Broadcom Forecasts Rapid Growth in Custom AI Accelerators

Broadcom’s custom AI opportunity

Broadcom’s custom XPU, networking ASIC, and related AI-cluster businesses are expanding rapidly. The company’s chief executive officer, Hock Tan, provided Wall Street with an especially ambitious forecast while discussing Broadcom’s third-quarter fiscal 2026 results.

Broadcom’s AI-related chip revenue is expected to grow from $2.6 billion in fiscal 2022 to $230 billion in fiscal 2028. This includes custom AI XPUs, AI networking, and other components used in AI systems.

Broadcom had already worked with Google for seven years to design and manufacture the company’s in-house TPUs through Taiwan Semiconductor Manufacturing Co. when the generative AI boom began. The company says it has secured the HBM memory, wafers, and packaging capacity needed to support its projected AI compute and networking revenue. Tan also said Broadcom could add capacity if demand exceeds current expectations.

Tan cautioned investors not to expect quarterly updates to the forecast. His objective was to emphasize the scale of expected revenue over the coming years and to indicate that Broadcom does not anticipate an imminent collapse in generative AI demand.

XPU customers and system costs

Broadcom has six XPU customers. Google, Anthropic, OpenAI, and Meta Platforms are confirmed customers. The remaining two may include ByteDance, which owns TikTok, and Apple. Amazon Web Services is using Broadcom for packaging and integration for Trainium 4.

Together, these customers represent approximately 30 gigawatts of compute capacity. This suggests a cost of about $12 billion per gigawatt for XPU systems, excluding CPUs, memory, and racks. Including those components, the estimated cost rises to approximately $15 billion per gigawatt.

For comparison, Nvidia’s Grace-Hopper systems cost about $18 billion per gigawatt. That figure includes CPUs, memory, GPUs, and the interconnects required for a rack-scale system. Nvidia’s Vera-Rubin systems are estimated at $40 billion per gigawatt.

Tan said Broadcom expects to ship “multi-tens of billions of dollars of TPUs annually over the next several years” to Google. He also highlighted the company’s ability to move TPU designs from product definition to production quickly without requiring resubmissions. Broadcom’s access to HBM allocations remains an important factor in its ability to support these deployments.

Anthropic and OpenAI deployments

Anthropic is deploying 1 gigawatt of Ironwood TPU 7 accelerators this year. The company plans to deploy 5 gigawatts of TPU 8i systems in 2027 and 10 gigawatts of the follow-on TPU 9 in 2028.

Tan said Anthropic is expected to become Broadcom’s largest XPU customer in 2027 and remain in that position in 2028. These figures may refer to Broadcom’s fiscal years, although Tan did not specify this when discussing the deployments.

OpenAI has selected Broadcom to manage the development and production process for its Jalapeno XPU. OpenAI is expected to purchase 1.3 gigawatts of Broadcom products for Jalapeno clusters in 2027. In 2028, Broadcom expects OpenAI to deploy another 5 gigawatts of Jalapeno ASICs and their successor, making OpenAI the company’s second-largest XPU customer and moving Google to third place.

The successor to OpenAI’s current chip is approaching tapeout, according to Tan. The two companies are also working on a third-generation XPU design.

For Meta Platforms, Broadcom will deliver three generations of MTIA devices between now and the end of fiscal 2027. Across those generations, Broadcom expects to serve approximately 3 gigawatts of Meta’s capacity requirements through the end of fiscal 2028.

Networking becomes a major AI business

Broadcom’s AI growth is not limited to XPUs. Networking is a larger part of the business, particularly through the Tomahawk 6 Ethernet switch ASIC.

The 100 Gb/sec and 200 Gb/sec SerDes versions of Tomahawk 6 are being adopted by hyperscalers and cloud providers that use Broadcom to develop their in-house XPUs. These customers are also using Tomahawk 6 for scale-out networking in AI clusters.

The Tomahawk Ultra version of the series reduces latency in the Ethernet stack, allowing it to support scale-up coherent memory interconnects between XPUs inside rack-scale systems. Adoption has begun at some customers, with deployment expected to increase through fiscal 2027 as new XPU products reach the market.

Broadcom’s 204.8 Tb/sec Tomahawk 7 Ethernet switch ASIC has also taped out. The product has not yet been formally announced. It uses 400 Gb/sec SerDes and provides 3.2 Tb/sec per port with eight lanes per port. Switches based on the chip are expected to appear in early calendar 2028, and possibly in late calendar 2027.

Possible expansion into CPU integration

Broadcom may eventually begin managing CPU development and increasingly focus on packaging, with limited design work. The company is already working with Fujitsu on its forthcoming Monaka Arm server CPU.

Fujitsu and the six XPU customers are using Broadcom’s 3.5D Extreme Dimension System in Package, or XDSiP, chiplet-stacking technology. XDSiP places stacked DRAM and stacked CPU chiplets into a single package.

The CPU total addressable market is projected to grow from $26 billion in 2025 to $220 billion by 2030. At the same time, hyperscalers and cloud providers are developing their own Arm server CPUs. Broadcom could therefore expand into CPU integration and development support.

Broadcom is unlikely to design its own Arm server processor for sale to these companies. The company previously sold its Vulcan processor to Cavium Networks. A more likely approach would be helping customers integrate CPU tiles with XPU tiles when they choose fixed CPU-to-XPU ratios.

An external CPU may also remain useful because it can provide a large memory address space, even if that memory is slower than HBM or other stacked memory. This could help XPUs that use less stacked memory because of limited capacity.

Fiscal third-quarter results

In the third quarter ended in early August, Broadcom reported $29.59 billion in total revenue. Revenue increased 85.5 percent year over year and 33.4 percent sequentially.

Operating income was $15.96 billion, up 2.71 times compared with the third quarter of fiscal 2025. Net income increased 3.17 times to $13.09 billion, equal to 44.2 percent of revenue.

Broadcom ended the quarter with $23.98 billion in cash and $59.42 billion in debt, including debt associated with the VMware acquisition.

Segment performance

Broadcom’s Semiconductor Solutions segment generated $20.84 billion in sales, an increase of 127.4 percent. Operating income was $12.71 billion, up 143.3 percent.

The Infrastructure Software segment, which includes VMware and other enterprise software businesses, generated $8.75 billion in third-quarter sales, up 29 percent. Operating income was $7.35 billion, representing 84 percent of segment revenue, and increased 40.7 percent.

An estimate attributed approximately $5.52 billion of the software segment’s revenue to VMware licenses and support, with the remaining $3.23 billion coming from other legacy software. Broadcom provides limited detail about the performance of its software businesses, so this breakdown is approximate.

AI and non-AI semiconductor revenue

AI-related semiconductor revenue totaled $16.74 billion in the quarter. That was 3.22 times the amount reported a year earlier and 54 percent higher sequentially.

Of the AI-related total, $12.22 billion came from AI compute and $4.52 billion came from AI networking. The figures illustrate the growing contribution of custom accelerators and networking products to Broadcom’s semiconductor business.