Broadcom's AI Bet Is Expanding From Chips to Entire Compute Systems

Broadcom is expanding its AI accelerator business as Google, Anthropic, OpenAI and Meta build custom silicon and demand more compute.
Broadcom's latest results show how quickly custom AI silicon is becoming a larger part of the infrastructure supporting frontier models. Its third-quarter fiscal 2026 revenue reached $29.6 billion, up 86% year over year, while AI semiconductor revenue rose 221% to $16.7 billion.
The company expects AI semiconductor revenue to reach $21.7 billion in the fourth quarter. Tan also expects the business to reach approximately $115 billion in fiscal 2027 and $230 billion in fiscal 2028.
The forecast is tied to a small group of customers building increasingly large AI systems. Google, Anthropic, OpenAI and Meta are all pursuing custom accelerators, giving Broadcom a role in designing chips around particular workloads. The company has been expanding that business throughout the year, after previously forecasting more than $100 billion in AI semiconductor revenue for 2027.
>
Custom Silicon Moves Into the Mainstream
Google provides the longest-running example of Broadcom's custom-silicon business. Tan says Broadcom is shipping Google's Ironwood TPU in volume and has begun production shipments of TPU 8i, Google's latest accelerator designed for inference workloads.
Google says TPU 8i is designed for post-training, inference and reasoning workloads. It has 288 GB of HBM, 384 MB of on-chip SRAM and delivers up to 80% better performance per dollar for inference than the previous generation.
Broadcom expects the Google relationship to extend across future TPU generations. Tan says a long-term agreement covers the development and supply of future TPUs and AI networking, with Broadcom planning to deliver multi-tens of billions of dollars of TPUs to Google annually over the next several years.
Anthropic is also becoming a major customer. Broadcom says it is deploying one gigawatt of Ironwood capacity in 2026, followed by five gigawatts of TPU 8i in 2027 and a potential additional 10 gigawatts in 2028. Tan expects Anthropic to become Broadcom's largest XPU customer in 2027.
OpenAI is developing its own path through Jalapeño, its first custom inference chip developed with Broadcom. OpenAI says the accelerator moved from initial design to manufacturing tape-out in nine months, with OpenAI models used to accelerate parts of the design and optimization process.
OpenAI's first measured results show Jalapeño producing more AI work per unit of power while returning responses more quickly. Broadcom expects a 1.3-gigawatt Jalapeño deployment in 2027 and sees more than five gigawatts across Jalapeño and its successor generation in 2028.
Meta is also expanding its MTIA accelerator program. Broadcom expects production shipments of its next MTIA accelerator in the fourth quarter, with three generations planned through 2027 and a deployment path of three gigawatts through 2028. Meta says MTIA is optimized for inference and recommendation workloads at scale.
The Accelerator Is Only Part of the System
The growth of custom accelerators creates a corresponding need for networking. Broadcom says its Tomahawk 6 Ethernet switches are deployed across AI hyperscalers, including customers that use Broadcom accelerators and those that do not.
Its Tomahawk Ultra platform is also beginning to enter scale-up applications. The system allows accelerators within a cluster to communicate over Ethernet, extending Broadcom's networking business beyond connections between separate systems.
Broadcom is expanding its optical manufacturing capacity alongside that networking business. Kawwas says the company is more than tripling capacity for EML, continuous-wave and VCSEL products year over year across facilities in the U.S. and Singapore.
The supply chain has already faced constraints around advanced manufacturing capacity. advanced manufacturing capacity Tan identifies leading-edge wafers, substrates, HBM and system memory among the components that could constrain AI capacity. Land, power and data center construction can also determine when planned capacity becomes operational.
Broadcom is adding capacity of its own. Tan says Broadcom will begin deploying substrate capacity from its Singapore facility in fiscal 2027, while Kawwas says the company is increasing capacity for optical components in the U.S. and Singapore.
Broadcom is supplying networking, optical components and substrates alongside the silicon used to process AI workloads.
Turning Compute Demand into Revenue
Broadcom's long-term forecast depends on how much of the compute capacity requested by customers can actually come online. Tan says the company has secured supply for its $115 billion fiscal 2027 and $230 billion fiscal 2028 outlooks, while customers still have to secure sites, power and completed data centers.
The distinction becomes clearer when measured in gigawatts. Tan says Broadcom expects its semiconductor content to remain in the range of $20 billion to $30 billion per gigawatt. More powerful accelerators consume more power, which means fewer chips can fit within the same amount of available power.
Broadcom is also helping two customers address the financing required to build that capacity. In June, it established the AI XPV Platform with Apollo and Blackstone to enable more than 20 gigawatts of compute infrastructure for Anthropic and OpenAI through 2028.
The first $35 billion tranche supports Anthropic's one-gigawatt deployment. Thuener says Broadcom may participate in future financing arrangements depending on the requirements of each transaction, while third-party financial partners independently underwrite and capitalize the assets.
Broadcom's customer commitments provide visibility well beyond the current quarter. Its business now spans custom accelerators, networking, optical components and substrates, with demand tied to AI systems that customers are planning years ahead.
Key Takeaways
- Broadcom's AI semiconductor revenue surged 221% to $16.7 billion, reflecting rapid industry growth.
- The company anticipates AI semiconductor revenue to reach $21.7 billion in Q4 2026.
- Broadcom projects a staggering $115 billion in AI revenue by fiscal 2027 and $230 billion by fiscal 2028.
- Major clients like Google, Anthropic, OpenAI, and Meta drive demand for custom AI silicon.
- Broadcom expands its role in designing chips tailored for specific AI workloads amid rising demand.