Broadcom draws a curve no one can falsify

$29.591 billion, an 86‑percent year‑over‑year increase. Four years ago, Broadcom’s full‑year revenue was roughly at this level; today this is merely one quarter’s figure.

After market close on September 2, Broadcom released its fiscal 2026 Q3 earnings report, a financial statement whose revenue mix has been rewritten.

What is the status of remaining businesses, with AI accounting for 56$?

Breaking down by segment: of the total $29.591 billion revenue, AI semiconductors contributed $16.7 billion, accounting for approximately 56% of total revenue and around 80% of the semiconductor segment. The remainder falls into two buckets.

The first bucket is non‑AI semiconductors, generating roughly just over $4 billion. This covers traditional strong categories including broadband access, storage connectivity, and switch chips, which are currently at a cyclical bottom. Investment in enterprise networking and telecom equipment has yet to show a meaningful recovery, with growth rates roughly flat or showing modest fluctuations. This segment contributes little to overall growth yet retains value in terms of product portfolio and customer relationships. The long‑accumulated technical capabilities such as SerDes and network interconnectivity within these product lines form precisely the engineering foundation enabling Broadcom’s custom ASIC business.

The second bucket is infrastructure software, the business segment formed after the VMware acquisition. Back‑calculated from reported revenue, its scale stands at about $8.7 billion. Its growth rate is far lower than that of the AI business, yet it is the company’s most stable division: subscription‑based revenue, high renewal rates, and robust cash flow. The market tends to label Broadcom an AI company. In reality, this business unit consistently generates steady cash flow, underpinning the company’s R&D investment and customer expansion in AI.

The conclusion is clear: the 86‑percent growth rate is single‑factor‑driven. The AI segment delivers the vast majority of incremental growth, while other businesses register low‑speed expansion or bottoming‑out conditions. This structure maximizes upside elasticity at the cost of rendering overall growth extremely sensitive to a single variable.

From $58 billion to $230 billion: real‑world underpinnings and hard constraints of the aggressive roadmap

Management’s projected trajectory: fiscal 2026 AI revenue guidance is raised from $56 billion to $58 billion (186% year‑over‑year growth). Subject to secured supply, AI revenue will double to roughly $115 billion in fiscal 2027, then double again to $230 billion in fiscal 2028, with approximately $350 billion in cumulative deliveries over the next two years.

The logic underpinning these figures stems from the long‑term contract‑locked nature of custom ASICs. Once customers lock in an architecture, switching suppliers would mandate new tape‑outs and qualification, incurring substantial sunk costs. Cooperation therefore features strong stickiness, granting multi‑year order visibility.

Furthermore, tangible upstream‑and‑downstream evidence corroborates this outlook. Dell posted fiscal 2026 Q2 revenue of $47 billion, up 58% year‑over‑year, with AI server backlog reaching $95 billion and full‑year revenue guidance lifted to around $192 billion. TSMC Co‑COO Hou Yong‑ching also stated that to accommodate AI capacity expansion, quarterly equipment procurement estimates have risen to roughly 1.9 times the December‑2025 forecast, alongside approximately 20 fabs under construction.

Translating this roadmap into actual shipment volumes requires overcoming three practical constraints first.

Advanced packaging capacity. Custom ASICs rely heavily on advanced packaging processes such as CoWoS. TSMC’s CoWoS capacity allocations for next year are largely pre‑booked, with lead times for some customers stretching beyond one year. Reports indicate that Broadcom and Meta are evaluating a shift from CoWoS to Intel’s EMIB solution for cost‑reduction purposes. MediaTek has publicly announced parallel adoption of both technologies, and SK Hynix is assessing substrates integrating Intel EMIB to combine HBM and logic chips. Packaging capacity represents the industry’s most tangible bottleneck, dictating the pace at which the roadmap can materialize.

Customer base structure. The $350‑billion‑worth projected demand derives mainly from a handful of large‑scale hyperscaler cloud providers. This structure brings high visibility and deep collaboration, yet lacks diversification. Any single participant delaying a chip generation, adjusting procurement cadence, or shifting toward in‑house‑developed packaging will exert a direct impact on production schedules.

Competitive boundaries. Leveraging platforms including NVLink Fusion, NVIDIA is shifting competition from individual GPUs to the rack level, while expanding its ecological reach through a $3.5‑billion convertible‑bond subscription in MediaTek. As general‑purpose solutions absorb flexibility advantages previously unique to custom chips in interconnect and integration, the dividing line between the two approaches will grow blurrier than today.

From an industrial perspective, the most valuable takeaway from this earnings release is not the 221‑percent growth figure. Instead, it lays bare the reality that bottlenecks in computing‑power supply have shifted from front‑end manufacturing to back‑end packaging. Key monitoring priorities going forward are the actual ramp‑up rhythm of advanced‑packaging capacity, the maturity of alternative solutions such as EMIB, and the generational advancement pace of cloud vendors’ in‑house chips. These three variables jointly determine the growth slope for the custom‑chip segment over the next three years.