Half Lemonade

The BBG Hopper GPU Rental Index with a big move down in late-October. Now down 28% Y/Y.

But that’s fascinating. How could OpenAI possibly develop AI technology related to semiconductors? Of course, since OpenAI is taking the lead in designing the ASIC with Broadcom, it’s undeniable that they already have some semiconductor design capabilities. But this sounds like OpenAI is entering the EDA territory — the realm of Cadence and Synopsys. How are they even able to do that?

AVGO holders, why do you think Google can’t drop Broadcom and develop TPUs with Marvell or fully in-house instead?

Elon is building a giant GPU cooler in Texas.

I own Vertiv $VRT (leader in power / cooling solutions for several end markets, but most importantly, data centers) and will continue to own. I’d imagine the stock continues to run into earnings on 2/21 after strong data center capex commentary from the hyper scalers and robust trends from other AI / data center beneficiaries like Super Micro $SMCI (there is a very linear correlation between cloud capex and Vertiv demand). Vertiv came public in February 2020 via SPAC. The sponsor was Dave Cote, the legendary CEO that turned around Honeywell $HON. This is like Satya Nadella forming a SPAC and buying an enterprise software company. These guys have domain expertise and domain dominance. This is a good (not great) company benefiting from a massively rising tide. Its also the biggest pure play available to ride that tide. My thesis is based on a few key points: market for data center power / precision cooling is exploding, gaining market share, and upside to numbers (at a reasonable valuation). I say good (not great) because while their solutions are certainly not commoditized, they are replicable (more similar to Super Micro than Nvidia $NVDA in terms of proprietary technology / margin profile) and they also had some execution hiccups around the pandemic (although under a different management team). I will further discuss these points below: The first point to mention is the durability of data center / AI capex, because everything else is downstream. Everyone from Microsoft $MSFT and Google $GOOGL to Super Micro and TSMC $TSM and Eaton $ETN / Trane $TT continues to be extremely optimistic about this trend continuing. Importantly, Microsoft spoke on their earnings call about inference use cases taking shape, showing the economic benefits of these outsized AI investments. This should put to bed the belief that there could be an “air pocket” in demand on the horizon. The transition from traditional CPU-focused servers to GPU-powered AI servers will take a long time and drive growth for the foreseeable future. My belief in this multi year wave is also underscored by record order pipelines and backlog – I expect orders to grow in excess of backlog again in 2024. Comments from management suggest current demand could take 3-4 years to work through. The data centers running those AI servers generate 5x more heat / power than traditional CPU servers and require 10x more cooling per square foot. ~75% of Vertiv’s business is power and cooling related infrastructure, with the balance largely being services like maintenance, installation, remote monitoring, etc. Vertiv has announced plans to double its production capacity for electrical switchgear and busway over the 2023-2025 period. This comes on top of a >100% increase in capacity over 2021-2023. And they recently opened a new manufacturing facility in India for thermal management products for domestic and international markets. This follows the opening of a thermal manufacturing facility in Mexico early in 2023. And at its recent investor day, management raised medium-term capex guidance to 2.0-2.5% of revenue, from an average of 1.6% over 2019-2022 period. This capacity expansion is obviously rooted in the belief that demand is durable and sustainable (similar to Super Micro’s comments of taking capacity to ~$30bn from low teens today). A general rule of thumb is that if volume of data goes up 10%, we generally will see 3-5% increase in the volume of gear to ensure adequate power supply. Many of these AI benefits are yet to flow through to Vertiv, particularly around liquid cooling. There are several reasons for this: The market for cooling infrastructure is gradually transitioning from air cooling (where Vertiv is currently the leader) to liquid cooling (which is Vertiv’s to lose). Super Micro’s CEO said on the last earnings call: “You are right. In these current 600 watt / 700 watt modules, people can still do well with air conditioning. And that's why people still are comfortable with our traditional air cooler. But when that system grows to 1,000 or even 1,000 watt per module, yes. I mean -- I think cooling becomes even much more critical.” Basically as power / compute demands increase, there is a need for more heavy duty cooling direct to the chip / rack. Super Micro CEO then said: “I anticipate that up to 20% or more of global data centers will transition to liquid-cooled solutions in just a few years.” So we are early days. Important to note that liquid cooling will largely be additive to Vertiv’s business – down the road ~1/3 of thermal cooling will still be from air cooling – liquid cooling does a great job taking heat away from the chip and rack but it still needs to be moved and expelled from the building. For example, traditional compute requires 6 computer room air conditioners vs. 10 for high density compute. As we move toward Nvidia’s next chip (B100), more cooling capability will be required. The industry was growing high single digits prior to liquid cooling – so growth with liquid cooling is obviously higher. Additionally, new AI data centers take time to permit, build, etc. and even existing data centers are only turned on with half – 2/3 capacity, which illustrates why the demand curve could be longer and more durable than the market appreciates. A few comments from management on the last earnings call: “still very early in AI opportunity – its barely started.” “AI will show up in more pronounced manner in 2024.” “More demand than capacity in industry.” Longevity of current order inflection: “Believe this is a long term trend – multi year cycle – this is just the beginning” As I mentioned, Vertiv is the largest pure play, with other competitors like Schneider $SCHN, Eaton and Stulz $STULZZ, as well as a long tail of independents (still ~50% of market). As hyper scalers continue to drive AI data center demand, there are only a few players that can service this demand. Scale and service excellence matter when dealing with hyper scalers (they want someone to hold their hand). Ecolab $ECL has a similar competitive advantage. Vertiv’s product offering is good, not great, but their ability to service customers and be a one stop shop / systems integrator for power, precision cooling and other infrastructure management systems is what sets them apart. For example, you don’t want several different teams in your data center, you want one team doing power, air cooling, liquid cooling, etc. This is why Vertiv has the prime position and this is why it is theirs to lose. There are also several startups working on liquid cooling technologies – these starts up could never service the needs of a hyper scale data center – so the natural move is for Vertiv to buy them. At ~2x net leverage and a cash generative business model, they are positioned to make this happen. We also cant overlook the fact that Vertiv has partnerships with Nvidia, Intel $INTC (bought Habana AI in 2019) and others. Jensen (Nvidia CEO) called out Vertiv as the world leader, saying their collaboration is growing in leaps and bounds and that their partnership is more important than ever. Numbers / estimates higher at a reasonable valuation = stock higher. Simple model summary and valuation snapshot below. These are base case estimates and I don’t expect them to guide to what I am expecting right away on the Q4 call. I also reserve the right to revise my estimates as the data suggests - likely higher. Starting with Q4 – management said to expect a similar growth rate in Q4 as in Q3. We saw 18% growth in Q3 and the street is at ~14% growth for Q4. Either people aren’t paying attention, or they don’t trust management. I trust that this management team will continue to be conservative with the hopes of creating a consistent beat / raise story. At the investor day in November 2023, management gave “very preliminary” 2024 guidance of 8-11% sales growth and operating margins 16.5-16.9%, with 83-87% FCF conversion. Given the “very preliminary” comment its clear that this is more a floor than a ceiling. The street is at ~10% growth for 2024 and has it slowing to HSD in 2025 and 2026. Important to note that Vertiv’s growth has largely been US driven (grew >40% in 2023). And while this is where most of the AI investment is taking place, international markets will eventually pick up – this is another advantage of Vertiv’s global footprint. Vertiv grew >20% in 2023, despite APAC down ~5% (macro weakness in China) and EMEA only up high single digits. Additionally, management has said incremental margins (margins on incremental revenue that flows through) are currently ~30%, but will move toward ~35% as the liquid cooling investments start to bear fruit. As you can see in my model summary, I can get to >20% operating margins in 2026 (managements long term guidance is 20%+ operating margins sometime in 2026-2028 period). Backlog is also at record levels ~$5bn, which tells you >3/4 of the year is already accounted for. And again, I expect orders to grow faster than sales through 2024, leading to another record backlog number on the horizon. Importantly, management spoke to taking ~5% pricing in their backlog, which tells me pricing could be up mid single digits again this year. Historically this has been a flat to down ~1% annually pricing business, but the demand environment has changed this and I expect it to last for several years before mean reverting. Management has said they expect to be price / cost positive on a go forward basis. Balance sheet is healthy, which will allow them to paydown debt, buyback stock and make bolt on technology acquisitions. Current debt structure is $850mm fixed rate bond 4.125% due October 2028 and a $2.1bn term loan due March 2027. The term loan is split between $1.1bn at variable rate currently ~8% (focus is to pay this down over next two years, which should allow them to accelerate capital deployment) and $1bn fixed ~4%. Vertiv also has authorization to buyback $3bn worth of stock over the next 4 years (~15% of market cap). The stock currently trades ~20x EV / operating profit (I’m using this metric because this is what they guide to). I’m assuming this multiple holds, which it will as long as numbers / estimates go higher. There is no pure play comp, but Eaton trades ~23x and Amphenol trades ~22x (also a low double-digit growth with >30% incremental margins). As I look out a few years, I can still get to meaningful upside, despite the recent run in the stock (see below). The key risk is that AI investment / capex slows. Additionally, Vertiv could lose market share if they don’t continue to innovate / buy the innovators. They could also make execution errors, which they did early in Covid under a different management team. Vertiv also serves certain telecom end markets, which have slowed post heavy 5G investments. However, this is a small part of the overall business (and getting smaller given data center growth).

I own Vertiv $VRT (leader in power / cooling solutions for several end markets, but most importantly, data centers) and will continue to own. I’d imagine the stock continues to run into earnings on 2/21 after strong data center capex commentary from the hyper scalers and robust trends from other AI / data center beneficiaries like Super Micro $SMCI (there is a very linear correlation between cloud capex and Vertiv demand). Vertiv came public in February 2020 via SPAC. The sponsor was Dave Cote, the legendary CEO that turned around Honeywell $HON. This is like Satya Nadella forming a SPAC and buying an enterprise software company. These guys have domain expertise and domain dominance. This is a good (not great) company benefiting from a massively rising tide. Its also the biggest pure play available to ride that tide. My thesis is based on a few key points: market for data center power / precision cooling is exploding, gaining market share, and upside to numbers (at a reasonable valuation). I say good (not great) because while their solutions are certainly not commoditized, they are replicable (more similar to Super Micro than Nvidia $NVDA in terms of proprietary technology / margin profile) and they also had some execution hiccups around the pandemic (although under a different management team). I will further discuss these points below: The first point to mention is the durability of data center / AI capex, because everything else is downstream. Everyone from Microsoft $MSFT and Google $GOOGL to Super Micro and TSMC $TSM and Eaton $ETN / Trane $TT continues to be extremely optimistic about this trend continuing. Importantly, Microsoft spoke on their earnings call about inference use cases taking shape, showing the economic benefits of these outsized AI investments. This should put to bed the belief that there could be an “air pocket” in demand on the horizon. The transition from traditional CPU-focused servers to GPU-powered AI servers will take a long time and drive growth for the foreseeable future. My belief in this multi year wave is also underscored by record order pipelines and backlog – I expect orders to grow in excess of backlog again in 2024. Comments from management suggest current demand could take 3-4 years to work through. The data centers running those AI servers generate 5x more heat / power than traditional CPU servers and require 10x more cooling per square foot. ~75% of Vertiv’s business is power and cooling related infrastructure, with the balance largely being services like maintenance, installation, remote monitoring, etc. Vertiv has announced plans to double its production capacity for electrical switchgear and busway over the 2023-2025 period. This comes on top of a >100% increase in capacity over 2021-2023. And they recently opened a new manufacturing facility in India for thermal management products for domestic and international markets. This follows the opening of a thermal manufacturing facility in Mexico early in 2023. And at its recent investor day, management raised medium-term capex guidance to 2.0-2.5% of revenue, from an average of 1.6% over 2019-2022 period. This capacity expansion is obviously rooted in the belief that demand is durable and sustainable (similar to Super Micro’s comments of taking capacity to ~$30bn from low teens today). A general rule of thumb is that if volume of data goes up 10%, we generally will see 3-5% increase in the volume of gear to ensure adequate power supply. Many of these AI benefits are yet to flow through to Vertiv, particularly around liquid cooling. There are several reasons for this: The market for cooling infrastructure is gradually transitioning from air cooling (where Vertiv is currently the leader) to liquid cooling (which is Vertiv’s to lose). Super Micro’s CEO said on the last earnings call: “You are right. In these current 600 watt / 700 watt modules, people can still do well with air conditioning. And that's why people still are comfortable with our traditional air cooler. But when that system grows to 1,000 or even 1,000 watt per module, yes. I mean -- I think cooling becomes even much more critical.” Basically as power / compute demands increase, there is a need for more heavy duty cooling direct to the chip / rack. Super Micro CEO then said: “I anticipate that up to 20% or more of global data centers will transition to liquid-cooled solutions in just a few years.” So we are early days. Important to note that liquid cooling will largely be additive to Vertiv’s business – down the road ~1/3 of thermal cooling will still be from air cooling – liquid cooling does a great job taking heat away from the chip and rack but it still needs to be moved and expelled from the building. For example, traditional compute requires 6 computer room air conditioners vs. 10 for high density compute. As we move toward Nvidia’s next chip (B100), more cooling capability will be required. The industry was growing high single digits prior to liquid cooling – so growth with liquid cooling is obviously higher. Additionally, new AI data centers take time to permit, build, etc. and even existing data centers are only turned on with half – 2/3 capacity, which illustrates why the demand curve could be longer and more durable than the market appreciates. A few comments from management on the last earnings call: “still very early in AI opportunity – its barely started.” “AI will show up in more pronounced manner in 2024.” “More demand than capacity in industry.” Longevity of current order inflection: “Believe this is a long term trend – multi year cycle – this is just the beginning” As I mentioned, Vertiv is the largest pure play, with other competitors like Schneider $SCHN, Eaton and Stulz $STULZZ, as well as a long tail of independents (still ~50% of market). As hyper scalers continue to drive AI data center demand, there are only a few players that can service this demand. Scale and service excellence matter when dealing with hyper scalers (they want someone to hold their hand). Ecolab $ECL has a similar competitive advantage. Vertiv’s product offering is good, not great, but their ability to service customers and be a one stop shop / systems integrator for power, precision cooling and other infrastructure management systems is what sets them apart. For example, you don’t want several different teams in your data center, you want one team doing power, air cooling, liquid cooling, etc. This is why Vertiv has the prime position and this is why it is theirs to lose. There are also several startups working on liquid cooling technologies – these starts up could never service the needs of a hyper scale data center – so the natural move is for Vertiv to buy them. At ~2x net leverage and a cash generative business model, they are positioned to make this happen. We also cant overlook the fact that Vertiv has partnerships with Nvidia, Intel $INTC (bought Habana AI in 2019) and others. Jensen (Nvidia CEO) called out Vertiv as the world leader, saying their collaboration is growing in leaps and bounds and that their partnership is more important than ever. Numbers / estimates higher at a reasonable valuation = stock higher. Simple model summary and valuation snapshot below. These are base case estimates and I don’t expect them to guide to what I am expecting right away on the Q4 call. I also reserve the right to revise my estimates as the data suggests - likely higher. Starting with Q4 – management said to expect a similar growth rate in Q4 as in Q3. We saw 18% growth in Q3 and the street is at ~14% growth for Q4. Either people aren’t paying attention, or they don’t trust management. I trust that this management team will continue to be conservative with the hopes of creating a consistent beat / raise story. At the investor day in November 2023, management gave “very preliminary” 2024 guidance of 8-11% sales growth and operating margins 16.5-16.9%, with 83-87% FCF conversion. Given the “very preliminary” comment its clear that this is more a floor than a ceiling. The street is at ~10% growth for 2024 and has it slowing to HSD in 2025 and 2026. Important to note that Vertiv’s growth has largely been US driven (grew >40% in 2023). And while this is where most of the AI investment is taking place, international markets will eventually pick up – this is another advantage of Vertiv’s global footprint. Vertiv grew >20% in 2023, despite APAC down ~5% (macro weakness in China) and EMEA only up high single digits. Additionally, management has said incremental margins (margins on incremental revenue that flows through) are currently ~30%, but will move toward ~35% as the liquid cooling investments start to bear fruit. As you can see in my model summary, I can get to >20% operating margins in 2026 (managements long term guidance is 20%+ operating margins sometime in 2026-2028 period). Backlog is also at record levels ~$5bn, which tells you >3/4 of the year is already accounted for. And again, I expect orders to grow faster than sales through 2024, leading to another record backlog number on the horizon. Importantly, management spoke to taking ~5% pricing in their backlog, which tells me pricing could be up mid single digits again this year. Historically this has been a flat to down ~1% annually pricing business, but the demand environment has changed this and I expect it to last for several years before mean reverting. Management has said they expect to be price / cost positive on a go forward basis. Balance sheet is healthy, which will allow them to paydown debt, buyback stock and make bolt on technology acquisitions. Current debt structure is $850mm fixed rate bond 4.125% due October 2028 and a $2.1bn term loan due March 2027. The term loan is split between $1.1bn at variable rate currently ~8% (focus is to pay this down over next two years, which should allow them to accelerate capital deployment) and $1bn fixed ~4%. Vertiv also has authorization to buyback $3bn worth of stock over the next 4 years (~15% of market cap). The stock currently trades ~20x EV / operating profit (I’m using this metric because this is what they guide to). I’m assuming this multiple holds, which it will as long as numbers / estimates go higher. There is no pure play comp, but Eaton trades ~23x and Amphenol trades ~22x (also a low double-digit growth with >30% incremental margins). As I look out a few years, I can still get to meaningful upside, despite the recent run in the stock (see below). The key risk is that AI investment / capex slows. Additionally, Vertiv could lose market share if they don’t continue to innovate / buy the innovators. They could also make execution errors, which they did early in Covid under a different management team. Vertiv also serves certain telecom end markets, which have slowed post heavy 5G investments. However, this is a small part of the overall business (and getting smaller given data center growth).