Broadcom's AI semiconductor revenue exceeds $16 billion in Q3

ASIC is making significant progress in the reports; next, we’ll see how dozens of gigawatts of demand turn into actual revenue.

Written by Jim, MSX MaiTong

Edited by: Frank, MSX Maotong

Last week's NVIDIA earnings report demonstrated that demand for AI computing power is far from peaking.

Last night, after Broadcom's earnings report was released, another trend became clearer.

Q3 AI semiconductor revenue reached $16.7 billion, exceeding the previous guidance of $16 billion; the next quarter is directly forecasted at $21.7 billion. More significantly, AI chip revenue is expected to reach approximately $115 billion in FY2027 and as much as $230 billion in FY2028.

So now it’s a bit late to debate whether ASIC qualifies as AI’s second growth curve—the money in AI has long since been spread beyond just GPUs.

At least at Broadcom, this line has already started appearing in the financial reports, and the subsequent volume is even larger than previously anticipated. What we truly need to monitor going forward is whether major projects from Google, OpenAI, Anthropic, and others can be delivered on schedule, and how much of that Broadcom will ultimately capture.

If everything goes smoothly, let’s keep the music playing and the dance going.

The most informative part of NVIDIA's earnings report is that, despite such a high baseline, cloud providers, AI companies, and model labs continue to increase their computing power.

However, when AI capital expenditures grow from tens of billions of dollars to hundreds of billions, and eventually reach annual levels of hundreds of billions or more, the procurement logic will inevitably change.

For example, large companies are beginning to ask more frequently: How much does it cost to complete the same AI workload? With the same 1 GW of power, how much effective computing power can actually be produced? If certain workloads are already highly stable, is it still necessary to use only the most expensive products?

This is the context in which ASICs have begun to grow increasingly important.

Of course, it's inaccurate to simply think of ASICs as "cheaper GPUs"; their true advantage comes from specialization.

Customizing a chip itself is not inexpensive, as it requires substantial upfront R&D investment, a long development cycle, and addressing a full suite of compatibility issues involving software, advanced packaging, networking, systems, and the supply chain.

Conversely, once a workload becomes sufficiently stable and its deployment scale grows from tens of MW to hundreds of MW, and ultimately to GW levels, improvements in unit cost, performance-per-power, and overall system TCO can be rapidly amplified.

So Google has consistently stuck with TPUs, Meta has continuously expanded its MTIA, and OpenAI has begun collaborating with Broadcom to develop its own processors—all recognizing that even migrating a portion of the most stable and largest workloads from general-purpose chips to custom chips could generate significant economic value.

A simple review reveals that this has also been enabled by several conditions maturing simultaneously, with the most critical factor being that inference is becoming an increasingly important incremental workload.

It is well known that model training is often conducted in stages, but with ChatGPT, Gemini, Claude, and an increasing number of agents now entering production, inference has become a continuous daily process.

Especially as query volumes continue to grow and model architectures and service models gradually stabilize, it is naturally better suited for hardware optimization targeting specific workloads. In June of this year, OpenAI officially launched Jalapeño, its first Intelligence Processor developed in collaboration with Broadcom, specifically designed for LLM inference.

More notably, this is not an isolated chip; OpenAI and Broadcom have explicitly defined it as the first generation of a multi-generational computing platform, with deployment planned to begin by the end of 2026 and expansion to gigawatt-scale in the future.

Meta’s roadmap is very similar, with four generations of MTIA being developed over two years, focusing on recommendation, ranking, and generative AI, with several products explicitly adopting an inference-first design approach. In April this year, Meta further expanded its collaboration with Broadcom, with the first phase of deployment exceeding 1 GW, and future plans to scale to multiple GWs; the multi-generation product collaboration between both parties will continue through 2029.

This indicates that today's large AI customers are moving beyond merely pursuing peak performance and entering a new set of rules: whether tokens can be produced more cheaply.

When AI truly begins to be commercialized, the cost of one million tokens, the amount of effective computing power generated per watt of electricity, and the total cost of ownership for an entire data center will become increasingly important.

The essence of ASIC is that large AI companies are beginning to try to reclaim control over these costs.

Once you understand this point, reviewing Broadcom's financial report will become much easier.

Last quarter, Broadcom's total revenue reached $22.187 billion, a 48% year-over-year increase; AI semiconductor revenue reached $10.8 billion, up 143% year-over-year.

The company's guidance for the next quarter is more aggressive, with total revenue of approximately $29.4 billion, and AI semiconductor revenue expected to reach $16 billion, representing over 200% year-over-year growth.

What does $16 billion mean? It’s roughly equivalent to 54% of Broadcom’s projected total revenue for the quarter.

In other words, if the guidance is fulfilled, even including infrastructure software businesses like VMware, AI semiconductors alone could account for more than half of Broadcom's revenue.

In other words, AI is directly transforming the company’s revenue structure.

However, there is a common misconception here: $16 billion cannot be directly equated to "ASIC revenue," as it also includes AI network products such as Ethernet switching chips, SerDes, PCIe, and optical interconnects.

Last quarter, network business accounted for nearly 40% of AI semiconductor revenue. Hock Tan also noted that this ratio may have approached a temporary peak and is more likely to return to around 30% in the long term.

So Broadcom is actually expanding two business lines simultaneously: Custom XPU and AI Networking, which is the biggest difference from many pure-play chip companies.

As AI clusters scale from thousands to tens of thousands, hundreds of thousands, or even larger numbers of chips, how the computing power is interconnected becomes increasingly important.

Broadcom is further breaking down AI networking into scale-up, scale-out, and scale-across: from high-speed interconnects within a rack, to large-scale networks within a data center, and finally to connections between multiple data centers.

As long as AI clusters continue to grow, if major cloud providers increase their development of proprietary ASICs, Broadcom can participate in the Custom XPU market; if GPU clusters continue to expand, the open Ethernet network market may also continue to grow.

This places Broadcom in an interesting position, as it is effectively betting that the complexity of AI infrastructure will continue to increase.

Of course, this doesn't mean the business has no competition.

Although Google has already signed a long-term agreement with Broadcom to co-develop future generations of TPUs and related components for next-generation AI racks, with the agreement extending until 2031, Google has recently expanded its custom chip collaboration with Marvell.

This at least indicates that large clients won't easily entrust their entire future architecture to a single vendor.

The market has long known that Broadcom's AI capabilities would be strong.

So what this earnings report was originally meant to determine—whether the $16 billion target could be achieved—has now been answered.

At least for now, the question regarding Broadcom is no longer about whether it has a second growth curve, but rather how long that curve can ultimately be—though this also means the market will become increasingly discerning going forward.

Previously, 100 billion itself was a surprise, but now that 115 billion and 230 billion are on the table, what matters next is how these figures will be delivered.

Can Google, Meta, OpenAI, and Anthropic’s projects scale to GW-level capacity as planned? After the first-generation chips enter mass production, can they continue securing second- and third-generation chips? Will their network infrastructure grow in tandem with cluster expansion?

These are more important than selling tens of billions of dollars more in a single quarter.

Another issue is becoming increasingly unavoidable: major clients won’t rely on just one supplier.

Google has already begun expanding its partnerships with other vendors, and Meta and OpenAI are also likely to maintain a multi-vendor approach. Therefore, Broadcom’s current advantage is real, but it has not yet reached the level of "customer lock-in."

From this perspective, proving its ability to remain a core part of these customers’ supply chains over the long term will be a decisive factor.

Regardless, Broadcom’s earnings report serves as a reminder that custom chips, networking, and connectivity—once seen as mere supporting elements—are gradually becoming central.

It also reminds the market that this round of AI infrastructure investment is likely far from reaching a stage dominated solely by存量 competition.

As large companies continue to build larger clusters, purchase more electricity, and compute more tokens, new bottlenecks will constantly emerge, and new profit pools will grow alongside them.

The AI business is gradually evolving from a single chip into an entire infrastructure—and that’s where the greater potential lies.