Friday, September 11, 2026
AI 인프라 · 뉴스 & 분석
반도체·하드웨어리포트
반도체·하드웨어 · 리포트

Broadcom은 맞춤형 AI 가속기 수요 급증을 활용하여 하이퍼스케일러를 위한 특화 실리콘 설계 및 네트워킹 분야에서의 역할을 확대하고 있다.

증가된 ASIC과 네트워킹 통합은 GPU 클라우드와 하이퍼스케일러들이 차세대 AI 클러스터와 인터커넥트 토폴로지를 어떻게 설계하는지에 직접적인 영향을 미친다.
업계 전문지Slicast · September 11, 2026 · 미국 · 출처: The Next Platform
중요도 80

Broadcom's dominance in custom AI accelerators is proving even more consequential than expected. When CEO Hock Tan outlined the company's third-quarter fiscal 2026 results to Wall Street, his revenue forecast for Broadcom's custom XPU and AI networking business underscored the magnitude of the opportunity ahead.

Between fiscal 2022—when the GenAI boom began—and the fiscal 2028 forecast, Broadcom's custom AI XPU business plus AI networking and related components will grow by two orders of magnitude, from $2.6 billion to $230 billion annually. To put that in perspective, the US military budget for the fiscal year just ended was $1 trillion; worldwide pizza consumption in 2024 reached $230 billion.

Broadcom had already been supporting Google's custom TPU designs through Taiwan Semiconductor Manufacturing Co. for seven years before the GenAI wave hit. The company has now secured the HBM memory, chip wafers, and packaging to fulfill this revenue stream. Tan stated that if demand exceeds current forecasts, Broadcom can add additional capacity. He cautioned Wall Street against expecting quarterly updates to the forecast but made clear that substantial revenue would flow in the coming years.

Across six primary XPU customers—Google, Anthropic, OpenAI, Meta Platforms, and two others from a potential pool including ByteDance and Apple, with AWS using Broadcom for Trainium 4 packaging and integration—there is roughly 30 gigawatts of total compute capacity. This suggests XPU systems cost approximately $12 billion per gigawatt before accounting for CPUs, memory, and racks; closer to $15 billion per gigawatt when those components are included. This compares favorably to $18 billion per gigawatt for Nvidia Grace-Hopper systems and $40 billion per gigawatt for Nvidia Vera-Rubin systems.

On the Google relationship, Tan stated that Broadcom would ship "multi-tens of billions of dollars of TPUs annually over the next several years" to the search giant. "We have consistently delivered the fastest time to market for TPUs from product definition to production without the need for resubmits," Tan said. "We believe these are very deep moats for any competitor to cross."

Anthropic is deploying 1 gigawatt of Ironwood TPU 7 accelerators this year, will deploy 5 gigawatts of TPU 8i systems in 2027, and plans to acquire 10 gigawatts of the follow-on TPU 9 in 2028. Tan said Anthropic would be Broadcom's largest XPU customer in 2027 and would sustain that position in 2028—meaning Anthropic is purchasing more TPU capacity than Google itself, despite Google being both a cloud provider with significant GPUs needs from rivals and an AI developer that might be expected to prioritize internal TPU procurement.

OpenAI has selected Broadcom to shepherd its Jalapeno XPU and is on track to purchase 1.3 gigawatts from Broadcom for Jalapeno clusters in 2027. Broadcom projects OpenAI will deploy an additional 5 gigawatts of Jalapeno ASICs and its successor in 2028, making OpenAI its second-largest XPU customer and pushing Google to third. Tan said the follow-on OpenAI chip is approaching tapeout, with the two companies now working on a third-generation design.

With Meta Platforms, Broadcom will deliver three generations of MTIA devices through the end of fiscal 2027, with line of sight on 3 gigawatts of total capacity across those generations through fiscal 2028.

Beyond XPUs, networking dominates Broadcom's AI expansion. The 100 Gb/sec and 200 Gb/sec SerDes variants of the Tomahawk 6 Ethernet switch ASIC are in high demand; hyperscalers and cloud providers using Broadcom-shepherded XPUs are deploying Tomahawk 6 for scale-out networking in AI clusters. The Tomahawk Ultra variant, which reduces latency in the Ethernet stack to enable scale-up coherent memory interconnects for XPUs within rackscale machines, is seeing surprising adoption with ramp progressing through fiscal 2027. The unannounced 204.8 Tb/sec Tomahawk 7 Ethernet switch ASIC, which uses 400 Gb/sec SerDes and delivers 3.2 Tb/sec per port with eight lanes per port, has just taped out and should appear in switches starting in early calendar 2028 or possibly late calendar 2027.

Broadcom is also likely to expand into CPU shepherding, particularly packaging and limited design work, as it is doing with Fujitsu for the forthcoming Monaka Arm Server CPU. Fujitsu and the six XPU customers are all using Broadcom's 3.5D Extreme Dimension System in Package (XDSiP) chiplet stacking, which integrates stacked DRAM and CPU chiplets into a single package. With the CPU total addressable market growing from $26 billion in 2025 to $220 billion by 2030, and with every hyperscaler and cloud builder developing its own Arm server CPUs, Broadcom is likely to deepen its CPU shepherding role. Integration of CPU tiles onto XPU tiles or large external CPU address spaces paired with memory-constrained XPUs could prove strategically valuable.

In the third quarter ended in early August, Broadcom reported $29.59 billion in total revenue, up 85.5 percent year-over-year and 33.4 percent sequentially. Operating income was $15.96 billion, up 2.71X compared to Q3 fiscal 2025. Net income grew faster still, rising 3.17X to $13.09 billion, representing 44.2 percent of revenues.

Tan's model of serving as a chip shepherd—managing designs, partnerships, and production for customers who each demand unique solutions—has proven more profitable than attempting to sell finished XPUs or CPUs to technology titans who refuse to standardize on solutions their rivals use, with the exception of x86 CPUs and Nvidia (and sometimes AMD) GPUs, tolerated only by cloud providers constrained by customer demand and GPU scarcity.

Broadcom ended the quarter with $23.98 billion in cash against $59.42 billion in debt from the VMware acquisition.

The Semiconductor Solutions group generated $20.84 billion in sales, up 127.4 percent, with operating income of $12.71 billion, up 143.3 percent. The Infrastructure Software group, which includes VMware and legacy enterprise systems software, reported $8.75 billion in sales, up 29 percent, and operating income of $7.35 billion, up 40.7 percent at 84 percent of revenues.

원문 보기