Broadcom's AI Revenue Soars 70% But Shares Fall on Margin Concerns and Client In-House Moves

TradingKey - Semiconductor giant Broadcom (AVGO) reported robust fourth-quarter fiscal 2025 earnings after market close on Thursday, December 11, U.S. Eastern Time, with revenue soaring 28.2% year-over-year to $18.02 billion. The company achieved new quarterly records for both revenue and EBITDA profit. Furthermore, AI chip revenue jumped 74% to $8.2 billion.

Following the earnings release, Broadcom's stock initially climbed 4% in after-hours trading on Thursday, but subsequently reversed course, dropping over 5% at one point. Analysts suggested this decline might stem from CEO Hock Tan's announcement of a $73 billion AI product backlog, a figure that reportedly fell short of investor expectations.

AI Revenue Boom at the Expense of Margins?

Based on its earnings report and conference call, Broadcom's AI semiconductor revenue surged 74% year-over-year in the fourth quarter,primarily driven by XPU (custom AI chips, or ASICs) and data center chips.Notably, the XPU business itself grew by over 100%. Google's TPUs, co-designed by Google and Broadcom, are not only utilized internally by Google but have also secured other clients, including Apple, which employs them for model training.

Future revenue from the AI semiconductor business also holds significant promise. Broadcom secured orders worth $10 billion and $11 billion from Anthropic in Q3 and Q4, respectively, for its latest-generation Ironwood TPU rack systems. Furthermore, Broadcom signed another order valued at $1 billion. CEO Hock Tan stated on the conference call that the company currently boasts an AI product order backlog totaling $73 billion.

However, despite this robust revenue growth, Broadcom issued a warning: the company's overall profit margins are declining due to AI product sales, with an expected 1% decrease in Q1 2026.This shift is primarily attributed to Broadcom's strategic repositioning from being a mere "chip supplier" to a provider of "complete rack systems."In essence, to attract and retain clients by offering more convenient, "one-stop" solutions, Broadcom has started selling full racks. These comprehensive racks include not only chips but also lower-margin components such as servers, power supplies, and cooling systems not manufactured by Broadcom itself, consequently dragging down overall profit margins.

It is noteworthy, however, that this "overall sales" model directly boosts revenue, as the selling price of a complete rack system is higher than that of individual chips. Moreover, this "one-stop service" significantly enhances customer convenience, fostering greater customer stickiness and facilitating long-term partnerships, thereby stabilizing revenue streams.

Broadcom is not the only supplier choosing to offer complete server solutions.Nvidiaalso provides comprehensive rack solutions, such as its DGX and HGX systems.Companies like AMD and HPEare also gradually moving towards offering integrated solutions, reflecting a broader trend within the semiconductor industry. Therefore, Broadcom's decline in gross margin does not necessarily indicate a reduction in overall competitiveness.

Tech Giants Turn to Self-Developed Chips: How Far Can Broadcom-Google Partnership Go?

A growing number of technology companies are opting for in-house chip development to reduce reliance on external suppliers, posing a key external risk for Broadcom.Analysts suggest Broadcom's after-hours decline on Thursday likely stemmed from market concerns over its long-term partnership with Google.

Broadcom has maintained a deep collaboration with Google since the latter began designing its first-generation TPU (TPU v1) in 2015. Google handled the high-level architectural design for TPUs, while Broadcom was responsible for the engineering implementation and physical design of Google's solutions. Although the two companies work closely, Broadcom investors are now concerned that Google might fully internalize its chip design efforts.

Broadcom CEO Hock Tan dismisses this possibility as unlikely. He implied that Google, as a leading developer of large language models,has a core mission to maintain its leadership in models and algorithms,and to compete with general-purpose chip giant Nvidia. Diverting efforts to in-house chip development to enhance engineering and physical design capabilities would be counterproductive, Tan suggested. Broadcom's value lies in executing aspects where Google lacks expertise, allowing Google to focus on top-level architectural innovation without grappling with large-scale, high-yield engineering challenges.

Of course, beyond fully in-house development, Google could also choose to bypass Broadcom and engage other chip companies for design. For instance, Google's TPU v7 adopted a hybrid development model,where Broadcom was only responsible for the core compute die and high-speed interconnect (SerDes) technology,while MediaTek handled parts such as the I/O modules and back-end physical design.

However, this move precisely underscores Broadcom's indispensability. Broadcom remains a leader in the semiconductor industry, possessing the yield optimization and engineering experience necessary to transform the most advanced chip designs (such as 3nm or 5nm process technology) into products.Broadcom also holds some of the most advanced technical IP in high-speed interconnect technology,a critical capability that Google's TPU chips heavily rely on and which other chip companies cannot replicate.

Despite Broadcom's "firmly established position," it must still account for the risks of declining bargaining power and eroding technological barriers.Google is already collaborating with another chip vendor, MediaTek, particularly for lower-performance chip versions like the TPU v7e. Broadcom's most significant technological barrier currently lies in its high-speed interconnect technology, yet competitors are rapidly catching up. Marvell has demonstrated its 3nm 224G SerDes IP, while Cadence's 224G SerDes IP has also shown robust performance and low bit error rates.