Over the past two years, the investment narrative around AI infrastructure has been almost synonymous with "buying GPUs." NVIDIA’s data center business is projected to reach $193.7 billion in fiscal year 2026, up 68% year-over-year and accounting for about 90% of the company’s total revenue. Hyperscale providers continue to set new records for capital expenditures—Morgan Stanley forecasts that the combined capex of the five largest hyperscalers will hit roughly $800 billion in 2026, climbing to $1.2 trillion in 2027.
However, as these hundred-billion-dollar capex plans move from blueprint to reality, market attention is shifting from "point compute power" to "system-level infrastructure." Training a trillion-parameter large language model requires not only the parallel computing power of tens of thousands of GPUs, but also ultra-fast, low-latency data transfer between those GPUs. The network layer—long regarded as just a "pipeline"—is now emerging as the key bottleneck that determines the actual compute utilization of AI clusters.
This structural shift is exactly where Broadcom (AVGO) finds its unique opportunity.
From "Buying GPUs" to "Building Data Centers": The Shifting Focus of AI Infrastructure Investment
To understand Broadcom’s AI narrative, you first need to grasp a fundamental transformation underway in AI data centers: the investment focus is expanding from individual compute chips to full data center architectures.
In the first half of 2026, Microsoft, Amazon, Google, Meta, and Oracle—all five major hyperscale cloud providers—collectively raised their capital expenditure guidance. Bank of America Securities analyst Vivek Arya’s team predicts that global hyperscaler AI capex will exceed $800 billion in 2026, up 67% year-over-year, and will break the $1 trillion mark in 2027. Goldman Sachs is even more optimistic, projecting that capex could reach $1.4 trillion under favorable conditions in 2027.
But not all of this massive spending is flowing into GPUs. As AI clusters scale from thousands to tens of thousands or even hundreds of thousands of cards, the share of investment going into network infrastructure is rising rapidly. A JPMorgan report notes that the AI ASIC market will reach about $60–70 billion by 2026, with a compound annual growth rate exceeding 40–50% in the coming years. At Fiber Connect 2026, Cisco stated that AI is pushing network architectures from the core to the edge, with bandwidth demand growth outpacing many vendors’ expectations—AI traffic now accounts for 5% of backbone network utilization, up from less than 1% just two years ago.
This structural change means the logic of AI infrastructure investment is shifting from "who has the most powerful GPU" to "who has the most complete and efficient data center architecture." In this system-level competition, Broadcom holds two irreplaceable positions.
Custom ASICs: Broadcom’s "Second Trump Card"
Broadcom is often perceived as a "network chip company," but its Q2 FY2026 financial report clearly reveals another growth trajectory: custom AI accelerator chips (ASICs).
On June 3, 2026, Broadcom announced its second quarter FY2026 results: total revenue hit $22.19 billion, up 48% year-over-year and a new record. AI semiconductor revenue soared to $10.8 billion, a 143% increase year-over-year—beating both company guidance and Wall Street estimates. Non-GAAP EPS came in at $2.44, topping analyst expectations of $2.40.
Even more noteworthy is the order backlog. Broadcom CEO Hock Tan revealed on the earnings call that Q2 AI semiconductor orders exceeded $30 billion, while actual shipments were just $10.8 billion. Additional data shows that AI chip contract backlog has reached $73 billion, with $53 billion from custom accelerators. This means customer purchase commitments far outstrip current delivery capacity, with order visibility now stretching into fiscal year 2028.
Broadcom’s ASIC model differentiates it from NVIDIA’s general-purpose GPU approach. While NVIDIA offers standardized compute products, Broadcom designs custom AI accelerator chips for six core clients, including Google, Meta, Anthropic, and OpenAI. The moat here is time-to-market: designing, validating, and deploying custom chips with Broadcom typically takes over two years, making it extremely costly for customers to switch suppliers.
JPMorgan projects that by 2027, Broadcom could capture about 60% of the AI server compute ASIC market. FY2026 AI semiconductor revenue is expected to reach $56 billion, up roughly 180% from FY2025; by FY2027, it could surpass $100 billion.
Networking Chips: The "Nervous System" of AI Clusters
If ASICs are Broadcom’s engine for growth, then networking chips are its defensive moat.
The expanding scale of AI training and inference is driving exponential demands on intra-data center data transfer efficiency. Over the past four years, cluster interconnect bandwidth has surged from 400 Gbit/s to 12.8 Tbit/s—a 32-fold increase. A single round of large model training now requires data interconnects at the terabyte or even petabyte scale. In this context, networking chips are no longer mere "pipes"—they are the critical link that determines whether compute resources can be fully utilized.
Broadcom’s AI networking portfolio spans the entire product matrix, from switch chips to optical interconnects. In Q2 2026, networking chips accounted for nearly 40% of Broadcom’s AI revenue. The company expects this proportion to stabilize at around 30% over the long term.
On the product front, Broadcom’s Tomahawk 6—the world’s first 102.4 Tbps Ethernet switch chip—has entered mass production. It supports 128 800G ports or 1.6T Ethernet capability. The company is also advancing development of 200-terabit switching technology. Additionally, Jericho3-AI, as an 800G switching silicon, enables the construction of AI fabrics interconnecting up to 32,000 GPUs.
At the OFC show in March 2026, Broadcom showcased its end-to-end AI infrastructure portfolio for gigawatt-scale AI clusters, emphasizing scalable, energy-efficient solutions. The company also announced a strategic partnership with OpenAI to jointly deploy OpenAI-designed AI accelerators, aiming to begin deployment in the second half of 2026 and complete it by the end of 2029.
Capex Shifts from GPUs to System Architecture: Broadcom’s Value Proposition
The capital expenditure structure of hyperscalers is changing. In 2026, global data center capex is expected to exceed $800 billion. Moody’s reports that hyperscalers plan to spend about $700 billion on AI data centers in 2026—nearly six times the 2022 figure.
This round of capex is driven by more than just training compute. IDC data shows that 91% of enterprises plan to increase data center interconnect bandwidth by more than 11% over the next 12 months to support AI, with 36% planning increases of over 51%, and 70% intending to double their GPU and switch environments. For the first time in 2026, AI inference traffic will account for more than two-thirds of total AI data center traffic. The network demands of inference are more distributed and sustained than those of training, posing new architectural requirements for data center networks.
This means that as hyperscalers deploy next-generation AI clusters, the share of spending on network infrastructure is rising systemically. For every GPU deployed, corresponding networking chips, switches, and optical interconnect components are required. Broadcom’s Tomahawk and Jericho series are core elements of this supporting ecosystem.
From a financial perspective, Broadcom’s AI revenue growth is accelerating rather than slowing: from 106% year-over-year in Q1 FY2026, to 143% in Q2, and guidance of over 200% for Q3. The company expects Q3 FY2026 revenue of about $29.4 billion, up 84% year-over-year. Adjusted EBITDA margin is an impressive 69%, with free cash flow representing 46% of revenue.
Market Reaction and Valuation Logic
Despite strong fundamentals, Broadcom’s share price fluctuated following its June 2026 earnings release. On the day of the report, the stock closed at $479.23, then dropped more than 13% after hours. The main reason: total revenue of $22.19 billion was slightly below Wall Street’s $22.27 billion estimate, and the company did not raise its full-year AI semiconductor revenue guidance.
This market reaction reflects investors’ lofty expectations for AI semiconductor companies—anything short of "perfect" can trigger short-term sell-offs. But over a longer horizon, Broadcom shares are still up nearly 38% year-to-date. Firms like Jefferies see the recent pullback as an attractive entry point.
JPMorgan analyst Harlan Sur set a price target of $580, among the highest on Wall Street. The key rationale: Broadcom’s AI business growth is highly visible and sticky, with long-term agreements from six core customers covering capacity plans through FY2027 and even FY2028.
Challenges and Risks
Broadcom’s growth prospects are not without challenges.
First, AI capex is highly dependent on the investment cycles of hyperscalers. If major customers slow their purchases, Broadcom’s growth could see a marked decline. Whether the current $700–800 billion annual capex level is sustainable depends on whether monetization at the AI application layer can keep pace with infrastructure investment.
Second, gross margins face structural pressure. Q2 FY2026 gross margin was 77.1%, down 230 basis points year-over-year, mainly because the lower-margin semiconductor business now makes up a larger share of total revenue. The company expects Q3 gross margin to fall further to about 74%. While this trend is a byproduct of rapid AI semiconductor growth rather than declining business competitiveness, it does impact the income statement.
Third, the competitive landscape is evolving. Marvell holds a 10–12% share of the AI ASIC market and is actively expanding its customer base. NVIDIA is also strengthening its networking portfolio (such as the Spectrum-X platform), aiming to compete in both InfiniBand and Ethernet. While Broadcom’s technological lead in Ethernet switch chips is unlikely to be challenged in the near term, competitive intensity is rising.
Conclusion
Competition in AI infrastructure is shifting from a "GPU arms race" to a new phase of "system-level architecture competition." As hyperscalers’ capex moves from hundreds of billions to trillions of dollars, point compute advantages are giving way to the comprehensive efficiency of entire data centers.
Broadcom holds two irreplaceable positions in this structural transformation: first, providing custom ASIC chips for the world’s largest AI model developers; second, supplying the critical networking infrastructure that determines actual compute utilization in AI clusters. The Q2 FY2026 AI semiconductor revenue of $10.8 billion—a 143% year-over-year increase—and more than $30 billion in quarterly order backlog are just snapshots of this long-term trend.
The "second layer" of AI infrastructure is becoming the new main battleground—and Broadcom has already secured the high ground.
FAQ
Q: How does Broadcom’s AI business differ from NVIDIA’s?
NVIDIA offers standardized, general-purpose GPU compute products, while Broadcom mainly designs custom AI accelerator chips (ASICs) for clients like Google, Meta, and OpenAI. In addition, Broadcom dominates the AI data center networking chip sector—a critical component for AI clusters that NVIDIA covers far less.
Q: What is Broadcom’s AI revenue guidance for fiscal year 2026?
Broadcom expects AI semiconductor revenue to reach $56 billion in FY2026, an increase of about 180% over FY2025. The company has also reaffirmed that AI semiconductor revenue will exceed $100 billion in FY2027, with growth momentum continuing into FY2028.
Q: What are AI networking chips, and why are they important for AI data centers?
AI networking chips serve as the "nervous system" connecting thousands of GPUs and accelerators in an AI cluster. Large model training requires frequent, massive data exchanges between GPUs, and the performance of networking chips directly determines whether compute resources can be fully utilized. As AI clusters scale up, investment in network infrastructure is rising rapidly.
Q: Who are Broadcom’s main AI customers?
Broadcom currently has six core custom chip clients, including Google, Meta, Anthropic, and OpenAI. These customers represent some of the world’s most aggressive investors in AI infrastructure and have signed long-term agreements with Broadcom, providing order visibility through 2028.
Q: How has Broadcom stock (AVGO) performed recently?
As of late June 2026, Broadcom shares are trading in the $370–$380 range. Despite a short-term pullback after earnings, the stock is still up nearly 38% year-to-date. Institutions like JPMorgan have set a target price of $580, viewing the current valuation as attractive.
