Broadcom has greatly outperformed the FAANGs over the past decade yet is rarely discussed as one of the market’s top-performing stocks. The ticker certainly does not participate in a catchy acronym. Half the battle with Broadcom is there are many revenue segments to analyze, some go through harsh cyclical downturns, and it acquires companies hand over fist. To put it plainly, this is not an easy stock to cover; there are no pithy ways to summarize the products and it doesn’t offer growth stock qualities.

We have revamped the Essentials Portfolio by including Broadcom. Stay tuned for trade alerts sent directly to your inbox, keeping you up-to-date with any adjustments to three key holdings.

Broadcom has many departments that have been strung together through acquisitions with a vision of consolidating Ethernet, ASICs and now virtualization under one company. The reason this company is sitting in plain sight is because it was a major winner from the mobile era, is partnered with Big Tech as we move in the AI era, and is putting together the pieces to participate in AI strategically by not needing to compete on GPUs.

Broadcom Financials

Broadcom is firing on all cylinders and the Q1 earnings report cemented the company as number two in terms of AI revenue. It’s not only the AI revenue that sets Broadcom apart, but also its developing software strategy with VMWare.

The headline numbers don’t help to translate underlying AI strength as Broadcom reiterated its full year guidance yet raised AI revenue. This is because some of Broadcom’s segments are coming in lower than expected, while AI is coming in higher than previously guided.

This comment kicked off the tone of the call: “I know we told you in December, our revenue from AI would be 25% of our full year semiconductor revenue. We now expect revenue from AI to be much stronger, representing some 35% of semiconductor revenue at over $10 billion.” This is up from $7.5 billion expected this year, and also up from a $6 billion run rate last quarter ($1.5B per quarter). The AI revenue is roughly 70% ASICs and 30% AI Networking.

In addition to stronger-than-expected AI revenue, Broadcom is expecting dramatic, sequential growth in software bookings, which are expected to grow about 70% QoQ. We need another quarter or two to verify if the rapid growth from VMWare Cloud Foundation will continue, but management implies it will continue to be strong. If so, Broadcom is quickly asserting itself as a leader in AI software as consolidated bookings are expected to add $1.2 billion QoQ from $1.8 billion this quarter to $3 billion next quarter.

Revenue and EPS:

Q1 revenue was $11.96 billion, beating estimates by $240 million, and representing YoY growth of 34%. Excluding VMWare, revenue growth was 11% for a 7 percentage point acceleration over the past two quarters, at 4% in the October quarter and 4.9% growth in the July quarter.
Q1 adjusted EPS was $10.99, beating estimates by $0.57. GAAP EPS was $2.84, compared to $8.80 in the year ago quarter.
Broadcom reiterated its fiscal year revenue guide of $50 billion and full year EBITDA guidance of 60%. This compares to an EBITDA margin of 63% to 65% in previous quarters.

Margins:

Q1 GAAP gross margin was 61.7%, compared to 67.4% in the year ago quarter. Amortization of acquisition-related intangible assets adversely impacted gross margin by ~1150bp in the quarter. Adjusted gross margin was 75.4%, compared to 73.8% in the year ago quarter.
Q1 GAAP operating margin was 17.4%, compared to 46% in the year ago quarter. The operating margin was mainly lower due to the increase of amortization of acquisition-related intangible assets, restructuring charges, and stock-based compensation. Adjusted operating margin was 57.1%, compared to 60.9% in the year ago quarter.
Q1 GAAP net margin was 11.1%, compared to 42.3% in the year ago quarter. The net margin was mainly lower due to the increase of amortization of acquisition-related intangible assets, restructuring charges, and stock-based compensation. Adjusted net margin was 43.9%, compared to 50.3% in the year ago quarter.

Cash Flows and Balance Sheet:

Q1 operating cash flow was $4.82 billion, representing a 40.3% margin.
Q1 free cash flow was $4.69 billion, representing a 39.2% margin. Excluding restructuring and integration spend of $658 million, free cash flow was 45% of revenue.
Cash, equivalents and short-term investments totaled $11.9 billion.
Debt totaled $75.9 billion. The debt increased from the $39.2 billion in the previous quarter due to the additional debt taken to finance the VMware purchase and the company also assumed $8.3 billion VMware’s debt. The average coupon-rate and years to maturity of fixed rate debt of $48 billion is 3.5% and 8.4 years, respectively. The average coupon-rate and years to maturity of floating rate debt of $30 billion is 6.6% and 3 years, respectively. In early March, the company repaid $2 billion of floating rate debt and intends to maintain this quarterly repayment throughout FY2024.

In Q1, Broadcom paid stockholders $2.4 billion of cash dividends based on a quarterly common dividend of $5.25 per share. The company repurchased $7.2 billion of common stock and eliminated $1.1 billion of common stock for taxes due on vesting of employee equity, resulting in the repurchase and elimination of approximately 7.7 million AVGO shares. The Q2 non-GAAP diluted share count is expected to increase to approximately 492 million as the shares issued including VMWare.

Days sales outstanding were 41 days in the first quarter compared to 31 days in the fourth quarter on higher accounts receivable due to the VMware acquisition. This is due to the accounts receivable from VMware having payment terms of 60 days compared to Broadcom’s 30 days.

The company ended the first quarter with inventory of $1.9 billion, up 1% sequentially.

Key Segments:

Software Revenue:

Management reiterated their software revenue guidance of $20 billion this year.

Q1 Software segment revenue of $4.6 billion was up 156% year-on-year and included $2.1 billion in revenue contribution from VMware. In the previous quarter, software was $1.97 billion. This implies 27% QoQ growth in software after stripping out VMWare. When asked about this, management said to not get too excited about this particular growth as it’s due to strong contract renewals. Instead, the CEO explicitly stated: “Yes, don't get too excited over that. So that has also accelerated, but that's not the star of this show, Stacy. Star this show is the accelerating bookings and backlog we are accumulating on VMware.” In fact, it was indicated that some of this could fall off in future quarters given the software guide was not raised.
What the CEO is referring to as the star of the show is the consolidated bookings in software, which grew sequentially from less than $600 million to $1.8 billion in Q1 and is expected to grow to over $3 billion in Q2. Per management: “Revenue from VMware will grow double-digit percentage. Sequentially, quarter-over-quarter, through the rest of the fiscal year.”

Management stated the rapid growth from the VMWare segment is because: “We are focused on upselling customers, particularly those who are already running their compute workloads with vSphere virtualization tools to upgrade to VMware Cloud Foundation, otherwise branded as VCF […] VMware and NVIDIA entered into a partnership called VMware Private AI Foundation, which enables VCF to run GPUs. This allows customers to deploy their AI models on-prem. And wherever they do business without having to compromise on privacy and data — in control of their data. And we are seeing this capability drive strong demand for VCF, from enterprises seeking to run their growing AI workloads on-prem.”

Semiconductor Revenue:

Semiconductor solution sales increased 4% YoY to $7.39 billion, a slight uptick from 3.3% growth in the prior quarter. Stronger-than-expected growth from AI more than offsetting the cyclical weakness in broadband and server storage. According to Bloomberg, this was a bit shy of expectations for $7.7 billion in revenue. This was most likely due to weak wireless, broadband, and server storage segments.

Q1 networking revenue of $3.3 billion grew 46% year-on-year, representing 45% of semiconductor revenue. Management stated the following: “For fiscal 2024, given continued strength of AI NAND working demand, we now expect networking revenue to grow over 35% year-on-year compared to our prior guidance for 30% annual growth.”
Q1 wireless revenue of $2 billion decreased 1% sequentially and declined 4% year-on-year representing 27% of semiconductor revenue. Wireless is expected to be flat YoY for FY2024.
Q1 server storage connectivity revenue was $887 million or 12% of semiconductor revenue, down 29% year-on-year. The company revised its server storage revenue to decline in the mid-20 percentage range compared to prior guidance for a decline in the high teens.
Broadband Q1 revenue declined 23% year-on-year to $940 million and represented 13% of semiconductor revenue. Broadcom revised its outlook for fiscal '24 broadband revenue to be down 30% year-on-year from prior guidance of down mid-teens year-on-year.
Q1 industrial resales of $215 million declined 6% year-on-year. Management stated that industrial resales will be down high single digits this year.

Broadcom’s Switch Products and Switch Fabric

Broadcom has a dominant market share of switching and routing semiconductors for hyperscalers and is seeking to maintain its market share, most especially as AI changes networking requirements. The Jericho3-AI launch was last April, which is a redesign intended to compete with Nvidia’s InfiniBand.

Broadcom has three switch products. The Tomahawk is the high-bandwidth switch platform, Trident is the platform with more features, and the new Jericho line combines the Jericho switch with routing ASICs. The Jerico3 was redesigned with deep packet buffers. Tomahawk and Trident are used in data centers yet are not optimized for AI workloads, especially when compared to InfiniBand.

Jericho has 160 switch ports dedicated to switch fabric, which allows multiple ASICs to be stitched together to support GPU clusters. The asymmetric split helps the chip overcome network congestion and network failures. According to Broadcom, Jericho3-AI performed 10% better than “alternative network solutions” — which is a clear reference to InfiniBand.

A few specs before we go deeper into how Broadcom’s solutions compare to Nvidia’s. As you’ll note, the specs for InfiniBand are superior with support for 64X 400Gbp or 128 200 GB/s ports compared to Jericho’s 36X 400 GbE and 72X 200GbE network-facing ports.

Jericho has 144X SerDes lanes and 106Gpbs PAM4 supporting 18X 800Gbe, 36X 400 GbE and 72X 200GbE network-facing ports. The Jericho3-AI allows for more than 32,000 GPUs to be linked for a massive AI training system and links directly to GPUs without the need for a server bus.
Tomahawk5 runs at 100 GB/second with PAM4 with aggregate bandwidth of 51.2 Tb/s
Compare this to Nvidia’s Quantum-2 InfiniBand which has support for 64X 400Gbp or 128 200 GB/s ports with 51.2 Tb/s and 66.5 billion packets per second.
Nvidia’s new Spectrum X Platform is an Ethernet solution that delivers 1.6X better networking performance than traditional Ethernet with 256X 200 Gbe ports or 16,000 ports for larger training systems.

On a similar note, and while we are on the topic, Marvell has some skin in the game too by offering a 51.2T switch called Teralynx 10 that offers ultra-low-latency at 80% cost savings. According to Moor Insight Strategy, Marvell is the supplier for AWS for “electro-optics, networking, security, storage, and custom-designed solutions.”

Marvell’s Nova 1.6 Tbps PAM4 electro-optics have eight 200G lanes that “double the networking bandwidth while reducing power and cost per bit by 30%.” According to the press release, “by doubling the bandwidth per lambda, the Nova-based modules reduce the number of lasers and related optical components by 50%.”

Marvell hopes to help data centers transition to 51.2 Tbps networking architectures by offering a platform that needs 32 optical modules instead of 64 optical modules.

AI Networking

A few years back, we discussed that Nvidia acquired Mellanox for the strategic synergy that InfiniBand and Ethernet can provide in boosting GPU performance. Without proper interconnection, GPU performance could be limited, and so Nvidia strategically wanted to create the best-case scenario of owning both markets for AI accelerators — and their fabric and interconnects.

Mellanox supports Virtual Protocol Interconnect (VPI), which allows the ubiquitous Ethernet to provide bandwidth as cheap as possible, and InfiniBand to deliver higher throughput and fewer bottlenecks during high loads. In 2019, the split in Mellanox’s revenue was about 40% InfiniBand and 60% Ethernet. By leveraging a hybrid of Ethernet and InfiniBand, Mellanox was able to take market share from Ethernet incumbents.

The acquisition went under review in China, with officials believing Mellanox’s market share at the time was about 55% to 60% of the global interconnect market and 80% to 85% of the Chinese market. This illustrates how popular Mellanox was before Nvidia acquired the company.

The outcome of the review was that Nvidia was required to decouple the sale of InfiniBand from the sale of its GPUs to where a customer could buy one but not the other at no penalty. Even if a customer can buy them separately, there are many cases where it’s not practical to do so, such as with the DGX and HGX systems which achieve optimal performance with the A100s/H100s and InfiniBand.

Nvidia has stated that InfiniBand increase the effectiveness of AI infrastructure by 20% to 30%. Remote Direct Memory Access (RDMA) reduces CPU overhead by offloading data movement to network adaptors. In addition, Ethernet has quality of service (QoS) flow control and advanced error handling mechanisms that increase its network efficiency capabilities.

Ethernet Vs InfiniBand

I wish I could make networking more conversational, but it’s pretty challenging to do that! Here are some bullet points on how the two compare; I’ve bolded the more important takeaways:

Benefits of Ethernet:

Raw bandwidth is a benefit with Ethernet hitting 51.2 Tb/s two years ago with support for 800 Gb/s port speeds. InfiniBand lags by topping out at 51.2 TB/s with 400 Gb/s port speeds. Although typical server nodes do no need the extra bandwidth, AI clusters come with 400 Gb/s NIC per GPU with some nodes having four to eight GPUs. By 2025, Dell’Oro believes switch ports for AI networks will be operating at 800 Gb/s and will further double to 1600 GB/s by 2027.Dell’Oro believes switch ports for AI networks will be operating at 800 Gb/s and will further double to 1600 GB/s by 2027.
smartNICs and AI-optimized switch ASICs help to reduce packet loss
Large pool of vendors whereas InfiniBand increases dependency on Nvidia. For this reason, AWS and Google Cloud have remained on Ethernet as they prioritize custom silicon.
Ethernet is the incumbent networking standard and most cloud providers have invested heavily here already. With Ethernet, providers don’t have to manage a new network stack.
Ethernet switching has evolved to where a new term has been coined “lossless” Ethernet. Even Nvidia is moving in this direction with their Spectrum X platform.

Benefits of InfiniBand:

Outperforms for AI/ML workloads due to low latency and by reducing packet loss. Data packets are sent in a serial approach so multiple channels of data can be sent simultaneously. This is much better for AI/ML than a parallel approach for internal data flow, which creates bottlenecks.
Has 3X to 4X lower latency than traditional Ethernet switches based on ASICs
Highly scalable, can support tens of thousands of nodes per subnet. InfiniBand is also cheaper as it requires fewer connections for reliability.
Its QoS and failover capabilities are a reason it’s adopted for high-performance computing environments.
Reduces CPU resources

So, why is Ethernet Making a Comeback?

Broadcom’s Jericho3-AI has some promising benchmarks that could help shift the dominant market share InfiniBand has in AI networking (or at least prevent a monopoly). These benchmarks showed the Jericho3-AI outperforming InfiniBand by 10%, which is substantial when dealing with AI systems as it’s enough to increase the collective operations of the system.

“Leveraging this unique functionality, the Jericho3-AI fabric provides at least 10 percent shorter job completion times versus alternative networking solutions for key AI benchmarks such as All-to-All. This performance improvement has a multiplicative effect on decreasing the cost of running AI workloads since it implies that expensive AI accelerators are used 10 percent more efficiently. The network, in effect, pays for itself.” — You can read the press release here.

This means bare metal can work more effectively. Per Broadcom: “because it can handle 800Gbps port speed (for PCIe Gen6 servers) and more, it is a better choice [than InfiniBand.” At high price points, all hyperscalers want to see their investments working at maximum clock times. This is achieved by better load balancing and congestion control to improve network latency, whereas InfiniBand reduces port and hop latency inside the switch. Broadcom calls their product differentiation “Perfect Load Balancing” and “Congestion-Free Operation.”

A note on PCIe

The maximum bandwidth supported by PCIe 5.0 is 400Gbps per port. By using 106Gbps PAM4 SerDes, ASICs can be tuned to support 100, 200 and 400 Gbps port speeds. To work around this, and to achieve 800Gbps, chip makers are building NICs directly into the accelerator.

Jericho3-AI supports 36 ports at 400 Gbps speed, and this can support Nvidia’s powerful DGX H100, which have 8 ports of 400 Gbps speed. In this case, four node racks are within Jericho’s capabilities. However, the Quantum-2 InfiniBand can handle 64 ports of 400Gbps, and so for Nvidia’s GPUs, it outperforms. Broadcom’s answer to this is that AWS and Google still prefer to not have vendor lock-in with Nvidia and use the Jericho3-AI to make use of their extensive Ethernet systems.

Overview of Custom Silicon (ASICs):

In addition to AI networking, Broadcom participates in the custom silicon market. ASICs are application-specific integrated circuits that are customized to perform a specific function for a specific application, hence the term “application-specific.” This is in contrast to GPUs which are more general-purpose. ASICs are expensive at the onset, yet become cheaper with volume production. We first published this graph in 2019 but the comparison still applies:

For now, custom silicon only makes sense for a company with deep coffers that has immensely popular applications – such as Google, Meta, Amazon. These companies use custom silicon to drive down costs on GPUs for their most popular applications. Across ASICs, the most well-known is Google’s tensor processing unit (TPU).

Google was one of the first to require low-power machine learning workloads for Search, YouTube and Google Maps. The compute intensive workloads were running on Nvidia’s GPUs for both training and inferencing until Google made their own processing unit, TPUs, to perform workloads at a lower cost and higher performance.

Performance between TPUs and GPUs is often debated depending on the current release (A100 versus fourth-generation TPU, for example). In some cases, TPUs have better performance per watt for power-constrained applications. Notably, some of this comes with the territory of being an ASIC, which is designed to do one specific application very well whereas GPUs can be programmed as a more general-purpose accelerator. In this case, the benchmarks where TPUs compete are object detection, image classification, natural language processing and machine translation — all areas where Google’s product portfolio of Search, YouTube, AI assistants, and Google Maps, for example, excels.

Notably, TPUs are used internally at Google to help drive down the costs and capex of its own AI and ML portfolio and they are also available to users of Google’s AI cloud services. For example, eBay adopted TPUs to build a machine learning solution that could recognize millions of product images.

Unless Google releases an internal technology as open-source, it won’t be adopted by the competitors. This is where Nvidia’s neutral position as traditionally a hardware company becomes a positive as it’s universally used by Amazon, Microsoft, Google — — and Alibaba, Baidu, Tencent, IBM and Oracle. Meanwhile, TPUs create vendor lock in (with a direct competitor) which most companies want to avoid. eBay is the exception here as the company needs Google-level object detection and image classification.

Why ASICs are Not a Near-Term Threat to GPUs

AI investors will need to get comfortable hearing about the battle between ASICs and GPUs. This debate has been going on since at least 2018, when Nvidia’s biggest threat was thought to be Google’s custom TPUs. There is some merit to these concerns as the largest customers for Nvidia’s GPUs have enough cash to make custom chips. There’s roughly $35B to $40B per Big Tech company per year that executives will naturally want to optimize to drive down costs.

To program ASICs is difficult, and they are application-specific, which means they cannot be reconfigured. Nvidia is wildly popular because GPUs are easy to program and are the best choice for a wide range of applications. Developers create the moat, which was our original Nvidia thesis. Therefore, I don’t believe there is much risk that Big Tech commercializes AI accelerators.

However, it’s quite plausible that someday Big Tech will reallocate capex toward more ASICs and fewer GPUs to where it will impact Nvidia. For now, demand outstrips supply, and there are long lead times for Nvidia’s GPUs. If a company like Google reallocates to more TPUs, another enterprise will certainly step up to fill those orders.

VMWare –Software-Defined Networks and Data Centers

In November, Broadcom closed its acquisition of VMware for $69 billion. VMWare is virtualization software that virtualizes compute and data centers. The software creates an abstraction layer, or a “hypervisor” which is the technical term for a computer or server that runs virtual machines called ESX. VMWare was the first company to virtualize x86 machines and was founded in 1998. Operating systems, such as Linux, Windows and MacOS, can share the same, virtualized resources by running on a x86 machine.

Over the past decade, VMware pivoted to offer a software-defined data center. On the Q4 earnings call, Broadcom discussed building essentially a private cloud on-premise, which has some advantages, such as lowering capex by pooling memory, security, networking and server resources. “Our strategy going forward is simply to enable global enterprises to run their applications across the other data centers as well as on public clouds by consuming VMware’s higher-value software stack.”

Later it was stated: “We are creating with VMware, the same experience of virtualization of the data center on-prem for those companies, which has workloads, by the way, that are already running VMware products that application that’s already written on VMware Cloud Foundation. This is then giving these enterprises the opportunity to have a hyperscaler on-prem. That’s the plan we’re doing, plain and simple.”

Where software defined networks have seen quite a bit of success is with 5G networks. SDNs separate the control plane on embedded switching systems. This allows the networks to be managed remotely. Products from different suppliers can be used without incompatibility issues. By using open APIs, the 5G market has benefited from a more neutral ecosystem by allowing products from different suppliers. This is because SDNs allow network functions to be programmed by APIs instead of proprietary interfaces.

In 2012, VMWare acquired Nicera to create VMware NSX, virtual networking and security software that virtualizes network components. The NSX products, including NSX-T data center, programmatically creates and manages virtual networks from Layer 2 to Layer 7, which is defined as switching, routing, access control, firewall and QoS. NSX Manager and transport nodes can be assembled in seconds for proof-of-concept deployments, deployments with up to 64 hosts, or large-scale environments. Software-defined networks have a natural synergy with Broadcom, the leader in networking hardware.

A few years ago, VMWare expanded to virtualize containerized workloads for Kubernetes clusters. This product is referred to as Tanzu. This was a necessary evolution to keep cloud native customers. Many Kubernetes clusters use something called a multi-tenancy solution, which is to have non-connected “tenants” use a common pool of resources. This can be hard to implement correctly, and also has limited functionality once it’s set up. Virtualized containers are similar to a single-tenancy solution by having its own API server, controller manager and storage for data. Yet, it’s similar to a multi-tenancy solution by using a common pool of resources. Per the Broadcom Q4 earnings call: “And to attract and keep these workloads across the environment, we are investing in a rich catalog of microservices tools. This will be our focus. And the noncore businesses of end-user computing and Carbon Black will be divested.”

All of this sounds good, but virtualization is not a wild success. The software defined data center market at one time was expected to reach $77 billion by 2020 but instead has only reached $28 billion as of 2023. There are many vendors in the space, which creates pricing wars.

Broadcom Technical Analysis

By Portfolio Manager Knox Ridley

We still believe AVGO is in a large uptrend, so the recent volatility was used to layer into this position. Even though we are seeing a bounce, there is a chance that AVGO’s correction is not over. This setup would see the current bounce fail, and then a breakdown below $1200. If this happens, we will look to add to AVGO between the $1150 – $1000 region. On the other hand, if we hold the $1200 level, we can keep pushing higher into the $1600 region. A breakout over $1365 would signal this scenario is likely playing out, which we would also use as a signal to buy into strength.

Conclusion

Broadcom belongs to our AI portfolio. Over the next few years, Big Tech is likely to diversify away from Nvidia’s GPUs, plus Ethernet networking continues to be upgraded for AI purposes, which means Broadcom should be on our radar. It’s not only the AI revenue that sets Broadcom apart, but also its developing software strategy with VMWare makes it an ideal choice.

Pro premium members receive deep-dive research on all the stocks in the portfolio and quarterly earnings kickoff webinars. In addition, the Advanced Market Signals Members receive regular technical and broad market analysis and weekly webinars from our Portfolio Manager, Knox Ridley. Learn more here.here.

Recommended Reading:

Category: Enterprise

Broadcom: $10B in AI Revenue This Year Plus Software is Rapidly Accelerating