Event: Broadcom released its FY25Q3 performance report.
– The company achieved a revenue of $16 billion for the quarter, a 22% year-on-year increase, exceeding analysts’ expectations of $15.84 billion. Among them, semiconductor solutions revenue was $9.2 billion, a 26% increase, surpassing the analyst expectation of $9.1 billion. AI revenue reached $5.2 billion, a 63% increase, while non-AI revenue was $4 billion, remaining flat quarter-on-quarter; infrastructure software revenue was $6.8 billion, a 17% increase. The company’s gross margin was 78.4%. Demand for AI business is strong, with current contract orders reaching $110 billion.
– The company guided that revenue for the fourth quarter is expected to be approximately $17.4 billion, a 24% year-on-year increase, exceeding analysts’ expectations of $17.05 billion; fourth quarter AI business revenue is expected to be $6.2 billion, a 66% increase, while non-AI revenue is expected to be $4.6 billion; infrastructure software revenue is expected to be around $6.7 billion, a 15% increase. The expected gross margin for the fourth quarter is 77.7%, down 0.7 percentage points, mainly due to the increased revenue share from XPU and wireless business.
Highlights from the earnings call:
1) Regarding ASIC
– Broadcom is developing a fourth major customer with considerable demand, with orders exceeding $10 billion, expected to be delivered in FY26Q3, thus driving FY26 (CY25M11-26M10) AI revenue beyond expectations; three other potential customers are also in development.
– With the expansion of computing clusters and differentiated competition, along with the maturity of AISC, the market share of ASIC is expected to gradually surpass that of GPU.
In this quarter, XPU business accounted for 65% of AI revenue, with continued growth in demand for ASIC from the three major customers. Additionally, Broadcom is collaborating with other potential customers to develop ASIC, and another potential customer placed a mass production order this quarter, bringing in $10 billion in value.
1) Regarding AI Networks
– Ethernet and GPU/ASIC are decoupled; similarly, Scale-across networks will not be bound to GPU/ASIC, so there is ample market space for Ethernet Scale-across; moreover, Ethernet is very mature, and the company believes there is no need to create new network protocols.
In terms of Scale up, the solution launched in collaboration with OpenAI based on open Ethernet technology can support customers using XPU to scale up to 512 computing nodes. For Scale out & Sale across, in June this year, the company launched an Ethernet-based 102Tb/s switch, simplifying the network architecture to two levels, thereby reducing latency and significantly lowering power consumption. Additionally, the next-generation Jericho 4 Ethernet structured router was launched, with a bandwidth of 51.2Tb/s, capable of supporting clusters across multiple data centers with over 200,000 computing nodes.
In terms of performance numbers: The revenue and profit figures for Q3 and Q4 are in line with expectations. AI revenue (custom chips ASIC & AI networks) for Q3 was $5.2 billion, in line with buyer expectations, while Q4 AI revenue guidance of $6.2 billion exceeds expectations, with buyer expectations at $5.9-6.2 billion. Including Q1/Q2 figures of $4.1/4.4 billion, the total AI revenue for this year is around $20 billion.
Continuity of AI ASIC: The earnings call mentioned that the 63% growth rate of AI revenue can be sustained for 10 quarters, implying that AI revenue for 2026/27 will start at least at $32.4/52.9 billion, with growth in 2026 expected to be higher than in 2025 (“fairly material improvement”). The company’s overall order backlog is $110 billion, which is a significant number and can confirm the continuity of orders for at least the next two years. In the last Q&A of the earnings call, it was implied that Broadcom’s 3-4 customers may exceed GPU in ASIC for many years to come. Observing the hardware, when the continuity remains strong and the slope does not decline, I believe it will exceed many people’s psychological expectations. There may be many potential players entering the market in the future, which we cannot predict, and this sector can only exceed expectations further.
Regarding customers: For the three major customers (Google, Meta, TT), Broadcom is gaining market share, which is also related to the recent adjustments in TPU, with META’s chain continuously revising upwards in Q2. The more important information from this earnings report is the fourth customer starting to ramp up, locking in a $10 billion order, which is also the reason for the upward revision of AI revenue for 2026, with this $10 billion order expected to be delivered in Q3 2026. The customer is speculated to be OpenAI, and it is believed that its supply chain will also bring many opportunities (OpenAI’s server outsourcing, PCB, optical modules, etc.). Among these four major customers, Broadcom’s share in their network segments is continuously increasing. The decrease in share outside these four major customers indicates that AWS/AMD/MSFT’s networking may be handed over to ALAB/MRVL and other companies. Regarding competition in the networking segment, Hock believes that Ethernet has better latency (less than 250 nanoseconds) compared to PCIe and Ualink, and even NVLink, and is open.
Hock also mentioned the scale across recently raised by NV, when a single data center site is limited by power, land, or other infrastructure constraints and cannot accommodate more than 100,000 GPUs or XPUs, customers typically deploy these devices across multiple data centers within a close distance (for example, within 100 kilometers). Using Jericho3 or the soon-to-be-released Jericho4 is sufficient, and there is no need to create new solutions. In the case of scale across, the content of Ethernet is higher.
Buyer surveys show that the market holds a certain optimistic view on the prospects of AI semiconductors and networking, but enthusiasm for software and non-AI semiconductors is relatively low; more importantly, investors need to be convinced that there is upward potential in the $60-90 billion AI serviceable addressable market (SAM) — especially at the current valuation levels: “We hope to receive clear comments and guidance to eliminate all doubts about the future growth strength of AI, networking, and software revenue.” At least in terms of AI, we have received strong order data, and more importantly, the company confirmed the fourth AI customer. Of course, the bears raised some doubts about the announcement of this customer, believing it may be premature or released in a way that allows for slight revenue adjustments, but this is still an important development; it implies (Hock also acknowledged) an upward revision of the growth rate for FY2026 to be higher than that of FY2025. You can scroll down to see the specific views of both bulls and bears, but overall, the bulls believe this is Hock’s consistent style — gradually releasing good news and reserving space for future quarters, while the bears believe that buyers have basically digested most of the good news regarding AI-specific integrated circuits. Even the bearish forecast of over $13 for FY2027 does not seem particularly pessimistic to me.
Harlan Sur’s insightful points about Broadcom:
– The $10 billion order from these four AI ASIC customers is more likely to be completed across two quarters in the second half of FY2026, and Hock expects to receive more follow-up orders thereafter.
– Harlan believes that SoftBank/Arm is a priority AI-specific integrated circuit project, but Hock stated in the call that the fourth AI-specific integrated circuit customer is more focused on the inference field; therefore, Harlan believes this customer is more likely to be OpenAI. (By the way, I think Hock’s reference to this fourth company as a “customer” may be somewhat clever, as there are still three quarters until its mass production ramp-up.)
– Hock confirmed that they will complete the first 2nm project tape-out using 3.5D packaging in the second half of 2025. This is a first in the industry — at the forefront of technology, far ahead of NVIDIA. The world’s most complex heterogeneous processor. Impressive.
– They responded that current M&A priorities are low (but no one questioned why they are not conducting stock buybacks — this is a question posed to me by several investors; unfortunately, this is a collective conference call, and each participant can only ask one question).
– Hock acknowledged that the demand for electro-absorption modulated lasers far exceeds their supply capacity. They plan to double EML supply within the next nine months.
Another noteworthy point is that OpenAI will begin mass production of self-developed AI chips in collaboration with Broadcom, with shipments expected next year.
OpenAI plans to produce its self-developed AI chips for the first time next year. The ChatGPT manufacturer is trying to meet the endless demand for computing power in the market and reduce reliance on chip giant NVIDIA.
According to several insiders, this chip, co-designed with American semiconductor giant Broadcom, will be delivered next year.
Broadcom CEO Chen Fuyang mentioned on Thursday that a mysterious new customer has committed to a $10 billion order.
OpenAI’s move follows the strategy of tech giants like Google, Amazon, and Meta, which have all designed self-developed chips specifically for running AI workloads. The industry is experiencing explosive growth in the demand for computing power required to train and run AI models.
According to people close to the project, OpenAI plans to use the chips for internal needs rather than providing them to external customers.
According to previous reports, OpenAI had preliminary cooperation with Broadcom last year, but the timeline for successful chip design and mass production had not been clarified.
During a conference call with analysts, Chen Fuyang announced that Broadcom’s custom AI chip business has won its fourth major customer, while the company’s earnings data exceeded Wall Street expectations.
Broadcom did not disclose the specific names of these customers, but insiders confirmed that OpenAI is indeed the new customer. Both Broadcom and OpenAI declined to comment.
Chen Fuyang stated that this deal has brought “immediate and considerable demand,” enhancing the company’s growth expectations, and starting next year, they will “strongly” ship chips for this customer.
Custom AI chips (referred to as “XPU” to distinguish them from ready-made chip suppliers like NVIDIA and AMD) are expected to gain an increasing share in the booming AI infrastructure market, a prospect that has driven Broadcom’s stock price up over 30% this year.
Following the positive earnings report, the company’s stock rose 4.5% in after-hours trading.
For example, Broadcom previously collaborated with Google to develop its customized “TPU” AI chip.
On the eve of Broadcom’s earnings report, HSBC analysts wrote that they expect Broadcom’s custom chip business growth to far exceed NVIDIA’s GPU business by 2026.
NVIDIA continues to dominate the AI hardware space, with large tech “hyperscale enterprises” still making up a significant portion of its customer base. However, compared to the astronomical growth figures seen in the early stages of the AI investment boom, its growth rate has slowed.
OpenAI CEO Sam Altman has repeatedly stated that stronger computing power is needed to meet the growing number of businesses and consumers using products like ChatGPT, as well as to train and run AI models.
The company was one of the early customers of NVIDIA’s AI chips and has since become a fervent consumer of its hardware.
Last month, Altman stated that given the demand growth brought by “[OpenAI’s latest model] GPT-5,” the company is prioritizing computing resources and plans to “double the scale of its computing cluster within the next five months.”
