Broadcom is redirecting significant AI capital expenditure and engineering resources toward custom silicon and networking solutions that extend beyond traditional GPU architectures.
Last week’s Nvidia earnings confirmed that demand for AI compute remains far from peaked. Following Broadcom’s earnings report, the broader picture is now coming into focus. Third-quarter AI semiconductor revenue reached $16.7 billion, surpassing prior guidance of $16 billion, with next quarter’s guidance raised sharply to $21.7 billion. More significantly, management projects approximately $115 billion in AI chip revenue for fiscal year 2027, and up to $230 billion by fiscal year 2028. Consequently, debates over whether ASICs represent AI’s second growth curve are becoming increasingly moot; AI capital expenditure is no longer flowing exclusively into GPUs.
For Broadcom, this trajectory is already reflected in its financials, with forward-looking projections exceeding earlier expectations. Going forward, the critical variables are whether major initiatives from Google, OpenAI, Anthropic, and other leading firms will launch on schedule, and what share Broadcom will ultimately capture. Provided execution remains smooth, the momentum continues.
1. GPUs Keep Racing Ahead, but Big Tech Is Also Making New Moves
The most significant takeaway from Nvidia’s earnings is that, despite an already elevated baseline, cloud providers, AI firms, and model laboratories continue to aggressively expand compute capacity. Yet as AI capital expenditure scales from billions to tens of billions, and eventually hundreds of billions annually, procurement strategies are inevitably shifting. Enterprise buyers are now asking sharper questions: What is the actual cost to execute a specific AI workload? How much usable compute can be extracted from a fixed 1GW of power? And when certain workloads stabilize, is it still economically rational to run them on premium general-purpose hardware? These questions form the foundation for ASICs’ growing strategic importance.
Characterizing ASICs merely as “cheaper GPUs” misses the point; their true advantage lies in specialization. Custom chip development requires substantial upfront R&D, lengthy cycles, and solutions to complex integration challenges across software, advanced packaging, networking, systems engineering, and supply chains. Conversely, once workloads stabilize and deployments scale from tens of megawatts to hundreds of megawatts, and eventually gigawatts, gains in unit cost efficiency, performance-per-watt, and total cost of ownership (TCO) compound rapidly. This explains why Google has long backed TPUs, Meta continues scaling its MTIA architecture, and OpenAI is co-developing custom processors with Broadcom. Each recognizes that migrating even a fraction of their most stable, high-volume workloads from general-purpose silicon to custom hardware unlocks significant economic leverage.
Broader industry tailwinds are accelerating this transition, most notably the rise of inference as the primary driver of incremental workload. While model training remains episodic, once models like ChatGPT, Gemini, and Claude—and a growing ecosystem of AI agents—reach full production, inference becomes a continuous, daily operation. As request volumes surge and model architectures mature, inference naturally aligns with workload-specific hardware optimization. In June, OpenAI officially unveiled its first co-developed Intelligence Processor with Broadcom, codenamed “Jalapeño,” designed specifically for large language model inference. Notably, OpenAI and Broadcom have framed this not as a standalone product, but as the first generation of a multi-generational compute platform. Initial deployment is slated for late 2026, with long-term targets reaching GW-scale capacity.
Meta follows a parallel trajectory, planning to advance four generations of MTIA over two years. Focused on recommendation, ranking, and generative AI, Meta’s new silicon explicitly adopts an inference-first architecture. In April, Meta deepened its partnership with Broadcom; the initial deployment phase has already surpassed 1GW, with subsequent phases targeting multiple gigawatts, supported by a multi-generational collaboration extending through 2029. This evolution signals a fundamental shift among the largest AI customers: moving beyond the pursuit of peak performance toward optimizing token economics. As AI transitions into full commercialization, metrics such as cost per million tokens, compute yield per watt, and enterprise data center TCO will dictate competitive advantage. Ultimately, ASIC adoption represents a strategic effort by major AI firms to reclaim direct control over infrastructure costs.
2. Broadcom's Real Bet Isn't Just ASICs — It's the Entire AI Cluster Getting Bigger
Contextualizing these trends clarifies Broadcom’s recent financial results. Last quarter, total revenue reached $22.187 billion, a 48% year-over-year increase, while AI semiconductor revenue surged 143% to $10.8 billion. Forward guidance is even more robust: total revenue is projected at approximately $29.4 billion for the current quarter, with AI semiconductor revenue expected to hit $16 billion, representing over 200% year-over-year growth. At $16 billion, AI semiconductors would account for roughly 54% of Broadcom’s total quarterly revenue. If realized, this segment alone would generate more than half of the company’s top line—a figure that already factors in infrastructure software operations like VMware. Simply put, AI is fundamentally restructuring Broadcom’s revenue composition.
However, a common misconception must be addressed: the $16 billion cannot be equated solely with ASIC revenue. It also encompasses AI networking products, including Ethernet switch chips, SerDes, PCIe controllers, and optical interconnects. Networking already represented nearly 40% of AI semiconductor revenue last quarter. CEO Hock Tan indicated this ratio may be approaching a cyclical peak, with a long-term equilibrium closer to 30%. Consequently, Broadcom is driving growth across two parallel fronts—Custom XPUs and AI Networking—a dual-engine strategy that distinguishes it from many pure-play semiconductor peers.
As AI clusters expand from thousands to tens of thousands, and eventually hundreds of thousands of accelerators, interconnect architecture becomes a critical bottleneck. Broadcom addresses this by segmenting AI networking into scale-up, scale-out, and scale-across tiers: high-speed intra-rack links, large-scale intra-datacenter fabrics, and multi-datacenter cross-connects. This positioning ensures Broadcom benefits regardless of hardware preference. If hyperscalers accelerate custom ASIC development, Broadcom supplies the custom silicon; if GPU-centric clusters continue to dominate, Broadcom captures share through open Ethernet networking ecosystems. In essence, Broadcom is wagering on the sustained escalation of complexity across the entire AI infrastructure stack.
Competition, however, remains a constant factor. While Google has entered a long-term agreement with Broadcom to co-develop future TPU generations and next-generation AI rack components—extending through 2031—it has also recently deepened its custom silicon partnership with Marvell. This diversification underscores a clear industry reality: major customers will not cede their architectural roadmaps to a single vendor.
3. After $100 Billion, What Comes Next?
Market expectations for Broadcom’s AI trajectory were already priced in. The central question ahead of this earnings release was never whether the $16 billion target would be met—the results confirm it has been. For now, Broadcom’s narrative has shifted from proving the existence of a second growth curve to demonstrating its longevity. This transition inevitably raises the bar for investor scrutiny. Where $100 billion initially served as a benchmark, the newly disclosed $115 billion (FY2027) and $230 billion (FY2028) projections reframe the challenge around execution.
The critical follow-ups are operational: Can initiatives from Google, Meta, OpenAI, and Anthropic scale to gigawatt-level deployments as scheduled? Will second- and third-generation custom chips secure continued orders following first-generation ramp? Can the networking division maintain growth velocity aligned with cluster expansion? These execution metrics carry far greater weight than marginal quarterly beat-and-raise scenarios.
Equally important is the structural reality of vendor diversification. Hyperscalers are actively mitigating single-source dependency. Google has already broadened its supplier base, and both Meta and OpenAI are structured to maintain multi-vendor ecosystems. Broadcom’s current market position is formidable, but it has not yet achieved customer lock-in. From this vantage point, whether Broadcom can prove it deserves to remain at t