The AI hardware sector is currently witnessing a “new king ascends” scenario—while the market is still debating NVIDIA’s growth rate, Broadcom (AVGO) has shattered expectations with a stunning earnings report: FY25Q3 AI semiconductor revenue surged by 63% year-on-year, with $110 billion in backlog orders weighing down like a giant stone,and they boldly stated that FY26 growth will far exceed this year’s 63% (JP Morgan even raised it to 125%!）. This move has transformed the “ASIC narrative” from a niche discussion in the tech circle to a hot topic in the capital market.
It is important to note that ASICs (Application-Specific Integrated Circuits), which were once considered a “backup” to GPUs a few years ago, have now become highly sought after by giants like Google and Meta. Broadcom has managed to carve out a space in NVIDIA’s territory, even being humorously dubbed the “invisible champion of the AI hardware industry.” What logic lies behind this? Where does the confidence of $110 billion in orders come from?
01
Broadcom’s “hard power”:
Not just a 63% growth rate, but the certainty of $110 billion in orders
When mentioning Broadcom, many still hold the impression of a “traditional chip giant,” but tracking its recent business trajectory reveals that this company has quietly transformed into a “key player” in the AI hardware race through ASIC technology. The Gelonghui Research Institute noted a detail in 2024: the share of Broadcom’s custom chips in Google’s data centers surged from 10% to 30%, leading to the judgment that “the trend of ASICs replacing GPUs is accelerating,” and Broadcom’s performance is now validating this judgment.
1. The “AI code” in the earnings report: growth exceeds expectations, orders secure the future
The figures from FY25Q3 earnings report all shout “AI is core”:
– Revenue exceeded expectations, with AI contributing significantly: total revenue of $15.95 billion (up 22% year-on-year, exceeding market expectations by $110 million); adjusted EPS of $1.69 (up 36%, exceeding expectations by $0.02). Among them, AI semiconductor revenue reached $5.2 billion, a staggering 63% year-on-year increase, accounting for nearly 50% of semiconductor business, and has been the undisputed growth engine for ten consecutive quarters.
– $110 billion in backlog orders, more reliable than mere talk: half of the backlog orders are from the semiconductor business, and the proportion of AI-related orders is significantly higher than non-AI. This means that even if market fluctuations occur, these orders can sustain growth for 1-2 years, which is much more substantial than those “pie-in-the-sky” forecasts.
– Q4 guidance is even more aggressive, maximizing profitability: total revenue is expected to reach $17.4 billion (up 24%, exceeding expectations by $350 million), with AI semiconductors projected to hit $6.2 billion (up 66%); adjusted EBITDA margin is 67%, exceeding expectations by 1 percentage point. This kind of “growth + profit” dual acceleration is rare among high-growth tech stocks.
The only minor hiccup is that the gross margin decreased by 0.7% quarter-on-quarter, but don’t panic—this is not due to a decline in profitability, but mainly because the proportion of XPU (custom AI accelerators) and wireless business has increased (these two have slightly lower margins), along with a slight decrease in the proportion of high-margin software business, which is a short-term structural issue and not a serious flaw.
2. Business breakdown: XPU snatching GPU’s lunch, network infrastructure as a moat
Broadcom’s ability to establish a foothold in the AI hardware sector relies on a combination of “custom chips + network infrastructure,” which is also the basis for its challenge to NVIDIA:
– XPU (custom AI accelerators): accounting for 65% of AI revenue, currently snatching GPU territory
Broadcom has customized at least two generations of XPU products for major clients like Google, Meta, and TikTok. The longer clients use them, the more familiar they become with the accompanying software tools, making them more willing to migrate computing tasks from GPUs to XPUs. Management stated frankly: “XPU surpassing GPU is not an overnight process, but the trend is clear.”
As the first player to implement the XPU model ASIC, Broadcom’s accumulation of core technology is not something that can be achieved overnight, with R&D investment increasing by over 15% for five consecutive years, creating a moat that is not easily breached.
– AI networking: from single racks to cross-data centers, positioning in all scenarios
As large model clusters expand to 100,000 nodes, networking becomes a bottleneck, and Broadcom’s solutions are tailor-made for this scenario:
– For single racks, the “Thorn 5” switch can support 512 XPU nodes, far surpassing the traditional NVLink’s 72 GPU nodes;
– For cross-rack, the “Thorn 6” reduces the network hierarchy from three layers to two, cutting latency by 30% and saving 25% on electricity costs;
– For cross-data centers, the newly released Jericho4 router can connect giant clusters of 200,000 nodes, a capability unique in the industry.
Some are concerned that Ethernet might be replaced by new technologies, but Broadcom directly poured cold water on this: “Ethernet has been around for 25 years, with latency reduced to below 250 nanoseconds, and it is open-source and inexpensive, so giants like Google have no need to switch.”
– Software business (VMware): a “cash cow” with a 93% gross margin
Software revenue reached $6.8 billion (up 17% year-on-year), accounting for 43% of total revenue, with gross margin soaring to 93% and operating margin at 77%, all thanks to VMware’s effective integration. The newly launched VCF9.0 has been purchased by 90% of its 10,000 large clients, and in the next two years, it can further tap into incremental revenue through value-added services like security and disaster recovery.
3. Non-AI business: U-shaped recovery, not hindering the main line’s rapid growth
Non-AI semiconductor business Q3 revenue was $4 billion (no increase quarter-on-quarter), and the recovery is indeed slow, but there is no need for excessive anxiety: broadband business increased by 8% quarter-on-quarter, which is a bright spot; wireless and industrial businesses remained flat, while enterprise networking and server storage continued to decline. Management described this as a “U-shaped recovery,” which may not see real improvement until mid-2026. Although it is currently lagging, it cannot overshadow the rapid growth of the AI business.
02
Why has ASIC suddenly become popular? Three major logics rewriting the AI hardware landscape
Broadcom’s strength essentially signals that the ASIC track has transformed from “laboratory technology” to a “money-making machine.” The logic behind this has long been laid out at several key points:
1. Large models have “split up,” creating opportunities for ASICs
In the early days, large models pursued “universality,” and GPUs became the first choice due to their strong ecosystem; however, that is no longer the case. Large models are now “splitting up”—financial models focus on risk control, medical models specialize in imaging, and each industry’s model requirements differ. At this point, the customization advantage of ASICs becomes apparent, optimizing for specific models and achieving 30%-50% higher computational efficiency than GPUs. Broadcom’s XPU is seizing market share in vertical fields based on this.
2. Giants have figured out the “cost equation”
AI computing costs for companies like Google and Meta now account for 25% of total operating costs. Whoever can help them save money will have a voice. ASIC solutions can reduce unit computing costs by 30%-40%. For instance, after using Broadcom’s XPU, Meta’s large model inference costs dropped by 35%. This is no small amount, and any company would place more orders.
3. $110 billion in orders = insurance for growth
Half of these orders are from the semiconductor business, and many are AI-related, with most being long-term contracts of 2-3 years. This effectively locks in growth for the next 1-2 years, providing a stark contrast to those manufacturers relying on quarterly orders to survive, making Broadcom’s strategy a reassuring sign for investors.
03
What to watch next? Three nodes determine market height
Broadcom is currently on a strong upward trajectory, but how the market will evolve depends on these key nodes:
1. Delivery of OpenAI’s $10 billion order: This order is due for delivery in Q3 2026. If it can be delivered early or in larger quantities, FY26 growth might even reach 70%, opening up valuation space.
2. The speed at which XPU is snatching GPU’s market share: Currently, XPU accounts for 15% of clients’ AI computing power. If this can be increased to over 25% in 2026, the logic of “ASIC replacement” will be firmly established, and the sector’s sentiment can rise to a new level.
3. The recovery pace of non-AI business: If enterprise networking and server storage can rebound in Q1 2026, it will create a “dual drive” of AI and non-AI, enhancing profitability resilience.
Risks cannot be ignored: if major clients cut AI budgets (for example, if Google suddenly slows down its investments), or if NVIDIA offers targeted low-priced GPUs, it could complicate matters for Broadcom. However, in the short term, the $110 billion in orders and technological barriers are sufficient to withstand these challenges.
04
Gelonghui Research Institute: From tracking to validation, the opportunities for ASICs have long been signaled
From capturing the AI computing market in 2023 to hinting that “ASIC may become the second growth curve” in 2024, the Gelonghui Research Institute’s tracking of the tech sector has never been about chasing trends. For instance, regarding Broadcom, we noticed a detail as early as Q2 2024: its second-generation XPU customized for TikTok had a reasoning speed 40% faster than GPUs, prompting us to suggest in our exclusive member circle that “this company is worth watching.”
This explosion of Broadcom once again proves that opportunities in tech stocks are always hidden within the chain of “technological breakthroughs → customer purchases → performance realization.” For investors, rather than getting caught up in “who will be the next NVIDIA,” it is better to look at practical companies like Broadcom that “have orders and growth certainty”—after all, in the capital market, real performance is always more reliable than stories.
Want to know what other potential stocks are in the ASIC track? Want to track the real-time dynamics of Broadcom’s order deliveries?
Scan to follow the Gelonghui Research Institute, where we use a full-chain research approach of “product tracking → order verification → financial modeling” to help you penetrate market sentiment and lock in the next wave of opportunities in AI hardware.
Note: The companies mentioned in this article are for case analysis only and do not constitute any investment recommendations. The market has risks, and investment should be cautious; please be sure to combine independent judgment before making decisions.