496: Cloudflare Q1, AMETEK 101, AWS + Crowdstrike, Las Vegas Sphere, Nuclear, Synesthesia, and Derek Sivers

"it must sound completely nuts"

Whenever you are stuck searching for the optimal plan, remember: Getting started changes everything.

—James Clear

🎶👂🏻🧠🎨 I mentioned my synesthesia in Edition #495.

My friend Saeah (📚🎹) sent me this video of a violinist describing her experience with synesthesia, accompanied by a computer-generated visualization to illustrate her perception.

I wish mine was like that! It must be so useful to “see” pitch.

I’m on the opposite side: I sometimes wonder if what I “see” doesn’t make it *harder* for me to figure out pitch because it’s often jumbled up and confusing.

To me, different instruments can have completely different shapes, textures, and colors. Some can be kind of “wide” and “blurry”, like a broad paintbrush stroke, while others are very “sharp” and “well-defined”.

Drums and bass may be pulsating in the background while the notes from a guitar solo may look like those sparkly things on kids’ birthday cakes.

I’ve also noticed that album cover art can affect what I see and “tint” the instruments with a certain palette (“this album sounds orange”).

Some instruments have a nice “look” — some piano notes can almost be like phosphorescent glass beads, while a very crunchy palm-muted metal riff can be more like an angular column of heavy black smoke.

It gets weirder. I also “see” colors in my mind’s eye for numbers, letters, words, days of the week, and months. “E” is like dark red wine, “O” is pale yellow, “5” is blue-gray, and “8” is dark green.

I know, it must sound completely nuts to someone who doesn’t experience this! 🤪

¯\_(ツ)_/¯

🌆👀🤖 Speaking of seeing things, I was thinking about how AI “sees” images.

If you open a JPEG file, it looks like this for pages and pages:

°I cÖzûíáÂd†€®
Çá∑≠G—<÷fil`‘∏{pƒ˝∂1¡i°UVW‹iEH0T⁄J}XÄW™Ã˘…´¶"“È√ßCà™õ¥˝[_ë¿ZÜÉKSÚ:œu˘éꈿÜ(4Ω¢µfl"Ñ©ïc©XØ^;~ıLj¶ı@ÿ≠!@pŸöƒæèÖ	{√CÈHíÊ∫9*o˜J®’fiã±4¬É]ô“ı瑪vr™¯Î7Îr¸_Ò0kiË∆ÒG»‚n]ñ€íÏœ›{MÒUAÂ2’ûNå≈û\/3áÖ7“ÍfNπ,¨ÍuÊú	÷ZÇs&OÃl≤tJ"Sz3◊=£œØïÕ(÷q(‡˜ã€1Ú%˘∑≥˜,8YQ–ƒ[©lãµÏ<2¿¿Ùw*kÉÉÓQVÅoòÄ≠hfNjÈE%qP(ÑrTej¨8ü‚jÌˇÄËaScÕ◊H4Ô3JÖi∫xˇgTc`0Ães§kWx@ñ{|¶%`NÀd-æÚÃÎÜ- :R„ºPwˬ‡∏T7zÔòˆ ≤S íâx≠‡Ú¡ÖÊh`L¿,⁄,i–—PûºÉßH[0JI!.ó˛{Ÿûo/Tzfls7Eé8úá=eæ`^ÅåIÉ≠Ãú‚‘‘¶∫!»Ô&j˙≤¸√fif˙Ã3fùeÜÅ’‡¢TÕ3º∂Mcî∞¸¯\4≠#5“0j[ä∏ÈÈ·QÍ3‹¯üöøá≥«_⁄Ô8Åî`.)ÇùòãP°ÌBFŒfl∏Sͱ(TÍ8B¨ºê®lß:”»|«V7#B‚ÄÊ#=sÚèúG%øB(+°ø
B…IÉ9Un"2¢rôHZ‘|Ê›x·rãBXÂr/Ç3<ˆ´Û)l·πç„∂åfiõ»;ȵÿ x¡ƒÛ[7ÌzÌ:£À oÌ/¥x˙Î[À√µ∂¸ã•
ß—fi3Úñ∫”,€Ï˜à˛™A“Éfipawî∑3s3CÇo—°ˆÉÂJ LjxÏ˚÷jÔLÄgˆüÇ

Each pixel’s color and location are encoded in there (I won’t get into lossy compression today…).

A multi-modal LLM can look at thousands of pages of this 👆 and “know” that it encodes a dog or a house or a ninja 🥷. The model has been trained on *trillions* of pages of similar data representing images and it has extracted the high-level patterns.

While it’s useful for us humans to think of these images in the way that our visual cortex renders them from the photons emitted by screens, for a LLM without a visual cortex 🧠, the image *is* the raw binary code from the decoded JPEG file.

Generative AI also works with this raw material to create new images.

There’s no point at which the model can “stand back and look at the image” like a human would 👁️. It has to create a new image that makes sense and looks good purely as thousands of pages of binary code. That code is then encoded as a JPEG and rendered on a screen by a device so that a human visual cortex can decode the photons and interpret them.

It’s a bit like a composer who has never heard a sound because they were born deaf, but despite this became so proficient at reading notes on the page that they can write new music based on all the written music they have seen previously, and when musicians play it, it sounds good to their ears.

It’s like Neo looking directly into the falling characters of the Matrix…

💚 🥃 🚢⚓️ If you’re getting value from this newsletter, it would mean the world to me if you become a supporter to help me to keep writing it.

If you think that you’re not making a difference, that’s incorrect. Only 2% of readers are supporters (so far — you can help change that), so you make a big difference.

You’ll get access to paid editions and the private Discord. If you only get one good idea per year, it’ll more than pay for itself!

---

🏦 💰 Liberty Capital 💳 💴

⛅️ What’s Going on with Cloudflare? 🤔

Cloudflare released its Q1 results a few days ago:

• Revenue: +30%

• Gross margin: 77.5% vs 75.7% YoY (non-GAAP: 80%. Impressive considering they aren’t a pure software company, and have a worldwide network of servers and fat pipes (280tbps!!!))

• Operating cash flow: +102.2% (73.6m vs 36.4m)

• Free cash flow: +156% (35.6m vs 13.9m — but of course, SBC 💰)

• 197,000 paying customers, +17% YoY

• 2,900 large customers (>$100k), +33% YoY, representing 67% of revenue.

• Dollar-based net retention: 115% (down 2% YoY ⬇️)

• Trivia: They have $1.7bn in cash on the balance sheet, so if they just invest in t-bill, they can make about half as much in interest as their operating income 🤔

That all looks pretty good to me, but something feels different about the company lately.

They spent the past few years constantly releasing new products and platform elements, which was exciting for nerds like me. They’re still shipping a lot of stuff, but their focus seems to have shifted to sales, an aspect of the business that they had been neglecting before — or maybe it’s that they *evolved* into a new phase after starting with an almost purely bottom up, self-serve approach.

They now have to graft a top-down enterprise sales engine to bag the bigger customers (hiring Mark Anderson was a great move for them).

CEO and co-founder Matthew Prince talked about this:

If I reflect back on the history of Cloudflare, it divides fairly neatly into 7-year eras. The first 7 years, 2010 to through 2017, were all about engineering and little else, figuring out if it was even possible to build the revolutionary network we have today.

The next 7 were all about product, taking that incredible engineering and packaging it up.

We continue to build on and improve that foundation. Looking forward to the next 7 years, we will continue to be the best in the world in engineering and product, but we'll add to that world-class sales and marketing.

The CFO added:

We are encouraged by the double-digit year-over-year improvement in sales productivity that we again delivered during the first quarter, which continues the upward trend from the trough of early last year from the productivity levels we consistently achieved in 2021 and early 2022. [...]

we are committed to the underlying unit economics of our business and make data-driven decisions to pace hiring. The improvement in sales productivity and other leading indicators of our business give us confidence to invest in additional go-to-market improvements and further expand sales capacity.

They certainly don’t seem to have any problem attracting talent:

We had 350,000 people apply to work at Cloudflare in Q1 for, I don't know, 250 jobs roughly. That's extraordinary.

I don’t even know how you can sift through 350,000 applications in one quarter. Maybe they just have AI do it 🤔 😬

That said, fixing sales won’t be easy:

I think we are still early in the journey of being really great at Enterprise [sales]

The last 18 months were all about us sort of cleaning up the sales organization. And now with Mark on board, it's all about now how do we get the real professionals on board, where we can become that strategic vendor for literally every large company in the world.

There is no large company that doesn't have a need for what it is that we sell.

It’ll be interesting to see if they can achieve their goal.

It’s not easy to change a company’s culture and DNA, but they seem to be taking this seriously at the very top of the company. They’re willing to fire bad performers, hire leaders with experience, and create a new incentive structure, so they have a shot.

On the Workers platform and their GPU rollout:

we crossed over 2 million active developers building applications on Cloudflare Workers.

Second, in April, we GA-ed a number of key products like D1, our serverless SQL database; Hyperdrive, which makes any traditional database perform like it's globally distributed; and Workers AI, which allows developers to run and tune AI models across our global network.

We're ahead of schedule rolling out GPUs across our network and now have them running in more than 150 cities globally, making us what we believe is the most widely distributed AI cloud by a huge margin.

Our next generation of servers that begin to roll out in Q2 have GPUs built in by default and will support faster inference and even larger, more complicated models. Developers are building incredible new applications using Workers AI, and we're making it increasingly easy for them.

We added support for Python, the second most popular programming language generally and the most common language for AI applications. We rolled out our partnership with Hugging Face, making it one-click simple to deploy most of their catalog of models to Cloudflare's network. And we added other bleeding edge models, including releasing Meta's Llama-3 production simultaneously the day it was announced.

Sounds like adoption is strong, and I hear good things from people using the platform. But it remains to be seen how profitable it ultimately is for them.

They’re probably taking the right approach of focusing first on getting devs in by giving away credits and leaving monetization for later. Platforms tend to be sticky and developers form habits that can last a long time (even if they move to a new job, where they may advocate for their favorite platform).

But the real sink or swim moment will be when they ask everyone to pay full price.

On capex intensity and additional capex for GPUs/AI:

as we sell more of our Zero Trust and SASE products, those are extremely high-margin products, and they don't require a significant additional amount of CapEx. That then frees up our ability to invest that CapEx in other areas, including in the AI space. I think we made some very smart decisions, specifically reserving space and the equipment that we deployed knowing that AI would be part of our story at some point in the future.

And so we really wanted to make sure that's the case. That means that as we deploy CapEx, it's literally not shipping an entire server to support AI, but shipping just the GPU cards that go into existing servers that are in the field. That reduces the amount of CapEx that has to be deployed. And again, it works because it is all running on one unified network. The fact that every server across Cloudflare's entire platform is capable of performing any function that we need, that has allowed us an enormous amount of flexibility in how we can deploy things and has helped us. [...]

Finally, I would say that inference is different than training. And so you need different resources for that. You don't need necessarily the most cutting-edge GPUs in order to do inference tasks. And so that has meant that we haven't had to chase down what is a — GPUs that have limited quantity. It's also meant that we can be much smarter about picking and choosing between different GPU vendors and matching workloads to whoever it is that can provide the best service.

The way they describe this, it doesn’t sound like they’re deploying a ton of big iron Nvidia DGX servers, which is probably smart. Why compete with hyperscalers and Jensen’s teacher’s pets like Coreweave when you can instead focus on low-latency inference, which plays better to Cloudflare’s strengths?

As more genAI gets deployed and all kinds of models (not all of them large) need inference, Cloudflare should be able to get its share of that, though how successful they are will depend on whether AI use cases that require proximity but can’t be offloaded to user devices take off or not.

I’m not so sure about that.

We were at 8% to 9% of revenue with network CapEx in the first quarter. We said the year will be close in the range of 10 to 12, and this includes the rollout of GPU capacity pretty much to every server and every location we have.

That’s not too bad compared to the capex number from Big Tech lately.

Speaking of Big Tech, Cloudflare differs in one important way when it comes to usage vs flat rate pricing:

[hyperscalers] tend to be purely usage-based models. And the good of usage-based models is that they react very quickly, both up and down. And so I think whereas we tend to be much more of a subscription-based model, and the good of that is that it has a lot more stability. But -- and that stability helps you when things are slowing down.

So we had significantly faster growth than the hyperscalers in the last few quarters, and we're now around where they are this quarter. And so I think we are a little bit more moderated than some of the more usage-based models. And I think that explains some of it.

I think the second thing that I'd add is that for some of the products, especially around the AI products, as Thomas said, we're really optimizing around adoption. And while we published pricing and that pricing was extremely well received, in addition to being very, very attractive for us from a margin perspective in the last months, that's something that is still not a place where we are heavily monetizing yet. We see plenty of opportunity there. And we think that when we decide the time is right to turn on some of those monetization, it will help contribute to our top line growth.

👋

📡 🛰️ 🛠️ AMETEK 101 ⚙️🔬

Back in Edition #475 and Edition #482 I wrote about AMETEK, a serial acquirer conglomerate. Here’s how I described them:

It was created in 1930 under the very cool name of ‘American Machine and Metal’ 🤘 — I kinda wish they had kept the name.

The way I think about them is kind of like a mix of Heico’s ETG segment + Roper’s pre-software stuff + pre-Life-Sciences Danaher + Mettler Toledo. Transdigm also has some similar elements, but big differences: they focus on sole-source parts and use higher leverage.

I know, that’s quite the Frankenstein’s monster… 🧟‍♂️

Business Breakdown recently had an episode on them and I thought it was quite good at providing an overview of their history, operating model, strengths and weaknesses, and the challenges that they face going forward.

If you’re curious about the company, this is a good place to start (and then for a deeper dive, check out the link to Edition #482 above):

• 🎧 Nael Fakhry - AMETEK: Industrial Excellence

Here are a few highlights:

there are a handful of end markets that dominate. The largest is med tech, that's a little more than 20% of total revenue. And then you have aerospace and defense, military, that's a little under 20%, about 18%, 19% of revenue. The power market, the serving utilities and others, that's another 10%. Automotive is another, roughly 10%. And then you have semis, R&D and then just other industrial.

Hard to be more diversified when it comes to end markets!

if you look up any one business unit within AMETEK, and there are over 40, you'll find that, that one business unit might sell 3,000, 4,000, 5,000 parts. So it's a really broad array. But I think the one unifying factor across products is that they are mission-critical. They're really important to the operation of some larger system, but the costs are actually usually really small relative to the overall cost of the system. [...]

They want to be in really small markets. So on average, the market size, whole market is about $200 million to $300 million in size. And they tend to have, on average, just a very broad, sweet number, but it's about 25% to 30% share on average in each market. They will actually avoid markets that are really large in size, say $1 billion or more, because they don't want to attract competition from the larger players that are seeking growth.

Riches in niches 💰

[AMETEK] never actually had a write-off of goodwill in the history of the company. So this is a company that dates back almost 100 years now, and they've obviously done hundreds of acquisitions over the history of the company.

And even Reading Alloys, that was a company that was purchased in 2008 for $110 million, viewed as not the most successful, unique business. They sold it for $250 million in 2020, 12 years later. So I would say management is really competent and thoughtful

Considering how hard it is to create value via M&A for most companies, this is pretty telling.

☁️🤝🦅 AWS Deepens Partnership with Crowdstrike 🔐

An existing collaboration goes deeper:

As part of the partnership, Amazon has unified its cybersecurity protection on the CrowdStrike Falcon platform, protecting the company from code to cloud and from device to data. Amazon is replacing a variety of cloud point products with Falcon Cloud Security, is using Falcon Next-Gen SIEM to secure big data logging and is deploying Identity Threat Detection and Response to prevent identity-based attacks.

Crowdstrike’s CEO had mentioned an ‘8-figure deal’ with a hyperscaler in a recent call. This is it.

As part of this deal, Crowdstrike is using Sagemaker and the Claude LLM for its own Charlotte AI:

CrowdStrike is expanding its use of Amazon Bedrock, including Anthropic’s Claude family of large language models (LLMs) and Amazon SageMaker to deploy custom models.

As friend-of-the-show Muji (💾) pointed out, each hyperscaler seems to have picked its horse in the security race; Google GCP has Palo Alto, Azure has Microsoft’s homegrown stuff, and AWS has Crowdstrike.

Yes, it’s adjusted for inflation 📊

In 2022 Americans under 25 spent 43% of their post-tax income on housing and education, including interest on debt from college—slightly below the average for under-25s from 1989 to 2019. Bolstered by high incomes, American Zoomers’ home-ownership rates are higher than millennials’ at the same age (even if they are lower than previous generations’).

A good reminder that vibes matter a lot.

It’s possible to feel poorer while being significantly richer than your parents or grandparents at the same point in their lives (although I think a lot of people subconsciously compare themselves to *older* cohorts because that’s what is visible).

There’s also the “good ol’ days” syndrome.

It’s easy to imagine that everything used to be better and that there was less uncertainty because looking back, we know how things turned out. But at the time, living through the 20th century, it was anything but easy or certain!

Let’s just look at the history books… 📖

---

🧪🔬 Liberty Labs 🧬 🔭

🫧 How the $2.3bn Las Vegas Sphere Works 🎪💰

What an engineering marvel! I’d love to go see a show there someday.

🇮🇹 Italy Reconsiders Nuclear Power ⚛️

Italy closed its nuclear plants in 1990 after the 1987 referendum on atomic energy following the Chernobyl disaster. Another referendum in 2011 effectively banned the operation of nuclear power plants within the country.

But that may be changing:

Environment and Energy Security Minister Gilberto Pichetto said Thursday that the government aims to pass the necessary legislation to make Italy's return to nuclear power possible by the end of the current parliamentary term. [...]

the country is also looking to boost its energy security following the war caused by Russia's invasion of Ukraine.

"Yes, we'll give it our all," Pichetto told Radio 24 when asked if the legislative framework for nuclear energy could be changed by the end of parliament. [...]

"A contribution from nuclear energy in our energy mix would help Italy a lot in meeting the net zero target by 2050," Energy Minister Gilberto Pichetto Fratin said at an event ahead of the G7 energy meeting. [...]

At the event, Pichetto expressed support for the development of so-called small nuclear reactors, which nuclear advocates say could reduce costs and help decarbonise highly polluting sectors such as steel production.

🔌⚛️💰 Is Nuclear Power “Too Expensive?” + What Could We Do? 🤔💡

Here’s something that I often hear and find misleading.

Opponents of nuclear call it “too expensive”. That’s confusing two things.

Nuclear power, when you look at the price per kWh over the lifetime of a plant’s operation, is generally not particularly expensive. It depends on the project, but in many countries, it’s some of the cheapest and most reliable power.

Even in places where it’s kind of middle-of-the-pack, it looks pretty good when you look at the whole picture (other cheaper sources pollute a lot more, or they are more volatile — maybe natgas is cheap for a while, but it could get expensive for long stretches and hurt the rest of the economy).

You have to compare apples to apples.

A solar farm does not compete directly with a nuclear power plant.

Its output varies hour-by-hour, day-by-day, week-by-week, and season-by-season. Operating characteristics are very different in Arizona vs Germany. “Solar” is not just one thing — every project is different. ☀️☁️🌤️

What could compete with a nuclear power plant is a Rube Golberg system of solar + batteries + idling natural gas plants.

If you want batteries to be charged when you need them, you need to overbuild solar to have a consistent surplus. But if you make the batteries so big that they can handle the whole long tail of weather (multi-week cloudy stretches) the overbuild required becomes so large and expensive that you get very little ROI on both the panels & batteries during most of the time when that buffer isn’t needed. So you probably end up with idling natural gas plants too, but these get expensive per kWh because if they’re only used to fill the gaps, their utilization rate is low.

So the *system* that can deliver reliable power is more complex and expensive than the solar farm alone. And it probably still burns fossil fuels.

What people really mean when they say nuclear is expensive is that the costs are front-loaded and are difficult for private entities to handle. Utilities would rather make multiple smaller bets than fewer big bets — if things go wrong, there’s more career risk with big bets, and for some of the biggest bets, there’s even bankruptcy risk involved.

But our civilization needs a lot of reliable clean power, so what can be done?

I think this is a great role for the government. Infrastructure and national security are natural domains for collective action.

Power grids are life-support systems and platforms on which the whole economy is built. Nuclear energy provides high energy security because you can easily stockpile multiple years of fuel in a small building (solar, wind, and batteries = relying on China).

What could be done? 👨🏻‍🔧👷🏗️

• Government loan guarantees to reduce interest expenses and risk to utilities. Once operating, nuclear power plants are plenty profitable, so those loans will be repaid. I’d much rather see gov’t take this risk than the gambles we’re taking with our grids, environment, and energy security right now.

• Regulatory reform. During the decades when nuclear power was unpopular, there was a regulatory ratchet that made it way too slow and expensive to build. This didn’t materially increase safety or take into account real-world trade-offs (ie. if nuclear isn’t built, what you get is more pollution and less energy security/more dependence on dictatorships and cartels). A lot of those rules could be streamlined or cut, making it faster and less expensive to build. The tech by itself isn’t inherently that expensive — making it so is a policy choice.

  (ie. NuScale designs SMRs — its DCA application cost $500m, was 12,000 pages long, and had an extra 2 million pages of supporting information. How can this actually be helpful or truly taken into account for a project? How can you usefully write 2 million pages of useful information about *any* project and how can anyone truly review it? Regulation this onerous is just a way to functionally shadow-ban something)

China is building nuclear power plants for about 1/3 the cost of those in the US and France, according to Bloomberg.

• Creating a long-term ongoing program for multiple plants would help the industry spin up the capabilities it needs and attract talent. One-off plants don’t allow going up the learning curve and optimizing supply chains. It’s not a theory: We had successful programs in many countries in the past (US, France, South Korea, Japan) and in places like China today. It’s obvious that if you build many plants in a row you can get good at it and bring down unit costs and delays. We did it once and if we went for it we can do it again, except hopefully even better since we can learn from the past and we’re richer and have more advanced technology.

Look at that 15-year run, going from almost no plants to 300GW+ at a time when GDP/capita and technological capabilities were a fraction of today’s. And during stagflation. Imagine the world today if that line on the graph had kept going instead of stalling out…

• Take into account wider long-term benefits: This would create *a lot* of high-quality construction and manufacturing jobs, and then thousands of local high-paying multi-decade jobs per operating plant (for 60-80 years!). This has a much bigger local economic impact than buying solar panels, wind turbines, and batteries from China, having relatively few people install them, and then very few relatively low-paid people working on maintenance. Just the economic activity and resulting tax benefits from a nuclear program would likely more than pay for whatever loan guarantees don’t work out.

Please don’t confuse my saying that it’s possible and that I want to see it happen with my implying that it’s easy or likely. I wish! But I can also see the challenges.

It’s frustrating to have invented incredible technology that could solve a lot of our problems and not use it. 😩

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

📒 Hell Yes or No by Derek Sivers 🤔💭💡

This was one of the 13 books I brought back from NYC earlier this month.

I’ve been savoring it, reading a couple of pages at a time, and then letting what I read swirl around the ol’ noggin’ for a while.

This is one of those books that has a high ratio of thinking-to-reading.

My all-time fave essay by Sivers is probably still There’s No Speed Limit, but I may find a contender for the throne in there.

I haven’t finished reading it, but I can already recommend it. You can buy it here.

And because Derek doesn’t do anything conventionally, has a cool model where you can get every digital format DRM-free for $15 or the physical book + every digital format DRM-free for $19 ($4 for every additional physical copy — get a bunch and give ‘em to all your friends and family!).