AI computing power expands to the "network layer": How Broadcom becomes the next stage winner?

Over the past two years, the investment narrative for AI infrastructure has been almost synonymous with "buying GPUs." NVIDIA's data center business reached $193.7 billion in fiscal 2026, up 68% year-over-year, accounting for about 90% of the company's total revenue. Capital expenditures by hyperscalers continue to break records—Morgan Stanley predicts that the combined capex of the top five hyperscalers will reach about $800 billion in 2026, climbing further to $1.2 trillion in 2027.

However, as hundreds of billions in capital expenditure shift from planning to reality, the market's focus is moving from "single-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 high-speed, low-latency data transfer between these GPUs. The network layer—once considered a mere "pipeline"—is becoming the key bottleneck determining the actual compute utilization of AI clusters.

This is where Broadcom's (AVGO) structural opportunity lies.

From "Buying GPUs" to "Building Data Centers": The Investment Focus of AI Infrastructure Is Shifting

To understand Broadcom's AI narrative, one must first understand a fundamental shift happening in AI data centers: the focus of investment is expanding from individual compute chips to the complete data center architecture.

In the first half of 2026, the five major hyperscalers—Microsoft, Amazon, Google, Meta, and Oracle—collectively raised their capital expenditure guidance. BofA Securities analyst Vivek Arya's team predicts that global AI capex by hyperscalers will exceed $800 billion in 2026, up 67% year-over-year, and surpass $1 trillion in 2027. Goldman Sachs is even more optimistic, forecasting that capex could reach $1.4 trillion in a bull case scenario by 2027.

But this massive spending is not all going to GPUs. As AI clusters scale from thousands of cards to tens of thousands or even hundreds of thousands, the share of spending on network infrastructure is rising rapidly. A JPMorgan report notes that the AI ASIC market will reach approximately $60 to $70 billion by 2026, with a compound annual growth rate staying above 40% to 50% over the next few years. At Fiber Connect 2026, Cisco stated that AI is pushing network architecture from the core to the edge, with bandwidth demand growth exceeding many suppliers' expectations—AI traffic now accounts for 5% of backbone network utilization, compared to less than 1% two years ago.

This structural shift means the investment logic for AI infrastructure is moving from "whose GPU is strongest" to "whose data center architecture is most complete and efficient." And in this system-level competition, Broadcom holds two irreplaceable positions.

ASIC Custom Chips: Broadcom's "Second Trump Card"

External perceptions often pigeonhole Broadcom as a "networking chip company," but its fiscal Q2 2026 earnings clearly reveal another growth curve: custom AI accelerator chips (ASICs).

On June 3, 2026, Broadcom reported fiscal Q2 2026 results: total revenue of $22.19 billion, up 48% year-over-year, a record high. Among this, AI semiconductor revenue reached $10.8 billion, up 143% year-over-year, exceeding both the company's own expectations and Wall Street analysts' forecasts. Non-GAAP EPS was $2.44, beating the analyst consensus of $2.40.

Even more noteworthy is the order backlog. Broadcom CEO Hock Tan revealed on the earnings call that AI semiconductor orders in Q2 exceeded $30 billion, while actual shipments were only $10.8 billion. Other data shows that the contracted order backlog for AI chips is as high as $73 billion, with $53 billion coming from custom accelerators. This means that customer commitments far exceed current delivery capacity, with order visibility extending into fiscal 2028.

Broadcom's ASIC model differentiates from NVIDIA's general-purpose GPU model. NVIDIA offers standardized compute products, while Broadcom customizes AI accelerator chips for six core customers, including Google, Meta, Anthropic, and OpenAI. The moat of this model lies in time cost—the entire process of designing, validating, and deploying a custom chip with Broadcom typically takes over two years, making customer switching costs extremely high.

JPMorgan estimates that Broadcom will capture about 60% of the AI server compute ASIC market by 2027. AI semiconductor revenue in fiscal 2026 is expected to reach $56 billion, up about 180% from fiscal 2025; in fiscal 2027, it is on track to exceed $100 billion.

Networking Chips: The "Nervous System" of AI Clusters

If ASICs are Broadcom's offensive engine, networking chips are its moat.

The scaling of AI training and inference imposes exponential demands on data transfer efficiency within data centers. Over the past four years, cluster interconnect bandwidth has surged from 400 Gbit/s to 12.8 Tbit/s—a 32x increase. The data interconnect demand for a single round of large model training reaches terabyte or even petabyte levels. In this context, networking chips are no longer mere "pipes"; they are the critical factor determining whether compute power can be effectively utilized.

Broadcom's AI networking portfolio covers a complete 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 around 30% in the long term.

On the product level, Broadcom's Tomahawk 6—the world's first 102.4 Tbps Ethernet switch chip—has entered volume production and shipment. The chip supports 128 800G ports or 1.6T Ethernet capabilities. The company is also advancing the development of 200-terabit switching technology. Additionally, Jericho3-AI, as an 800G switching silicon, can build a large AI fabric connecting up to 32k GPUs.

At the OFC 2026 trade show in March, Broadcom showcased an end-to-end AI infrastructure portfolio for gigawatt-scale AI clusters, emphasizing scalable and energy-efficient solutions. The company also announced a strategic partnership with OpenAI to jointly deploy OpenAI-designed AI accelerators, targeting deployment to start in the second half of 2026 and complete by the end of 2029.

Capital Expenditure Shifting from GPUs to System Architecture: Broadcom's Beneficiary Logic

The capital expenditure structure of hyperscalers is changing. In 2026, global data center capex is expected to exceed $800 billion. A Moody's rating report shows that hyperscalers' AI data center spending plans for 2026 are around $700 billion, nearly six times the spending in 2022.

The driver of this capex cycle is not just training compute power. IDC data shows that 91% of enterprises will increase data center interconnect bandwidth by more than 11% to support AI over the next 12 months, with 36% increasing by over 51%, and 70% planning to double their GPU and switch environments. AI inference traffic exceeded two-thirds of total traffic for the first time in 2026. The network demands of inference are more distributed and persistent than training, placing new requirements on data center network architecture.

This means that when hyperscalers deploy new-generation AI clusters, the share of spending on network infrastructure is systematically rising. For every GPU deployed, corresponding networking chips, switches, and optical interconnect components are needed. Broadcom's Tomahawk and Jericho series are core components in this supporting system.

Financially, Broadcom's AI revenue growth is accelerating rather than decelerating: from 106% year-over-year growth in fiscal Q1 2026, to 143% in Q2, to guidance of over 200% in Q3. The company expects fiscal Q3 2026 revenue of approximately $29.4 billion, up 84% year-over-year. Adjusted EBITDA margin is as high as 69%, and free cash flow accounts for 46% of revenue.

Market Reaction and Valuation Logic

Despite strong fundamentals, Broadcom's stock price fluctuated after its June 2026 earnings release. On the earnings day, the stock closed at $479.23 but fell over 13% in after-hours trading. The main reasons were total revenue of $22.19 billion slightly below Wall Street's expectation of $22.27 billion, and the company's decision not to raise its full-year AI semiconductor revenue guidance.

This market reaction reflects investors' high expectations for AI semiconductor companies—any metric falling short of "perfect" can trigger short-term selling. However, over a longer timeframe, Broadcom's stock is still up nearly 38% year-to-date. Institutions like Jefferies view the recent pullback as an attractive entry opportunity.

JPMorgan analyst Harlan Sur issued a target price of $580, one of the highest on Wall Street. The core logic supporting this valuation is that Broadcom's AI business growth enjoys high contract visibility and customer stickiness, with long-term agreements from its six core customers covering capacity planning through fiscal 2027 or even fiscal 2028.

Challenges and Risks

Broadcom's growth outlook is not without challenges.

First, AI capex is highly dependent on the investment cycles of hyperscalers. If major customers slow down procurement, Broadcom's growth could see a significant downturn. Whether the current annual capex scale of $700 to $800 billion is sustainable depends on whether the monetization ability of the AI application layer can keep pace with infrastructure investment.

Second, gross margin faces structural pressure. In Q2 2026, gross margin was 77.1%, down 230 basis points year-over-year, mainly due to the rising share of lower-margin semiconductor business in total revenue. The company expects Q3 gross margin to decline further to about 74%. This trend is a byproduct of rapid AI semiconductor revenue growth rather than a decline in business competitiveness, but it will indeed impact the income statement.

Third, the competitive landscape is evolving. Marvell holds a 10% to 12% share in the AI ASIC market and is actively expanding its customer base. NVIDIA is also strengthening its networking product line (e.g., the Spectrum-X platform), trying to push forward on both InfiniBand and Ethernet fronts. While Broadcom's technological leadership in Ethernet switch chips is difficult to shake in the short term, competitive intensity is rising.

Conclusion

Competition in AI infrastructure is entering a new phase—from "GPU arms race" to "system-level architecture competition." As hyperscalers' capital expenditure moves from hundreds of billions to trillions of dollars, single-point compute advantages are giving way to the overall efficiency of the complete data center.

In this structural shift, Broadcom holds two irreplaceable positions: first, providing custom ASIC chips to the world's largest AI model developers; second, supplying the critical networking infrastructure that determines the actual compute utilization of AI clusters. Its Q2 fiscal 2026 AI semiconductor revenue of $10.8 billion, up 143% year-over-year, and a quarterly order backlog exceeding $30 billion are just interim milestones in this long-term trend.

The "second layer" of AI infrastructure is becoming the new main battlefield, and Broadcom has already secured the high ground.

FAQ

Q: How does Broadcom's AI business differ from NVIDIA's?

NVIDIA provides standardized general-purpose GPU compute products, while Broadcom customizes dedicated AI accelerator chips (ASICs) for clients like Google, Meta, and OpenAI. Additionally, Broadcom dominates the AI data center networking chip space, an area NVIDIA covers less but is indispensable for AI clusters.

Q: What is Broadcom's AI revenue guidance for fiscal 2026?

Broadcom expects AI semiconductor revenue of $56 billion in fiscal 2026, up about 180% from fiscal 2025. The company also reiterated that AI semiconductor revenue will exceed $100 billion in fiscal 2027, with growth momentum extending into fiscal 2028.

Q: What are AI networking chips, and why are they important for AI data centers?

AI networking chips are the "nervous system" connecting tens of thousands of GPUs and accelerators in an AI cluster. Large model training requires frequent massive data exchanges between GPUs; the performance of networking chips directly determines whether compute power can be fully utilized. As AI cluster scale expands, the investment share of networking infrastructure is rising rapidly.

Q: Who are Broadcom's major AI customers?

Broadcom currently has six core custom chip customers, including Google, Meta, Anthropic, and OpenAI. These customers represent the most aggressive AI infrastructure spenders among large tech companies and have signed long-term agreements with Broadcom, with order visibility extending to 2028.

Q: How has Broadcom's stock (AVGO) performed recently?

As of the end of June 2026, Broadcom's stock traded in the $370–$380 range. Despite short-term pullbacks after the earnings release, it is still up nearly 38% year-to-date. JPMorgan and other institutions have issued target prices of $580, viewing the current valuation as attractive.