Last week, Broadcom released its fiscal Q2 results and delivered another blinder of a report, once again showcasing its sheer quality and proving it is one of the most significant beneficiaries of the AI boom. It is proving the doubters wrong, and, incredibly, its outlook only continues to strengthen, with this growth story proving much more durable than expected, and still underestimated. Once again, Broadcom forces analysts and me to raise medium-term expectations – there are just no companies better positioned for the next 3-5 years.You see, the key here is that Broadcom sits at the absolute epicenter of the AI infrastructure buildout — not on one side of it, but both sides simultaneously. It designs custom accelerators that hyperscalers use to run their AI workloads, and manufactures the networking silicon that connects those accelerators into the massive clusters that frontier AI demands. Every dollar a hyperscaler spends scaling its AI infrastructure flows, in part, through Broadcom, and as those hyperscalers collectively commit hundreds of billions to AI capital expenditure through the end of the decade, Broadcom’s revenue visibility only deepens.What makes this truly exceptional is that neither business is speculative; XPU revenues are locked in under multi-year contracts with the world’s most powerful technology companies, while the networking franchise is protected by a technological lead that competitors are years away from closing. Growth is accelerating, margins are expanding, and the contracted backlog is growing faster than revenue.Broadcom is simply in a sensational position.Today, I will update my Broadcom thesis (as I do every quarter, given it is one of the key positions in my personal portfolio) by going over the fiscal Q2 results, breaking down the growth drivers, reviewing its financials, and eventually updating my medium-term financial model, fair value, and view on the stock.Let’s delve in!This is a paid-exclusive analysis. Want full access? Consider upgrading to InvestInsights Premium for $9.50 a month or $95 a year! This gets you:6–8 Research Reports per month, 2x more than free subs do, including 3-4 Deep Dives and 2 mid-week analyses (that comes down to just $1 per analysis)Full visibility into my personal portfolio, including allocations, transactions, and conviction levelsAn overview of target prices and ratings, updated weeklyInvestInsights Premium is designed to help you find best-in-class compounders, identify deep-value opportunities, avoid costly mistakes, and stay focused on fundamentals when markets get noisy, with full transparency into how I make decisions over time.Subscribe nowFinancial & Performance ReviewLet’s just jump straight into the Q2 numbers, because there is plenty to cover.Broadcom reported total fiscal Q2 revenue of $22.2 billion last Wednesday, topping the consensus by $70 million and beating my own expectations. Granted, it is a very minimal beat of just 0.3%, despite the numbers topping Broadcom’s own consensus, with expectations already fairly high.Yet, that doesn’t make the result any less impressive. This $22.2 billion in Q2 revenue reflects an incredible 48% YoY growth, a huge acceleration from high-twenties growth in prior quarters. Despite its significant run-rate, Broadcom is delivering the strongest organic growth in its history.Breaking down the results by segment, let’s start with the most important one – Semiconductor Solutions – which accounts for 68% of group revenue and is the undisputed growth driver. Remarkably, semiconductor revenue in Q2 hit a record $15 billion, with growth accelerating from 52% in Q1 to 79% YoY in Q2, driven by unparalleled AI demand.You see, 49% of Broadcom’s revenue comes from AI-related applications, generated by two primary operations – XPUs and Networking. And this business is booming as hyperscalers race to build ever-larger AI clusters to meet computing demand, which requires both purpose-built custom accelerators to replace costly general-purpose GPUs and the high-performance Ethernet switching fabric to connect hundreds of thousands of those chips, driving incredible growth for Broadcom’s solutions.AI revenue in Q1 totaled $10.8 billion, up a blinding 143% YoY, comfortably above guidance. Yet, even more impressively, AI-related orders in Q2 totaled over $30 billion, nearly 3x revenue, underscoring that demand is far outstripping supply, which bodes well for future growth.To provide some background for those unfamiliar, Broadcom’s XPUs (Accelerator Processing Units) are custom-designed AI chips that Broadcom builds in close collaboration with specific hyperscale customers — think Google’s TPU, Meta’s MTIA, and now OpenAI’s custom silicon — tailored precisely to each customer’s unique AI workload rather than being general-purpose chips sold to anyone.So, it’s custom silicon. Unlike Nvidia’s GPUs, which are off-the-shelf accelerators designed to handle a wide range of tasks across thousands of customers, XPUs are purpose-built to run specific inference or training workloads far more efficiently, delivering better performance-per-watt for that narrow use case at the cost of flexibility.The key is that Nvidia’s GPUs are deliberately engineered for universality, which means a hyperscaler running a specific inference workload, say serving billions of ChatGPT queries or Google Search results, is paying for enormous silicon real estate and energy consumption dedicated to capabilities it simply doesn’t need.A custom XPU strips all of that away and redirects every transistor toward the exact mathematical operations that the hyperscaler’s model actually performs, resulting in dramatically better performance-per-watt and performance-per-dollar for that narrow but massively scaled use case. At Google or Meta’s scale, where a single percentage point improvement in inference efficiency translates into hundreds of millions of dollars of annual infrastructure and energy savings, that specialization advantage compounds into an overwhelming economic case for custom silicon over off-the-shelf GPUs, which is precisely why every major hyperscaler has now committed to building their own, with Broadcom as the indispensable design partner making it possible.Broadcom’s role is to provide the chip architecture, SerDes interconnect IP, and deep process-node expertise, while the hyperscaler contributes the microarchitecture specification and workload requirements. Couldn’t hyperscalers do this in-house? Well, time-to-market and product performance today are important to maintain a critical edge. Broadcom’s existing expertise, IP, and technological lead, built up over decades, is the best way to do that.Broadcom today provides these XPUs for six hyperscalers, including Google, OpenAI, Anthropic, Meta, and two unnamed customers, all of which have signed multi-year deals with Broadcom worth tens of billions each.This XPU business generated 60% of AI revenue in Q2, or nearly $7 billion. This is the headline-grabbing business currently.However, it’s the networking business that remains largely underappreciated, yet Broadcom’s competitive advantage, rooted in decades of accumulated IP and technological dominance, is much stronger in Networking and far less concentrated among a handful of customers than in XPUs. It generated nearly $4 billion in revenue in Q2, or roughly 40%, up from 30% in Q1 – it outgrew the XPU business in Q2 amid unprecedented demand. In management’s words, “Demand for XPUs and networking is simply insatiable”.The key insight here is that AI models don’t run on a single chip but across thousands, increasingly tens of thousands of chips simultaneously. Every parameter update during training and every token generated during inference requires GPUs to constantly exchange data with each other. As GPU clusters scale to hundreds of thousands of processors, interconnect speed becomes the critical binding constraint. GPUs are only as powerful as the data pipeline feeding them.For the first three decades of enterprise data center design, the network was an afterthought, a plumbing layer that moved packets between servers. The rise of large language models and large-scale deep learning training has fundamentally changed this, with the industry witnessing a “connectivity-first” revolution in which high-performance interconnects are outpacing GPUs as the next strategic play. AI infrastructure is now limited more by the network than by compute, as clusters are becoming compute-rich but bandwidth-limited.By current estimates, for every $1 spent on AI accelerators, hyperscalers spend an additional $0.40–$0.60 on networking infrastructure. And Broadcom is the dominant force here.According to estimates from JPMorgan, Broadcom holds roughly 80% market share in the data center/AI Ethernet switching and routing chipset market, with its Tomahawk (scale-out) and Jericho (scale-across/WAN routing) families being the backbone of virtually every major Ethernet switch on the market.And while Broadcom is facing tough competition from the likes of Nvidia and Arista, its market share isn’t eroding, protected by technological dominance. With performance being the most important metric for hyperscalers, going with a suboptimal product simply isn’t an option. Therefore, Broadcom is protected by its superior technology and neutral position.Impressively, Broadcom seems to be at least one generation of technology ahead of the competition. Take scale-out, which refers to expanding computing capacity by adding more XPUs or GPUs horizontally. Broadcom has been shipping the only 100-terabit Ethernet switch on the market, the Tomahawk 6, for over a year, and it will start shipping its next-gen 200-terabit product this quarter. Meanwhile, Nvidia is expected to release the equivalent Spectrum-X1600 at 102.4 Tbps only in the second half of 2026, leaving its technology roughly a year behind Broadcom’s. At the current level of innovation, that is light-years.The significance of this lead is huge, with Broadcom practically the only viable solution for hyperscalers. The 200 vs 100 terabit lead means you can move twice as much data in the same amount of time, or the same amount of data in half the time.A 200 Tbps switch means a cluster of GPUs or XPUs spends less time waiting for data to arrive from neighboring chips and more time actually performing useful computation, which translates directly into faster training runs, lower inference latency, and the ability to efficiently coordinate a larger number of chips simultaneously. Practically speaking, going from 100 to 200 Tbps either lets you run your existing cluster twice as efficiently, or lets you scale to roughly twice as many chips before the network becomes the bottleneck.I think it’s clear how important that is and how Broadcom’s technological lead allows it to maintain its dominance.In scale-up within racks, which refers to connecting a small number of chips tightly together inside a single server or rack using ultra-high-bandwidth, ultra-low-latency interconnects, so they behave almost like one giant chip rather than many separate ones, Broadcom also leads with 200G and 400-Gig SerDes. And in CPOs (co-packaged optics), it is the de facto industry standard, with its Jericho3 and Jericho4 fabric solutions.Yeah, Broadcom is hugely dominant in networking, and amid the current computing build-out, that generates huge amounts of revenue.Ultimately, the beauty of Broadcom is that it benefits from both sides of the equation, generating heaps of revenue from compute through its XPUs and from networking infrastructure. And that is what drives current mind-boggling growth. Critically, the XPU business is protected by multi-year contracts (with Google running through 2031), while the networking business is protected by a significant technological lead, making this growth quite durable.Broadcom is simply in a sensational position, honestly, one of the best for the next 3-5 years, with growth likely ramping even higher over the next year and running into the 2030s, going by current estimates.Besides AI, Broadcom’s business is strong. Ultimately, AI is “just” 49% of total revenue, with Broadcom still a highly diversified play. Don’t underestimate the importance of this. For one, Broadcom has two big AI plays, not relying on a single trick, and it still has a non-AI semiconductor business and software business stacked on top of that.Non-AI semiconductors generated $4.2 billion in Q2 revenue, up 6% YoY, with strong bookings at $6 billion, indicating that this business, long pressured by cyclical headwinds, is finally recovering. Broadband, server storage, and enterprise networking together were up, partially offset by a seasonal decline in wireless.Meanwhile, Q2 software revenue was $7.2 billion, up 9% YoY. Again, bookings are strong, with ARR growing 17% YoY, which is a promising indicator for future growth. Crucially, Broadcom’s software business doesn’t seem to be facing any disruption from Agentic AI.Overall, Broadcom delivered an astonishing quarter; there is no way around that.With that, let’s move to the P&L.