Broadcom: The Silent Giant in AI Chip Revolution – Putting Numbers To It

Endless waffling has been done regarding competition for Nvidia’s dominance in AI, and potential for competitors. While much of this discussion has centered around AMD, with some sidebars on Intel and Cerebras, the fact is that in 2023, no one besides Nvidia will generate even $1B revenue from chips that run large language models.

Scratch that, there is one other player.

Often overlooked is that Broadcom is the second largest AI chip company in the world in terms of revenue behind NVIDIA, with multiple billions of dollars of accelerator sales. This is primarily driven by Google’s aggressive TPU ramp as part of its self-described “Code Red” in response to the Microsoft + OpenAI alliance that is challenging Google’s world leadership in AI.

Today we will be detailing how many dollars will go from Google to Broadcom from this gargantuan ramp. This will cause a massive shift in Broadcom’s business profile that no one is really talking about.

Google isn’t the only one of course. Meta also makes their in-house AI chips with Broadcom, but Meta doesn’t deploy too many of these… yet. Aside from a massive TPU ramp which we will detail further below, there are numerous positive developments taking place within Broadcom. This includes their other custom silicon products with firms like Meta, their world leading networking business, which is the 2^nd largest spending bucket for a AI infrastructure after accelerators, continued strength in wireless chips despite weak smartphone sales due to Apple content growth, and the upcoming VMWare integration. There’s a lot here, as this is one of the more financial number focused reports we’ve put out.

Before we discuss the Google TPU and networking growth, we feel there needs to be a dive into Broadcom’s culture, leadership, and how it has managed to nurture multiple crown jewel products/core IP out of a disparate portfolio of many individual companies.

It should also be noted, Hock Tan, the Broadcom CEO, is objectively a top 3 fabless semiconductor CEO, alongside Nvidia’s Jensen Huang and AMD’s Lisa Su. Despite this, there’s this notion that Broadcom acts like a strip-mining private equity roll-up operator that is constantly raising prices, cutting R&D, raising debt, to acquire more companies. This is undeserved FUD, but it causes many people to be reluctant of positivity towards Broadcom.

Broadcom Undeserved FUD

Broadcom has a bad rap among many for its heavy focus on M+A, a perception that is mainly held by those that either do not understand the underlying strategy or do not agree with it, so they discount it.

In the archetypal roll up strategy gone wrong, an acquirer will purchase one or many leading franchises with strong products and slash headcount, particularly in research and development, driving margins up, leading to a huge increase in cashflows for the company. The products then coast for several quarters or even years until the next product cycle or industry pivot. The company then loses their product leadership, market share, and suffers declining revenue and margins, leading to a flameout in cash cow business lines. Private equity firms love this model, and they will raise huge amounts of debt to fund their holding in the company, hoping to exit before the flameout.

Many perceive Broadcom as this sort of operator, but after numerous acquisitions over 17 years, this scenario has yet to play out. They have outlasted bearish prognostications for an eventual blow up.

Broadcom’s Strategy Simplified

The strategy is simple – Broadcom acquires companies that sell market leading products with sticky customers, recurring revenue, and high margins but have excessive operating expenses and are generating below potential profit and cash flow.

Broadcom then cuts costs deeply, eviscerating layers of middle management, cutting sales and marketing functions down to those needed to directly support individual products, and almost completely eliminating general and administrative costs in favor of utilizing Broadcom’s existing corporate platform resources.

Research and development is a different story. Broadcom does eliminate science projects with unclear near-term return on investment as well as common research and development functions not directly driving revenues, but it leaves product teams intact. With layers of middle management cut out and a multitude of committees eliminated, product teams can obtain approval for plans directly from senior management and can execute them with greater alacrity. By loading overhead costs onto the product groups’ P&L and holding managers accountable for the group’s results, Broadcom has further driven a culture of efficiency and in many cases their market share has grown.

This practice is well suited to semiconductor products, where the customer set is well defined, and where revenue growth is mainly driven by content growth and spec upgrades rather than rapid growth in the customer base. In the wireless business for example, where most of Broadcom’s sales are to Apple, there is little need for a traditional sales organization, with orders for parts and volume taken by staffers and pricing negotiated at the upper management level.

The result is a very lean organization with much higher margins and cash flow generation that is focused on core products and franchises and carrying out the R+D necessary to drive future generations and advances in those products and fields, with M+A bringing in new products and pivoting the company towards new directions.

This seems easy enough to understand, but why does the community continue to exhibit symptoms of FUD?

For analysts, the issue is that the company is so broad, that subject matter expertise can be fragmented. Most semiconductor analysts do not have experience analyzing software companies and hence treats Infrastructure Software as a black box and applies a conglomerate discount. The software analyst will box Broadcom into the Semiconductor sector and will not look at the company at all.

On the buy side, many analysts have expertise and experience across both semiconductor and software  – so this is less of an obstacle. However – many can be uneasy with Broadcom as they see a Semiconductor company pivoting to play in infrastructure software. With the VMWare acquisition looming, many are unsure if the pivot to software will be successful.

Analysts always like to use their toolkits to understand companies, but our contention is that Broadcom’s strategy is much more generalized. It is a platform company focused on technology that acquires companies that sell market leading products with sticky customers, recurring revenue, and high margins but have excessive operating expenses and are generating below potential profit and cash flow.

The perception of high debt loads is also another source of pushback, with the acquisition of VMWare taking Broadcom up to 2.9x debt/LTM Adjusted EBITDA. While Broadcom does load up debt when they acquire companies, we think the track record of growing free cash flow while deleveraging quickly post-acquisition AND paying shareholders considerable dividends while conducting buybacks mitigates this concern.

A brief historical overview of Broadcom and its roll-up strategy

Broadcom is a portfolio of 16 semiconductor franchises and two main software franchises that have come together to form today’s five semiconductor segments and the infrastructure software segment. The best way to explain each of the segments is to trace the lineage of these franchises and how they came to form present day segments.

Wireless: The Apple Division

Broadcom’s story starts with HP’s semiconductor division. Since the birth of the semiconductor industry up until the 80s and 90s, most hardware companies were vertically integrated with their semiconductor divisions and foundries. HP was no different. Buried inside HP was its test and instrumentation segment which itself contained an internal chip division focused on RF components such as filters as well as passive components such as resistors and capacitors. The diamond in the rough was RF filters, a device that allows an antenna to tune into a specific frequency band while filtering out all the unwanted radio signals on other bands.

In 1999, HP decided to spin out businesses that were not tied to its core Enterprise Servers, Software, PC and Printers divisions. Agilent was spun out in 1999, containing HP’s test and measurement business. The semiconductor division within Agilent was further spun out in 2005 to a KKR and Silver Lake-led private equity consortium. By late 2005, the company was rebranded as Avago.

Hock Tan, our protagonist, took the helm as Avago’s CEO in 2006, setting the stage for the evolution of today’s Broadcom. His first major move was in 2008 when he purchased Infineon’s Bulk Acoustic Wave (BAW) business for $30M, bringing the Film Bulk Acoustic Resonator (FBAR) Filter product into the fold – a deal that arguably ranks as one of the most profitable M&A transactions of all time.

In contrast to the 90s and 2000s, when having a phone with the 850Mhz, 900Mhz, 1800Mhz, 1900Mhz and 2100Mhz radio bands was sufficient to connect to most cell networks globally, the explosion of mobile data consumption required a greater number of frequencies and much wider frequency bands to the point where the list of bands supported by the iPhone 14 Pro (We counted 69 bands) would fill up a third of a page. The FBAR filter thus went from an expensive high specification part that was not particularly needed to an indispensable part of modern smartphones, earning an estimated $3-4B of revenue annually.

Today, other than RF front end modules and filters, Broadcom also sells Wi-Fi, Bluetooth and GPS SoCs (from a later acquisition). The fact that most of the segment’s sales are to Apple often raises concerns of Broadcom’s bargaining power, customer concentration risk, and possibility of Apple commoditizing products by playing it off against other vendors.

The track record does not support this concern. Broadcom’s wireless business has grown revenue at a 10-12% CAGR from FY16 through FY22, a lot faster than iPhone unit sales (in the 200-240M range for nearly the past 10 years), due to chip content increase and specification upgrades, for example filter upgrades to support increasing numbers of bands at 5G and Wi-Fi technology upgrades such as the ongoing Wi-Fi 6E transition which adds the 6GHz spectrum.

Broadcom’s negotiations with Apple are also of an entirely different dynamic to your typical Apple supply chain bottom feeders – with Hock Tan said to personally negotiate supply and pricing agreements.

The market’s downbeat reaction to Apple’s latest earnings report due to macro worries should not be cause for excessive caution at Broadcom’s wireless division. Other than a tailwind from content gains and spec upgrades to WiFi 6E within the iPhone, Apple itself is likely bucking a sluggish smartphone market in the developed world by focusing on growing sales in emerging market countries, and most importantly, continuing to gain share against Android based smartphone.

Over the years, the success of the filters and broader wireless business turbocharged Broadcom’s M&A strategy, providing it with the strong cash flow needed to pursue ever larger acquisitions.

Storage and Custom Silicon: The LSI Acquisition

In late 2013, Avago made its first major acquisition, purchasing LSI for its networking, storage and nascent custom silicon business. LSI’s largest business was in the Storage Area Network (SAN) product. A Storage Area Network is a high-speed network that provides access to a number of storage devices, making Solid State Drive (SSD), Hard Disk Drive (HDD) arrays or tape libraries accessible to any device on a network as if the storage drive was a local drive. LSI also had capabilities in using the Fibre channel protocol to implement SANs. Broadcom’s subsequent acquisition of Brocade Communications in 2016 solidified the storage segment by adding Brocade’s Fibre Channel and SAN directors to the segment.

The more interesting businesses that came in from LSI were the billion dollar PCIe switch business and the ASIC custom silicon products division that was formed from an earlier combination of LSI Logic and Agere’s ASIC groups. A custom ASIC business is where a product group designs chips on behalf of an external customer, leveraging and incorporating their core IP and design capabilities. From less than 20% of LSI’s revenue at acquisition, Custom Silicon has grown into a $2-3B+ annual revenue business.

Of course, this is accelerating massively due to generative AI. Custom Silicon products today include AI chips such as Google’s TPUs, Meta’s MTIA, as well as YouTube video encoding chips, custom routing/switching silicon. Custom Silicon falls under Broadcom’s Networking segment, but we argue it deserves its own segment. More on that later.

Networking, Broadband and Avago’s 2015 Purchase of Broadcom

One of Hock Tan’s biggest moves was Avago’s 2015 acquisition of Broadcom. Prior to the acquisition, Avago’s networking products were an order of magnitude smaller than Broadcom’s. Broadcom was known as the bluest of blue chips among semiconductor companies, with a strong business franchise in broadband and the leading networking business, where it sold best in class ethernet switching chips predecessors to the current Tomahawk family, businesses that today form the bulk of present-day Broadcom’s networking business.

Other than the Jericho family of router silicon, present-day Broadcom’s networking division includes Ethernet routing and switching silicon for use in data centers, Fiber optic communication components, and a custom silicon business that designs AI chips such as Google’s TPU and Meta’s MTIA as well a host of other customized routing/switching chips.

Most products that form the current Broadband segment were added through the Broadcom acquisition. These products include Digital Subscriber Line (DSL), Passive Optical Network (PON), DOCSIS Cable Modem components and system on chips (SoCs), as well as Set Top Boxes and Wi-Fi access point SoCs, enabling the buildout of faster home/consumer broadband internet access networks. This segment is also accelerating due to the unprecedented DOCSIS 4.0 capex cycle and subsidies for home fiber.

Software: The Final Frontier?

In 2018, just as investors and analysts were starting to see the fruits of Broadcom’s strategy and take a more constructive stance towards it, Broadcom shifted the landscape once again by acquiring CA Technologies, a company that focuses on Mainframe IT management and monitoring software and application development tools that essentially help manage, connect and coordinate disparate IT systems within an enterprise. Though CA’s revenue had been rangebound for years, a hallmark of its business is that it sold leading products with sticky customers, recurring revenue, and high margins.

Broadcom followed up with its acquisition of Symantec’s Enterprise business in 2019. Symantec’s business focuses on Endpoint Protection, Secure Web Gateways and Data Loss Prevention. It holds the highest market share among its competitors in each of these businesses. Though revenue was generally stable at Symantec in 2019, like CA, product gross margins were high, customers were generally enterprises that tended to subscribe to Symantec’s products on a recurring basis.

The combination of both companies, now known as Broadcom’s Infrastructure Software segment, delivered an impressive 92% non-GAAP gross margin and a 72% non-GAAP operating margin in the most recent earnings report.

Broadcom’s journey with software continues as its acquisition of VMWare is in the process of closing. More on that later in the article, with our accretion and profit estimates.

Networking and Custom Silicon – the Crown Jewel

Broadcom’s custom silicon ramp is the real focus for today. Most people don’t realize the magnitude of it. Today we will put dollars out the door to Google and in the door to Broadcom. Furthermore we will go through the rest of their networking growth and the accretion model from VMWare as well.

Broadcom’s should split its Networking segment into two parts – a pure play networking segment and a Customs Silicon and ASICs segment. We estimate that in FY22, as much as 35% of segment revenue was from Custom Silicon (eg TPUs, Video encoding chips, and custom routing/switching silicon), and 65% of networking segment revenue was from networking and switch silicon (eg Tomahawk, Jericho, Trident families).

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

What makes this segment the crown jewel is not directly just Broadcom’s competency in either networking and routing chips or custom silicon, but their dominant SerDes IP enabling higher speeds from the chip.

It is also why Broadcom is winning in a host of custom silicon and compute offload designs including Meta and Google’s AI ASICs, routing and switching custom silicon including Cray and many others.

While this has been a good business for Broadcom, Google’s big TPU ramp later this year makes it one the most compelling businesses in the world.

First, let’s spend some time going through the core networking and switching products which have been growing in their own right over a number of years and benefiting considerably from the surge in AI investments.

The largest portion of Broadcom’s networking products are its routers and switches silicon. Networking products serve Carrier Networks and Data Centers.

Carrier networks carry internet traffic long distances, provide connectivity to data centers as well as the backbone for consumer internet access, and Broadcom sells mainly routing silicon such as the Jericho and Qumran family.  Routers are typically used to connect different networks to one another – particularly across long distances.

A router typically consists of a CPU for overall management, line cards that interface via the physical layer to connect to other networks (eg via fiber optic cables), and a switching fabric which consists mostly of just a series of Serializer/Deserializers (SerDes – a device that encodes data for transmission at extremely high speeds) to connect the line cards to one another.

Broadcom’s Jericho family of chips provide silicon for both the Line Card, where most of the compute intelligence is needed, as well as the switching fabric. Broadcom sells silicon to major router companies which then sell the complete router device to telecom companies or hyperscalers. They also sell directly to some hyperscalers.

Data Centers needs are focused on switching data traffic within the data center – directing traffic from each individual rack unit, to a top of rack switch to a leaf switch and ultimately to a spine layer that connects multiple leaf switches. In contrast to routers, switches direct traffic within a single network to connect end devices together – for instance within a data center.

In contrast to a router, an ethernet switch is heavily centered on the switch silicon, which provides much of the intelligence and inceitherates a number of SERDES on chip and connects via copper traces to the input/output ports which may be ether copper wire based or fiber optic based.  Broadcom’s Tomahawk family of switching chips serve as the main intelligence within the switch. Its Tomahawk 5 switch can handle up to a total of 51.2 Tbps, which would work out to 64 ports at 800 GbE or 128 ports of 400 GbE or 256 ports of 200 GbE.

In a typical data center deployment, there is a hierarchy of switches. At the top of a rack (TOR), a switch will connect all the individual rack units (RU) together, with the TOR connecting to a leaf switch that connects that individual racks to other racks on that leaf. A spine switch will then connect to multiple leafs, creating a fabric that provides a diversity of routes for each leaf switch to connect to another leaf switch.

Of course as part of the mix is the InfiniBand standard, which is at this point only sold by NVIDIA (ie from its Mellanox acquisition). NVIDIA’s Quantum InfiniBand and Spectrum Ethernet products compete with Broadcom’s switching and routing silicon. Right now Nvidia is using GPU allocation as a big weapon to increase networking content. Generally they are somewhat behind Broadcom, but not by a massive amount. Their Infiniband solution has some real advantages, but there’s also real disadvantages.

As we mention in our note, there are inherent problems with InfiniBand’s credit-based flow control, leading to resource exhaustion, backpressure propagation, and a slow response to changing network conditions. Ethernet is also a significantly larger market, is usually ~2 years ahead of InfiniBand in terms of max radix/speed. Though some GPU buyers might feel pressured to deploy an InfiniBand architecture, plenty like Amazon are opting to stick with Ethernet.

And that’s not all – Broadcom is bringing its core expertise to bear in products such as being the first to introduce co-packaged optics (CPO). They also have a host of other networking components such as DSPs and other transceiver components, Active Electrical Cables, and retimers.

Where AI Comes In

AI networks by nature operate on huge models where compute must be distributed among thousands or tens of thousands of GPU node that need to communicate with each other. In contrast to switching traditional traffic, AI traffic tends to be bursty, synchronized, and mostly planned due to how various parallelism strategies work (pipeline, data, and FSDP).

This leads to significant congestion issues as this surge of traffic can overwhelm the capacity at various parts of the switch network, causing flow collisions, link failures resulting in delays and packet loss. Some AI models take 20-30 days to run, and thus congestion time can add significantly to time and cost. Some of the training models Meta has run have suffered from as much as 50% of time spent on compute waiting for network to deliver data.

Broadcom’s solution space to this problem is with intelligence at the switch silicon layer. Broadcom has two discrete approaches built around its Jericho-3AI routing chip and its Tomahawk 5 switching chip.

Jericho3-AI focuses on a switch schedule approach, managing traffic using a credit system, where a sending switch must be granted credits by a destination switch before packets may be sent over the broader spine fabric of the network. Scheduling of data transmission happens at the switch, with endpoints (eg the GPUs themselves).

Tomahawk 5 focuses on an end-point scheduled approach, whereby the switches interface with the individual NICs to switch traffic on a load aware basis.

Networking chips clearly benefit from volume in the training networks being built out, but they also bring much more value to the table in solving this problem. Put simply, the combination of two factors are why AI needs a strong network solution and needs Broadcom in particular. It is no coincidence that Broadcom is the second largest user of CoWoS, from its routing/switching products as well as from its AI Custom Silicon.

TPUs – Incredible Ramp

As we noted on the weekend, Google’s Gemini has an incredible ramp. Due to the LLM craze.

Aug 28, 2023

Google Gemini Eats The World – Gemini Smashes GPT-4 By 5X, The GPU-Poors

Dylan Patel, Daniel Nishball

Google’s infrastructure is incredibly custom, and very good.

Apr 12, 2023

Google AI Infrastructure Supremacy: Systems Matter More Than Microarchitecture

Dylan Patel, George Cozma, Gerald Wong

Likewise, other firms like Meta would like to also make chips for AI that are customized for their needs, and they also use Broadcom.

May 19, 2023

Meta Custom Silicon: What’s Old Is New

Dylan Patel, Gerald Wong

With some of the more than dozen custom silicon projects identified, we see the revenue opportunity within AI Custom Silicon alone at $3B of revenue in FY23, and we think custom silicon for switching/routing and other related compute offload could be up to $1.9B of revenue for a total of $4.9B in the entire Custom Silicon business within networking.

We think that the TPU ramp will only start in earnest in the second half of FY23, with AI Custom Silicon reaching about a $2B per quarter run rate entering FY24, for a total of $8-9B in revenue by FY24.

Below we present our product-based bottom-up estimates for Broadcom’s custom silicon business in the next few quarters. This is based on both CoWoS allocations, Google’s ramp in units, and ethernet/Jericho demand. We are modeling both TPUv5e and the unannounced TPUv5.

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

In total, we have networking exiting FY24 at a $4.2B quarterly run rate vs a $2.6B rate in 2Q FY23, with AI custom silicon making up almost half of that quarterly revenue.

With minimal incremental SG+A and R+D spend required to ramp up sales of custom silicon products, custom silicon (we assume lower GPMs but lower SGA costs, hence custom silicon OPMs the same as networking segment average) could contribute more than half of the networking segment’s operating profit by FY25 and contribute as much as 24% of the company’s earnings, up from only 2-3% in FY21.

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

VMWare

VMWare’s integration into Broadcom’s platform will also drive meaningful profit growth in the software business. VMWare’s business provides software that enables virtualization of server hardware, allowing enterprises to both improve utilization of in-house servers and implement a cloud computing environment within an enterprise.

Though revenue growth has been slower since 2021 after growing at a 10-12% CAGR for many years, VMWare’s products still form the backbone of many enterprises’ cloud and server architecture and hold an overwhelming 70% market share of installed base for server virtualization.

Despite healthy 80-85% gross profit margins, VMWare’s relatively high S+GA expenses at 38% of revenue are a key reason why the company earned a below potential non-GAAP operating margin of 28% in the twelve months ended October 2022, far below Broadcom’s Infrastructure Software non-GAAP operating margin.

In summary, VMWare sells market leading products with sticky customers, recurring revenue, and high margins but has excessive operating expenses and has been generating below potential profit and cash flow.

Does that sound familiar?

Before moving to accretion, we also want to remind our readers about the recent announcement of the VMWare and NVIDIA partnership that will focus on hosting deployments of NVIDIA’s AI enterprise suite running on VMware vSphere sitting on top of NVIDIA GPUs as well as traditional CPUs. VMWare could stand to benefit over time as NVIDIA’s Cloud Strategy and AI software solutions gain traction. We will be writing more on what we think will be the most impactful use case for NVIDIA’s AI cloud strategy in an upcoming post.

VMWare Integration – Apply the Broadcom Method

We estimate that Broadcom’s strategy will drive a 20 point improvement in non-GAAP operating margins at VMWare, even if revenue growth decelerates during the integration process.

We expect the usual cuts to SG+A, but contrary to how detractors at times portray an M+A strategy, we expect Broadcom to increase Research and Development spend by up to $2bn as per Hock Tan’s recent statements at VMware Explore Las Vegas.

VMware has tremendous potential. Our plans to invest an additional $2 billion annually in R&D and the deployment of VMware’s solutions will help us to better unlock customer value. Our customers and partners will have access to a stable, growing, and powerful multi-cloud platform underpinned by world-class security, enabling them to accelerate innovation for all their applications.

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

We think that Broadcom will do the following:

1. Cut general and admin costs from 8-9% of sales to only 3% by FY25 as many of the overhead and platform costs can be eliminated completely as VMWare’s corporate costs move onto Broadcom’s platform.

2. Streamline and consolidate sales operation under one roof – reducing sales and marketing from 33% of revenue to 10% by FY25.

3. Increase Research and Development spend by $1.4B vs FY22. R+D as a percentage of sales would increase from 24% to 35% given our revenue estimates.

4. Shift to a higher percentage of compensation paid through stock-based compensation (from 14% in FY22 to 26% by FY26 – in line with the average for Broadcom). This would not change the economic compensation costs to shareholders but would improve operating cash flow and mean greater flexibility with regards to cash flow.

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

Broadcom’s management has made a fairly upbeat forecast that the VMWare would be accretive in year one. This sounds dramatic because it is – we expect Broadcom to slash hard and fast, with combined Sales and G+A costs dropping from 45% of revenue in Apr-23 quarter to 22% by Apr-24 quarter.

Non-GAAP operating margins at VMWare could improve from 28-31% in FY21-22 to 46% by FY25, contributing about $5.6B of post-tax non-GAAP operating income to the $26B we project for Broadcom ex-VMW in FY25.

As alluded to early, Broadcom’s expectations for accretion are bold. In its call announcing the acquisition, Broadcom stated that it expects the deal to be “accretive out of the gate” and that it will “get very accretive as we get through our integration process”.

In year one – on a non-GAAP basis, we estimate that the acquisition could result in a 3% non-GAAP EPS accretion, with this figure rising to 6% by FY25 as the integration enters the second year out of a three year integration timeframe.

Edit: We have updated for more accuracy after the quarter, but are leaving the old pre-earnings figures as a record. Click in to see the two tables.

Broadcom Revenue by Segment

Putting it all together AI will be 20-21% of EPS up from 3% last year.