AMD confirmed a fundamental bottom with a beat this quarter and a beat for next quarter. Analysts were expecting revenue growth of 6.8% and AMD reported growth of 8.9%. Adjusted EPS marginally beat at $0.69 versus $0.68 expected. The results were aided by data center revenue reaching a record as growth accelerated to the triple digit range, at 115% YoY for $2.84 billion, marking a 35-percentage point acceleration from 80% in Q1. CEO Lisa Su said AMD’s “AI business continued accelerating and we are well positioned to deliver strong revenue growth in the second half of the year led by demand for Instinct, EPYC and Ryzen processors.”

We had stated on our most recent webinar and in the write-up AMD’s Future Looks Bright that we want to give AMD the space to fill the very big shoes an Nvidia contender has to fill. Data center growth in Q2 of 115% is a clue that AMD is a serious contender on AI accelerators. The only other companies that have posted triple digit growth from AI in a standalone segment are Nvidia and Super Micro. To be fair, some of the revenue is from EPYC CPU processors, but the majority of the growth is coming from Instinct GPUs. We can sparse out the growth as Instinct drove $1B in revenue this past quarter, with EPYC contributing $1.84 billion. Without Instinct, data center revenue would have grown 41.5% versus 115% with Instinct.

With AMD down (-6%) YTD while Nvidia is up 109%, the market continues to communicate “not good enough.” Yet, what makes these numbers intriguing is we are at the bottom for this company (not a top – this is key), with fundamentals improving and accelerating from here.

Revenue and EPS:

AMD’s revenue accelerated to 8.9% YoY in Q2, up from 2.2% last quarter due to data center revenue accelerating significantly this quarter.

• Q2 revenue was $5.84 billion, up 8.9% YoY and 6.6% QoQ from $5.47 billion last quarter.
• Adjusted EPS of $0.69 beat estimates by $0.01, representing YoY growth of 19% and QoQ growth of 11%.
• GAAP EPS of $0.16 missed estimates by $0.02 but represents 700% YoY growth and 129% QoQ growth as margins bottomed and turned up this quarter (this high growth is due to   
• For Q3, AMD guided revenue to be $6.7 billion, +/- $300 million, for YoY growth of approximately 15.5% at midpoint. Analysts were expecting Q3 revenue to be $6.61 billion, for YoY growth of approximately 14.1%, so Q2 and Q3 beat/raised by 1 to 2 percentage points.

Key Segments:

AMD reported record data center revenue in Q2, with growth accelerating to the triple digits on strong CPU and GPU demand and the “steep ramp” of Instinct GPUs.  The company stated it’s AI accelerator the MI300 contributed $1 billion in revenue.

Data center revenue was $2.83 billion, up 115% YoY and 21% QoQ. For reference, in Q1, AMD reported data center revenue growth of 80% YoY and 2.4% QoQ, so this is a rather sharp acceleration in just one quarter. Per management this was “driven by the steep ramp of Instinct MI300 GPU shipments and a strong double-digit percentage increase in EPYC CPU sales.” There is expected to be strong growth next quarter in the DC segment.

Client revenue was $1.49 billion, up 49% YoY and 9% QoQ, driven by sales of Ryzen processors. This was a solid print, as Client rebounded from a (6%) QoQ decline in Q1. This was “driven by strong demand for our prior generation Ryzen processors and initial shipments of our next-generation Zen 5 processors.” There is expected to be strong growth next quarter in the Client segment.

Gaming revenue was $648 million, down (59%) YoY and (30%) QoQ as the segment continues to weigh on growth. Management stated the gaming market remains soft and sales will decline further in the second half of the year, and will decline double-digit percentage next quarter.

Embedded revenue was $861 million, down (41%) YoY but up 2% QoQ as inventory levels normalize. Management guided last quarter for Embedded revenue growth to be flat, so the 2% sequential increase is slightly better than expected. Management expects Embedded to gradually recover in H2, and this segment will be up next quarter.

Margins:

AMD’s margins improved throughout, with data center driving a sequential improvement in GAAP operating margin.

• GAAP gross margin was 49% in Q2, up from 46% last year and 47% in Q1. Adjusted gross margin was 53%, in line with management’s guidance and up from 50% last year and 52% in Q1. Management said higher data center revenue was a primary driver of the gross margin expansion in the quarter.
• For Q3, AMD guided for adjusted gross margin of 53.5%, a slight 50 bp QoQ expansion; management has previously pointed to increasing data center mix as a gross margin tailwind.
• GAAP operating margin was 5% in Q2, up from 0% last year and 1% in Q1.
• This was driven largely by data center, which saw operating income rise more than 37% QoQ and 405% YoY to $743 million, for a 26.2% segment operating margin (up from 23.1% in Q1).
• Adjusted operating margin was 22%, up from 20% last year and 21% in Q1.
• Based on management’s expenses guide, adjusted operating margin is expected to come in just above 25% in Q3, a 300 bp QoQ expansion.
• GAAP net margin was 5%, up from 1% last year and 2% in Q1. Adjusted net margin was 19%, up from 18% last year but flat with Q1.

Cash and Debt:

• Operating cash flow was $597 million in Q2, a 10% margin. OCF rose more than 14% QoQ and 56% YoY.
• Free cash flow was $439 million, an 8% margin. FCF rose nearly 16% QoQ and 73% YoY as a result of higher operating cash flow generation.
• Inventory was $4.99 billion, an increase of 7.3% QoQ.
• Cash and equivalents totaled $5.43 billion, while debt totaled $1.72 billion. The company retired $750 million in debt with existing cash this quarter. The company will close Silo AI next quarter for $665 million in cash.

The company returned $352 million to shareholders, repurchased 2.3 million shares with $5.2 billion in share authorization remaining.

Earnings Call:

$4.5B in AI Revenue for FY2024, up from $4B

AMD’s AI accelerator, the MI300, is the fastest ramping product in AMD’s history. I said previously that this is saying a lot as it’s ramping faster than EPYC CPUs, which took a shocking amount of market share from Intel in the data center.

Per a previous write-up:

“My take is that the glass is 30% full and will likely exit the year half-full. Per the call, one analyst’s math is for $900M in GPUs next quarter. If we take $2.4 billion for the DC segment this quarter and assume strong double-digit growth, that puts us at a $3B data center segment next quarter (roughly). If this analyst’s math is correct, this means within two quarters of shipping; GPUs will be 30% of DC segment in Q2. I can’t think of another company that has ramped this fast outside of Nvidia.”If this analyst’s math is correct, this means within two quarters of shipping; GPUs will be 30% of DC segment in Q2. I can’t think of another company that has ramped this fast outside of Nvidia.”

This quarter, AMD seconded this by shrugging off rumors there may be issues with qualifying the MI300: “I think there's a lot of noise in the system. I wouldn't really pay attention to all that noise in the system. I mean this has been an incredible ramp. And I'm actually really proud of what the team has done in terms of just definitely fastest product ramp that we've ever done to $1 billion here in the — over $1 billion in the second quarter and then ramping each quarter in Q3 and Q4.”

EPYC took about 10 years to reach $1.7B in quarterly revenue. AMD will likely reach this quarterly revenue by 2025, or in less than two years with Instinct 300 Series GPUs.

The next MI300 Series release will be the MI325 due out this year with double the memory, and the highly anticipated MI350 will be out early next year to compete with Nvidia’s Blackwell. From there, AMD will continue with a one-year product road map. Look for rack scale systems in the MI350 release next year, which is critical for AMD to keep pace with Nvidia on Blackwell at the hyperscaler and Tier 2 OEM level.

AI Software

The Silo AI acquisition is big news as it will boost AMD’s ability to compete with Nvidia at the enterprise level. We covered the acquisition on our pre-earnings writeup here. Per management: “It's a great acquisition for us. 300 scientists and engineers. These are engineers that have experience with AMD hardware and are very, very good at helping customers get up and running on AMD hardware. And so we view this as the opportunity to expand the customer base with talent like Silo AI, like Nod.ai, which brought a lot of compiler talent. And then we continue to hire quite a bit organically.”

At the Big Tech level, AMD announced that Microsoft has announced the general availability of the MI300X instances. The Azure virtual machines combine AMD’s RocM software platform for “leadership-inferencing price performance.” Hugging Face has adopted the Azure instances “to deploy hundreds of thousands of models on MI300X GPUs with one click.”

On the developer side, Meta’s Llama 3.1 model is supported by MI300 accelerators, Stable Fusion announced they are working with MI300s for their image generation LLM, and AMD supports Flash Attention-2, an algorithm used to enhance efficiency for Transformer models.

RocM is AMD’s attempt to remove the CUDA roadblock the company faces in competing with Nvidia. We’ve covered this here in AMD is Ready to Rival on AI Acceleration. The following update was shared in terms of the progress that’s being made: “the exciting part of this is that the ROCm capability has really gotten substantially better because so many customers have been using it. And with that, what we look at is out-of-box performance, how long does it take a customer to get up and running on MI300. And we've seen, depending on the software that companies are using, particularly if you're based on some of the higher-level frameworks like PyTorch, we can be out-of-the-box running very well in a very short amount of time, like, let's call it, very small number of weeks. And that's great because that's expanding the overall portfolio.”

UALink: Standardizing GPU Interconnects

AMD is being tapped by a consortium of AI acceleration companies, such as Broadcom, Intel, Cisco and Big Tech to assist in creating an Ultra Acceleration Link (UALink) open standard for GPU interconnects to reduce dependency on Nvidia’s NVLink. NVLink is a GPU interconnect that scales GPUs into pods with their own data and computational domain. AMD is being tasked to create an open standard that will serve as an alternative to Nvidia’s NVLink based on AMD’s Infinity Fabric.

The takeaway is AMD is not only viewed as a runner-up to Nvidia, but is actively sought after by the industry to stave off its monopoly. If you read between the lines, this is an important nod to AMD’s capabilities.  Look for more updates in Q3.

AI PCs and Zen 5 EPYC Processors

A major part to AMD’s AI story is laptops, desktops and edge devices.  I can’t emphasize this enough!

The Ryzen AI 300 laptops and the Ryzen 9000 series for desktops are powered by the 5th generation of the Zen architecture. The Ryzen AI 300 laptop has a XDNA 2 neural processing unit (NPU) that is designed for Microsoft Copilot+ AI software. This will deliver 50 TOPS of AI performance. To put this into perspective, the Macbooks with the M4 chip from Apple – considered the most advanced AI laptop on the market – is capable of 38 TOPS of AI performance.

The laptops are already on the market as of now and the desktops will hit the market in August. Management stated investors can expect a strong H2: “As we go into the second half of the year, I think we have better seasonality in general, and we think we can do, let's call it, above-typical seasonality given the strength of our product launches and when we're launching. And then into 2025, you're going to see AI PCs across sort of a larger set of price points, which will also open up more opportunities.”

AMD’s Zen 5 architecture will have 128 cores and 256 thread count and will double the chiplets from eight to 16. The cache is getting a massive upgrade to 512 MB, which was not possible on the Zen 4 architecture at this core and thread count. 

In the data center, Turin EPYC processors will have 192 cores and 384 threads. Per the opening remarks: “We publicly previewed Turin for the first time in June, demonstrating our significant performance advantages in multiple compute-intensive workloads. We also passed a major milestone in the second quarter as we started Turin production shipments to lead cloud customers. Production is ramping now ahead of launch, and we expect broad OEM and cloud availability later this year.” Management stated they believe Turin will help them “continue to grow market share” in the second half of the year.

Conclusion:

At the close of the opening remarks, Lisa Su stated the company is “well positioned to grow revenue significantly in the second half of the year” and “our data center business is on a steep growth trajectory.” These are the words of a company at a fundamental bottom.

There is no doubt, this company ticked every box we have on our checklist this evening. We don’t chase price, rather we look for quality companies. This often means we are early to a move in either direction. You can expect this to be a leading position of ours into the foreseeable future as we patiently wait to see how this bottom unfolds, especially come 2025 for AMD.

Damien Robbins, Equity Analyst for the I/O Fund, contributed to this analysis.

Recommended Reading:

• AMD Q2 Pre-Earnings: The Future Looks Bright
• Hewlett-Packard Enterprise: Sleeper Stock with AI Potential
• Marvell: Tons of AI Potential Obscured by Underperforming Segments

Every quarter, we objectively review our portfolio for our earnings kickoff webinar to determine the strongest and weakest stocks. AMD topped our fundamental checks going into this earnings season due to its expected YoY and QoQ acceleration. The company is expected to post annual revenue growth of 13% for FY2024, accelerating to 28% growth in FY2025.

On a quarterly basis, the upcoming quarter is expected to be the bottom, with growth of 6.8% for June, growth of 14% in the September quarter, with further acceleration into the December quarter at 23% and the March quarter at 33%.

The future looks bright as the Client segment is expected to stabilize and the data center segment is expected to do its thing. Where there’s a will, there’s a way, and AMD is slowly making cracks in Nvidia’s Empire. The single most important announcement this quarter across our portfolio companies was the announcement of AMD’s acquisition of Silo AI. AMD plans to hit Nvidia where it hurts, which can be summarized in two words: open source.open source.

I fully expect AMD’s data center potential to take time to materialize, for the market to go through periods of doubting the stock, and for that to create immense opportunity for our portfolio. Those who have been with us for a while know that we are an incredibly patient analyst team; in fact, we first called AMD an AI stock about 4.5 years ago and the company is only now reporting actual AI revenue for the first time 2024. Per our analysis AMD: 2020 Premium Research:

“Nvidia remains my top AI choice as there is a better moat with the developer platform CUDA (in my opinion). AMD is my second choice in artificial intelligence and I find it fortunate the selloff has given me a second opportunity to build a position at a reasonable valuation.”

I can’t tell you exactly when the stock performance will match it’s AI potential, but I am uniquely skilled at finding semiconductor bottoms. We are at a fundamental bottom for AMD, and I doubt we return to this low of growth for the company for a very long time. 

Financials:

These numbers will be updated Tuesday night with a report hitting your inboxes after hours. For now, here’s a preview of what to expect:

Revenue:

Management guided for revenue of $5.7 billion +/- $300 million, for growth of 6.4% at the midpoint. Analyst consensus is for revenue of $5.72 billion for growth of 6.8%.

• September quarter is expected to report 14.1% for revenue of $6.6 billion.
• December quarter is expected to report 23.2% for revenue of $7.6 billion
• March quarter is expected to report 33% for revenue of $7.28 billion (December quarter is higher due to PC sales).

The rebound is also seen on a fiscal year basis where FY2023 reported growth of (3.9%) for revenue of $22.7 billion.

• FY2024 ending in December is expected to report growth of 12.6% for revenue of $25.5 billion
• FY2025 is expected to report growth of 27.6% for revenue of $32.6 billion
• FY2026 is expected to report growth of 18.4% for revenue of $38.6 billion

Key Segments:

Last quarter, AMD reported data center revenue of $2.34 billion, up 80% YoY and up 2.4% QoQ. The guide for GPUs was originally $2B coming into this year, and the company is now guiding for $4 billion. Per management: “Expect data center segment revenue to increase by double-digit percentage, primarily driven by the data center GPU ramp.”

Another key point is that AMD’s Client segment is expected to increase sequentially. Last quarter, Client reported $1.37 billion, which was up 85% YoY yet down 6% QoQ. This segment has seen nothing but bloodshed for many quarters. Consider that in 2022, AMD peaked at $2.8 billion in quarterly revenue for the Client segment. Management’s guidance communicates that last quarter was the bottom: “Client segment revenue to increase sequentially.” Client segment revenue to increase sequentially.” This is key for AMD’s price action as Client is too big of a hit to offset GPUs ramping.

Gaming continues to weigh on results, reporting $922 million last quarter. Per management: “Based on current demand signals, gaming revenue expected to decline by significant double-digit percentage sequentially.”

Embedded revenue of $846 million was down (46%) YoY and (20%) QoQ. This segment is tied to automotive weakness. Per management: "Given the current embedded market conditions, we're now expecting second quarter embedded segment revenue to be flat sequentially with a gradual recovery in the second half of the year."

Zooming out, this is what management stated to expect for FY2024: “Sequentially, we expect data center segment revenue to increase by double-digit percentage, primarily driven by the data center GPU ramp. Client segment revenue to increase. Embedded segment revenue to be flat. And in the Gaming segment, based on current demand signals, revenue to decline by significant double-digit percentage.”

EPS:

Last quarter, AMD reported adjusted EPS of $0.62 with consensus seeing adjusted EPS turning up from here; meaningfully so in late 2024 and early 2025. This quarter is expected to report $0.68 in adjusted EPS for growth of 17.4% YoY. Over the next two to three quarters, we will see nearly a doubling in adjusted EPS.

• September quarter is expected to report adjusted EPS of $0.94 for growth of 34.7% YoY.
• December quarter is expected to report adjusted EPS of $1.24 for growth of 60.4% YoY.
• March quarter is expected to report adjusted EPS of $1.15 for growth of 85.3% YoY.

Margins:

AMD’s gross margin last quarter was 47% versus Nvidia’s 78%. In 2022, before the AI boom, AMD’s gross margin was 45% versus Nvidia’s 65%. This is one of the reasons Nvidia has historically had a premium valuation. AMD undercuts Intel on price, and this is the strategy with Nvidia going forward, as well.

• AMD reported a gross margin of 47% last quarter.
• Management guided for adjusted gross margin of 53%, and if reported, will be a 100 bps improvement from last quarter. This will also mark the highest adjusted gross margin in two years. This will represent adjusted gross profits of $3.021 billion.
• Last quarter, AMD reported a GAAP operating margin of 1% for operating profits of $36 million. This is very low as AMD had a GAAP OM of 22% in FY2021.
• The adjusted operating margin guide is for 21%, which if reported, will be flat QoQ.
• Net margin last quarter was 2% for GAAP net profits of $123 million. We’ve seen up to a 26% GAAP net margin in FY2020.
• Adjusted net margin was 19% for adjusted net income of $1.01 billion.

Cash:

Last quarter, AMD reported $521 million in operating cash flow for a OCF margin of 10%. This was a nice 400 bps uptick from the previous quarter, which reported a 6% OCF margin. We’ve seen up to a 25% OCF margin for AMD in Q2 2021.

Last quarter, AMD reported free cash flow of $379 million for a FCF margin of 7%. The company has cash and short-term investments of $6.03 billion and debt of $2.46 billion.

Stock based compensation is 7% of revenue.

Silo AI Acquisition

We had stated on our most recent webinar that we want to give AMD the space to fill the very big shoes an Nvidia contender has to fill. A good example of AMD playing the long-game is the acquisition of Silo AI for $665 million. This is Europe’s largest AI-related acquisition, sizably larger than the acquisition of DeepMind by Google for $400 million in 2014.

The company is known for its pool of AI talent, with experience in training large language models on AMD Instinct GPUs. These custom, open source LLMs called Poro and Viking are multilingual and can be customized and applied to many end-markets. Poro is a 34 billion parameter model that is cross-lingual for Europe’s 24 official languages and offers AI sovereignty by allowing companies or countries to create proprietary models. Viking is a 7 billion parameter model and highlights Silo AI’s unique approach in developing smaller LLMs for Nordic languages. These low-resource languages lack large training data sets. There are also 13B and 33B parameter Viking models, but the point is to not have to use hundreds of billions of parameters or even trillion+ parameter models being developed by OpenAI and Deep Mind/Google’s Gemini. Instead, Silo AI rivals LLMs such as Mistral and Meta’s Llama in English, yet processes multiple Nordic languages and programming code.

The official announcement for the 7B Viking LLM provides a clear message on why Silo AI was acquired by AMD: “With a purpose-built software layer to train models on AMD, Silo AI and TurkuNLP possess unmatched experience with training on AMD at scale, having shown that their theoretical predictions for throughput scaling materialize in weak and strong scaling experiments. As one of the seminal initiatives on AMD GPUs, this shows how it’s possible to achieve good throughput on the AMD-based LUMI, training the models with their open source training framework and utilizing up to 4096 MI-250X GPUs simultaneously.”

Our original thesis on Nvidia centered around the CUDA moat. This moat is fully in tact today, and has helped Nvidia enjoy unrivaled pricing power. Silo AI greatly speeds up AMD’s open-source software effort, which is the critical piece to AMD’s strategy as CUDA is closed-source and proprietary.

As the Forbes article points out, Hugging Face has partnered with AMD to run AI models on Instinct GPUs. Meta and OpenAI have ordered AMD’s new GPUs. From there, these companies can also open source their frameworks and models to help speed up time to market for smaller teams.

These large R&D departments are sophisticated enough to circumvent CUDA and program custom silicon or program competing GPUs if the total cost of ownership (TCO) presents a compelling reason. AMD’s MI300s go for $15,000 and as low as $10,000 when sold in bulk. Meanwhile, Nvidia’s GPUs go for an average of $35,000.

I first covered this in March of 2020 when our analysis pointed out: “It’s estimated that for every $1.00 in Rome chip sales, Intel loses $2.25 on average in Intel Xeon SP sales. The savings are then deployed to buy more Rome chips, which can further depress Intel’s revenue.”$1.00 in Rome chip sales, Intel loses $2.25 on average in Intel Xeon SP sales. The savings are then deployed to buy more Rome chips, which can further depress Intel’s revenue.”

In the July of 2023 I/O Fund analysis: “AMD is Ready to Rival on AI Acceleration” it was pointed out:

“From there, AMD undercuts Intel on price, which becomes a virtuous cycle as driving down costs means more chips will be bought from AMD. […] In the past, AMD advertised up to 20% Capex savings compared to Intel based on Epyc processors delivering more performance from a single chip compared to Intel’s dual-processor powered by two CPUs. Big Tech has capex budgets into the tens of billions. Although it’s not specifically disclosed exactly how much goes toward AI acceleration, we know that Big Tech is driving forward Nvidia’s GPU sales at $8 billion per quarter or $35 billion to $40 billion per year.$8 billion per quarter or $35 billion to $40 billion per year.

Here is the thesis in a nutshell: If a competitor can deliver 20% savings on this kind of budget with similar performance, then it will turn heads. We can geek out all day long on the computing performance of Nvidia’s H100 GPU, however, if the MI300s drive down total cost of ownership through low unit pricing, better power efficiency and reducing the number of GPUs required, then hyperscalers will line up to support this.

What Google, Amazon, Microsoft, Meta and large enterprises want most of all is to build incredible AI systems but at a manageable cost. This goes back to the virtuous cycle. The more they save, the more they can build.”

As Big Tech becomes pressured over their capex spend, it will only be natural that AMD is evaluated as an option to help alleviate this massive AI infrastructure spend.

Conclusion:

The future is bright for AMD. This company has what it takes to make cracks in the Nvidia GPU data center Empire. For our purposes, we think AMD could take up to 20% market share of the GPU data center, although it would take up to a decade for this to materialize. We are basing this estimate on what AMD has achieved in gaming GPUs, and the market share dynamic 80/20 on data center CPUs with Intel.

Equally important, AMD is a frontrunner on the Client side for AI. The company spans both x86 and Arm architectures for AI devices, and has the Xilinx acquisition waiting in the wings once automotive heats up again.

In our webinar, we had stated that a great tech story should have financials to match. AMD ticks this box, as well, with a rebound into the second half of this year and early next year. You will get an update from us post-earnings that looks at the gritty details of the report. But let me emphasize well ahead of time that we are in no rush for AMD’s AI story to materialize, as our sights our firmly on the horizon.

Recommended Reading:

• AMD Q1 Earnings: GPU Revenue Outlook Raised to $4B
• Hewlett-Packard Enterprise: Sleeper Stock with AI Potential
• Lam Research FQ4 Earnings Preview: Eyes on 2025 Outlook
• Marvell: Tons of AI Potential Obscured by Underperforming Segments

Marvell is one of the earliest semiconductor stocks we’ve covered on our site, dating back to November of 2019 when we first covered application-specific integrated circuits (ASICs), commonly known as custom slicing. Despite having a clear AI story, Marvell has lagged other AI stocks over the past year:

Our last position in Marvell was bought at $56.90 in June of 2023 and closed at $77.19 in March of 2024. Not bad, but not great. At the time of closing the stock, it was becoming clear that Broadcom was the stronger near-term story when my last analysis stated: “As I left the Marvell call and moved along to join the Broadcom earnings call, there is no doubt which company is stronger right-here, right-now. It’s Broadcom. Marvell has a strong product story but it’s in a sea of AI whales that are ramping quickly.”

Despite closing Marvell, the stock remains on the list of our top 10 ideas. There is abundant AI potential buried by other segments that are in a steep, cyclical trough. As we look on the horizon, CY2025 has the makings of a solid comeback for this often-overlooked AI stock.

Marvell’s Fiscal Q1 2025 Financials:

For fiscal Q1 ending in April, Marvell reported revenue growth of (-12.2%) for revenue of $1.16 billion. This marginally missed estimates by (-0.1%). The revenue growth is the lowest since we’ve tracked the stock, dating back to 2021. According to analyst consensus, this should mark the bottom with sequential growth of 8% next quarter.

For fiscal Q2 ending in July, management guided revenue of $1.25 billion, at the midpoint, representing a decline of (-6.8%). According to consensus, this will be the first quarter to report sequential growth of 7.8%, and the sequential growth is expected to continue.

Here is some more information on the rebound that is materializing:

• Fiscal Q3 ending in October is expected to report (-0.9%) YoY for $1.41 billion, which will represent QoQ growth of 12.8%.
• Fiscal Q4 ending in January is expected to report 11.3% YoY for $1.59 billion, which will represent QoQ growth of 12.7%
• Fiscal Q1 ending in April is expected to report 39.76% YoY for $1.62 billion, which will represent QoQ growth of 1.9%.

When looking on a fiscal year basis, it’s easy to see that Marvell’s stock is struggling due to cyclical segments. This is not unique to Marvell as the recovery in consumer electronics, automotive, and telecom has taken longer than anticipated. As a reminder, these segments surged during the pandemic and during a long period of quantitative easing. Now, semiconductor companies are collectively weathering a deep trough that began in CY2022.

For FY2025 ending in January, analyst consensus is for (-1.88%) on revenue of $5.4 billion – yet, twelve months ago, FY2025 estimates were for growth of 17.8% for revenue of $8.40 billion. What’s interesting is that AI is doing better than expected, and it’s the other segments that created the twelve-month disparity.

For FY2026, the estimates have gone up but not due to higher revenue, rather due to lower comps. The growth rate of 32.6% for FY2026 is expected on revenue of $7.17 billion. This is a bit lower than the $7.52 billion expected for this fiscal year twelve months ago.

What this situation represents is a lack of confidence in both management’s tone and analysts’ financial modeling in predicting when consumer-driven segments will see a sustained recovery.

Marvell is not GAAP profitable due to recent acquisitions and the related costs, and also stock-based compensation sits at 11% of revenue.

In the most recent quarter, the company reported GAAP EPS of ($-0.22) which missed estimates of (-$0.25). Adjusted EPS of $0.24 was in line. According to analyst estimates, this is expected to be the bottom with sequential growth beginning in the July quarter.

For the upcoming quarter ending in July, the company is expected to report adjusted EPS of $0.30, representing a YoY decline of (-9.78%).

Here is what the rebound looks like on the bottom line. We can reasonably assume Marvell will be GAAP profitable again sometime during FY2026. Notably, this depends on the other segments as growth in AI accelerators (custom silicon) weighs on margins.

On a fiscal year basis, Marvell is expected to see the following:

• FY2025E adjusted GAAP EPS of $1.40 for a decline of (-7.5%)
• FY2026E adjusted GAAP EPS of $2.46 for growth of 76%
• FY2027E adjusted GAAP EPS of $3.33 for growth of 35%

Key Segments:

For the past two quarters, the data center has been growing rapidly, and has reached a historical high. This is notable given Marvell completed a large data center-focused acquisition a few years back (Inphi) which provided immediate, accretive data center revenue.

Data center revenue in the current quarter was $816.4 million, up 87% YoY and up 7% QoQ. This is on the heels of another historic data center quarter of $765.3 million, up 54% YoY and up 38% QoQ. The data center outperformance comes from electro-optics and interconnect products, whereas custom silicon saw “initial shipments” in the quarter. Looking to next quarter, management expects data center to grow in the mid-single digits “as our custom AI silicon continues ramping.” It was mentioned on the call that optical interconnects are up against a tough comp, and thus, will be flat QoQ but will still perform well YoY.

“I'd say in the short-term, the way to think about the optical business into July is we're modeling it right now and our guide is flattish to slightly up. And the reason for that is we outperformed pretty big both in Q4 and Q1. […] So as we look into July, we're modeling it to be flat to slightly up, it may do better, let’s see order trend come in. But year-over-year will be very strong because also in the second half to your point, those traditional standard cloud infrastructure build-outs and upgrades are going to happen.”

Of this, the company is expected to exit the year with a minimum of $1.5 billion in AI revenue in FY2025. About a year ago, we had published that Marvell was on track to report 14.4% in AI revenue when the company doubled its AI expectations to $800 million, up from $400 million.

With the current update of $1.5 billion provided in April at the AI Investor Day, the company is now on track to report 27.8% in AI revenue. Per management comments, the $1.5 billion is a “floor” and there was discussions in the Q&A on the likelihood the FY2025 exit rate will be higher in the next few months.

Brace yourself, however, as the other segments are deep in the red:

• Carrier infrastructure was down (75%) YoY and down (58%) QoQ for $72 million. Carrier infrastructure is expected to be flat sequentially next quarter. According to commentary, the recovery in this segment is harder to predict than the others. The company is shipping a new 5nm DPU product next year that is expected to help expand 5G market share.
• Enterprise networking was down (58%) YoY and down (42%) QoQ for $153 million. Enterprise networking is also expected to be flat sequentially. The recovery is expected to begin in the second half of this fiscal year.
• Consumer was down (70%) YoY and down (71%) QoQ for $42 million. This segment has been weighed down from a soft gaming market, yet Marvell’s primary customer is expected to rebound and the segment is expected to double on a sequential basis.
• Automotive was down (13%) and down (6%) QoQ for $78 million. This segment is expected to be flat sequentially yet will resume growth in the second half of the fiscal year.

Margins:

• Gross margin in the current quarter of 45.5% is low and there were questions on the call about this (see below). Ultimately, the AI story weighs on Marvell’s gross margins but does not affect the operating margin. The guide for next quarter is gross margin of 46.2%. This will represent gross profits of $577.5 million.
• Adjusted gross margin of 62.4% in the most recent quarter with a guide of 62% next quarter is low compared to the historic adjusted gross margin in the 65% range. Next quarter, adjusted gross profits are expected to be $775 million.
• GAAP operating margin last quarter was (13.1%) and this is certainly a blemish in the report. Next quarter, GAAP operating margin is expected to be (8.8%) for a GAAP operating loss of $110.5 million.
• Adjusted operating margin of 23.3% last quarter is lower than the historic adjusted OPM in the mid-30% range. Adjusted operating margin in the upcoming quarter is expected to be 25.6% for adjusted operating profit of $320 million.
• Net margin last quarter was (18.6%) and adjusted net margin was 17.8% for adjusted net profits of $206.7 million.

Cash Flow:

Cash flow for Marvell is decent yet the debt-to-equity ratio is high.

In the most recent quarter, the operating cash flow was $324.5 million for a margin of 28%. The free cash flow was $232.5 million, for a margin of 20%. This is lower than usual due to annual employee cash bonuses. Inventory was $826 million, decreasing $38 million from the prior quarter. On a year-over-year basis, inventory has been reduced by $200 million or 20%. Days sales outstanding decreased 8 days to 69 days.

The company has $847.7 million in cash on the balance sheet and has $4.15 billion in debt. The company’s net debt to EBITDA ratio is 1.8X and the gross debt to EBITDA ratio is 2.27X.

In the recent quarter, the company returned $52 million to shareholders through cash dividends. The company also repurchased $150 million of our stock during the first quarter, an increase of $50 million from the prior quarter with expectations to increase repurchases in Q2.

Quick refresher on Marvell’s Products:

Marvell offers 200-gig, 400-gig and 800-gig PAM-based electro-optics. The 800-gig is the primary interconnect for AI deployments. The company is qualifying a 1.6T solution with 200-gig per lane for the next leg up in AI acceleration. For the 1.6T solution, Nvidia will be a lead partner. Here’s a video on Marvell and Nvidia’s partnership on optical interconnects.

Electro-optics help to increase data rates and has replaced NRZ data transmission due to doubling the bit rate. Hyperscalers require high bandwidth and port density. PAM4 connects networking ASICs and machines, like servers and AI machines. Digital-based PAM4 uses analog-to-digital converters to clean up the signal in the digital domain before converting it back to analog to transmit.

Artificial intelligence and machine learning drive demand for the 800-gig PAM to increase the speed of input-output and to process the data flows. This doubles the throughput (bandwidth) due to an 8x100Gpbs optical transceiver for inside and between AI clusters.

In the most recent earnings call, Marvell discussed their plans to compete in the PCIe Gen 6 retimer interconnect market. PCIe 6.0 will be the first to use PAM4 signaling technology. Marvell is sampling eight and 16 lane PCIe 6.0 retimers with customers, which will help data center compute fabrics scale. Per management: “AI applications are driving data flows and connections inside server systems at significantly higher bandwidth, driving the need for PCIe retimers to meet the required connection distances at the faster speeds.”

Marvell also offers data center interconnect (DCI) products, which connects data centers over various distances to transfer data, content and critical assets. COLORZ silicon photonics increase the speed of data movement while keeping power and cost low. The 400 gig ZR and 800 gig DCI products with coherent DSP (digital signal processor) extends the reach to 1,000 kilometers.

Teralynx are ethernet switches with the 800 Gb/sec Teralynx 10 built for cloud data center and AI fabrics. The company also provides Ethernet controllers and PHY transceivers, and is a competitor to Broadcom on switch ASICs in that regard. Teralynx and Broadcom’s Tomahawk will be in lock-step for the release of 1.6Tb switch ASICs.

Custom silicon refers to ASICs or application-specific integrated circuits that are customized to be “application-specific” with the benefit of becoming cheaper with volume production. ASICs are expensive at the onset, yet become cheaper with volume production. Custom silicon is attractive to Big Tech as cash is not an issue with these companies for ASICs very high startup costs (well into the millions). Big Tech also immensely popular applications to justify the non-recurring engineer (NRE) costs in developing chips for a specific purpose.

Across ASICs, the most well-known is Google’s tensor processing unit (TPU). Yet, there is a vast array of custom silicon that has hit the market since TPUs were first introduced in 2016 for the TensorFlow framework. Amazon was second to diversify with custom silicon for AI workloads with Graviton and Inferentia in 2018, and the more recent Trainium announced in 2020. Last year, Microsoft announced the 5nm Maia 100 AI chip to reduce dependency on data center GPUs, and a Cobalt 100 Arm-based CPU to increase the performance on Azure-based virtual machines for scaling web applications, microservices and open-source databases. We covered in our 2019 Marvell analysis that Microsoft was pursuing FPGAs (Xilinx), but FPGAs have now been replaced with ASICs, which is what Marvell and Broadcom offer.

Discussions on AI Revenue:

Naturally, there were questions on the guidance for $1.5 billion in AI revenue exiting the fiscal year. Regarding the current quarter, one analyst is modeling for $500 million per quarter in AI revenue.

When pressed, management hinted this is the minimum number to work with: “And then, the whole thing in flex meaningfully in the second half and I'd say from a full year perspective, the way to think about it, maybe some additional color would be, we talked about a floor of $1.5 billion for AI revenue for Marvell for this fiscal year with about two-third in electro-optics and a third in custom. . And we see now both of those exceeding that number.”

Where the market could get excited is if Marvell’s custom silicon surprises to the upside. As of now, it’s expected to contribute one-third of the $1.5 billion quoted above next year. Marvell’s custom AI silicon business is beginning to ramp and investors will see more evidence of this in the second half of this year. Per the opening remarks: “Our custom compute AI programs are beginning to shift in the first half of this fiscal year and we are expecting a very substantial ramp in the second half of this year, followed by a full year of high volume production in fiscal 2026.”

The market for custom silicon is expected to grow from $7 billion in CY2023 to $40 billion in CY2028 at a 45% CAGR. My comment is that this is probably too low; as the market is too nascent to accurately predict a higher number.

Outside of custom silicon, Marvell’s management expects the aggregated data center opportunity to grow at a CAGR of 29% from $21 billion to $75 billion. There was a comment that management predicts they will double their market share from 10% to 20%: “We see a massive opportunity ahead with the data center TAM expected to grow from $21 billion last year to $75 billion in calendar 2028 at a 29% CAGR, we have numerous opportunities across compute, interconnect, switching and storage, as a result, we expect to double our market share over the next several years from our approximately 10% share last fiscal year.”

That’s quite a statement as it implies Marvell’s data center segment will grow from $3.2 billion today to $15 billion in the next several years. Given they tied the statement to the 2028 projection, then this implies a 368% growth rate on the data center segment over the next 4 years.

Notably, the statement was repeated in the Q&A: “So we articulated the AI day, a very robust custom silicon TAM in excess of $40 billion going out into 2028 time frame and that TAM growing very significantly. And yes, we — I think your numbers are about right in terms of the share. We're going to end up with near-term and then Raghib articulated our goal to drive that in the custom silicon area to 20%. So you got to draw a line kind of from here to there in terms of the opportunity.”

Looking forward, management stated the floor for next fiscal year on AI revenue is $2.5 billion, up from the $1.5 billion, as custom silicon is expected to see its first full year of volume.

On the topic of custom silicon expanding next year, this will weigh on the gross margin, yet it will help to drive a strong operating margin, primarily due to non-recurring engineering costs.  

Conclusion:

We have another attempt at Marvell in the works, and we are looking closely at timing. On one hand, we may be too early and have to deal with a couple of earnings reports that are duds until we get to the rebound in 2025. On the other hand, Marvell may start to move quicker than current consensus is forecasting as it’s participating in a few explosive trends.

How the market perceives the non-AI segments until we get a material recovery in these segments is anyone’s guess. In a risk-on environment, these segments will be dismissed and the number of times a management team mentions the words “AI” on an earnings call is all that matters. In a risk-off environment, Marvell’s unfortunate exposure to telecom and gaming will mute the upside.

To put it simply, we are cautiously optimistic on Marvell. Over the next few months, we plan to revisit if we see a break above $79.

I/O Fund Advanced Members receive trade alerts and weekly webinars to discuss our entries and exits. Learn more here.Learn more here.

Recommended Reading:

• Q3 2024 Earnings Kickoff Webinar Replay
• Arista Networks: Ethernet AI Networking Opportunity
• Revisiting TSMC: Sales Update, Buy Plan
• Nvidia Q1 Earnings: “We will see a lot of Blackwell revenue this year.”

Recently, the I/O Fund team wrote an article in Forbes: “AI Power Consumption: Rapidly Becoming Mission-Critical” where it was stated: “Over the past few months, multiple forecasts and data points reveal soaring data center electricity demand, and surging power consumption. The rise of generative AI and surging GPU shipments is causing data centers to scale from tens of thousands to 100,000-plus accelerators, shifting the emphasis to power as a mission-critical problem to solve.”

As the analysis points out, eventually we will see million-plus accelerator data centers. Thus, AI’s electricity demand is forecast to surge, especially in the data center. Morgan Stanley’s base case is calling for 500% increase in power demand over the next three years. Wells Fargo is projecting AI power demand to grow 550% by 2026, before rising another 1050% by 2030, from 8 TWh to 652 TWh in a seven-year period.

Liquid Cooling plays an important role in reducing the heat that AI systems generate. We first covered Liquid Cooling in a Super Micro analysis in May of 2023, yet would like to double-click on this trend for our premium members as it’s becoming what we consider to be “the third realm of competition.” Liquid cooling technology has been around for decades, yet this technology is becoming mission critical due to the increasing levels of compute power from AI accelerators, starting with the GB200 systems and B200 GPUs.

Per the previous analysis: “In 2022, Supermicro stated that liquid cooling is being used in 10% of supercomputers but will grow to be used in the “vast majority” in order to offset the heat generated by power-consuming components.”

Although the GB200 will ship end of this year, and the B200 will fully ship in early 2025, vendors are scaling their liquid cooling capacity now. The capacity investments are being made right now, and we can see evidence of this in Super Micro’s most recent earnings report with an increase in inventory. We also find hints of this in Dell’s earnings report, with the company also reporting an increase in inventory as these leading AI server companies wait for Nvidia’s Blackwell to ship.

Last week, we clearly outlined that AI power consumption is a problem the industry must work diligently to solve. For our premium members, we take this further to discuss how liquid cooling is at the forefront of driving down energy requirements for AI systems. Below, we look at the beneficiaries of this important trend and what we are looking for in managing our positions – including our plans for re-entering Super Micro, how we view Dell given the weak earnings report, plus some brief analysis on Hewlett Packard Enterprises (HPE), Vertiv, to name a few.

Brief Overview of Liquid Cooling:

Liquid molecules are closer together than air molecules, which results in higher heat transfer. This results in liquid removing 4 times more heat than air. The heat capacity of water or glycol is higher than air, so the amount of heat being transferred is higher. Most servers today are air-cooled, yet AI necessitates a shift to liquid cooling are GPUs are already at 700W of power and are moving toward 1000W of power.

AI/ML require massive amounts of data processing, and as future generations of CPUs and GPUs are released, these systems will exceed air cooling capacity. Liquid cooling also solves throttling, which occurs when CPUs and GPUs overheat and are throttled back to avoid damage to the chip. In the case of high-performance computing, liquid cooling reduces total cost of ownership as air cooling requires air conditioning and server fans to run constantly.

Cooling data center servers is responsible for 40% of the data center energy consumption. According to Dell, enclosed DLC solutions can save up to 23% of energy compared to traditional air-cooled racks. McKinsey places this number at 27% savings when there is 75% liquid cooled and 25% air cooled servers.

There are a few different approaches to liquid cooling, such as:

• Direct Liquid Cooling: DLC uses liquid-cooled cold plates that in direct contact with GPUs and CPUs. The cold plates transport heat away from the processors. The process of circulating liquid directly over the components is also known as Direct to Chip Cooling, and is a closed loop system, or also known as a self-contained cooling system.
• Immersion Cooling: The system is immersed in liquid for cooling. The immersion tank has a coolant distribution unit, including a pump to circulate dielectric fluid to extract heat from the servers.
• Rear Door Heat Exchanger: Uses a specialized rear door to the rack where coolant absorbs the heat.
• In-row cooling refers to solutions designed to cool and distribute air in a data center aisle. When combined with row or rack containment, the in-row coolers capture 100% of the IT-generated heat.

In addition to lowering power consumption, benefits from liquid cooling includes higher server density as the need to create space for airflow is removed. Liquid cooling also eliminates hot spots with more even distribution. The lower temperatures from liquid as opposed to air also extends the life of the server (removes throttling).

Nvidia’s Blackwell is Hot

Nvidia’s A100 released in 2021 operates at 300W, the H100s released in 2023 operate at 700W. The Blackwell architecture is a catalyst for liquid cooling as it nears 1000W, specifically the GB200 systems and the B200. This represents a 40% increase from the previous generation. Tom’s Hardware makes the argument that: “we can only refer to the basic rule of thumb with heat dissipation, which says that thermal dissipation typically tops out around 1W per square millimeter of the chip die area” and that “When it comes to high-performance AI and HPC applications, we need to consider the performance measured in FLOPS and the power it takes to achieve these FLOPS and cool the resulting thermal energy. What matters for software developers is how to use those FLOPS efficiently. What matters for hardware developers is how to cool the processors producing those FLOPS.”

Therefore, as computing power increases with each GPU generation, which is measured in FLOPS, cooling the GPUs is the crux, as it becomes a larger thermal dynamics issue that must be solved.

In terms of timing, the B100 is due out first and will be primarily air cooled. The B200 systems and chipsets will be the first release to be primarily liquid cooled. This is due to consuming upwards of 1,000 watts, which is too hot to be air cooled. The B200 doubles the transistor count compared to the H100 and provides 20 petaflops of AI performance compared to the H100s 4 petaflops. The resulting 3X leap in training performance and 15X leap in inference performance is shifting the focus to ways that power consumption can be lowered. Note: We’ve covered Nvidia’s upcoming Blackwell extensively, please see resources below.

The B200 will technically be out in late 2024 as a system that combines either 36 GPUs or 72 GPUs with Nvidia’s in-house Arm-based Grace Hopper CPUs. These superchips are called the GB200 NVL36/NVL72 and will operate as one supercomputer, allowing for trillion-plus parameter models to be trained.

The B200 chipset will ship in early 2025. As stated in a previous analysis on Nvidia, the B200 chipset will offer “a second-generation transformer engine that supports 4-bit floating point (FP4) with the goal of doubling the performance and size of models the memory can support while maintaining accuracy […] This is helpful because AI models are moving toward neural nets that lean on the lowest precision and yet still yield an accurate result. In this case, 4 bits double the throughput of 8-bit units, compute faster and more efficiently, and they require less memory and memory bandwidth.”

Nvidia offers SuperPODs that combine eight H100 GPUs connected with NVLink with the DGX SuperPOD connecting 32 nodes of eight GPUs for a total of 256 H100 GPUs. As the B200s come out, these SuperPODs will scale to provide more than 1 exaflop of AI compute power and move from FP8 precision to FP4. The DGX GB200 SuperPODs will connect up to tens of thousands of GB200 superchips with shared memory.

This section is included to help cement the inevitability that liquid cooling is becoming what Damien Robbins, equity analyst at the I/O Fund dubbed the third realm of competition.

Recent Commentary on Liquid Cooling:

At BofA’s Global Technologies Conference in early June, analyst Vivek Arya questioned Nvidia’s VP of Accelerated Computing Ian Buck about the increasing power requirements per GPU. Since liquid cooling’s catalyst begins with Nvidia’s second release in the Blackwell generation (GB200 and B200), it’s important for us to examine what is being said before we look more closely at the beneficiaries.

Vivek Arya:

How is the outlook around Blackwell as we look at next year? First of all, do you think that because of the different — the power requirements that are going up significantly, does that constrain the growth of Blackwell in any way?

Ian Buck, Nvidia:

“Data centers don't drop out of the sky. They're big construction projects. [Customers] need to understand what is a Blackwell data center look like and how is it going differ than Hopper. And it will. The opportunity we saw with Blackwell was to transition to a denser form of computing, to put 72 GPUs in a single rack, which has not been taken to scale before. … In Taiwan, for example, the people that are building the liquid cooling infrastructure, the power shelves, the WIPs, which is the cables that go down into the bus bars. The opportunity here is to help them get the maximum performance through a fixed megawatt data center and at the best possible cost and optimized for cost. By doing 72 GPUs in a single rack, we need to move to liquid cooling. We want to make sure we had the higher density, higher power rack, but the benefit is that we can do all 72 in one NVLink domain.

Connect them all up with copper instead of having to go to optics, which adds cost and adds power. And every time you add cost and power, you're just taking away from a number of GPUs you can put in your 10-, 50-, 100-megawatt data center. So that is driving us towards reducing cost, increasing density.

So when you look at a Blackwell, you may say, well, it's really hot, that's actually going to be significantly improving the total throughput of a fixed power data center. So there's a strong economic and technological driver to transition to more denser and more power efficient and more — and next-generation cooling technologies than just air.

Water is a fantastic mover of heat. Your house is built with insulation that is nothing more than just trapping air. Air is actually an insulator. It's not a good transfer to heat, but water is excellent at it. If you ever jumped from a 70-degree pool from a 70-degree air, it feels really cold.

That's because water is sucking the heat right out of you. It's really good at moving heat around. And that efficiency goes right to more GPUs, more capabilities and denser, more capable AI systems.”

–End Quote

Buck’s response did not directly touch upon increasing power requirements, but hinted that this shift to Blackwell and beyond, with larger racks and denser compute, will be built upon liquid cooling, which will allow these more powerful and power-hungry GPUs to operate efficiently at scale.

Nvidia CEO Jensen Huang also discussed liquid cooling in Q1’s earnings call: “the Blackwell platform has expanded our offering tremendously. The integration of CPUs and the much more compressed density of computing, liquid cooling is going to save data centers a lot of money in provisioning power and not to mention to be more energy efficient. And so it's a much better solution.”The integration of CPUs and the much more compressed density of computing, liquid cooling is going to save data centers a lot of money in provisioning power and not to mention to be more energy efficient. And so it's a much better solution.”

Super Micro: Leader in Liquid Cooling

Super Micro’s ascent in the server market has been breathtaking. In 2021, the company ranked #6 among server companies with $2.5 billion in revenue compared to Dell’s $14 billion in server revenue and HPE’s $12 billion. Fast-forward and SMCI is on a direct path to reaching $25 billion in revenue over the next year. That’s a 10X revenue increase in about 3-4 years’ time.

Super Micro expects liquid cooling to be rapidly adopted over the next year and a half. The company is deploying three of the “world’s largest DLC liquid-cooled” systems in the current quarter, ending in June. The Nvidia HGX AI supercomputers with liquid cooling are expected to “potentially” save customers up to 40% of energy costs compared to air-cooled systems.

SVP and CFO, David Weigand, explained at BofA’s conference:

“we have started to ship liquid cooling at really at scale, at larger volumes in this core. And so, there's no question that that industry is coming up to speed with the reality of where we're going, which is the fact that power is constrained all around the world. And then — and therefore, when you build these large data centers, you're going to have to think twice now about using liquid cooling, because by you using liquid cooling, you can not only — we say that it's free with a bonus, and that's because it's free because you're not only having to put in smaller chillers, you don't have to use air conditioning.

If you have really liquid cooled racks, you can put more dense racks and more racks into a data center, so it's more efficient if you're using liquid cooling. And so, it's really the cutting-edge companies right now that are putting in liquid cooling, liquid cooled racks into their data centers so that it's — the huge use of power right now is going to really drive liquid cooling. As much as the fact that all the GPUs and CPUs are running at higher wattage as they go over 1000, it's going to start to become painfully obvious.”

–End Quote

SMCI’s management has stated that liquid cooling will cost more as it takes longer to assemble and test, and the company plans to charge for this. It’s also expected that SMCI will be the first to ship liquid cooled AI systems before its competitors.

At a recent investors conference, management stated the company has a rack capacity of 5,000 per month and a liquid cooled rack capacity of 2,000 racks per month. This means that SMCI plans to utilize direct liquid cooling (DLC) in 30% of its racks, to start.

According to the CFO, the Malaysia site will offer the opportunity to “double eventually our worldwide capacity” and will offer both air cooled and liquid cooled servers. With this in mind, Super Micro CEO Charles Liang expects direct liquid cooling (DLC) adoption to reach 15% in the next 12 months and 30% over the next two years, a rapid shift up from 1% of the market. The CEO updated this on xAI, stating they now foresee DLC adoption growing from <1% to 30% in a year.

Super Micro is Seeing Higher Inventory Levels and Lower Cash Levels

Per our recent SMCI earnings write-up, inventory increased to 92 days compared to 67 days in the previous quarter. The company’s Q3 closing inventory was $4.1 billion, which increased by 67% quarter-over-quarter from $2.5 billion in Q2 due to the “purchase of key components.” Our post-earnings analysis explained that the increase of inventory and key components was partly related to liquid cooling.

The CEO stated: “Two reasons we had to increase inventory: One is because Q4, I mean, June quarter, we will have a strong revenue growth; a second reason because we're preparing for high-volume liquid cooling. Again, we have more than 1,000 of 100k watt, I mean, liquid cooling rack we have to ship to customers in Q4. And liquid cooling as you know, is pretty new. So we had to prepare enough inventory so that we can deliver liquid cooling rack scale product to customer on time or with minimal lead time. So both factor, indeed, is a positive factor. And with our economic scale continuing to grow, indeed, our inventory average [ daily ], indeed, will slightly improve.”

Super Micro has further discussed its plans to fund its capacity expansion efforts through short-term debt, or perhaps by diluting shareholders. In the last earnings writeup, I called cash Super Micro’s “Achilles heel” as the seemingly invincible company has one drawback – in order to keep growing, they must build more facilities, which requires more cash. Liquid cooling also necessitates more components, which means inventory levels will rise. This combination is putting a wrench in SMCI’s cash flow.

Ruplu Bhattachary:

David, let's talk about working capital. So, to support growth, you need to support a lot of working capital. When do you make the determination, or how do you make the determination that you need to raise more capital? And how should investors think about your trade-off between using more debt or doing another equity raise?

David Weigand:

Yeah. We've had to do a couple of raises in the past six months, because we saw the permanent level of our business going up higher. So it wasn't temporary. Now, remember, as a manufacturer, if we sell a billion dollars, an additional billion dollars in a quarter, we have to. And remember, when I first started, we were doing about $3.5 billion a year, and now we're doing well, more than that per quarter.

So, when you're increasing by a billion dollars in a quarter, you've got to go out and you've got to, let's say even if the margin is 20%, you have to buy 80% of materials and you have to carry those through inventory. You have to carry them through accounts receivable until they convert to cash. And so it becomes an immediate problem. And we've had some very large customers come that I've sat across the table from, and they say, we have two questions. Do you have the capacity, do you have the capital to take this project on? And so we had to go out and get some more permanent capital so we could answer that question always, yes. So, we finished out with $2 billion at the end of last quarter.

We think that the things that we've done in terms of raising the visibility of the company, raising the profits, raising the sales, have been good for our shareholders, and we want to continue to balance that because we don't like dilution. We previously used to repurchase shares that's still in our tool bag. But right now, it's about being able to deliver against our backlog. And so therefore, we will get as much capital as we need to in order to do that. And so, it's really about whether you see, with short term debt, we can address temporary increases, but if we see sustained orders such as we have seen, then we're going to have to do some more permanent debt raises like we've done with the convertible bonds and also with the common stock equity raise.

–End Quote

The rise in inventory due to liquid cooling components plus Super Micro needing to increase capacity to further meet AI server demand may lead to a lower entry price, which we will gladly take.

Barron’s published just today that SMCI is the top performing stock in the S&P 500 for the first six months of the year. This is on the heels of being the second-best performing stock of 2023, ending the year a tad bit higher than Nvidia. We’ve participated since mid-2023 for a roughly 300% return in less than a year. We have plans to re-enter Super Micro which can be found here.

Dell

Dell has a long way to go to catch up with Super Micro as the company reported $2.6 billion in AI-optimized server revenue and AI server backlog of $3.8 billion. This represents 7.6% of Dell’s revenue. Compare this to Super Micro with over 50% of its revenue from AI and this number is surely higher today.

Due to Dell’s scale, it will take some time before Dell sees this level of AI concentration, as Client revenue is a large portion of Dell’s overall revenue. Therefore, for Dell to become a full-fledged AI stock, it will need AI PCs to participate. It’s only a matter of time before AI PCs provide the next leg up for AI investors, with our best guess being 2025 on the early side and late 2026 on the late side.

Dell may have a long way to go to see the levels of concentration that Super Micro has, but AI also has a long way to go. In our Dell write-up, the base case is for 15% of Dell’s revenue to be from AI, yet the more likely outcome is we will see something in the 30% range by 2027. This does not factor-in AI PCs which will rapidly accelerate this percentage once the trend is in play.

It's doesn’t require much speculation to think Dell will be the runner-up when SMCI reaches capacity. Jensen Huang of Nvidia recently stated: “Everybody who is building these chatbots and Generative AI, when you are ready to run it, you need an AI factory and nobody is better at building end-end systems of very large scale for the enterprise than Dell.” you need an AI factory and nobody is better at building end-end systems of very large scale for the enterprise than Dell.” Last week, we saw Elon Musk’s xAI announce the AI project is ordering servers from both Super Micro and Dell.

Dell’s Power Edge Servers with Liquid Cooling

Dell’s Power Edge servers are designed for AI and HPC (high-performance computing) workloads. In September, the servers were launched with support for four H100 Tensor Core GPUs with liquid cooled GPUs with higher efficiency due to liquid cooling and higher GPU capacity per rack.

In May, Dell announced a new Power Edge server “L” version with liquid cooling and eight Blackwell Tensor Core GPUs. The eight GPUs communicate seamlessly with NVLink across memory and cores, which helps to support the training of large language models. Independent industry analysts have described the new Power Edge Server XE9680L as “the densest rack scale architecture in the industry.” The ”L” version is expected in the second half of this year and offers “33% more GPU density per node.” The air-cooled version can support 64 GPUs whereas the liquid cooled rack scale design supports 72 GPUs.

Upcoming AI Releases for Dell

Dell has a few more important AI release coming out this year. AI factories integrate Nvidia’s AI Enterprise software to allow companies to go-to-market quickly on AI workloads. The goal is to reduce setup time for AI development by up to 86%. The fully integrated solution combines Dell’s Hardware with Nvidia’s infrastructure and software.

As of now, Nvidia has three software businesses: Nvidia AI Enterprise, Omniverse and newly-announced Nvidia NIM. Dell’s AI factories set up Nvidia’s road map for both AI Enterprise software and NIMs, which provides models as optimized containers for generative AI application development. You can think of NIMs as something similar to an app store, to where developers can develop and market AI apps.

We’ve stated numerous times that Nvidia’s AI software revenue will rival the company’s GPU revenue. This is one of many examples where the stage is being set. In this case, Dell will ship fully integrated systems to enterprises, startups and SMBs who want to skip critical steps to deploy quickly.

Dell NativeEdge is another recent announcement, and is a software platform that reduces the amount of resources required for deploying an AI application at the edge. The platform targets the immense amount of operations work that is needed for when AI applications are deployed across many endpoints and devices. The most obvious first customer will likely be the Federal Government or hospitals and other industries that manage very large data sets at the edge.

Dell is Reporting Higher Inventory, Too

There were comments on the call that inventory is higher-than-usual at Dell, as well. If the higher-than-usual inventory levels at both Dell and SMCI are due to building out liquid cooling systems, then we will likely see higher inventory again this quarter. Inventory should alleviate in Q4 to Q1 when Blackwell’s GB200s and B200s ship. There are some cases where higher inventory is a good thing, such as when companies prepare for a spike in demand. It’s likely these companies are preparing for a spike in demand on DLC systems, rather than the opposite, which is that inventory is building because demand is waning.

Here is what Dell’s CFO stated:

“Our cash conversion cycle was negative 47 days, flat sequentially, with higher inventory related to our AI business, offset by strong collections performance.”

Here was a discussion on the earnings call relating to the higher inventory, and why this may be a bullish indicator for determining demand over the next six months and beyond:

Amit Daryanani

[…] And then, Yvonne, could you also just clarify, the inventory was up dramatically in the quarter and it's somewhat unusual for it to be up in this quarter. So just talk about what's driving that and is it AI pre-builds or strategic inventory? Any help on that would be great as well. Thank you.

Yvonne McGill

Sure. So let me start with inventory, because I think that's pretty straightforward. So inventory was up and I would say slightly, for 25 days, really representing about a $1.2 billion increase quarter over quarter. We mentioned inventory was up slightly as we ramp our AI server business. So I think it's nothing substantial. I don't know, Jeff, if you have anything to add on inventory.

Jeff Clarke

No, but we didn't go out and make any strategic purchases. Some of the terms of the AI gear we need to deploy means we take ownership of it. We did and we have it in backlog. We'll ship it as those customer orders are fulfilled. That was the driver. We weren't out buying strategic or making strategic investments of inventory across the large component basins.

Vertiv

Super Micro, Dell and Vertiv are three stocks with fantastic returns this year. SMCI is up about 200% (down from a high of about 300%), Dell is up 83% (down from a high of 118%) and Vertiv is up 82% (down from a high of 117%).

Vertiv offers power management and thermal management to data centers and telecom companies, such as Alibaba, AT&T, China Mobile, Tencent and Verizon. The company was formed in 2016 after spinning off from Emerson, and reported $6.8 billion in revenue last year. The company is considered one of the larger players in data center technologies, in terms of power management and thermal management, with 24,000 employees and 30 manufacturing facilities. Vertiv’s thermal management technologies include liquid cooling for servers and racks.

The data center accounts for 75% of Vertiv’s business with communications networks and commercial/industrial facilities at 25% of revenue. Most recently, their management team stated that AI-related projects have doubled in the past two months.

“The ramp-up of production of liquid cooling globally continues as planned, and I'm happy to report we have production underway already at two of the three plants we shared with you we were planning to activate in 2024. We are on track with the capacity ramp-up as shared in February. We continue to see strong momentum with AI-related orders. While we are not disclosing specific detail on our liquid cooling orders, or more broadly AI-related orders, we did see the pipeline for AI projects more than double in the last two months.”

Vertiv offers many thermal management solutions. Among them is the Liebert XDU, which is a compact unit that sits in the row near the rack or on the perimeter. The liquid-to-liquid cooling distribution unit (CDU) functions as a heat exchanger between the data center and IT equipment, and is used in all forms of liquid cooling: direct-to-chip, rear door heat exchange and immersion. The Liebert XDU offers a secondary fluid cooling loop so that alternative cooling fluids can be used alongside water.

In 2023, Vertiv acquired a company called CoolTera after partnering with the company for three years to add advanced cooling technologies to its thermal management portfolio. One of the main areas of need for data centers and colocation sites is to convert air-cooled equipment to liquid cooled equipment. Retrofitting existing air-cooled infrastructure is an area where Vertiv specializes, as opposed to only providing thermal solutions for new servers and racks.

In May of 2023, Nvidia selected Vertiv to design a cooling system that secured a $5 million grant from the COOLERCHIPS program. In 2024, Vertiv joined the Nvidia Partner Network with a statement that Vertiv is “collaborating to build state-of-the-art liquid cooling solutions for next-gen NVIDIA accelerated data centers powered by GB200 NVL72 systems.”

In late 2023, Vertiv announced a partnership with Intel to supply air-cooled and liquid-cooled servers for the Gaudi3 AI accelerator.

This is a thematic analysis on liquid cooling, and thus, we have not done a deep dive into Vertiv’s financials. Briefly, the company reported revenue of $1.63 billion in Q3, up 7.76% YoY yet down sequentially from $1.86 billion. The operating margin of 12.6% expanded from 9.6% for operating profit of $206 million. The adjusted operating profit was $249 million. Net margin decelerated from 3.3% to (-0.36%). Cash flow was $101 million in the most recent quarter and the company repurchased $600 million for share repurchases in the quarter,

Hewlett-Packard Enterprise (HPE)

HPE is another commoditized hardware company that is seeing a revival due to its large portfolio of liquid cooling technologies and patents. The company has over 300 patents related to direct liquid cooling (DLC) with four of the world’s top 10 supercomputers featuring liquid cooled servers from HPE. According to HPE, their Apollo DLC System reduces fan power by 81%.

The HPE Cray EX Liquid-Cooled Cabinet offers liquid-cooled cabinetry that provides DLC to all the components in a compact design. This is for CPUs and GPUs in excess of 500W, and can help to reduce the interconnect cabling systems that are required, which further reduces operational expense. As a reminder, Cray is a supercomputer built by HPE that ranks as the #1 and #2 supercomputers in the world. Therefore, the liquid cooling for these systems is quite advanced as the #1 Cray supercomputer contains hundreds of thousands of AMD EPYC CPUs and 37,000 AMD Instinct GPUs. The cooling technology for Cray features a bladed cabinet that allows for the mixing and matching of various CPUs and GPUs, and allows for easy upgrading as new generations of CPUs and GPUs are released. At one point, a system the size of Cray was reserved for only supercomputers, but the AI market is driving forth 24,000-plus GPU clusters today and Broadcom believes we will see million-plus GPU clusters by 2027.24,000-plus GPU clusters today and Broadcom believes we will see million-plus GPU clusters by 2027. HPE’s experience on liquid cooling the Cray supercomputers will be helpful as GPU clusters continue to scale.

HPE held a recent conference with Nvidia’s Jensen Huang, who appeared at a recent HPE conference to showcase the strength of the partnership between the two companies. There was a string of announcements, the primary one being that HPE and Nvidia are partnering on a private cloud (presumably to compete with Broadcom’s VMWare integration). You can find more information here in the press release on Nvidia-HPE announcements.more information here in the press release on Nvidia-HPE announcements.

In the most recent earnings report, HPE provided the following color related to AI sales: “Our cumulative AI system product and service orders since Q1 2023, rose approximately $600 million sequentially to $4.6 billion. I am very pleased with our AI system product revenue more than doubled sequentially to over $900 million. This strong revenue growth allowed us to make progress against our backlog, which is now $3.1 billion.” The company also stated that enterprise is “north of 15%” of the AI orders, which is a key market for both HPE and Dell (as opposed to predominately hyperscalers like SMCI).

The stock has risen about 20% YTD, quite a bit less than SMCI’s outperformance, and is lagging Dell and Vertiv considerably, as well.

Conclusion:

As pointed out in our AI power consumption write-up, AI power demand is forecast to rise at a rapid rate. GPU demand is showing no signs of slowing as Big Tech continues to spend billions on AI infrastructure, with each GPU generation seeing higher peak power consumption. The industry is quickly taking steps to address this, and power consumption, or more specifically, power efficiency per chip, looks to be emerging as the third realm of competition.

The first two realms of competition are raw computing power and memory; both have been extensively covered for our premium members. Now, we turn toward keeping an eye on the AI power consumption space as new winners will emerge now that power consumption has become mission critical.

As of now, our plans are to jump aboard the AI bullet train again (i.e., Super Micro) at key levels and to also follow our trading plan on Dell. If we decide any others are a good fit, then you will surely get a deep dive into those stock names.

To view our recent Advanced Market Signals webinar with SMCI and DELL trading plans, click here.click here.

Damien Robbins, Equity Analyst at the I/O Fund, contributed to this article.

Resources:

• Super Micro FYQ3: Cash is the Achilles Heel
• Super Micro Q3 Pre-Earnings: Puts and Takes for the AI Bullet Train
• Dell Q1 Pre-Earnings: It’s All About the QoQ AI Revenue Growth
• Dell Q1 Earnings: AI Server Shipments up 113% QoQ, Margins Contract
• Lam Research: Eyeing Strong 2024 Exit Boosted by Memory Rebound
• AI Power Consumption: Rapidly Becoming Mission-Critical