, which offers enhanced memory and bandwidth. OpenAI's first-generation custom accelerator, Jalapeno, shipped in the third quarter, with plans to deploy 1.3 GW in 2027 and over 5 GW by 2028, making it Broadcom's second-largest XPU customer. Collaboration on a third-generation XPU is already underway. Anthropic is set to deploy 1 GW this year, 5 GW in 2027, and an additional 10 GW in 2028. Meanwhile, Meta is ramping up production of its custom MTIA accelerators, with expectations to deliver three generations of products by the end of 2027, reaching a deployment scale of 3 GW.

Tan delivered a powerful summary of the custom chip advantage over traditional GPUs, referencing OpenAI's recent announcement that its Jalapeno chip outperforms Grace Blackwell Ultra in latency, throughput, and power consumption, performing comparably to Vera Rubin GPU on OpenAI workloads. He explained that jointly developing a chip optimized for specific LLM workloads allows customers to surpass any general-purpose GPU, and crucially, achieve this at half the cost.

Long-Term Forecast: AI Revenue to Double Consecutively

Responding to the surging demand, Broadcom broke from its typical quarterly guidance pattern, providing ambitious targets for fiscal 2027 and 2028 directly to the market. Tan stated that supply for 2027 is secured to double AI revenue to about $115 billion, while acknowledging that actual demand exceeds this estimate and the company will work to expand supply. He also expressed clear visibility into 2028, projecting another doubling of AI semiconductor revenue to $230 billion.

Furthermore, Tan conveyed strong confidence in achieving earnings per share exceeding $30 in fiscal 2028. When pressed by analysts on the supply constraints behind the $230 billion target, he humorously noted that whatever supply figure he mentions, analysts will likely report a larger number, so the company strives to remain conservative. He clarified that the guidance is based on the current leading-edge wafer, substrate, and HBM memory supply chains, along with customer data center site readiness by 2028.

Addressing Growth Challenges: Supply Chain and Margin Dilution

Despite the ambitious vision, the market has raised concerns in two key areas. The first is gross margin dilution. Due to the increasing proportion of XPUs and high-bandwidth memory (HBM) in AI hardware, Broadcom's consolidated gross margin declined sequentially to 75% in Q3, with a projection of 73% for Q4. However, CFO Amie Thuener emphasized that strong revenue growth is generating significant operating leverage, with operating margins expected to remain stable at around 66% in Q4. Tan directly addressed the concern, advising investors to "stop focusing on gross margin" and instead look at what truly matters: the resulting operating margin.

The second concern is the physical bottleneck of deploying computing capacity, specifically land, power, and shell (LPS). Tan acknowledged that LPS is the key determinant of deployment timelines, which is why Broadcom is not only providing chips but also partnering with Apollo and Blackstone to establish the "AI XPV platform" to provide financing support for OpenAI and Anthropic. He calculated that each gigawatt of computing capacity deployed could generate $30 billion in annual recurring revenue, calling it an excellent business model.

Broadcom's Moat in Networking and Manufacturing

On the networking side, Broadcom maintains a deep competitive moat. Its 100T Tomahawk 6 switching chip has been highly successful, and the 200T Tomahawk 7 has already been taped out. More significantly, the Ethernet-based Tomahawk Ultra is being widely adopted for scale-up interconnects in both XPU and GPU clusters, breaking down traditional proprietary protocol barriers. To address packaging substrate constraints, the company announced it will begin production at its new Singapore facility starting in fiscal 2027.