--- title: "AVGO (Trans): Pure-play chips; customers like Google are diversifying supply chains." type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/41463849.md" description: "Below is Dolphin Research's Trans of $Broadcom(AVGO.US) FY26 Q2 earnings call. For our earnings take, please see 'AVGO: Giants vs. Giants — Is the ASIC camp splitting?'Core highlights from the quarter include: 1) shareholder returns and 2) Q3 outlook. The company paid $3.1bn in cash dividends in Q2, equivalent to a quarterly cash dividend of $0.65 per common share, and guides Q3 consolidated revenue of $29.4bn (+84% YoY)..." datetime: "2026-06-04T04:10:30.000Z" locales: - [en](https://longbridge.com/en/topics/41463849.md) - [zh-CN](https://longbridge.com/zh-CN/topics/41463849.md) - [zh-HK](https://longbridge.com/zh-HK/topics/41463849.md) author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)" --- # AVGO (Trans): Pure-play chips; customers like Google are diversifying supply chains. **Dolphin Research compiled the following transcript of**$Broadcom(AVGO.US) **FY26 Q2 earnings call. For the earnings take, see '**[**Broadcom AVGO: Titan vs. Titan, Is the ASIC Camp Splitting?**](https://longbridge.cn/en/topics/41463124?channel=SH000001&invite-code=294324&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=23482b80-75b9-4b8a-a144-c4c32cc9402d)**'.** **I. AVGO Earnings Highlights** 1\. **Shareholder returns**: Q2 cash dividends paid totaled $3.1bn. This equates to a quarterly cash dividend of $0.65 per common share. 2\. **Q3 guide**: AVGO guides for consolidated revenue of $29.4bn (+84% YoY). Within that, **semis revenue is ~$20.5bn (+124% YoY), AI semis $16.0bn (\>+200% YoY), non‑AI semis ~$4.5bn (+12% YoY)**, and infrastructure software ~$8.9bn (+31% YoY). AVGO expects Q3 operating margin (OPM) to hold around ~67%. Adj. EBITDA margin is guided at ~68%. 3\. **Margin trends**: Q2 GPM was 77.1%, down 230bps YoY, mainly as semis mix increased within the portfolio. However, OPM improved 200bps YoY to 67.3% with largely flat opex, and adj. EBITDA reached $15.2bn, or 69% of revenue, above the 68% guide. Q3 consolidated GPM is expected to dip to ~74%. Management emphasized this is not a structural change in semis profitability, but a mix shift between semis and software, and advised investors to model semis and software margins separately. 4\. **Cash flow & balance sheet**: Q2 FCF hit a record $10.3bn, at 46% of revenue. Capex was $231mn, with quarter‑end cash at $19.6bn (vs. $14.2bn in Q1). Inventory stood at $4.3bn, with days rising to 86 (vs. 68 in Q1), positioning for a H2 acceleration in AI semis. 5\. **Tax & share count**: Q3 and FY26 non‑GAAP tax rate is expected at ~16%, reflecting the global minimum tax and geographic mix. Q3 non‑GAAP diluted shares are ~5.94bn, excluding potential buyback impact. **II. AVGO Earnings Call Details** **2.1 Management highlights** 1\. **Overall results and AI semis** a. Q2 revenue hit a record $22.2bn (+48% YoY), topping guidance on strong AI semis. OPM reached a record 67%. b. AI semis revenue was a record $10.8bn (+143% YoY), above prior outlook. Networking contributed nearly 40% of Q2 AI revenue. c. **AI semis bookings topped $30bn in the quarter, while shipments were $10.8bn, with 'insatiable' demand for XPUs and networking**. Supply remains tight relative to demand. d. AVGO expects **H2 FY26 AI semis revenue to double vs. H1**. Q3 AI semis are guided to accelerate to $16.0bn (\>+200% YoY). For FY26, AI semis are guided to $56bn, up ~180% vs. FY25. **Management reiterated FY27 AI semis will exceed $100bn**, and expects further growth into FY28 on momentum across six core customers. 2\. **AI core customers (six customers)** a. **Google**: In Apr, AVGO announced a long‑term agreement to develop and supply multi‑gen TPUs and AI networking. The relationship is strategic and large in scale, with AVGO citing a significant lead in technology and execution vs. alternatives. b. **Anthropic**: AVGO will provide over 1 GW of Broadcom TPU compute in 2026. In Apr, the parties signed to **add 5 GW of next‑gen TPU compute from 2027**. c. **OpenAI**: Chips have been delivered, with volume production expected by end‑2026. **There is a 1.3 GW contracted commitment for 2027**, part of the previously announced plan to deploy 10 GW by 2029. d. **Meta**: In Apr, AVGO announced a collaboration to deliver multi‑gen MTIA XPUs, with 3 GW planned by end‑2028 under the agreement. The first 1 GW order, including XPUs and networking, has been received, with deliveries starting in H2 2027. e. **Two additional customers**: Shipments are expected to start by end‑2026 and ramp in 2027. AVGO has received $600mn in purchase orders so far. 3\. **Networking** a. AVGO maintains at least a one‑generation lead in networking technology and products. On scale‑up (in‑rack), industry‑leading 200G/400G SerDes enables direct‑attach copper, driving co‑packaged copper with Ethernet and PCI Express switches. b. On scale‑out (inter‑rack), AVGO is the only vendor with a 100T Ethernet switch in shipment for over a year. The next‑gen 200T switch will tape out this quarter. c. AVGO is the de facto standard in CPO, 1.6T DSPs, and CW/EML lasers. It also leads inter‑DC connectivity with Jericho3 and Jericho4 fabric solutions. 4\. **AI XPU platform (with investors)** a. Strategy: Combine AVGO's leading tech with partners that have the strongest balance sheets, delivering ample compute at the lowest cost and power for frontier AI labs like Anthropic and OpenAI, at scale. The goal is to ensure sufficient capacity as these labs ramp training and inference. b. AVGO is building the AI XPU platform with leading investors including Apollo and Blackstone. The platform targets deploying over 20 GW by 2028, with an initial $35bn program now being launched by Apollo. 5\. **Non‑AI semis** a. Q2 revenue was $4.2bn, up 6% YoY. Bookings topped $6.0bn in the quarter, signaling a path toward a full cyclical recovery. b. Broadband, server storage, and enterprise networking all grew, partly offset by seasonal declines in wireless. Q3 non‑AI semis are guided at ~$4.5bn, up 12% YoY. 6\. **Infrastructure software** a. Q2 software revenue was $7.2bn, up 9% YoY and in line with guidance, with ARR up 17% YoY. Q3 software revenue is guided to ~$8.9bn, up 31% YoY. b. AVGO just released VMware Cloud Foundation 9.1, focused on infra efficiency, security, and enterprise AI inference workloads. Amid robust global server demand, on‑prem private cloud deployments of VCF 9.1 are very strong. c. The release adds heterogeneous compute support across GPU and CPU architectures (AMD, Intel, NVIDIA), enabling enterprise cloud customers to run AI Kubernetes and traditional virtualization workloads in a unified private cloud environment. **2.2 Q&A** **Q: Your FY26 AI revenue outlook implies H2 is 2x H1 and should theoretically exceed $60bn, but the full‑year guide is $56bn. How do we reconcile this?** **Also, given strong project momentum, a broader customer base, and multiple GW‑scale deals starting next year, is the 18‑month AI backlog (H2 FY26 to H1 FY27) already $200bn or higher?** A: On FY26, the math is essentially 2x against 2x. H1 AI revenue is around $19bn, and if H2 doubles as we indicated, you land close to the ~$56bn we guided, so the numbers line up well. On the longer‑dated question into FY27, your analysis is careful and, yes, we expect to sustain this momentum. **Broadly speaking, FY27 is close to 2x FY26, and you should see FY27 comfortably exceed $100bn, consistent with last quarter's outlook**. We continue to expect FY27 to be above $100bn. If anything changes based on current execution, it is more likely to be stronger than weaker, though we will not guide FY27 every quarter. For now, we will say FY27 stays above $100bn, aligned with the trajectory we see exiting H2 FY26. **Q: On the long‑term agreement with Google (8‑K disclosure), the market worries about your share at that customer. Can you discuss confidence in the deal, whether there is upside, and whether it is a fixed dollar amount or involves share?** A: It is a very, very strong agreement that reflects the strength of the relationship, given the products, multi‑gen roadmap, and IP embedded in the program. To your specific question, it is a very large dollar commitment, very substantial in magnitude. At the same time, we accept the reality that, as AI compute development and consumption rise, even a partner like Google will diversify supply to a degree. Still, their commitment to us is very meaningful in dollar terms. **Q: What is driving the GPM step‑down? Is it structural between XPUs and networking, or a rack‑level vs. chip‑level mix? Will this trend persist next year?** A: (Kirsten) As semis grow faster than software, consolidated GPM will compress somewhat, though it remains accretive given strong operating leverage. OPM will trend up over time. Within semis, we have consistently said ASICs, TPUs and parts of wireless carry lower margins, so as TPUs accelerate, blended margins are pressured. That said, connectivity and AI networking margins are very attractive and help offset. (Hock) Structurally, semis margins remain very stable and solid. **The dilution is a portfolio mix effect between software, non‑AI, and fast‑growing AI semis**. **Q: Is the rack‑level vs. chip‑level question settled now?** A: No racks, it is chips. (Kirsten) **We do chips, period**. (Hock) We make chips. **Q: Competitors suggest dollars per GW TAM will rise sharply this decade, not just from infrastructure but also compute and networking content, similar to what Jensen has said about total infra moving from ~50 to 100, with higher compute intensity. Do you see the same and could this accelerate your dollars per GW?** A: On power, keep in mind dollars per GW of rack content do not accelerate as much as that claim suggests. Each chip’s power is rising, so fewer chips are needed, even as ASPs go up. As a result, dollars per GW (in the multi‑billion per GW range) are relatively stable. The GW count itself will keep accelerating, and that is what our commentary points to, with required capacity (in GW) growing very rapidly. Even for Anthropic and OpenAI, for whom we are building this platform to ensure adequate compute, their capacity in GW terms already exceeds what we expected six months ago. That is just two customers. We have not yet included consumption outside the XPU platform, such as Google's internal workloads and other customers. **Putting it all together, total GW should keep growing in FY27 and FY28, and we expect a notable step‑up in FY28 vs. FY27**. **Q: On supply, can you 'write a $20bn check' like some peers to secure incremental wafers and HBM? Are you considering other foundries for flexibility?** A: Securing supply is not only about money, although money helps. We are confident we have locked in supply for FY26 and FY27 for the categories you mentioned, and we are now working on FY28 and FY29. (Follow‑up: If customers ask for incremental supply, can you get it?) Over the past few months, customers have come back for incremental volumes and we expect that to continue. Overall, yes, we can. **Q: Your GW shipment targets by customer for next year versus last quarter’s 'near 10 GW in FY27'—any change? Can you outline the cadence for FY27, more weighted to H2, and is it more, less, or the same GW?** A: Great question. **We previously guided to about 10 GW in FY27, and that remains very solid**. We plan to ship 10 GW in FY27, unchanged. The cadence is H2‑weighted, which sets up an interesting trajectory into FY28, where we expect more GW off that H2‑heavy base. **Q: How should we think about networking? It was ~40% of AI revenue this quarter. As custom projects ramp toward year‑end and early next year, will that mix fall back or stay at the high end? When does optics and CPO become material?** A: It is a great but tough question with many variables. As more customers shift to XPUs, XPUs will consume a lot of our networking content, which boosts consumption and is positive for us. It also means we can sell networking into non‑XPU use cases, which will dilute the growth rate. The 40% print reflects a 'stars‑aligned' moment where we shipped a lot of networking into non‑XPU, while **XPU growth let us bundle networking into our XPU programs, reaching 40%**. I think that is likely the peak mix for networking as a share of AI revenue. A more reasonable long‑term mix for networking within AI revenue is closer to ~30%. **Q: In the recent Anthropic deal you used Broadcom chips as a 'backstop'. Will we see more of these, and how will you finance them—continue using chips?** A: I need to correct that. The Anthropic deal we disclosed in the 8‑K is us supplying compute to Anthropic using GPU chips we developed, which we won on merit. It is not a backstop in that sense. We are the party providing chips and compute to Anthropic. **Q: You have emphasized focus on hyperscale XPU platforms. With many derivatives emerging around interconnect and storage, are any more niche projects attracting your interest?** A: No, I do not think so. Our model is straightforward: we develop XPUs and custom accelerators for customers who are largely LLM developers, for both training and inference, while building a suite of critical components that allow those XPUs and GPUs to cluster and perform better. We deliver this technology in chip form—compute accelerators we call XPUs—and the networking silicon that connects them, including switches, PCIe interconnects, DSPs, lasers, NICs, and routers. That remains our semis model. To help some LLM players access large‑scale GW‑level compute for their expanding models, we partner with firms that have the best balance sheets. The platform provides a vehicle to fund chips for LLM players who otherwise might struggle to access our lowest‑power, lowest‑cost technology at scale. **Q: Looking at GW by customer, two consumer‑facing AI CSPs (Anthropic, OpenAI) have large longer‑term GW commitments. As AI reaches enterprises and consumers, could this drive a second wave of massive demand?** A: It is an interesting view and you are likely right that enterprise AI consumption is still relatively early. We also see that enterprise token consumption is often purchased from the same API platforms of the players we discussed—Anthropic, OpenAI, and Gemini—making those big platforms the source of token pull. Most token consumption is tied to those LLMs. As these LLM players productize frontier models—whether ChatGPT 5.5 or Gemini 3.5—they ultimately draw on the same compute capacity we supply to them. So even as enterprise demand picks up—where enterprises buy AI tokens for their workloads and for productivity users much like consumers do—they are buying from the same few 'big houses'. This is what drives the large‑scale compute ramp we are seeing into 2027 and, now, into 2028 as well. This is shaping into a fairly sustainable and steepening demand trajectory. We expect it to continue. **Q: Does this change the dynamic we discussed earlier? Google already offers GPU cloud, which opens XPU access to smaller firms that do not meet the self‑developed ASIC bar. They could use Broadcom technology via those platforms—why not?** A: It is possible, but the reality is most AI compute is delivered as SaaS, with APIs pulled from the cloud—whether Bedrock, Vertex, Azure, or first‑party—and ultimately provided from the cloud. In the end, most demand stems from a few frontier model developers and the products they deliver to consumers and enterprises globally. We supply compute to that source of demand, rather than selling XPUs to individual firms or banks to build their own software stacks and apps. I am sure a few enterprises are doing that, but not many, and it is still early days. For now, most demand comes from the frontier labs creating code assistants and vertical engineering apps—products from the same small group building frontier models—rather than from 100,000 companies buying XPUs or GPUs directly. **Q: AI bookings of $30bn this quarter are large versus shipments for this and next quarter. Why is backlog so high now?** A: That reflects massive compute demand. The six customers realize there are lead times and planning is required to secure capacity. It is not just wafers for chips and ensuring HBM/DRAM availability. They also need power and the physical enclosures and infrastructure to deliver that power, which all require planning ahead. We are seeing orders placed not for immediate delivery, but with acceptance that several elements must be lined up before delivery. They are placing large orders early to lock in, which gives us unusual visibility in semis. Our visibility now extends into 2028; three months ago it was roughly through 2027. This is also why we are building the XPU platform, to pre‑plan and stand up capacity for those frontier model customers with very large token consumption you see in some financial disclosures. We benefit from substantial lead times now. It is not just our components that must be ready, but also power and nationwide connectivity across U.S. infrastructure to distribute inference to consumers and enterprises. **Q: You previously discussed ~$1.5–2.0bn per GW. Next year’s implied ~10 GW means well over $10bn. You also said dollars per GW varies by program—how should we think about its evolution over time?** A: Our content per GW will increase, simply because XPU content is rising materially in price, particularly as we add not only SRAM but also embed CPU cores and build these into multi‑die with large HBM. Content per GW will trend up on a generation‑by‑generation basis. It will not increase every month or every quarter, but with each new generation the dollars per GW rise. That is the trajectory we expect. **Risk disclosure and statement:**[**Dolphin Research Disclaimer and General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### Related Stocks - [DXYZ.US](https://longbridge.com/en/quote/DXYZ.US.md) - [AVGO.US](https://longbridge.com/en/quote/AVGO.US.md) - [ASIC.US](https://longbridge.com/en/quote/ASIC.US.md) - [META.US](https://longbridge.com/en/quote/META.US.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) ## Comments (1) - **奥马哈的旗帜 · 2026-06-04T09:57:02.000Z**: That makes sense.