Broadcom Inc. (AVGO) - Deep Dive Research Report

Sector: Technology (Semiconductors & Infrastructure Software) | Listing: Nasdaq Global Select Market, ticker AVGO | Fiscal year end: Sunday closest to 31 October | Report date: 31 July 2026

Reporting-period reconciliation. Broadcom reports quarterly, roughly four to five weeks after quarter-end. FY2025 ended 2 November 2025. Q1 FY2026 ended 1 February 2026 and was reported 4 March 2026. Q2 FY2026 ended 3 May 2026 and, on the expected cadence, results were due in the first week of June 2026 - confirmed released 3 June 2026 . Q3 FY2026 ends 2 August 2026; management stated on the Q2 call that Broadcom plans to report Q3 FY2026 after market close on Wednesday 2 September 2026 , which is a future date. The six most recent reporting periods used throughout this report are therefore Q1 FY2025 through Q2 FY2026.

1. What the company does

Broadcom does two things that look unrelated and are held together by one man's capital-allocation philosophy.

First, it designs chips it does not manufacture. Broadcom sells silicon that moves data and, increasingly, silicon that does AI arithmetic. On the data-movement side it sells the switch chips inside the racks of virtually every large data centre, the network interface cards that plug into servers, the optical DSPs and lasers that drive fibre links, the retimers and PCI Express switches, and the SerDes (serialiser/deserialiser) circuits that everything else is built on. On the compute side it co-designs custom AI accelerators - it calls them XPUs - for a handful of the largest technology companies on earth. Google's TPU is the flagship example. Broadcom does not own the product; the customer does. Broadcom supplies the hard parts and takes a margin on the finished package.

Second, it sells enterprise infrastructure software. Broadcom has two reportable segments: semiconductor solutions and infrastructure software. The software half is essentially a portfolio of acquired franchises - VMware, CA Technologies, Symantec's enterprise business, Carbon Black - that sit at the bottom of corporate IT stacks and are extremely painful to remove. VMware Cloud Foundation (VCF) is the flagship: the virtualisation and private-cloud layer that abstracts servers, storage and networking into a pool that applications consume.

The founding story, which is really an acquisition story

The company's DNA is not Silicon Valley startup. It is a corporate carve-out. The semiconductor products group of Hewlett-Packard became part of Agilent, and in 2005 was sold to KKR and Silver Lake and renamed Avago Technologies. Avago listed in 2009. Hock Tan became CEO in 2006 and has run it ever since.

What followed was one of the most consequential acquisition campaigns in modern technology: LSI (storage and networking silicon, 2014), Broadcom Corporation (2016), Brocade (Fibre Channel storage networking, 2017), CA Technologies (mainframe and enterprise software, 2018), Symantec's enterprise security business (2019), and VMware (2023). In 2016 Avago acquired Broadcom Corporation for roughly $37 billion and adopted its name; the combined company would be called Broadcom.

The goal was not growth at any cost - it was consistent, durable cash flow that could support more acquisitions and shareholder returns. After the Broadcom deal, the pace increased.

The VMware transaction was worth $61 billion when announced in May 2022 and pushed the chipmaker deeper into software , closing in November 2023 after an unexpectedly long 18-month process that culminated in approval in China .

The playbook is consistent and has been applied to every acquisition: buy a franchise with a defensible installed base, cut everything that is not the core, focus the sales force on the largest customers, convert perpetual licences to subscriptions, raise prices, and harvest the cash. With VMware, Broadcom eliminated perpetual licences, moved customers onto a subscription model, pushed toward longer-term contracts, raised the minimum licensed cores per order from 16 to 72, required customers to buy a full bundle under the VCF banner, significantly reduced the number of resellers, and focused attention on the top 10,000 customers out of a base in excess of 300,000 . The original 84 product SKUs have been consolidated into roughly half a dozen offerings, with VMware Cloud Foundation and vSphere Foundation carrying every capability that used to be licensed separately.

That is the historical Broadcom: a cash machine assembled from other people's businesses. What has changed since 2023 is that one line item inside the semiconductor segment - custom AI accelerators and the networking to wire them together - has grown so fast that it now dominates the company's identity, its revenue mix, its margin structure and its risk profile.

The core value proposition

There are two distinct ones.

For a hyperscaler or a frontier AI lab: you have a workload you understand better than anyone else, and you are paying a general-purpose GPU vendor a very large margin for silicon that is not optimised for it. Broadcom will co-design a chip that is. You define the architecture and own the intellectual property; Broadcom supplies the physical implementation, the high-speed interfaces, the memory interface, the advanced packaging, the foundry relationship, and - critically - the supply. Hock Tan put the last point plainly: "Our ability to assure supply in these times of constrained capacity in leading-edge wafers, in high bandwidth memory, and substrates ensures the durability of our partnerships."

The economics are compelling for the customer. Goldman Sachs analyst James Schneider noted that the Google-Broadcom TPU is rapidly closing the performance gap with Nvidia, estimating a roughly 70% reduction in cost-per-token as the technology evolves from TPU v6 to v7 - and in a world where AI inference costs can severely impact a balance sheet, that efficiency is a powerful pull toward custom silicon.

For an enterprise IT organisation: you have thousands of applications running on VMware. Broadcom's pitch is that VCF is the layer that lets you run traditional virtual machines, containers and now AI workloads on the same private-cloud substrate. Management's framing: "VMware Cloud Foundation, VCF, is the essential software layer in data centers integrating CPUs, GPUs, storage, and networking into a common high-performance private cloud environment. As the permanent abstraction layer between AI software and physical chips, silicon, VCF cannot be disintermediated or replaced."

What makes the product hard

An XPU is not "a chip." It is a system-in-package that has to solve four brutal problems simultaneously.

Getting data on and off the die. Modern accelerators are bandwidth-starved. Broadcom's SerDes - the analogue circuits that push serial data down a copper trace or into an optical module - are the company's crown jewels, refined over roughly two decades across switch generations. Broadcom's 200G and 400G SerDes are what enable direct-attached copper for scale-up within racks, driving co-packaged copper with Ethernet and PCI Express switches.

The company markets the industry's lowest bit-error-rate SerDes, which reduces link flaps and accelerates job completion time. A link that flaps in a 100,000-accelerator cluster can stall an entire training run.

Attaching high-bandwidth memory. HBM stacks sit millimetres from the compute die on a silicon interposer. This requires HBM PHY and controller IP, thermal engineering and 2.5D/3.5D advanced packaging capacity. SemiAnalysis's TPUv7 analysis describes Broadcom charging Google on a system-in-package basis including HBM, effectively reselling packaged HBM from suppliers such as Samsung and SK hynix as part of a 3.5D XDSiP solution rather than Google procuring HBM directly - an arrangement consistent with external commentary on Broadcom's role in HBM controllers, PHY IP and advanced packaging for hyperscaler accelerators. This detail matters enormously for margins, and Section 8 returns to it.

Securing capacity. Leading-edge wafers, CoWoS-class packaging slots and HBM allocation are all scarce. Asked on the Q2 FY2026 call about securing incremental wafer and HBM supply, Tan said Broadcom is comfortable with supply for 2026 and 2027 and is working on securing supply for 2028 and 2029, and has been able to meet incremental customer demand.

Wiring it all up. A single accelerator is useless. The value is in the cluster, which means switching silicon at bandwidths that did not exist three years ago. Tomahawk 6 provides 102.4 Tbps of switching capacity per chip, supporting up to 512 ports at 200 Gbps or 1,024 ports at 100 Gbps, with options for both copper and co-packaged optics.

The product in action: a walk through the Google TPU

Step one, architecture. Google's silicon team specifies what it wants: matrix-multiply array dimensions, sparsity handling, memory capacity, interconnect topology, power envelope. Google owns this IP.

Step two, co-design. Broadcom brings the pieces Google does not want to build from scratch: SerDes, HBM interface, chiplet-to-chiplet links, physical design and floor-planning, timing closure, design-for-test. TPU v7 (Ironwood) is manufactured on TSMC's N3P process in a dual-chiplet design co-developed with Broadcom and MediaTek, and features two TensorCores with doubled 256x256 MXU arrays plus four SparseCores.

Each chip delivers 4,614 FP8 TFLOPS with 192 GB of HBM3E memory at 7.37 TB/s bandwidth.

Step three, tape-out and manufacture. Broadcom takes the design to TSMC, manages the mask set, yields, HBM procurement and 3.5D packaging at the OSAT, and delivers a tested package.

Step four, the network. Broadcom then sells Google the switching and interconnect silicon that turns thousands of TPUs into a cluster: Tomahawk switches for scale-out, PCIe and co-packaged copper for scale-up inside the rack, Sian DSPs and lasers for the optics, Jericho fabric routers to stitch buildings together.

Step five, repeat for seven generations. Broadcom has confirmed six major XPU customers, including Google, which remains the longest-standing partner, with seven generations of co-designed TPUs since 2014.

Step six - new as of FY2026 - finance the deployment. Broadcom has moved beyond selling components into orchestrating the capital that buys them.

"Our strategic vision is to bring together Broadcom's leading technology and investor partners with the strongest balance sheets to deliver at scale sufficient compute capacity at the lowest cost and power for the leading AI frontier labs, including Anthropic and OpenAI. To deliver this vision, we are creating the AI XPU platform."

• Hock Tan, Q2 FY2026 earnings call, 3 June 2026

That single sentence is the most important strategic statement Broadcom has made in a decade. It describes a company that has stopped thinking of itself as a component vendor.

2. Business segments

Broadcom reports two segments. In FY2025 the split was roughly 58% semiconductor solutions and 42% infrastructure software: consolidated revenue grew 24% year over year to a record $64 billion, AI revenue grew 65% to $20 billion, driving semiconductor revenue to a record $37 billion for the year, while strong adoption of VMware Cloud Foundation drove infrastructure software revenue growth of 26% to $27 billion.

That mix is being rewritten quarter by quarter. In Q2 FY2026 Semiconductor Solutions revenue was $15 billion, representing 68% of total revenue and up 79% year on year, with AI semiconductors alone contributing 49% of total revenue and networking accounting for nearly 40% of AI semiconductor revenue.

Because the semiconductor segment now contains two businesses with radically different growth rates and economics, this section treats them separately.

2.1 Semiconductor Solutions - AI custom accelerators (XPUs)

What it does. Co-designs and supplies custom AI compute silicon for a defined set of very large customers. Broadcom describes this as providing advanced technology and intellectual property platforms for customers

• the polite framing of what is in practice a deep multi-year engineering partnership. The customer owns the architecture and the product; Broadcom owns the implementation and the supply chain.

The named roster as of the most recent call: Tan said on the Q2 FY2026 earnings call that the company has six core custom chip customers, including Google, Meta, Anthropic, and OpenAI.

Meta, ByteDance and Fujitsu round out the confirmed customer list, with analysts identifying Apple and Arm/SoftBank as potential future engagements.

The core capability. Three things, none of which can be bought.

The first is a two-decade SerDes and physical-layer library, continuously re-proven across switch generations that ship in enormous volume. A switch chip is the most demanding SerDes application in the industry; winning there produces IP that transplants directly into accelerators.

The second is advanced packaging and HBM integration know-how at production scale, plus the commercial relationships to secure interposer capacity and HBM stacks. This is why Broadcom can bill at the system-in-package level.

The third is a track record of not missing tape-outs. For a customer committing a gigawatt of data-centre capex to a chip, a slipped generation is a catastrophe. Broadcom took Tomahawk 6 to production in less than three quarters from initial sampling, which its own SVP called "a testament to its innovation at scale." That reputation is the moat that actually decides deals.

Why it exists separately. It does not, formally - it is reported inside Semiconductor Solutions. But it is managed and discussed as a distinct business, with its own revenue disclosure line, its own backlog disclosure, its own customer count, and now its own financing vehicle. Management separates "AI semiconductor revenue" from "non-AI semiconductor revenue" on every call precisely because they behave nothing alike.

Competitive position. Dominant, and increasingly so. Market share estimates place Broadcom at 70%-plus of the custom AI accelerator design services market, up from the 60%-80% range Bloomberg Intelligence flagged earlier in 2026, and the concentration is increasing as Broadcom's customer wins compound.

Broadcom and Marvell together control roughly 95% of the ASIC co-design market. Where it loses: on price against MediaTek's cost-optimised approach, and structurally when a customer decides to internalise physical design.

How it fits into the group. It is the entire growth story and the reason the stock exists in its current form. It is also the segment diluting consolidated gross margin.

Revenue mix. AI semiconductors contributed 49% of total company revenue in Q2 FY2026.

2.2 Semiconductor Solutions - AI and data-centre networking

What it does. Sells the silicon that connects accelerators to each other. Three tiers: scale-up (inside a rack), scale-out (rack to rack within a building), and scale-across (building to building). Broadcom's networking connectivity solutions manage the movement of data in data centre, service provider and enterprise networking applications, offering an open, flexible, standards-based Ethernet NIC and switching solution to resolve connectivity bottlenecks, particularly in AI data centres where compute bandwidth and cluster sizes grow rapidly.

Customers are the hyperscalers directly, plus the system vendors and ODMs that build boxes around Broadcom silicon - Alpha Networks, Celestica, DNI Emerging Technologies, Micas Networks, Quanta, Wistron and Wiwynn all showcased Tomahawk 5, Tomahawk 6 and TH6-Davisson based platforms at the 2025 OCP Global Summit .

The core capability. Being one full generation ahead on bandwidth per chip, and betting correctly that open Ethernet would beat proprietary interconnect for AI. Tan's claim on the Q2 FY2026 call: "While we have significant IP and execution leadership in XPUs, networking is key to building scalable XPU and GPU clusters. Here in networking, we have at least one generation of technology and product leadership."

Tomahawk 6 entered volume production in March 2026 as the industry's first 102.4 Tbps Ethernet part; the companion Jericho 4 fabric chip at 51.2 Tbps began shipping the prior August and is designed to interconnect over one million XPUs across data centres, while Nvidia's competing Spectrum-X1600 was not expected in volume until the second half of 2026.

Broadcom has also deliberately open-sourced the surrounding standards to widen the ecosystem it sits at the centre of. Tomahawk 6 is designed to be part of an open scale-up Ethernet ecosystem; Broadcom announced the Scale Up Ethernet framework at OCP Dublin in April 2025, made it freely available, and is sharing the technology with open standards development organisations including OCP.

Jericho4 is fully compliant with Ultra Ethernet Consortium specifications, ensuring interoperability across open standards-based Ethernet AI fabrics and allowing integration with a broad ecosystem of UEC-compliant NICs, switches and software stacks.

Why it exists separately. Different customer set (system vendors and ODMs as well as hyperscalers), different economics (merchant silicon at high margin versus SiP pass-through at lower margin), and a longer history. It is the business that produced the SerDes IP the XPU business now runs on.

Competitive position. Broadcom has historically owned the data-centre switch-silicon market. Broadcom has been estimated to hold roughly 55% of the data centre Ethernet switch chip market, with Marvell around 20% and Nvidia, through Mellanox, around 15%. But the competitive picture at the system level has shifted sharply. According to IDC, Nvidia's share of the data centre Ethernet switching market climbed from less than 4% two years ago to 21.5% in the first quarter of 2026.

Nvidia is selling Ethernet as part of a broader AI factory architecture rather than as a networking platform, a distinction that could reshape competition, because winning AI infrastructure deals may require more than matching Nvidia on switch speeds or pricing. That is a direct challenge to the merchant-silicon model.

How it fits. It is the attach-rate engine. Every XPU Broadcom ships pulls switches, NICs, retimers, DSPs and lasers with it, and it also attaches to Nvidia GPU clusters, which is why AI networking grew even when XPU shipments were lumpy.

Revenue mix. Networking accounted for nearly 40% of AI semiconductor revenue in Q2 FY2026

• roughly a fifth of total company revenue.

2.3 Semiconductor Solutions - the non-AI franchises

What it does. Broadcom offers semiconductor solutions in five major end markets: Networking Connectivity, Wireless Device Connectivity, Servers and Storage Systems, Broadband and Industrial. Stripping out AI, this is a collection of mature, high-share, cash-generative product lines:

• Wireless device connectivity. Leading-edge connectivity solutions for the wireless device market, including RF front-end modules and filters, Wi-Fi/Bluetooth combination chips, custom touch controllers and inductive charging devices.

The RF devices selectively filter, amplify and route RF signals; filters enable wireless systems to support many subscribers simultaneously by ensuring multiple voice and data transmissions do not interfere, and the portfolio includes multi-chip front-end modules integrating transmit/receive switching and filtering across multiple frequency bands, filter modules and discrete filters, all using Broadcom's proprietary film bulk acoustic resonator (FBAR) filter technology. This is overwhelmingly an Apple business.

• Broadband access. Complete SoC platform solutions for DSL, cable, passive optical networking and wireless LAN, for both customer premise equipment and central office; CPE devices are used in broadband modems, residential gateways and Wi-Fi access points and routers, enabling service providers to deploy G.fast, DOCSIS, PON and Wi-Fi. Plus set-top box SoCs.
• Servers and storage systems. SAS and RAID controllers, host bus adapters, PCI Express switches, and Fibre Channel storage-area networking (the Emulex and Brocade franchises).
• Industrial. Factory automation, renewable energy and automotive electronic solutions including optocouplers, industrial fibre optics, industrial and medical sensors, motion encoders, light emitting diode devices, and Ethernet ICs , used across industrial automation, power generation and distribution, medical systems, defence and aerospace, and vehicle subsystems including EV powertrain and infotainment .
• Enterprise and service provider networking. Trident campus switching silicon, Qumran routing silicon for telco backhaul.

The core capability. Process-specific manufacturing that most fabless peers cannot do. FBAR filters are not CMOS logic; they are acoustic resonators built in Broadcom's own specialty fab. The Fort Collins site is a 200mm RF fab , and Apple already helps support more than 1,100 jobs in Broadcom's Fort Collins FBAR filter manufacturing facility . Owning that process, and the yield learning inside it, is why Apple has stayed for decades.

Why it exists separately. Acquisition history, entirely. These are the Avago, Broadcom Corporation, LSI, Emulex and Brocade franchises. They serve consumer and industrial cycles, not data-centre capex cycles.

Competitive position. High share in narrow niches, with genuine competition. FBAR competes with Qorvo and Skyworks (now merging) and increasingly with Apple's own integration ambitions. Broadband competes with MaxLinear and MediaTek. Storage controllers compete with Marvell and Microchip. Industrial is fragmented.

How it fits. It is the ballast, and right now it is flat. In Q4 FY2025, non-AI semiconductor revenue was up 2% year over year with broadband recovering, wireless flat, and all other end markets declining as enterprise spending remained weak.

The non-AI semiconductor business - broadband chips, wireless connectivity, storage controllers, enterprise networking - is roughly flat year over year; all of the semiconductor segment's growth came from AI, and if hyperscaler AI capex pulled back hard there is no comparable growth engine sitting underneath it.

That said, the wireless leg just got a large, long-dated underpinning. Broadcom is part of Apple's American Manufacturing Program, and a new agreement marking Apple's largest AMP commitment to date will enable Broadcom to expand and modernise its Fort Collins facilities with a $1.5 billion capital investment, producing advanced RF components including FBAR filters and advanced wireless connectivity technologies.

Announced 8 July 2026, the multiyear agreement is expected to exceed $30 billion and will result in production of more than 15 billion U.S.-made chips, with Fort Collins serving as the production hub for custom ASIC silicon, RF components including FBAR filters, and next-generation wireless connectivity for future Apple device generations through at least 2031.

Revenue mix. Roughly 19% of total company revenue in Q2 FY2026, and guided to approximately $4.5 billion in Q3 FY2026, up 12% from a year earlier .

2.4 Infrastructure Software

What it does. Sells and supports the software layers at the base of enterprise IT. The portfolio spans private cloud (with VMware vDefend and VMware Avi Load Balancer available as advanced services for VCF), application development and data services via Tanzu solutions, and a mainframe software portfolio including AIOps and automation, database and data management, DevX and DevOps, cybersecurity and compliance, and foundational open mainframe solutions.

The FY2024 annual report groups the software estate into Private Cloud (VMware Cloud Foundation, VCF Edge, vSphere Foundation, Private AI, Live Recovery, Telco Cloud Platform), Tanzu (Tanzu Platform, Tanzu Data Solutions, Tanzu CloudHealth, Tanzu Labs), VeloCloud (SD-WAN and SASE, VeloRAIN, Fixed Wireless Access), and Application Networking and Security. Symantec Enterprise Security and Carbon Black provide the cybersecurity leg. Broadcom delivers technology from Symantec, Carbon Black, CA Technologies and VMware.

The core capability. Not technology - commercial architecture. Broadcom knows how to take a sprawling software estate, collapse it into a handful of mandatory bundles, remove buyer optionality, lengthen contract terms, and route the whole thing through a shrunken direct-sales motion aimed at the largest accounts. It has now executed this three times (CA, Symantec, VMware) with reproducible results.

Why it exists separately. Completely different customers (enterprise CIOs, not silicon architects), completely different cost structure (near-software-industry gross margins), completely different sales motion, and completely different cyclicality. It also serves a specific corporate purpose: it is the counter-cyclical cash generator that funds the semiconductor R&D and the dividend.

Competitive position. Strong incumbency, real churn risk. Alternatives exist - Nutanix, Microsoft's Hyper-V and Azure Local, Red Hat OpenShift, Proxmox, open-source KVM stacks - and Broadcom's pricing has motivated customers to evaluate them. "Broadcom's VMware strategy prioritizes monetizing the existing customer base over expanding it," in one analyst's characterisation, "an approach that has already generated strong short-term financial results." Broadcom is investing to keep the platform relevant: it has poured significant resources into VCF, as evidenced by the release of VCF 9.1, described as an AI- and Kubernetes-native private cloud platform with integrated security , and the release adds heterogeneous compute support across GPU and CPU architectures including AMD, Intel and Nvidia platforms, enabling enterprise cloud customers to run AI, Kubernetes and traditional virtualised workloads on a common private cloud environment . On the agentic frontier, Broadcom announced VMware Tanzu Platform agent foundations, a secure-by-default agentic runtime intended to move enterprise developers beyond siloed AI experiments into scalable, governed production on VCF .

How it fits. The margin engine and the cash anchor. Software gross margin reached 93% and operating margin improved to 78% in Q4 FY2025, up from 72% a year earlier, and management expects low double-digit growth for infrastructure software in FY2026, supporting durable high-margin cash flows to fund accelerated AI investments across silicon and packaging.

The growth phase of the VMware repricing is largely done, though. Q1 FY2026 infrastructure software revenue was in line with guidance and up just 1% year on year. Management's answer is expansion into the tier below the giants: on the March call Tan noted the largest 10,000 companies are finding success and value with VCF, adding "We are now looking at whether the next 20,000, 30,000 midsized companies see it the same way. Stay tuned."

Revenue mix. Infrastructure Software revenue was $7.2 billion in Q2 FY2026, up 9% year over year and in line with guidance

• roughly 32% of total revenue.

2.5 The AI XPV Platform - a new leg that is not a segment

In June 2026 Broadcom did something no component supplier has done at this scale: it built a financing vehicle to buy its own products.

Broadcom announced the establishment of the AI XPV Platform with Apollo and Blackstone's Credit & Insurance business as initial anchor investors; the Platform is designed to enable more than 20 gigawatts in compute capacity using Broadcom's XPUs and networking solutions customised for leading frontier AI labs, including Anthropic and OpenAI, through 2028.

It launched with an initial tranche of $35 billion led by Apollo, in partnership with Blackstone, to facilitate Anthropic's previously announced capacity expansion of more than 1 gigawatt of compute infrastructure expected to deploy in Fluidstack-based sites beginning in mid-2026.

Apollo describes the initial transaction as the product of a deeply collaborative relationship with Broadcom, designed to deliver committed, certain capital across a multi-year draw schedule.

Why this exists: frontier AI labs have enormous compute appetites and, unlike hyperscalers, no operating cash flow to match. The binding constraint on Broadcom's order book is increasingly the customer's balance sheet, not Broadcom's supply. The XPV Platform substitutes private credit for that missing balance sheet. It is asset-backed financing: chips, networking, power and long-term customer commitments bundled into one vehicle institutional investors can fund at scale.

Strategically it is elegant. It also imports credit risk into a semiconductor company's demand curve, which Section 8 treats as a first-order risk.

Segment summary

3. Products and business detail

3.1 The full catalogue

Custom compute (XPU) and custom silicon. Broadcom's custom silicon products include AI chips such as Google's TPUs and Meta's MTIA, as well as YouTube video-encoding chips and custom routing and switching silicon. Active or announced programmes span Google TPU (through seven-plus generations), Meta MTIA, OpenAI's accelerator, Anthropic's TPU-based compute access, ByteDance, Fujitsu, and two customers Broadcom has not named. OpenAI signed a multi-year collaboration in October 2025 for 10 gigawatts of custom accelerators, with first deployment targeting the second half of 2026 using both 3nm and 2nm designs.

Ethernet switching. Tomahawk is the scale-out and scale-up workhorse; Trident serves enterprise campus; Jericho and Qumran serve deep-buffer routing and fabric. Broadcom's dominance is driven by the widespread adoption of its Tomahawk, Trident and Jericho product families, extensively used in hyperscale cloud data centres and spine-leaf architectures. The current line-up:

• Tomahawk 6 - the family includes an option for 512 200G or 1,024 100G SerDes on a single chip, enabling AI clusters with extended copper reach and efficient use of XPUs and optics with native 100G/200G interfaces . It complies with Ultra Ethernet Consortium specifications and supports modern AI transports, congestion signalling and telemetry for large distributed training, across scale-up, Clos, rail-only, rail-optimised and torus topologies.

Multiple deployments are planned with more than 100,000 XPUs using Tomahawk 6 for both scale-out and scale-up interconnect.

• Tomahawk 6-Davisson - Broadcom's third-generation co-packaged optics Ethernet switch, offering 102.4 Tb/s .

• Tomahawk Ultra - a lower-latency variant targeted at scale-up fabrics.

• Jericho4 - a purpose-built ethernet fabric router for distributed AI infrastructure, designed to interconnect over one million XPUs across multiple data centres , where a single system scales to 36,000 HyperPorts each operating at 3.2 Tb/s, with deep buffering, line-rate MACsec, and RoCE transport over 100km-plus distances . The rationale: as models grow, infrastructure requirements exceed the power and physical limits of a single data centre, and distributing XPUs across facilities requires a new class of router optimised for very high-bandwidth, secure, lossless transport across regional distances.

• Trident4 X11 / X11c - enterprise and campus switching silicon.

Network interface cards. Thor 2 and Thor Ultra. Thor Ultra is an 800G NIC that doubles Thor performance and adds congestion control for clusters scaling beyond 100,000 XPUs. Specifications that matter: available in standard PCIe CEM and OCP 3.0 form factors; 200G or 100G PAM4 SerDes with support for long-reach passive copper; PCI Express Gen6 x16 host interface; line-rate encryption and decryption with PSP offload relieving the host/XPU of compute-intensive tasks; secure boot with signed firmware and device attestation; a programmable congestion-control pipeline; and packet trimming and congestion signalling support with Tomahawk 5, Tomahawk 6 or any UEC-compliant switch.

Broadcom re-engineered RDMA for AI workloads, adding packet-level multipathing, out-of-order data placement, hardware-based retransmissions and programmable congestion control to prevent bottlenecks in scale-out environments. The design logic is worth understanding: "Tomahawk 6 lets customers stay within two tiers instead of three, which simplifies the network and cuts complexity... As clusters grow from 100,000 to half a million XPUs, the NIC becomes the control point for congestion control and reliability at scale."

Optical and physical layer. Broadcom's end-to-end Ethernet AI platform includes Thor NICs, Agera retimers, Sian optical DSPs, co-packaged optics, and software development kits, delivering a complete solution for next-generation AI infrastructure. Management's own claim on positioning: "And in CPOs, which is co-packaged optics, 1.6-terabit DSPs, CW and EML lasers, we are the de facto standard in the industry. To extend AI clusters across data centers, we remain the industry leader with our Jericho 3 and Jericho 4 fabric solutions... enabling the world's largest deployments of multiple hyperscalers."

Wireless. FBAR filters, filter modules, multi-band RF front-end modules, Wi-Fi/Bluetooth combination chips, custom touch controllers, inductive charging devices. An FBAR filter separates radio waves in mobile devices and wireless systems, permitting desired signals such as 5G or Wi-Fi to pass while blocking disruptive frequencies.

Broadband. DOCSIS cable modem SoCs, PON and DSL SoCs, residential gateway and Wi-Fi access point SoCs, set-top box SoCs, central-office line cards.

Servers and storage. SAS/RAID controllers, Fibre Channel HBAs, Brocade Fibre Channel SAN switching, PCI Express Gen5/Gen6 switches. Jabil showcased an Arm-based AI inference solution integrated with a Broadcom PCIe switch and Ethernet NIC

• an illustration of how the storage-era PCIe franchise has found new life in AI servers.

Industrial. Optocouplers, industrial fibre optics, industrial and medical sensors, motion encoders, LEDs, industrial Ethernet ICs.

Software. VMware Cloud Foundation (9.0/9.1), VCF Edge, vSphere Foundation, NSX, vSAN, Aria, VMware Private AI Foundation, VMware Live Recovery, Telco Cloud Platform, vDefend, Avi Load Balancer, Tanzu Platform and Tanzu Platform Agent Foundations, Tanzu Data Solutions, Tanzu CloudHealth, VeloCloud SD-WAN/SASE and VeloRAIN, Symantec Enterprise Security, Carbon Black, and the CA mainframe estate.

3.2 What makes it hard to make

Three constraints define Broadcom's engineering problem.

Reticle and package limits. A 102.4 Tbps switch or a dual-chiplet accelerator with 192 GB of HBM pushes against the physical limit of what a reticle and an interposer can hold. Going bigger requires chiplets, which requires die-to-die interfaces, which requires yet more SerDes expertise.

Signal integrity at 200G per lane. At these rates copper traces behave like transmission lines and every via is a reflection. This is analogue engineering that cannot be simulated away; it is learned across silicon generations. Broadcom's bit-error-rate advantage is the accumulated output of that learning.

Thermal and power. Co-packaged optics exist because pluggable optical modules consume too much power at 1.6T. Moving the optical engine into the switch package saves watts and creates a completely new set of reliability and serviceability problems that Broadcom is now on its third generation of solving.

There are no formal regulatory certifications gating these products in the way pharmaceuticals or aerospace parts are gated. The gating function is customer qualification: months of interoperability testing, thermal validation, and at-scale burn-in before a hyperscaler will accept a part into production. That qualification process is the real barrier, and it is why incumbency compounds.

3.3 Manufacturing and delivery

Broadcom is fab-lite. It operates a "fab-lite" business model focused on designing and supplying complex semiconductor solutions and infrastructure software for large enterprise customers.

Advanced digital logic goes to TSMC - TPU v7 on N3P, OpenAI's accelerator spanning 3nm and 2nm nodes. Advanced packaging goes to OSATs; the FY2024 10-K names Amkor Technology among Broadcom's suppliers and partners alongside Foxconn Technology Group . HBM comes from the memory oligopoly and, per external analysis, is often procured and repackaged by Broadcom rather than by the customer.

Broadcom does own fabs, but for processes foundries do not offer. Fort Collins, Colorado is the FBAR filter facility - a 200mm RF fab , now receiving a $1.5 billion capital expenditure investment to expand and modernise . Broadcom will produce advanced radio frequency components including FBAR filters and advanced wireless-connectivity technologies at its 4380 Ziegler Road facility in Fort Collins.

Employees there often work 12-hour rotating shifts in cleanroom or manufacturing capacities.

Logistically, one detail in the accounts explains the geographic revenue table and is worth internalising: for the majority of Broadcom's products, title and control transfer to customers in Penang, Malaysia, and the products are then transported to customer-specific locations. Penang is the pivot of the physical supply chain.

Capital intensity is remarkably low for a company of this scale, which is the structural reason the free cash flow conversion is so high. Capital expenditure in Q2 FY2026 was just $231 million against $10.5 billion of operating cash flow. Broadcom sells silicon and software; it does not build data centres. Management said so explicitly when pressed: asked about rack-scale versus chip-scale dynamics, Tan and CFO Kirsten Spears clarified that Broadcom is "in the chip business only."

3.4 Geographies

FY2025 revenue by shipment or delivery location: United States $16,506 million; China including Hong Kong $11,155 million; Singapore $10,796 million; Taiwan $6,451 million; other foreign countries $18,979 million. China's share has been declining: in fiscal years 2025 and 2024, 17% and 20% of net revenue respectively came from shipments or deliveries to China including Hong Kong.

The interpretive caution is important. These geographic delivery locations are not necessarily indicative of the location of end customers or where end customers sell devices containing Broadcom products; Broadcom believes a substantial portion of products shipped to China including Hong Kong is included in devices sold by its end customers in the United States and Europe. Taiwan crossing the 10% disclosure threshold for the first time in FY2025 is a direct fingerprint of the AI build: that is where the ODMs assemble.

Many of Broadcom's semiconductor customers design products in North America or Europe that are then manufactured in Asia; to serve customers globally Broadcom has developed relationships with large global electronic component distributors complemented by regional distributors, and also sells to OEMs and their contract manufacturers.

3.5 Milestones that changed the business

• 2005-2009: HP/Agilent semiconductor carve-out becomes Avago under KKR and Silver Lake; IPO 2009.

• 2014: LSI acquisition brings storage and custom-silicon capability.

• 2016: Avago acquires Broadcom Corporation for roughly $37 billion and takes its name.

• 2017-2019: Brocade, CA Technologies, Symantec enterprise - the software pivot.

• November 2023: VMware closes after an 18-month process, in a transaction worth $61 billion at announcement.

• July 2024: ten-for-one forward stock split, with split-adjusted trading commencing 15 July 2024.

• June 2025: Tomahawk 6 launches, delivering 102.4T Ethernet with an open, scalable fabric as an alternative to InfiniBand.

• August 2025: Jericho4 begins shipping and sampling to customers.

• October 2025: OpenAI and Broadcom announce a collaboration for 10 gigawatts of custom AI accelerators, with Broadcom deploying racks of AI accelerator and network systems targeted to start in the second half of 2026 and complete by end of 2029.

Days later at OCP, Broadcom introduces Thor Ultra, the industry's first 800G AI Ethernet NIC.

• March 2026: Tomahawk 6 enters production volume, moving from initial samples to production deployment in under three quarters.

• April 2026: long-term agreements signed with Google (multi-generation TPU), Anthropic (five additional gigawatts) and Meta (multi-generation MTIA).

• June 2026: the AI XPV Platform is established with Apollo and Blackstone, designed to enable more than 20 gigawatts of compute capacity through 2028.

• July 2026: Apple signs a multi-year agreement valued at over $30 billion, its largest single AMP commitment, generating more than 15 billion U.S.-made chips and including a $1.5 billion expansion of Fort Collins.

4. Customers

4.1 Who actually buys

Broadcom's customer base is barbelled to an extreme degree: a handful of accounts that matter enormously, and a very long tail that matters mostly for margin.

The disclosed concentration is stark. During fiscal years 2025, 2024 and 2023, one customer accounted for 32%, 28% and 21% of net revenue respectively; revenue from this customer was included in the semiconductor solutions segment. One customer accounted for 44% and 18% of net accounts receivable as of 2 November 2025 and 3 November 2024 respectively.

Broadcom believes aggregate sales to its top five end customers, through all channels, accounted for approximately 40% of net revenue for each of fiscal years 2025 and 2024, expects to continue to experience significant customer concentration, and warns that the loss of, or significant decrease in demand from, any of its top five end customers could have a material adverse effect. Channel structure: sales to distributors accounted for 48% of net revenue in each of fiscal 2025 and 2024.

Named or reliably reported accounts, by product line:

• XPU: Google (TPU), Meta (MTIA), OpenAI, Anthropic (TPU-based compute), ByteDance, Fujitsu, plus two undisclosed. Six core custom chip customers as stated by Tan on the Q2 FY2026 call.

• AI networking: the same hyperscalers, plus Nvidia-cluster operators, plus system vendors and ODMs (Arista, Celestica, Quanta, Wistron, Wiwynn, Micas, Alpha Networks, Jabil, ASRock Rack).

• Wireless: Apple, dominantly.

• Broadband: cable and telco operators via CPE OEMs.

• Software: a base in excess of 300,000 customers, with sales attention concentrated on the top 10,000.

4.2 Who inside the customer decides, and on what criteria

For an XPU programme, the decision sits with the customer's infrastructure and silicon leadership - the VP of infrastructure, the chief architect, the head of the internal silicon team - with sign-off from the CFO because the commitment is measured in gigawatts and billions. The criteria are performance per dollar of total cost of ownership, performance per watt (because power is the binding constraint on data-centre expansion), roadmap credibility across multiple generations, and supply assurance. The sales cycle is measured in years: architecture engagement, tape-out, silicon bring-up, qualification, then volume. Tan's own description of the sequencing on the OpenAI programme illustrates the timeline: "For OpenAI, we have delivered silicon, and we are on track for production late 2026... We have a contractual commitment to deploy 1.3 GW in 2027 as part of the larger 10 GW that by 2029 agreement we announced last year."

Note also that Broadcom is selective about who it takes on. The company continues to engage closely with additional prospects but is selective in qualifying new customers, focusing on large-scale LLM developers, and management has suggested that a small number of customers will represent the vast majority of near-term AI XPU growth. This is a supplier that rations its own engineering capacity - a rare posture and a genuine sign of pricing power.

For AI networking, the buyer is the network architecture team, and the criteria are bandwidth per chip, radix (port count, which determines how many switching tiers you need), power per bit, and ecosystem interoperability. Standards compliance is a purchase criterion here in a way it is not for XPUs, which is why UEC and OCP participation matters commercially, not just reputationally.

For wireless, the buyer is a single handset OEM's silicon and RF team, and the criteria are filter performance in a shrinking board area, yield at volume, and now supply geography. The Apple AMP structure is instructive - Apple's funding of Broadcom's Fort Collins facility aims to secure access, and obtaining chips from the Fort Collins plant is Apple's primary interest given its AMP pledge, with Broadcom's proficiency in custom silicon the next priority.

For software, the buyer is the CIO or infrastructure director, and the decision is increasingly a renewal negotiation rather than a purchase. Criteria are migration cost avoidance, application compatibility, compliance and sovereignty requirements, and support obligations. VMware customers used to negotiate hypervisor licensing; they now negotiate a multi-product subscription bundle, and the bundle has been reshaped four times since close.

4.3 Why they choose Broadcom

Four specific reasons, none of them generic.

Because Broadcom can actually deliver the silicon on schedule. Tomahawk 6 sampling to production in under three quarters is a proof point customers cite internally.

Because Broadcom controls the supply chain the customer cannot. Wafer allocation, packaging slots, HBM. This is why Tan's "assure supply" line is the single most commercially loaded sentence in the transcripts.

Because networking and compute come from the same vendor. For Broadcom, the OpenAI collaboration reinforces the importance of custom accelerators and the choice of Ethernet for scale-up and scale-out networking , and Charlie Kawwas frames it as "custom accelerators combine remarkably well with standards-based Ethernet scale-up and scale-out networking solutions to provide cost and performance optimized" infrastructure . A customer buying both from one supplier avoids the integration risk that sits between two.

Because the customer wants to own the architecture. This is the deepest reason and the one that distinguishes Broadcom from Nvidia. OpenAI's stated rationale: "By building our own chip, we can embed what we've learned from creating frontier models and products directly into the hardware, unlocking new levels of capability and intelligence."

Tan's version of the same argument: "New frontier models and techniques put unexpected pressures on AI systems. It's difficult to serve all clusters of models with a single system design point... it is hard to imagine that a general-purpose accelerator can be configured and optimized across multiple frontier models."

4.4 Switching costs

For XPUs, high but not absolute. The customer owns its architecture and can in principle port it. What it cannot easily port is the physical implementation: Broadcom's SerDes, HBM PHY, chiplet interfaces and packaging flow are embedded in the design database. Moving suppliers means re-doing physical design, re-qualifying, and almost certainly losing a generation - roughly 12-18 months in a market where 12 months is a competitive lifetime. It also means renegotiating wafer and HBM allocation from a weaker position.

The evidence that these switching costs are real but finite is Google itself. TPU v7 is a dual-chiplet design co-developed with Broadcom and MediaTek.

Google's strategic objective in partnering with MediaTek is reported to be moving toward a model in which the silicon partner charges only for what it adds value on, with Google paying closer to bill-of-materials cost; The Information's 2023 reporting framed Google's long-term goal as breaking from Broadcom on TPU supply to reduce unit costs, while Reuters coverage of the MediaTek partnership highlights lower per-chip pricing and diversification as primary motives. And Tan himself conceded the point on the June 2026 call - he acknowledged that Google would likely draw on multiple chip suppliers . That admission is the reason the stock fell.

For networking silicon, switching costs are moderate: SDK and software-stack investment, board redesign, and qualification, but merchant silicon is by design substitutable.

For wireless, switching costs are high on the filter side because FBAR is a proprietary process with specific performance characteristics that a competitor's bulk-acoustic-wave filter does not exactly replicate, requiring RF re-tuning of the entire front end.

For software, switching costs are the highest in the entire company and the reason the repricing worked. Migrating thousands of VMs off vSphere means re-platforming applications, retraining operations staff, rebuilding automation and disaster recovery, and accepting an outage risk that most CIOs will pay a great deal to avoid.

4.5 Contract structures and revenue predictability

Broadcom has shifted decisively from purchase-order business to multi-year contracted commitments, and the disclosures now read more like an infrastructure company's than a chipmaker's.

Multi-year agreements with leading industry customers underpin long-term expectations of $100 billion-plus in annual AI semiconductor revenue starting in fiscal 2027. The specifics, as laid out on the Q2 FY2026 call:

Tan said Broadcom entered into a long-term agreement with Google in April to develop and supply multiple generations of TPUs and AI networking; has agreements involving TPU-based compute access for Anthropic including more than one gigawatt in 2026 and another five gigawatts beginning in 2027; for OpenAI has delivered silicon and remains on track for production in late 2026 with a contractual commitment to deploy 1.3 gigawatts in 2027 as part of a larger 10-gigawatt agreement by 2029; and for Meta expects to deploy 3 gigawatts through the end of 2028 under a partnership involving MTIA XPUs, with an initial 1-gigawatt order set to begin delivery in the second half of 2027.

For the other two customers, shipments are expected to begin late 2026 and accelerate into 2027, with purchase orders totalling $6 billion received to date.

The backlog disclosures put a number on visibility. At Q4 FY2025 the order backlog for AI switches exceeded $10 billion, Broadcom had also secured record orders on DSPs, optical components like lasers, and PCI Express switches, and these components combined with XPUs brought total order on hand in excess of $73 billion, almost half Broadcom's consolidated backlog of $162 billion, with the $73 billion AI backlog expected to be delivered over the next eighteen months. Two quarters later the demand-supply gap had widened dramatically: "During the quarter, bookings for AI semiconductors were over $30 billion against the $10.8 billion we shipped."

On the software side, revenue predictability comes from subscriptions and total contract value. Bookings continued to be strong in Q1 FY2026, with total contract value booked exceeding $9.2 billion and annual recurring revenue growth of 19% year on year. One accounting nuance worth knowing, because it makes software revenue lumpier than pure ARR would suggest: for infrastructure software, whether or not a customer has the right to terminate causes variations in revenue recognised in each period.

Against all of this, the company is candid about the volatility concentration creates. "From time to time, some of our key semiconductor customers place large orders or delay orders, causing our quarterly net revenue to fluctuate significantly. This is particularly true of our products used in AI and wireless applications as fluctuations may be magnified by the timing of customer deployments, as well as product launches."

Is the concentration a risk or a reflection of quality? Both, and the honest answer is that the ratio has shifted. When the top-five was 40% spread across Apple, Google and a set of OEMs, it read as quality. With one customer at 32% of revenue and 44% of receivables, and that same customer publicly dual-sourcing, it reads as risk. Section 8 treats it as such.

5. Competitive landscape

5.1 The structure of the industry

Broadcom sits in a peculiar position: it is the dominant supplier in a market whose entire premise is that its customers want independence from the dominant supplier.

The AI compute market has bifurcated. Nvidia sells a vertically integrated stack - GPU, interconnect, networking, software - to everyone. Broadcom sells the ability to not buy that stack, to a small number of companies large enough to justify designing their own silicon. This creates a competitive dynamic in which Broadcom's custom chips compete with Nvidia for hyperscaler wallet share, while Nvidia's CUDA ecosystem and broad customer base provide a moat that custom chips cannot replicate for the thousands of smaller AI companies that lack the scale to justify custom silicon.

Within the custom-silicon layer, the structure is a duopoly with a widening gap. Broadcom and Marvell together control an estimated 95% of the custom AI ASIC co-design market , but Broadcom holds roughly 70% share of the custom AI ASIC market and runs multi-billion-dollar hyperscaler programmes , while Marvell commands 20%-25% share anchored by AWS Trainium and Microsoft Maia design wins, positioned as the secondary option hyperscalers choose when they want a dual-source relationship - a structural incentive to keep Marvell in the game - and Marvell's own earnings have explicitly flagged the competitive risk that customers develop fully in-house design capabilities.

The important structural insight is that the hyperscalers are not really competing with Broadcom, even though they design their own chips. The hyperscalers aren't building these chips in isolation: they define the product requirements and own the output, but the actual silicon co-design work - architecture, IP, packaging, interconnect - runs through Broadcom and Marvell.

Every hyperscaler is "building their own AI chips." Most of them aren't actually building anything alone.

5.2 Named competitors

5.3 Barriers to entry - how high, honestly

Very high in XPU co-design, and rising. A new entrant needs: multi-generation-proven SerDes at 200G-plus per lane; HBM PHY and controller IP; production-qualified 2.5D/3.5D packaging flows; a foundry relationship strong enough to secure leading-edge capacity in a shortage; HBM supply relationships; and a delivery record credible enough that a customer will bet a gigawatt of capex on it. Tape-out cost alone at 2nm runs to hundreds of millions of dollars, and a failed spin costs the customer a year. The oligopolistic structure is driven by high R&D costs and the need for cutting-edge semiconductor manufacturing capabilities that create substantial barriers to entry. The only credible new entrants are companies that already have most of these pieces - MediaTek, Alchip, GUC - which is exactly who is entering.

High but eroding in data-centre switching. Bandwidth leadership is a real barrier, but Nvidia has demonstrated that a well-capitalised vertically integrated player can take share fast by changing the unit of purchase from a switch to an AI factory. IDC explicitly identifies competitive responses from Cisco, Arista and Broadcom-ecosystem vendors to Nvidia's leading position as one of three primary risks for the market in 2026.

Very high in FBAR. Proprietary process, own fab, decades of yield learning, single dominant customer with a validated design-in.

High in infrastructure software, but of a different kind. The barrier is not technical; it is the migration cost sitting on the customer's side of the table.

5.4 Where Broadcom is strong and where it is exposed

Strong: custom accelerator co-design at the top of the market, where it has roughly 70% share and six of the most important customers in AI; networking bandwidth leadership by roughly a generation; the optical and physical-layer components that everyone needs regardless of whose accelerator wins; an installed software base that funds everything; and now a unique capability to arrange the capital that finances deployment of its own products.

Exposed: on the single largest customer, which is actively dual-sourcing; on gross margin, because selling a system-in-package including HBM mathematically dilutes percentage margins as volume grows; on the unit of competition, because Nvidia is selling systems while Broadcom insists it is "in the chip business only"; on the creditworthiness of frontier labs that are now a large share of its forward order book; and on the absence of any second growth engine if AI capex decelerates.

There is no point pretending this is a fortress with a single moat. It is a company with a very strong position in a market that is being contested by the best-capitalised technology company in the world, while its own biggest customer looks for alternatives. The position is excellent; it is not safe.

6. Industry

6.1 What drives demand

Demand for Broadcom's growth products is a function of one number: how many gigawatts of AI compute the world's largest technology companies and best-funded AI labs decide to build. Gigawatts, not units - the industry has adopted power as its unit of account because power, land and grid interconnection are the binding constraints.

The demand chain runs: model capability improves → inference usage grows → token economics determine the cost of serving that usage → the cheapest silicon per token wins the marginal deployment → deployment requires power and networking. Broadcom is positioned at two nodes of that chain (the accelerator and the network) and now at a third (the financing).

The economic driver behind custom silicon specifically is the cost-per-token argument. If a purpose-built accelerator serves a known workload at materially lower cost per token than a general-purpose GPU, the largest workload owners will build one. Per-chip, Ironwood's 4,614 TFLOPS sits close to Blackwell's approximately 5,000 FP8 TFLOPS, but SemiAnalysis estimates that TPUs achieve higher sustained model FLOP utilisation of roughly 90% for transformers versus 70% to 80% for GPUs. Utilisation, not peak FLOPS, is where custom silicon wins.

On the networking side, the driver is cluster size. As clusters grow from tens of thousands to hundreds of thousands to a million accelerators, and then span multiple buildings, the network stops being plumbing and becomes the performance-limiting component. Broadcom's Ram Velaga argues Ethernet is open, resilient and economical, making it the ideal technology for scaling AI clusters across data centres, unifying operational tools to enable scale-up, scale-out and scale-across architectures.

For the non-AI franchises, demand drivers are entirely different and currently subdued: smartphone unit volumes and RF content per phone, telco and cable operator capex cycles for broadband, and enterprise server refresh for storage connectivity.

For software, the driver is enterprise IT budget and, increasingly, the desire to run AI workloads on-premises for data-sovereignty and cost reasons.

6.2 Industry size and growth

The custom-ASIC share of AI compute is inflecting. Custom ASIC server shipments are projected to reach 27.8% of the total AI server market in 2026, the highest share since 2023 according to TrendForce, growing 44.6% year over year, nearly three times the 16.1% growth rate projected for GPU-based AI servers in the same period. Longer term, global shipments of AI server compute ASICs are projected to triple by 2027 compared with 2024 levels, on track to surpass traditional GPU shipments by 2028, with Counterpoint forecasting that top AI hyperscalers will deploy more than 40 million custom chip units cumulatively during 2024-2028 in what the firm calls the custom XPU era.

Nvidia remains the incumbent by a wide margin: Nvidia still holds approximately 70% of the AI chip market, but that share is projected to erode as Google, Amazon, Meta, Microsoft and OpenAI invest billions in purpose-built chips.

Networking is growing faster than most people appreciate. The broader Ethernet switch market grew 39.8% year over year to $15.4 billion in Q1 CY2026, with data centre switching accounting for $10 billion of that total, up 61%.

Within the data centre segment, 800 Gb/s switches alone accounted for 35.8% of revenues , and 200G and 400G switches accounted for a further 34.1%, meaning these speeds make up 70% of global data centre Ethernet revenue.

Downstream, the switch-silicon market itself is concentrated: the top five players - Broadcom, Marvell, Nvidia, Intel and Cisco - collectively held over 65% market share as of 2025.

6.3 Where Broadcom sits in the supply chain

Broadcom occupies the design-and-integration layer between the workload owner and the manufacturing base. Upstream it depends on TSMC for leading-edge wafers, on advanced-packaging capacity (TSMC's CoWoS-class offerings plus Amkor and ASE), and on SK hynix, Samsung and Micron for HBM. Downstream, Broadcom's silicon goes to ODMs (Quanta, Celestica, Wistron, Wiwynn, Jabil, Foxconn) who build boards, racks and systems, and to the hyperscalers who operate them.

This position is enviable in one respect and vulnerable in another. Enviable because Broadcom captures margin on a large bill of materials without owning fabs or building data centres - hence the very low capex intensity. Vulnerable because behind Broadcom's growth lies a critical challenge: securing sufficient manufacturing capacity at TSMC. Broadcom does not control the two scarcest inputs in the industry.

There is also a subtler point about value capture. Under the model where Broadcom builds the TPU and Google builds the complete system, Broadcom secures higher profits in the silicon component while Google captures the system economics.

As the TPU's co-designer, Broadcom earns a high margin on the silicon which is the largest component of system bill of materials, while still leaving room for Google to earn very good margins. Broadcom deliberately does not go up the stack - "we are in the chip business only" - which caps its share of the total AI infrastructure dollar but keeps its capital intensity near zero.

6.4 Regulation and policy

Three policy vectors matter.

Export controls. US restrictions on advanced AI compute exports shape which customers Broadcom can serve and where products can ship. China including Hong Kong was 17% of FY2025 net revenue by delivery location, down from 20% in FY2024 . Management has been notably unwilling to reassure on this: asked about prospective impacts of changing export controls on AI product shipments, Hock Tan said "Nobody can give anybody comfort in this environment."

Onshoring policy. US industrial policy is actively reshaping where Broadcom manufactures. Apple has been working with the federal government and businesses across the nation to help create an end-to-end silicon supply chain in the United States, Broadcom is part of Apple's American Manufacturing Program launched to accelerate US manufacturing, and the new agreement marks Apple's largest AMP commitment to date.

Apple's CEO framed the investment as part of Apple's $600 billion US spending pledge and thanked the President for supporting domestic chip production. Policy is now a revenue driver, not just a compliance cost.

Open standards. There is no licensing regime here, but the Ultra Ethernet Consortium and Open Compute Project function as a quasi-regulatory layer determining interoperability. Broadcom has chosen to lead these bodies rather than resist them, correctly judging that an open Ethernet ecosystem is the only structure in which a merchant-silicon vendor beats a vertically integrated one. The shift toward open Ethernet architectures, with technologies like SONiC driving demand for merchant silicon, is expected to grow the open networking ecosystem at a 28% CAGR.

6.5 Cyclicality

Semiconductors are cyclical; the two halves of Broadcom are currently in opposite phases of different cycles.

The non-AI franchises have been in a downcycle. In Q3 FY2025 non-AI demand continued to be slow to recover, with revenue flat sequentially.

Non-AI semiconductor revenue fell 9% sequentially in Q1 FY2025 with mixed recoveries across broadband, server storage, enterprise networking, wireless and industrial. Recovery has been gradual and partial.

The AI business is not a classic semiconductor cycle at all - it is a capital-project cycle, driven by a small number of very large budget decisions. That makes it less sensitive to consumer demand and far more sensitive to sentiment among a handful of boards and their financiers. It is also lumpy in a way that ordinary chip demand is not, because a single gigawatt commitment can shift a quarter.

Software is the least cyclical piece: subscription revenue from customers who cannot easily leave.

6.6 Industry tailwinds and headwinds

Tailwinds: custom silicon taking share from merchant GPUs at nearly three times the growth rate; Ethernet displacing proprietary interconnect in AI fabrics; scale-across architectures creating an entirely new product category (distributed AI routers); inference workloads growing faster than training and being more cost-sensitive, which favours custom silicon; sovereign AI programmes adding a new buyer class; and private credit's willingness to fund AI infrastructure as an institutional asset class.

Headwinds: power, land and grid interconnection as hard physical limits on deployment; CoWoS-class packaging and HBM as persistent bottlenecks; the concentration of global AI capex in a small number of balance sheets, several of which are pre-profit; memory price inflation raising system costs; and the possibility that model-architecture change strands silicon designed for today's transformer workloads. IDC also flags macro uncertainty including tariff risk and regional economic volatility as a watch item that could temper investment.

7. Growth triggers

Every item below is drawn from one of the six concalls or from a company announcement discussed on those calls.

Custom accelerator programmes and customer additions

• Google multi-generation TPU long-term agreement signed April 2026, covering multiple TPU generations and AI networking (Q2 FY26 concall, 3 June 2026). Tan said Broadcom entered into a long-term agreement with Google in April to develop and supply multiple generations of TPUs and AI networking.

• Anthropic: more than 1 GW of TPU-based compute in 2026, plus a further 5 GW beginning 2027 (Q2 FY26 concall, 3 June 2026).

"In April, we entered into an agreement to enable Anthropic to access another five gigawatts of next-generation TPU-based compute beginning in 2027."

• OpenAI accelerator into production late 2026, with a contractual 1.3 GW deployment in 2027 inside the 10 GW / 2029 agreement (Q2 FY26 concall, 3 June 2026; the underlying 10 GW deal announced 13 October 2025). "For OpenAI, we have delivered silicon, and we are on track for production late 2026... We have a contractual commitment to deploy 1.3 GW in 2027 as part of the larger 10 GW that by 2029 agreement we announced last year."

The announcement specified Broadcom deploying racks of AI accelerator and network systems targeted to start in the second half of 2026 and complete by end of 2029.

• Meta MTIA: 3 GW through end-2028, with an initial 1 GW order beginning delivery in 2H 2027 (Q2 FY26 concall, 3 June 2026). "For Meta, in April, we announced a partnership to deliver multiple generations of MTIA XPUs. Under this agreement, we expect to deploy 3 GW through the end of 2028. The initial order for 1 GW, which includes XPUs and our networking, has been received and will start delivery in the second half of 2027." Repeated theme: the Meta roadmap was defended on the prior call too - "Meta's custom accelerator MTIA roadmap is alive and well" (Q1 FY26 concall, 4 March 2026).

• Two further XPU customers begin shipping late 2026 and accelerate into 2027, with $6 billion of purchase orders already received (Q2 FY26 concall, 3 June 2026). "For our other two customers, we expect shipments to begin late 2026 and accelerate into 2027. To date, we have received purchase orders totaling $6 billion."

• Sixth core XPU customer added, roster expanded from five (Q1 FY26 concall, 4 March 2026). Broadcom's key strategic message was that its custom AI accelerator ramp is progressing across a growing customer base, expanding from five named customers to a sixth, with management highlighting ongoing work on Google's TPU roadmap including the seventh-generation TPU referenced for FY2026, and positioning subsequent generations as a driver into FY2027 and beyond.

• Fifth XPU customer secured via a $1 billion order for late-2026 delivery (Q4 FY25 concall, 11 December 2025).

"I am pleased today to report that during this quarter, we acquired a fifth XPU customer through a $1 billion order placed for delivery in late 2026."

• Fourth XPU customer qualified with over $10 billion of orders, mainly for 2H FY2026 delivery (Q3 FY25 concall, 4 September 2025). "Last quarter, one of these prospects released production orders to Broadcom. And we have accordingly characterized them as a qualified customer for XPUs. And in fact, has secured over $10 billion of orders of AI racks based on our XPUs. And reflecting this, we now expect the outlook for fiscal 2026 AI revenue to improve significantly from what we had indicated last quarter."

Delivery was specified as mainly in the second half of fiscal 2026.

• Two additional hyperscalers selected Broadcom for custom accelerators, with tape-outs scheduled (Q1 FY25 concall, 6 March 2025). "It is, therefore, no surprise to us since our last earnings call that two additional hyperscalers have selected Broadcom to develop custom accelerators to train their next-generation frontier models... there are now four more who are deeply engaged with us to create their own accelerators," with Broadcom "on track to tape out their XPUs this year."

Networking product ramps

• Tomahawk 6 (102.4 Tbps) sampling then production. Guided at Q1 FY25 - "we have tapped out our next generation 100 terabit Tomahawk 6 switch running 200G SerDes at 1.6 terabit bandwidth. We will be delivering samples to customers within the next few months" (Q1 FY25 concall, 6 March 2025) - and by Q4 FY25 the AI switch order backlog exceeded $10 billion with Tomahawk 6 "the first and only one of its capability out there" continuing to book at record rates (Q4 FY25 concall, 11 December 2025). Repeated across at least three of the six calls.

• Thor Ultra 800G AI Ethernet NIC introduced October 2025, complementing Tomahawk 6 for clusters scaling beyond 100,000 XPUs (announced 14 October 2025; referenced in Q4 FY25 commentary on record component orders).

• Jericho4 scale-across fabric shipping, designed to interconnect over one million XPUs across multiple data centres (shipping announced August 2025; positioned in Q2 FY26 concall, 3 June 2026 as "we remain the industry leader with our Jericho 3 and Jericho 4 fabric solutions... enabling the world's largest deployments of multiple hyperscalers" ).

• Scale-up copper and co-packaged optics ramp within racks (Q2 FY26 concall, 3 June 2026). "For scale up within racks, we enable direct attached copper based on our industry-leading 200G and 400G SerDes, driving co-packaged copper with Ethernet and PCI Express switches."

• 1 GW-plus of Anthropic capacity deploying in Fluidstack-based sites from mid-2026, funded through the new platform (announced 9 June 2026, following the Q2 FY26 call).

The AI XPV Platform

• AI XPV Platform launched with Apollo and Blackstone: 20 GW-plus of compute capacity through 2028, first tranche $35 billion (previewed on Q2 FY26 concall, 3 June 2026; formally announced 9 June 2026). The Platform is designed to enable more than 20 gigawatts in compute capacity using Broadcom's XPUs and networking solutions customised for leading frontier AI labs including Anthropic and OpenAI through 2028, launching with an initial tranche of $35 billion led by Apollo in partnership with Blackstone.

"We are at a historic inflection point where the demand for AI compute is fundamentally reshaping the global economic landscape." - Hock Tan

Revenue trajectory guidance

• FY2026 AI semiconductor revenue guided to $56 billion; Q3 FY2026 to $16 billion; second half FY2026 to double the first half (Q2 FY26 concall, 3 June 2026). "In the second half of 2026, we expect AI semiconductor revenue to double from the first half we shipped this year. Consistent with this trend, in Q3, we expect AI semiconductor revenue to accelerate to $16 billion, up over 200% year-on-year. For the full year 2026, we expect to achieve AI semiconductor revenue of $56 billion, up approximately 180% from fiscal 2025."

• FY2027 AI semiconductor revenue reiterated in excess of $100 billion, with growth continuing into FY2028 (Q2 FY26 concall, 3 June 2026). "We expect this momentum to continue into fiscal year 2027 and reiterate our AI semiconductor revenue guidance to be in excess of $100 billion... We expect AI semiconductor revenue growth to continue in fiscal 2028, based on the following initiatives with our six core customers." Repeated: the FY2027 target evolved from a "$60 billion to $90 billion" serviceable addressable market from three hyperscale customers in FY2027 (Q1 FY25 concall, 6 March 2025) to a firm revenue guide above $100 billion by Q2 FY26.

• 2028 expected to grow substantially over 2027 on gigawatt demand from the six-customer roster (Q2 FY26 concall, 3 June 2026). "If you ask about 2027 or 2028, that will continue to grow. We expect, in fact, 2028 to be a substantial growth from what we are forecasting in 2027."

Software expansion

• VCF 9.1 released with heterogeneous compute support across AMD, Intel and Nvidia platforms (Q2 FY26 concall, 3 June 2026). "This release adds heterogeneous compute support across GPUs and CPU architectures, including AMD, Intel, and NVIDIA platforms, enabling enterprise cloud customers to run AI, Kubernetes, and traditional virtualized workloads on a common private cloud environment."

• VCF go-to-market extension from the top 10,000 accounts into the next 20,000-30,000 midsized companies (Q1 FY26 concall, March 2026). Tan noted the largest 10,000 companies are finding value in VCF and added: "We are now looking at whether the next 20,000, 30,000 midsized companies see it the same way. Stay tuned."

Non-AI and wireless

• Non-AI semiconductor revenue guided to grow again, to approximately $4.5 billion in Q3 FY2026, up 12% year on year (Q2 FY26 concall, 3 June 2026) - the first clear inflection after several quarters of flat-to-down commentary.

• Apple multi-year agreement exceeding $30 billion and Fort Collins $1.5 billion expansion, running through at least 2031 (announced 8 July 2026, after the Q2 FY26 call). More than 15 billion U.S.-made chips, with Fort Collins as the production hub for custom ASIC silicon, RF components including FBAR filters, and next-generation wireless connectivity through at least 2031.

Trigger summary

8. Key risks

8.1 One customer is roughly a third of revenue, and that customer is shopping

Mechanism. One customer accounted for 32% of net revenue in FY2025, up from 28% and 21% in the two prior years , and 44% of net accounts receivable at fiscal year-end 2025. If that customer moves a meaningful share of a TPU generation to a second supplier, Broadcom loses revenue at a rate no other business line can offset, and it loses it in a step function rather than gradually, because a generation is either won or not.

The mechanism is already visible. TPU v7 is a dual-chiplet design co-developed with Broadcom and MediaTek.

On the June 2026 call Hock Tan formally acknowledged the possibility of Google diversifying its TPU supply, and that acknowledgment was the most significant near-term driver of the stock decline following the call. The strategic logic on Google's side is explicit: moving toward a model in which the silicon partner charges only for what it adds value on, with Google paying closer to bill-of-materials cost.

Calibration. High probability of gradual share loss; low probability of abrupt loss. Broadcom's own framing is that absolute dollars keep growing even as share slips, and one read of the situation is that Google-related market share competition risk exists but the possibility of the absolute revenue base being impaired is limited. That is probably right for the next two years and unknowable beyond.

8.2 The revenue mix is structurally diluting gross margin

Mechanism. Broadcom bills XPUs at the system-in-package level, which appears to include HBM procured and repackaged by Broadcom. HBM is expensive and carries thin margin as a pass-through. As XPU revenue scales, low-margin content becomes a larger share of the mix and percentage gross margin falls mechanically, even if gross profit dollars rise strongly.

The numbers show it happening. Gross margin was 77.1% of revenue in Q2 FY2026, down 230 basis points year on year, as semiconductor became a larger proportion of the product mix , and consolidated gross margin was guided to decline to approximately 74% in Q3 FY2026 as AI semiconductors become a larger portion of total revenue.

The company faces pressure on semiconductor margins due to the lower margins of ASICs and TPUs, with the rapid growth in AI semiconductor revenue causing a decline in consolidated gross margin. This was flagged as far back as Q2 FY2025: CFO Kirsten Spears noted that "XPU margins are slightly lower than the rest of the business other than Wireless."

Management's counter is that operating leverage offsets it: "Q2 operating margin was a record 67%, and adjusted EBITDA was a record 69% of revenue, which was above our guidance. Even as our revenue scales up massively, driven by AI, our operating and EBITDA margins remain strong and stable."

Calibration. High probability, moderate drag. This is arithmetic, not a surprise. The risk is not the margin itself but that investors valuing the company on a software-like margin structure re-rate it as the mix shifts. Broadcom's shares fell more than 12% the session after Q2 FY2026 results as investors focused on declining gross margins of approximately 74%.

8.3 Customers can insource the part Broadcom sells

Mechanism. Broadcom's value-add is physical design, IP and supply orchestration. A hyperscaler that hires enough physical-design engineers can shift to a customer-owned-tooling model, where it pays a partner only for manufacturing and licenses IP à la carte. Broadcom's transcripts define the term explicitly, which tells you it comes up in investor conversations: "COT (Customer Owned Tooling): A business model in which customers design their own chips, leveraging a semiconductor partner only for manufacturing." Marvell has publicly acknowledged the same risk in its own filings.

Calibration. Low probability in the next three years for the frontier labs (they are bandwidth-constrained on engineering talent and want to spend it on models, not on timing closure), materially higher for Google and Amazon, which have had silicon organisations for a decade.

8.4 Broadcom's forward order book now depends on frontier-lab creditworthiness

Mechanism. A growing share of committed gigawatts comes from Anthropic and OpenAI - companies with enormous compute appetites and no comparable operating cash flow. Broadcom has addressed the funding gap by building the AI XPV Platform, but that solves the funding problem, not the credit problem: it transfers the risk to private-credit investors whose willingness to keep drawing depends on their view of the labs' prospects. This is asset-backed financing: chips, networking, power and long-term customer commitments bundled into one vehicle. If frontier-lab funding conditions tighten, the draw schedule slows, and Broadcom's backlog stops converting on schedule.

Tan's own language shows how quickly the gigawatt commitments have been escalating, which cuts both ways: "We are seeing that particularly to the point where for even two of our customers, we're talking about, which is Anthropic and OpenAI... We're talking about capacity as measured by gigawatt power that are way ahead of what we have expected, say, six months ago." Forecasts that move that far in six months can move in both directions.

Calibration. Moderate probability, high severity. This is the newest and least-tested risk in the business, and it is the one least covered by Broadcom's historical track record.

8.5 Supply is secured only through 2027

Mechanism. Wafers, advanced packaging and HBM are all allocated well in advance. Broadcom's commitments to customers now run to 2028 and 2029, but its own supply is not yet locked that far. "We are comfortable with our supply for 2026 and 2027 and are working on securing supply for 2028 and 2029." That is an honest answer, and it is also an admission of an open exposure. Separately, the $30 billion of AI bookings against $10.8 billion shipped in Q2 FY2026 indicates potential supply-chain strain.

Calibration. Moderate probability of episodic constraint, low probability of a structural shortfall given Broadcom's leverage with TSMC. Watch inventory: inventory rose to $4.33 billion as the company secured supply to support AI demand, with days of inventory increasing to 86 from 68 in the prior quarter. Building inventory to secure supply is prudent when demand is real and expensive when it is not.

8.6 The guidance bar has been raised so high that meeting it is no longer enough

Mechanism. Broadcom now guides multiple years out with specific dollar figures. When it does not raise them, the market reads a miss. This happened in June 2026: Broadcom's stock was down about 15% the day after Q2 FY2026 results on disappointment that Tan didn't raise the full-year target of $100 billion in AI chip sales. The internal arithmetic also drew scrutiny: "On this fiscal year, AI sort of 2x growth, second half over first half. That would put AI revenues over $60 billion with sequential growth in fiscal Q4. You gave us this $56 billion number, which is only 1.5x, half over half growth with Q4 AI actually being down sequentially. If you could just help us square the numbers there."

One analyst characterisation was that the forward AI guide, while a beat on official consensus, landed below what the most aggressive institutional models had built in - and that there is very little tolerance for any dimension of a miss.

Calibration. High probability of recurrence, moderate severity. This is an expectations risk rather than a business risk, but for a company whose cost of capital and acquisition currency is its own equity, expectations risk is business risk.

8.7 Non-AI semiconductors provide no downside cushion

Mechanism. If AI capex decelerates, there is no second engine. The non-AI semiconductor business is essentially flat, and if hyperscaler AI capex pulled back hard there is no comparable growth engine sitting underneath it; the semiconductor segment would lose its primary growth engine with nothing comparable to replace it. The Apple AMP agreement improves this materially on a multi-year view but does not change the near-term arithmetic.

Calibration. High-impact if AI capex slows, but conditional on that trigger.

8.8 The VMware franchise is being harvested, not grown

Mechanism. Growth in infrastructure software has come from repricing an installed base, not from winning new customers. Once conversion of the top 10,000 accounts is complete, the growth rate reverts to whatever the underlying base does. Q1 FY2026 infrastructure software revenue was up just 1% year on year , which is what that reversion looks like. By Q1 FY2025 Broadcom had owned VMware for over a year, price increases and subscription conversion were already underway, so this was the first clean apples-to-apples read on VMware under the new model - and it grew 1%. Meanwhile, the customer relationship has costs: changes including an invite-only channel programme and the termination of perpetual licences caused price rises of up to 600% for some customers. Customers whose renewals are painful evaluate alternatives, and the midmarket push into 20,000-30,000 smaller companies is aimed at buyers who are far more price-elastic and far less locked in than a Fortune 500 IT organisation.

Calibration. High probability of a low-growth software segment; moderate probability of visible churn. The segment's role as the margin and cash anchor is what is at stake.

8.9 Key-person and finance-leadership transition

Mechanism. Broadcom is unusually dependent on one executive's judgement about which businesses to buy, which to cut, and which customers to serve. "The board and I have agreed that I will continue as the CEO of Broadcom through 2030 at least. These are exciting times for Broadcom, and I'm very enthusiastic to continue to drive value for our shareholders."

The board granted Tan a PSU award in September 2025 to extend his leadership through fiscal 2030, and states that it continues to focus on and be actively engaged in CEO succession planning. The award is aggressively AI-linked: the contract extension brings hundreds of millions of dollars in stock compensation if the company reaches $120 billion in AI product sales by 2030, with 610,521 shares due if AI revenue hits $90 billion by fiscal 2030 and 300% of the payout at $120 billion. That alignment cuts both ways: it locks in the CEO and it creates an incentive to chase an AI revenue number.

The CFO seat also just turned over. On the Q2 FY2026 call Broadcom announced that Kirsten Spears would retire on 12 June 2026 after 12 years, introducing incoming CFO Amie Thuener.

Calibration. Low probability, high severity for the CEO risk. The CFO transition is routine but arrives at the most complex financial moment in the company's history, with a $35 billion off-balance-sheet-adjacent financing platform being stood up.

8.10 Balance sheet and geopolitical exposure

Broadcom carries substantial acquisition debt: the weighted average coupon and years to maturity of gross principal fixed-rate debt of $67.1 billion were 4% and 7.2 years as of Q4 FY2025. At current cash generation this is comfortably serviced, but it constrains flexibility if the AI cycle turns while a large acquisition is in flight.

On geopolitics, 17% of FY2025 net revenue was delivered to China including Hong Kong , and management has declined to offer reassurance on export-control direction. Tariffs and export rules are a live variable that Broadcom cannot hedge.

9. Walk the talk

The six concalls used: Q1 FY2025 (6 March 2025), Q2 FY2025 (5 June 2025), Q3 FY2025 (4 September 2025), Q4 FY2025 (11 December 2025), Q1 FY2026 (4 March 2026), Q2 FY2026 (3 June 2026). The most recent is 58 days before the date of this report.

Starting with the oldest call

On 6 March 2025 Hock Tan set two markers. The near-term one was ordinary quarterly guidance: Q2 revenue of $14.9 billion, up 19% year on year, AI revenue of $4.4 billion up 44%, and adjusted EBITDA at approximately 66% of revenue. The long-term one was a framework, deliberately labelled as a market opportunity rather than a forecast:

"These R&D investments are very aligned with the roadmap of our three hyperscale customers as they each raise toward 1 million XPU clusters by the end of 2027. And accordingly, we do reaffirm what we said last quarter that we expect these three hyperscale customers will generate a Serviceable Addressable Market or SAM in the range of $60 billion to $90 billion in fiscal 2027."

He also made a specific product promise on Tomahawk 6 - samples to customers "within the next few months"

• and a specific engineering commitment on two new hyperscaler engagements: "We are on track to tape out their XPUs this year."

Note the careful construction of the SAM. It was framed as a market, from three named customers, and explicitly excluded the pipeline: "And to be clear, of course, these four are not included in our estimated SAM of $60 billion to $90 billion in 2027." That is a man building himself room.

Q2 FY2025: guidance met, and a conspicuous refusal to update the long-term number

The Q2 number came in at $15 billion, up 20% year over year against a $14.9 billion guide, with AI semiconductor revenue over $4.4 billion, up 46% against the $4.4 billion guide, and adjusted EBITDA at 67% of revenue, above the 66% guided. Met or beaten on all three.

For Q3, he guided $15.8 billion consolidated revenue, up 21%, and AI semiconductor revenue of $5.1 billion, up 60%, with EBITDA at or above 66% , and extended the AI growth framing: Broadcom expects sustained annual growth of roughly 60% in its AI semiconductor segment through FY2026.

The revealing moment was what he would not say. Asked directly whether the 60% growth rate would carry into FY2026 in line with the SAM, he engaged with the logic but then declined to reset the 2027 number: Tan provided no update on the 2027 AI revenue opportunity, emphasising that forecasts rest solely on factors and customer activity currently visible to Broadcom , and management stated that near-term growth forecasts do not include potential future contributions from prospects beyond active customers, with updates provided only when revenue conversion is certain.

This is the single most important behavioural pattern in the six calls, and it repeats: Broadcom guides only what is contracted. It does not forecast pipeline.

Q3 FY2025: beat, then a step change

Q3 came in at a record $16 billion, up 22%, with adjusted EBITDA at a record $10.7 billion up 30%, and AI semiconductor revenue of $5.2 billion up 63%

• ahead of both the $15.8 billion and $5.1 billion guides.

Then came the fourth XPU customer with over $10 billion of orders, and with it a rare public upgrade to the forward view: "And reflecting this, we now expect the outlook for fiscal 2026 AI revenue to improve significantly from what we had indicated last quarter." Note the pattern: he did not upgrade until orders were in hand.

Q4 guidance: $6.2 billion AI semiconductor revenue up 66%, infrastructure software of approximately $6.7 billion up 15%, adjusted EBITDA at 67% , on consolidated revenue of $17.4 billion up 24%.

He also used the call to remove the largest overhang on the equity story:

"The board and I have agreed that I will continue as the CEO of Broadcom through 2030 at least. These are exciting times for Broadcom, and I'm very enthusiastic to continue to drive value for our shareholders."

Q4 FY2025: all three guided lines beaten

Total revenue was a record $18.0 billion, up 28% year on year and above guidance on better-than-expected growth in AI semiconductors as well as infrastructure software

• against $17.4 billion guided. AI came in at $6.5 billion, up 74% against $6.2 billion guided. Software came in at $6.9 billion, up 19%, exceeding the previous outlook of $6.7 billion .

Alongside came the disclosures that reframed the company: total order on hand in excess of $73 billion, almost half of a consolidated backlog of $162 billion, with the AI portion expected to be delivered over the next eighteen months , plus a fifth XPU customer, an AI switch backlog above $10 billion, and a 10% dividend increase.

Q1 FY2026 guidance: $19.1 billion consolidated revenue up 28%, semiconductor revenue of approximately $12.3 billion up 50%, AI revenue doubling to $8.2 billion.

Q1 FY2026: beaten again

"In our fiscal Q1 2026, total revenue reached a record $19.3 billion, and that's up 29% year-on-year and exceeding our guidance on the back of better than expected growth in AI semiconductors. This top line strength translated into exceptional profitability with Q1 consolidated adjusted EBITDA hitting a record $13.1 billion, which is 68% of revenue." Guided $19.1 billion and 67% EBITDA; delivered $19.3 billion and 68%. AI came in at $8.4 billion, up 106% , against $8.2 billion guided.

Q2 guidance: AI semiconductor revenue of $10.7 billion, revenue growth of 47% to $22.0 billion, adjusted EBITDA of 68%. Software guided to approximately $7.2 billion, up 9% year on year.

He also used the call to knock down a specific market rumour rather than let it fester: "Meta's custom accelerator MTIA roadmap is alive and well."

Q2 FY2026: the guidance was met, and the market objected anyway

"In our fiscal Q2 2026, total revenue reached a record $22.2 billion, up 48% year-on-year, above our guidance, on strength in AI semiconductors. Q2 operating margin was a record 67% and adjusted EBITDA was a record 69% of revenue, above our guidance." Guided $22.0 billion and 68% EBITDA; delivered $22.19 billion and 69%. AI came in at $10.8 billion against $10.7 billion guided. Software came in at $7.2 billion, up 9%, in line with guidance.

Six for six on the quarterly guide. And the stock fell.

The reason is instructive. Three things happened on that call. He reiterated rather than raised the FY2027 AI figure: "We expect this momentum to continue into fiscal year 2027 and reiterate our AI semiconductor revenue guidance to be in excess of $100 billion." He conceded that Google would likely multi-source. And he guided gross margin down to approximately 74%. Two disclosures pushed the selloff deeper: Tan acknowledged Google would likely draw on multiple chip suppliers, and he warned that surging AI semiconductor sales were weighing on overall gross margins.

He also declined, again, to be pinned down beyond what is contracted. Asked by JPMorgan's Harlan Sur about the FY2027 trajectory, Tan said the company remains "very much on track, if not stronger," but declined to provide a more specific quarterly guide for fiscal 2027.

Promise versus outcome

Assessment

On the one-quarter horizon, this is among the most reliable management teams in large-cap technology. Six consecutive quarters of meeting or modestly beating its own revenue, AI-revenue, software-revenue and EBITDA guidance is not luck; it is a company that guides what it has already booked. The product commitments have also been kept, and in the case of Tomahawk 6 beaten - sampling to volume production in under three quarters on a chip of that complexity is genuinely exceptional execution.

On the multi-year horizon, the pattern has been systematic conservatism. The $60-90 billion FY2027 SAM from three customers, reaffirmed twice in early 2025, has been overtaken by a $100 billion-plus FY2027 revenue guide from six customers. The "roughly 60% AI growth through FY2026" framing from June 2025 became approximately 180% growth. Management repeatedly refused to include unconverted pipeline in forecasts, and pipeline repeatedly converted.

There are two legitimate criticisms.

First, the sandbagging has been so systematic that investors began extrapolating beyond the guidance, and management has done little to discourage it. When the Q2 FY2026 guide implied a sequential Q4 decline in AI revenue and an analyst asked him to square the arithmetic, the answer did not fully resolve the discrepancy. A management team that has trained the market to expect upside cannot then be surprised when a merely-met guide is treated as a disappointment.

Second, and more substantively, the FY2027 number has changed category. In March 2025 it was a "serviceable addressable market" - an opportunity size. By June 2026 it is "our AI semiconductor revenue guidance." That is a promotion from estimate to commitment, and it removes the room Tan has historically been careful to keep. He is now on the record with a specific figure for a fiscal year that depends on frontier-lab funding, TSMC allocation for 2027-2028, and Google's supplier decisions - three things he does not control.

Plain verdict: this is management that does what it says on the horizon it controls, and has consistently under-promised beyond it. The credibility is earned and substantial. The new risk is not that they will miss the next quarter; it is that they have now committed publicly to a number two years out, in a business where the demand-side inputs move by gigawatts in six months.

10. Shareholder friendliness index

Dividends. Broadcom has raised the dividend in each of the last three fiscal years. In FY2023 dividends per share to common stockholders were $18.40, versus $16.40 in FY2022 and $14.40 in FY2021

• equivalent to $1.84 split-adjusted following the ten-for-one forward stock split effected in July 2024 . In FY2024 the quarterly rate was $0.53 per share on a split-adjusted basis paid on 30 September 2024, totalling $2,484 million , and was then increased by 11% to $0.59 per share for FY2025 - a $2.36 annual rate, paid at $0.59 in each of the four FY2025 quarters ( March 2025 , June 2025 , September 2025 and December 2025). For FY2026, "based on increased cash flows in fiscal year 2025, we are increasing our quarterly common stock dividend by 10% to $0.65 per share for fiscal year 2026"
• a $2.60 annual rate, confirmed by declarations in March 2026 and June 2026 . Coverage is comfortable and the payout ratio reveals nothing alarming: the annual dividend runs at about 43% of the past year's earnings per share , and dividend payments in Q2 FY2026 came to $3.1 billion, about 30% of free cash flow.

Buybacks and dilution. Two overlapping authorisations sit in the three-year window, plus recent activity. Historically, boards authorised $10 billion in December 2021 and a further $10 billion in May 2022, and all $20 billion was utilised prior to expiration on 31 December 2023. Then in April 2025 the board authorised a new $10 billion programme through 31 December 2025 , subsequently extended to 31 December 2026, with no shares repurchased in the fiscal quarter ended 2 November 2025 and $7,550 million of the authorised amount still available as of that date

• implying roughly $2.45 billion executed under it during FY2025. Execution then jumped and collapsed within two quarters: "Consistent with our commitment to return excess cash to shareholders, we returned $10.9 billion in the first quarter through $3.1 billion of cash dividends and $7.8 billion of stock repurchases" (Q1 FY2026), followed by $600 million in Q2 FY2026, a 92% reduction in a single quarter . MoatMap's trailing-90-day window (since 2 May 2026) captures three repurchase disclosures dated 9 June 2026 totalling 41 million shares - approximately 16 million shares at ~$153 (the FY2025 activity), 23 million at ~$341 (Q1 FY2026, ~$7.85bn) and 2 million at ~$300 (Q2 FY2026, $600m) - which reconciles to the company's own quarterly figures rather than representing new June-2026 buying. On dilution, share count has crept up rather than down: 468,140,569 shares outstanding as of 24 November 2023 (equivalent to about 4,681 million post-split) versus 4,741,273,799 shares outstanding as of 28 November 2025
• roughly 1.3% net growth over two fiscal years, meaning buybacks have offset but not exceeded equity-compensation issuance.

Verdict: Returns Capital - three consecutive double-digit dividend increases and roughly $10.8 billion of repurchases across the last five reported quarters, though the sharp Q2 FY2026 buyback pullback and the slight net rise in share count mean the return is real but not aggressive.

11. Insider activities

Source basis: the MoatMap cross-market disclosure block (US venue, data current 31 July 2026, covering the trailing 12 months) is the spine. The last ~2 weeks have been cross-checked against SEC Form 4 filings; the most recent Form 4 for Broadcom was accepted on 15 July 2026 , which corresponds to the 10 July transaction already in the MoatMap set. No material filings after 10 July 2026 were located.

Recent material transactions

Buys - reading the signal

There is essentially nothing to read. Over the trailing 12 months the MoatMap set records 1 buy against 148 sells and 13 other, and the single buy is not attributable to a named executive or director of material seniority within the block. A longer-window aggregation of Form 4 history is directionally consistent: across 247 Form 4 filings, Broadcom insiders reported 15 open-market purchases totalling $49.7 million and 232 sales totalling $2.7 billion, a net selling position of $2.6 billion , with one insider buy in the last 90 days.

No cluster buying. No CEO, CFO or board purchase. No open-market buying by any executive during a drawdown in which the stock fell 16%-18% from its high. That last point is the one that carries information: shares fell more than 16% during June 2026, the worst stretch since March 2025, and closed at $377.75 on 1 July, nearly 18% off the highs from a month earlier , and no insider stepped in. That is not a bearish signal on its own - many companies' executives never buy - but it is the absence of a bullish one at precisely the moment it would have been most meaningful.

Sells - working out the why

Henry Samueli (co-founder, Chairman). By far the largest seller. Samueli has offloaded more than $651 million in stock over the past three months and filed to sell another $71 million on 29 June 2026. The reason is not disclosed in the filings located, but two pieces of context are relevant and documented. First, this is founder diversification at scale by a director who has held the position since the 2016 Broadcom Corporation combination. Second, and materially, Samueli operates a pledging arrangement that the board approved as a specific exception to policy: Samueli pledged 16,175,000 shares of Broadcom common stock, which at that time represented approximately 18% of the total shares he beneficially owned and 0.3% of total outstanding shares; the board was advised of the potential risks including the possibility the pledge could result in a forced sale of the pledged shares, and the Nominating and Corporate Governance Committee and board concluded the arrangement does not pose a material risk to Broadcom or its stockholders. The laddered price-tier execution across 11 bands within a single day, combined with the offsetting zero-price conversion rows, is the signature of a pre-arranged programmatic disposition rather than a discretionary decision to exit.

Mark David Brazeal (Chief Legal & Corporate Affairs Officer). The most active insider in the window by transaction count , selling in a highly regular 25,000-share cadence: 25 June, 8 July and 10 July 2026, following a pattern established well before - on 23 June 2025 he sold 50,000 shares at an average $253.285, and continued to hold 415,956 shares directly, including 336,250 restricted stock units. Reason not disclosed in the filing footnotes located; the metronomic size and timing are consistent with a pre-scheduled Rule 10b5-1 plan, though a plan reference was not confirmed in the available filings. Timing was awkward: his 2026 sales exceeded $23 million, and shares subsequently dipped to close near $377.75 by 1 July, just below his sale price. The relevant nuance is that the great majority of his remaining position is unvested RSUs, so his economic exposure is compensatory rather than discretionary.

Hock Tan (CEO). A 300,000-share sale worth roughly $101.3 million earlier in 2026. Reason not disclosed. Context that matters: the board granted Tan a PSU award in September 2025 to extend his leadership through fiscal 2030, and information about his post-vesting holding requirement is disclosed in the proxy

• meaning a substantial portion of his forward economic interest is locked to AI revenue outcomes in FY2028-FY2030 regardless of what he sells today.

Gayla Delly and Justine Page (Directors). Small sales of 1,890 and 1,602 shares respectively, in each case immediately following non-market "other" rows of identical or related size. This is standard director vest-and-sell behaviour on annual equity awards. Not a signal.

Net assessment

Insiders are decisively net sellers, and have been for the entire trailing 12 months: 148 sells against 1 buy in the MoatMap window. The activity is highly concentrated - Samueli accounts for the overwhelming majority of dollar value, and Brazeal for the majority of transaction count. Nothing about the pattern is new; Broadcom insiders have been persistent net sellers for years, and virtually all of the stock being sold was created by equity compensation or by the 2016 merger rather than purchased in the market, which makes selling the mechanical default rather than a statement.

Two observations temper the benign reading. First, the June 2026 concentration is notable: Samueli's largest tranche landed on 17 June, two weeks after a call that knocked roughly 15% off the stock, and Brazeal sold three times in the following month. Second, and more telling, nobody bought. In a 16%-18% drawdown following a quarter in which management met every guided line, not one director or officer stepped into the market. A CEO or CFO purchase in that window would have been an unambiguous statement; its absence is a quieter one.

Plain-language read: mild concern. Not a red flag - the pattern is long-established, the largest seller has a documented pledging arrangement and a founder's diversification rationale, and the CEO's forward compensation is locked to 2030 AI targets. But sustained one-directional selling by the Chairman and the most active officer, combined with zero conviction buying during a material drawdown, is not the insider profile of a management team that believes the stock is mispriced downward.

12. Scenarios

Bull case

The gigawatt commitments convert on schedule, and the schedule turns out to have been conservative. OpenAI's accelerator reaches production in late 2026 as promised, the 1.3 gigawatts contracted for 2027 lands, and the 10-gigawatt agreement pulls forward rather than slipping. Meta's initial MTIA gigawatt begins delivering in the second half of 2027 and the follow-on generations arrive on the multi-year cadence Broadcom has signed for. Anthropic's five gigawatts from 2027 ramps into the Fluidstack sites funded by the XPV Platform, and Apollo and Blackstone, having seen the first tranche perform, come back for the second and third. The two unnamed customers turn their $6 billion of purchase orders into repeat programmes. The six-customer roster becomes eight.

In this world Broadcom's networking franchise turns out to be worth more than the accelerator franchise. Because scale-across is a genuinely new product category, Jericho4 sells into every multi-building deployment regardless of whose silicon is doing the arithmetic, including Nvidia's. Tomahawk 6 holds its generation lead through the 200G SerDes cycle and Broadcom converts co-packaged optics from a technology demonstration into a volume standard. The attach rate on every gigawatt rises, because power constraints force operators to spread compute across sites and every additional site multiplies the interconnect content.

Google's diversification to MediaTek happens but proves to be a share loss inside a much larger pie, so Broadcom's absolute TPU dollars keep rising even as its percentage of the programme falls. Meanwhile the parts of the company nobody has been paying attention to quietly reassert themselves: the Apple agreement locks in the wireless franchise through 2031 with $1.5 billion of Apple-underwritten capacity at Fort Collins, non-AI semiconductors resume growth as enterprise IT spending recovers, and VCF 9.1's heterogeneous compute support turns the private-cloud franchise into the default place enterprises run on-premises AI, opening the 20,000-30,000 midmarket accounts Tan has been probing.

The valuation-relevant part of this story is not the revenue. It is that operating margin holds near 67% while gross margin falls, proving the operating-leverage argument, and that free cash flow conversion stays in the mid-forties as a percentage of revenue on capital expenditure of a few hundred million dollars a quarter. Broadcom ends up looking like a royalty on global AI infrastructure with almost no capital intensity - and with the AI XPV Platform, a company that has learned to manufacture its own demand.

Base case

Broadcom does roughly what it has said. FY2026 AI semiconductor revenue arrives near the $56 billion guided, with the second half roughly double the first. FY2027 comes in above $100 billion but not dramatically above, because the constraint shifts from demand to supply and to the pace at which frontier labs can actually energise data centres. Backlog conversion stays front-loaded, as it has been, and the eighteen-month delivery windows management quotes keep compressing to twelve.

Google multi-sources more of each TPU generation to MediaTek, and Broadcom's share of that programme declines gradually. The absolute dollars still grow, but the concentration disclosure improves for the right reason and the wrong one simultaneously: the top customer's percentage falls because the others grow, not because Google grows. Meta, OpenAI and Anthropic together become a larger block than Google by FY2028.

Gross margin drifts down through the mid-seventies and then the low seventies as HBM-inclusive system-in-package revenue dominates the mix, while operating margin holds because operating expenses do not scale with pass-through content. The market spends most of the period arguing about which of those two margins is the right one to capitalise, and the stock is volatile around results as a consequence, exactly as it was in June 2026.

Infrastructure software grows in the high single digits to low double digits - enough to keep funding the dividend and the R&D budget, not enough to matter to the growth narrative. The midmarket VCF push produces something, but not transformation. Non-AI semiconductors recover to modest growth, helped by the Apple agreement. The dividend rises roughly 10% a year. Buybacks are opportunistic and lumpy, sized to whatever free cash flow is left after dividends and whatever acquisition Hock Tan is contemplating.

Nothing breaks. Nothing dramatically exceeds. Broadcom remains the most important company in AI infrastructure that does not make a GPU, and it remains dependent on other people's capital-allocation decisions in a way it cannot hedge.

Bear case

The bear case does not start with AI demand collapsing. It starts with the financing.

Frontier-lab funding conditions tighten - a disappointing model release, a shift in private-credit risk appetite, a repricing of data-centre asset-backed paper. The XPV Platform's multi-year draw schedule slows. Anthropic's five gigawatts from 2027 becomes three, then two, and the timing slides right. OpenAI's 1.3-gigawatt 2027 commitment is honoured but the path to ten gigawatts by 2029 quietly stops being discussed on calls. Because Broadcom has committed publicly to an FY2027 revenue figure above $100 billion, the slippage is not absorbed quietly; it becomes a guidance cut, and the guidance cut is the first in the Hock Tan era.

Simultaneously, the Google relationship deteriorates faster than expected. MediaTek's cost-optimised approach wins a majority of a TPU generation rather than a minority, and Google - having built the internal capability across seven generations - moves to a customer-owned-tooling model where it licenses Broadcom's SerDes and HBM IP and pays for manufacturing rather than for co-design. A third of Broadcom's revenue reprices from a co-design margin to an IP-licensing margin. Amazon and Meta, watching, ask for the same terms.

The margin story then compounds in the wrong direction. Gross margin has already fallen from the high seventies toward the low seventies on mix. If revenue growth decelerates while the mix damage persists, the operating-leverage defence stops working, because operating expenses were built for a $100 billion AI business. Meanwhile there is no cushion underneath: the non-AI semiconductor franchises are flat, and infrastructure software - having grown 1% in a clean comparison quarter - is a monetisation story with its growth phase behind it. Enterprises who accepted 600% price increases because migration was too painful discover, three renewal cycles later, that Proxmox and OpenShift have improved and that the migration is now merely expensive rather than impossible. Churn appears in the software base at exactly the moment its cash flow is most needed.

Layer on the structural threats. Nvidia's shift to selling AI factories rather than switches - already visible in its move from under 4% to 21.5% of data-centre Ethernet switching in two years - continues, and Broadcom's insistence that it is "in the chip business only" leaves it selling components into a market that has started buying systems. Export controls tighten further against the 17% of revenue delivered into China. And the $67 billion of acquisition debt, entirely comfortable at peak cash generation, becomes a constraint when cash generation is not at peak and the CEO wants to buy his way out of the problem the way he always has.

The final element is the one nobody models. Hock Tan's compensation vests on AI revenue between FY2028 and FY2030. He has been given every incentive to sign gigawatt commitments and none to be cautious about who signs them back. In a downside scenario, the AI XPV Platform is not remembered as financial engineering that unlocked demand. It is remembered as the moment a semiconductor company took on customer credit risk it was not equipped to underwrite.

13. Further reading

• Google TPUv7: The 900lb Gorilla In the Room - SemiAnalysis, 2026-01-06 [paid]
• Broadcom's Google TPU Revenue Explosion, Networking Boom, VMWare Integration - SemiAnalysis, 2025-10-29 [paid]