Tomasz
00:00:00.000
I think you're pointing out a really interesting question, which is One, most of AI today is single player. There are very few examples of multiplayer AI. Two, single player AI for the individual unequivocally increases productivity Three, individual productivity improvements don't necessarily confer overall productivity improvements to enterprises. And so how, if, when will we make the leap to multiplayer productivity where we actually see it compound at the enterprise scale? This is DevOps Paradox, episode number 374. Does AI make companies more productive?
Darin
00:01:34.148
Viktor, I'm hearing all sorts of ramblings when I'm talking to my friends at enterprises especially. They're starting to use AI. Sometimes they want to, sometimes they don't. But assuming it's the ones that they want to, now they're starting to get pressure from their bosses like, "Okay, great. We're paying this extra money for you to use AI. We're expecting two to three at least times productivity out of you every day." What do you say to that?
Viktor
00:01:58.008
That's the ultimate goal. I'm guessing. I just haven't seen it yet on scale, I haven't seen Google producing three times more than they were normally doing or, Netflix having three times more TV series. I haven't seen it yet
Darin
00:02:22.424
What do you think about that statement I just had about that expectation from management saying, "Hey, now you've got AI. Give me two or three X extra or you're gone"?
Tomasz
00:02:34.294
I think it's probably largely true. That's the expectation. I think reality's probably… the studies I've seen are 25 to maybe 150% on the high end in terms of overall productivity, but really closer to 20 to 50% today. And in coding, if you measure on lines of code, sure, no problem. I can hit that. In other disciplines uh, it's much harder
Darin
00:02:55.078
Well, With code we could always hit it. Hey, we need to speed something up. The first thing you do before you speed something up is you slow it down, so then you can pull the while loop out to speed things up. We can't do that again anymore. think code can get us to that point, but I don't know that it's going to be performant code
Viktor
00:03:14.230
if you look at it, we did that many times before. Many things are infinitely faster than they were, like, 100 years ago, right? There's probably no even need to name them. Almost everything. The question and confusion is whether such goals can be done in a matter of months, not decades, Especially in larger organizations
Tomasz
00:03:38.102
That's true. I write a, a blog almost every day, and I've been trying to use AI to accelerate that, right? So let, let's just ground this in an example for a second. a year ago it was fine-tuning on different kinds of models. I've tried skill distillation and, and I finally got a system that works where I'll write a blog post with AI and then I'll have the AI record every single change that I make change a comma, change a semicolon, write it down distill a skill and a knowledge from it and then run through it. And what I've noticed is if I look at the total number of edits that I make to a blog post roughly about 130 per post. The number was surprisingly high and throughout the evolution of this system its basically been roughly constant which was a surprise like how we thought the purpose of AI was actually lets reduce the total number of edits and then I went through and I graded for each of the last ten years fifteen posts and the result is the quality has surged this year so I'm not producing more posts but the posts I've produced are significantly higher quality But I can't measure that in terms of productivity and maybe it's not just twenty five percent to fifty percent productivity gain Its ultimately an improvement in the underlying output What do you think of that?
Viktor
00:04:56.238
I mean, The manager would… I'm guessing you don't have a manager above you because manager would, would tell you kind of, "So you were not doing it well before."
Tomasz
00:05:04.688
Yeah that's, the critique while I was doing it as well as an unassisted human. It-- I think it's… I was trying to make this analogy with chess and I imagine it probably also exists in sports which is there was a certain level of chess you could attain without AI within whatever four or five hours worth of training and then after the AI chess era all of the elos or the relative scoring rankings increased because everybody could be significantly stronger And it's not that we spent less time training, it's just we achieved a higher level and so the productivity gain led us to higher more strength
Viktor
00:05:40.613
the problem with this whole conversation is that when we look at what we do as individuals, I have no doubt that we are more productive. Whether it's 50% or 500% more whether that's in quality or quantity, doesn't matter. We are much, much more productive. I have no doubt about that. What I do have doubt is whether that translates to such increased productivity when working within the restrictions of a company, Because th- there is a big difference on whether I'm more productive and whether my team, my company is more productive as a whole. Which I have no doubt that they will reach that as well, but I feel that those are very different types of challenges, right?
Tomasz
00:06:26.746
Yeah, that's an excellent point. is the company more effective as a result of all of these individuals being more productive? Right now, probably not, because we're optimizing… If we think about it as a system, we're optimizing each individual point in isolation rather than optimizing the system. That'd be one way of looking at it.
Viktor
00:06:46.401
But then that also brings the question is the-- are we going to reach those levels by working in teams of humans now in this concept just to be clear right? Or is the future of okay I'm going, to do this i-I don't need you people your people are slowing me down even if you people are working with agents and you are also individually more productive Those hand inevitable handoffs might be what is actually slowing things down And we m-might not notice that slowdown that much without AI but now they're becoming almost self evident Like if when writing blog posts righT that you mentioned before If you would have another four people working with you Would that actually result in you churning articles faster
Tomasz
00:07:42.931
No, I don't think so. Not in the case of blog posts. I think you're pointing out a really interesting question, which is One, most of AI today is single player. There are very few examples of multiplayer AI. Two, single player AI for the individual unequivocally increases productivity Three, individual productivity improvements don't necessarily confer overall productivity improvements to enterprises. And so how, if, when will we make the leap to multiplayer productivity where we actually see it compound at the enterprise scale?
Tomasz
00:08:21.399
Well, You could argue the software factory is the first instantiation of multiplayer collaborative productivity the GitHub falling over and the reimagination of the source control systems is a demonstration of that. And the fact that, CI/CD infrastructure is now the limiting factor in these software factories, that suggests that the business is operating at a different level of production utilization than it has been in the past That might be a good, might be the first example of a large scale multiplayer, much more productive system
Viktor
00:08:57.661
Yeah, Yeah. You, when you say large scale, you're meaning more like large scale of what the output of what they're doing is serving, not necessarily large scale in terms like we have 100,000 people in this organization. How are 100,000 people more productive? Should they be?
Tomasz
00:09:17.327
Ideally they would be you know. But I-- You know, look at like shipping velocity, Rippling is a really good example of, I think how much faster, and I'm not involved in the business, I'm just watching them from the sidelines. But how much faster they're able to ship features or even N-Nvidia's acquisition of the poolside software or the model factory is a good example of how some of these systems I think are really valuable and can produce a lot more. You get Laguna, I think was from start to end it's 120 billion parameter model close to state-of-the-art was 45 days. you know, that's a large scale hundred person maybe a thousand person order of organization that's more productive as a result of AI. Maybe we're just starting to see it
Darin
00:10:02.660
I'm just trying to get my head around what that new factory would look like. I came up in the '60s and '70s working in sheet metal fabrication, and we had raw coils come in on the left side of the thing, and we had the final product coming out the right side. Classical manufacturing. And we rubber-stamped that thinking back to theory of constraints, The Goal, Goldratt, all the g-- you know, all the classical literature. Yes, I'm that old. now I still see a factory, but what's inside the factory isn't the typical flows But I don't know where it's going Are you seeing anything new? People are saying it's just not good enough. We can complain about Git all day long, but it's the best we've got right now. are some reasonable guesses? I'm not gonna say theories 'cause I, I don't know there are any, but what could we do?
Tomasz
00:10:53.260
I don't know. I mean, I've never worked on a manufacturing line. But The way I imagine it is let's say you're producing a car. It used to be that you needed five people to rivet the chassis together and then you did ano- another five people to put in the drivetrain and another 25 people to put in the engine And now we're at a place where only one person needs to do each of those. And so you have this reduction by a factor of let's say four or three because we started with that number. And so then you have people that are now responsible for each of the individual components or the subassemblies. And at some point those jobs will be abstracted up through harnesses and meta harnesses as the fences around agents become better and the total number of people required to manage a software factory will diminish. But it's incremental . It's not, It's not like you just buy a robot and all of a sudden it works
Darin
00:11:47.872
That's the thing though is the robots were replacing the people on the line, but the line was still practically intact. The line did not shift a lot. I'm wondering if the line is going to shift a lot when it comes to AI and knowledge work in a factory
Viktor
00:12:03.444
In this context, when you say line, you mean not agents, you're talking about the tools that they are using?
Darin
00:12:08.806
The tool, the… Yeah, it could be w- all the above. We build a, a manufacturing facility, it has four walls. But what's inside those four walls going forward, there will probably be hints and ghosts of what used to be there, but completely different
Viktor
00:12:26.091
that's something that I wrongly guessed maybe a year ago, maybe two. Back then, I was almost certain that with what is happening right now we will see new programming languages, new tools, new things that are going to be more optimized or better for agents to work with. I was completely wrong on that one, or I'm not following the right channels and I'm not aware of them, right? But basically nothing changed from the tooling perspective, right
Darin
00:13:10.753
They're using… Have not changed . They're still using Bash. Yeah, it's it's I use bash, they use bash. It's like, okay, what's the difference? The other thing that sort of is in this same area is what do we do with the juniors or the entry-level work? what I didn't set up at the very beginning is Tamas is a founder at Theory Ventures. Yes, we have a VC person on. What are you seeing from companies that you're looking at, whether you invested in them or not, how are they dealing with entry level? Or are they even looking for entry-level people today?
Tomasz
00:13:43.467
they start with senior level. So a brand new company is often starting with a senior engineer and a senior collection of senior engineers and an architect, and they design the factory they don't hire juniors until maybe there are about 15 to 20 people on the team, roughly and the factory is working then there's a question of how to train those juniors because the training is radically different It's not like I can lop off a series of PRs to review and assign it to a junior Because the junior doesn't yet have the experience to be able to opine effectively on that PR especially in this era and so that is not solved We have seven AI engineers on staff and we hire technical interns and w-we are trying to understand the best way of structuring the education So I don't have a great answer for you there I think startups are primarily recruiting senior architects and then Trying to train juniors where they can just by watching over the shoulder
Darin
00:14:52.854
Couldn't you argue that the startups of today shouldn't exceed 15 to 20 people ever?
Tomasz
00:14:58.454
I think to achieve the same level of productivity as five years ago, yes. The expectations on product breadth are now so much broader that they inevitably will grow almost back to the same levels
Darin
00:15:13.991
Okay, let's hold right there. Why do you believe or why do you think, why do you state product breadth is expected to be much wider than before? Because I don't wanna buy, sorry SAP, I don't wanna buy another SAP that does everything and the kitchen sink. I'm looking for point solutions. Why do you think we still need those, my word, kitchen sink type solutions?
Tomasz
00:15:37.043
two reasons. The first is The buyer can vibe code a point solution or has the perception that they can vibe code a point solution, and so there may not be as much perceived value. We can debate whether or not it's real or perceived, but perceived value in a point solution. And then the second is the competitive dynamics associated with it. So if you have two 15-person teams, each of which are building a point solution, and one wants to grow faster than his competitor or her competitor, they will naturally expand into an adjacent category in order to capture more budget because the cross sell, selling a second product to an existing customer, is much easier than selling another new product to a new customer. And so I think both of those forces contribute to expansion
Viktor
00:16:20.023
How much will we be able to make… When I say "we," I mean in software industry to make those deals in the future, 100k a million recurring revenue in a year per customer. Why would companies be giving that money? I understand if it's something that has a moat like okay, so I'm gonna going to give money to AWS because, I cannot build all that physically. But do most of the other solutions, you know, software solutions why would they give that money to begin with as a company? Before I was calculating because it would cost me five times or ten times or hundred times more
Tomasz
00:17:03.426
okay, so first we observe it. Five years ago s- let's say in a classic software era, average deal size for a Series A company, very early business, maybe 10 or 15 people. If you sell a $20,000 or $30,000 deal, that's pretty good. Now we are seeing software companies and infrastructure companies, and not inference businesses, closing $10 million deals within the first or the second quarter of launch because one, the enterprise can't build it. Two, the cost of being late to a market is extreme because of the growth rates. And so the willingness to pay is equally extreme. And then three, you know, if you have four beer brewers, they're all trying to do more with less. And if they can have two or three times the productivity because the AI system can automate their back office in a way, they will pay huge, money for that because they can win more market share, and with more market shares, more market cap. So I think it's a true dislocation in the way that we're working as people who are willing to pay premium
Viktor
00:18:14.834
you 100% on uh, i- if you actually sell AI systems in a way, because that's what they need to do everything else. I was more referring to everything else I'm giving you, I don't know like, uh, a system that will do what X, whatever X is, Not an AI system. I might have built it with AI system, but it's not an AI system, uh, Itself. So it's not a tool that you will give me money so that you can build all the other tools. I'm talking about buying all the other tools. Why would I do that?
Tomasz
00:18:46.935
Sorry, I'm having a hard time following. Maybe I… Could you ground it in an example?
Darin
00:18:50.015
let let me give one. Let… So I think he might be going something like planned maintenance software or I'm thinking shop control because that's my background, but a- anything like that, right? It's just something that's specific to an industry for productivity within the industry
Tomasz
00:19:13.335
Yeah, I think one, the maintenance costs associated with that software are higher than
Darin
00:19:19.487
no, just it would become m- higher in tokens. Yes. This is where… and also before we finish answering this, you had said the phrase inference business, and I'm assuming by that you meant like the Anthropics and OpenAI's or is… Does inference business mean something
Tomasz
00:19:34.201
No, that's right. Yeah, Fireworks, Baseten together. I mean, You'd even put the uh, Crusoe, CoreWeave those,
Darin
00:19:40.063
anyway that-that's, that's what that is so that's beyond stupid money So we're not talking about beyond stupid money We're just talking normal money I'm trying to build software there's vendors out there for every type of specific HVAC service solutions all those things All those services exist today why would I want to even bother with that? Is it more that it gets into regulatory It's like yeah I could build my own Let's say I had an HVAC company I can build my own servicing software but I'm not doing anything about PCI or PII or any of the other stuff Is that the reason why I would stick with a solution
Tomasz
00:20:17.998
Not really. I mean, So in HVAC we can take that example. You know there's a publicly traded company called ServiceTitan and they are classic software company started about 10 years ago and they would manage initially your advertising and then somebody would click on a Google Ad and it would manage recording your information and what you needed And then t- held all that information into CRM and then it had dispatch technology send a truck to your house with a technician. that system is still really important just the way that Salesforce is really important because you need that database of everything that's going on inside of your business. And what's now starting to happen is the automation on the voice side Automation of predictive maintenance; y-you need a new f- air filter and then the automation of inventory management So what needs to be on which trucks? Which components need to be on which trucks? And it's not that…I don't think the amount of money people pay for the core platform will increase meaningfully The willingness to pay remains roughly the same that's true for Salesforce as well You get the core growth on the sales part of Salesforce's business They just increased prices about 10% per year and that business grows about 10% per year What's really growing is the agentic. And so if yo--You probably need one to sell the other for it to be effective and that's why you buy both but your willingness to pay on the first one doesn't increase
Darin
00:21:43.448
You've been talking to, I'm assuming, a lot of startups looking for money. What are they building? Put them in big buckets. Are they trying to do some inference and other people are doing other silly things that are going to get Sherlocked by somebody else very fast? What, What is the market right now?
Tomasz
00:22:03.303
Yeah, okay, so databases used to be the biggest market in software. It's now inference. There's a lot happening in inference. You have specialization of inference, really fast inference, medium fast inference, slower but less expensive inference on your laptop, inference for voice, inference for robots. That's one really big market, and everything to do associated with inference optimization all the way from the chip to the memory to the network connections inside data center. So there's that whole world in and around inference, and that's just having an AI calculate the next token. There are-- is a huge volume of people building software around AN- AI apps. If I wanted to build an agent and it needs to call somebody on the phone or ecosystem there Or if I need a way of storing a whole bunch of information about a company so I can inform an AI agent, that's really important. The fastest growing ecosystem today in startups is around producing data to teach AI what's called reinforcement learning environments And that's creating a simulation of a computer where an AI can point and click and manipulate Excel or VS Code as a way of producing information to improve the models. It's incredibly fast growing market. And then you have the layer of applications, so agents for sales, agents for marketing. And maybe the very last market I'll say which is also exceptionally fast growing is security And that's we're writing so much more software. A lot of it needs to be far more secure, and it needs to be secured much more quickly than it has in the past
Tomasz
00:23:47.972
Well, Okay, so the last major platform shift was cloud and, you know, AWS launches in 2004, and no enterprise will touch it until about 2010, 2011. So there's 10 years where the new platform is really the domain of startups, and startups use it because they don't want to buy servers, and they don't wanna rack them, and they don't wanna manage them, and they don't… They're willing to pay a premium for it. And then at some point, the enterprise decides eight to 10 years later, AWS is more secure and it's easier and it's faster. So that takes 10 years to get to scale. ChatGPT three-five, we're talking like less than two years for the biggest enterprises to start spending hundreds of millions or billions of dollars in adopting this platform. That's completely different, it starts at the Fortune 500. That's where those budgets are. So that is huge difference. Like, uh, 10 years ago, if you were a startup, you were selling to small companies out of the gate. Today, you can sell to the biggest companies in the world literally in the first year, maybe even the first quarter
Viktor
00:25:00.576
so how does that match then VC world? If, If I can start making big deals after seed investment, for example, right? Why do I need A, B, C, D and so on and so forth, right? Is it that I'm selling for big money but actually my expenses are infinitely greater than they were in previous instances or
Tomasz
00:25:26.630
No. So we have these letters to describe these financial products, right? Seed, call it three to five pre-seed, a million seed, three to five million Series A, 15 to 25 million. for a long time startups wanted those financial products and it was okay to sell those financial products as a standard. There's a Standard Series A. And now because these businesses are growing really fast, a standard seed for a company that can close a $10 million deal in its first year doesn't make sense. It doesn't. Or if you're in Neo lab and you need $100 million dollars to train a model three times, standard seed is not viable. So the market has changed to custom or bespoke financial products. n-now you can raise a seed of 500 to a billion dollar, 500 million to a billion dollars with just a team because you're embarking on a huge project and the market will give it to you if you have the right pedigree. That's what's really changed. And so the precede and the seed markets are radically different even the series A market valuations have increased ownership that we can purchase as a fraction of the total ownership has decreased. I'll stop there but has changed and many of these products are now bespoke in a way that they weren't again five years ago
Viktor
00:26:49.803
does that also change dynamic that if I'm a startup and I'm making big deals that actually I become profitable before I go public type of situation?
Tomasz
00:27:00.288
Well, You need to be much bigger today to be public than you were five or 10 years ago, in the late '90s, to go public, you needed 25 million of trailing revenue with a few quarters of profitability. Today, to go public, you need to be 500 million to a billion at the smallest, and we can talk about why. So you need to achieve a radically different level of scale to do that. And then I think the second reason that you don't see this like push-- You do see some push to profitability. Midjourney would be a really good example of this where little bit of venture money, and then all of a sudden the company is just printing cash and the founders wanna operate it that way. Many of the founders who raise venture capital, the goal is to build as big of a business as possible rather than to generate the profits because businesses have this arc where you invest in a business, you burn a lot of cash, you achieve a certain level of scale, and then you slow for whatever reason, and the market changes from valuing you based on growth to value based on profits, and then ultimately, business dies over time. And so if you're founding a company, like many people, not everybody, but a lotta people just want to keep getting it bigger and bigger before it crests
Darin
00:28:09.482
But again, that sounds like we've played this too many times before. Do we see those build build, build Crest people? They just haven't done that before, so they think that's the way to do it instead of looking at history of all the littered companies that never made it
Tomasz
00:28:24.783
know many people start businesses two or three times that way and they reach-- their goal is to reach higher and higher crests and take more risk as they pursue each success-successive cycle. And I think, the boom bust cycle is an essential part of capitalism. I think it's a feature not a bug, we will over invest and lots of companies won't make it but the rate of business evolution and genetic mutation in the way that we build businesses move so much faster for it. And that's why I think the US has and continues to be an incredible engine of innovation and growth because we have these boom bust cycles
Darin
00:29:02.493
You said the phrase a few minutes ago, "right pedigree," in order to get funding. What does that mean today?
Tomasz
00:29:09.123
It depends. But if you are building a Neo lab, so that's a new AI model company, you probably need a bunch of PhDs in the field with some experience coming out of a company that's done something similar. that's the right pedigree. I think in the world of applications, that's a demonstrated ability to sell into a market or market yourself, position yourself into a market. So it's context specific.
Viktor
00:29:36.528
Neo Labs are something I cannot wrap my head around. I have impression as if the whole bui- business of Neo Labs is that we are serving the gap in the market until that gap is filled and then we will not exist type of situation right? Because eventually we will not have shortage of capacity, right? if that happens, if we stop having shortage of capacity why would anybody use Neo labs than big guys
Tomasz
00:30:07.708
It's true, right? And the analogy there would be the hyperscalers in compute where Microsoft, Google and Amazon Are effectively the most r-- I mean only real… Uh, digitalOcean bears mentioning but there are not very many other companies. I think in this case the difference is There will be an extended shortage at least through 2030 and that latency provides an opportunity for large number of or larger number of companies to reach public scale and so it should be more fragmented
Viktor
00:30:39.303
And that we will have the shortage because production cannot ramp up Because I, I understood that among other things production is not increasing drastically because nobody wants to invest in those big factories because that's potentially a temporary demand, But if what you're saying is a permanent demand, then one of them is wrong, right? Either, either new clouds are wrong or hardware manufacturers are wrong thinking that this is not worthwhile three years of investment and by that time there will be no demand in a way
Tomasz
00:31:18.413
right. And so you, you can see this within the supply chain. So two years ago GPUs-- prices of GPUs doubled and then we realized that those GPUs needed a lot of CPUs and they needed CPUs because when you use them not as a search engine but you used them as an agent you need CPU's .And then all of a sudden we realize that actually these systems consume a lot of RAM and so the prices of RAM went up and then we didn't have enough RAM so we started to use solid state disk to write whatever was in RAM to solid state disc ,and then the prices of those nearly tripled and then said okay "We can't afford solid-state disk anymore." Until it's right to hard disks.That's now the price is of hard dis… See how there's like cascading chain? And it all starts because TSMC which is the f- fabricator of many GPU's has done Viktor exactly what you said, which is they have said," We don't want to massively expand manufacturing lines because we have seen the cycle before and if we over invest there will be glut and we'll see a collapse ."That's not true at every layer of the stack. So SK Hynix has just invested I think more hundreds of billions in developing a new massive fab that's about a third the size of the Empire State Building and they are one of the three major producers of high bandwidth memory which is a component of GPUs. And some people are pushing harder but if the whole ecosystem or the whole supply chain does not increase altogether you'll still have these points where you have a rate limiting factor in the ultimate supply chain. So I think we'll be in this for a long time. And then we haven't even talked about power copper and concrete
Darin
00:33:06.495
about it because is that, is that 2030? Are we gonna magically have all that stuff solved in 2030 as well?
Tomasz
00:33:13.149
No, I mean in the US it's tough because… so you look at the price-- The most expensive component of building a data center is concrete. It's just linearly increasing. It's now over 3x what it was two years ago. And I'm not even including the land. The second hard part is actually finding where to build a data center. So initially Texas was a great place to build. Governor has now blocked new data center builds. New York has blocked it. Arizona from my understanding is slowing down. Pennsylvania is actually a really great place to build it wherever there was an old coal factory that were connected to the electrical grid and so there are lots of old coal factories can be repurposed but there are fewer- There's fewer and fewer places where you can actually build data centers And to give listeners a sense of perspective roughly this is order of twenty gigawatts worth of data center capacity today in the United States and the vast majority of that is in Loudoun County in Virginia and the goal --the projections are somewhere between two and three hundred gigawatts by the end of 2030 so we're trying to 10x or more
Darin
00:34:11.796
to 15x what we're going to get. let me do basic math instead of all this crazy math. Uh, That's four years away. Because I've been in manufacturing, I know how long it takes to build things. If it's slow getting the materials I need to build, it's going to slow things down. To actually believe in four years we have 15x our data center capacity, I don't know what drugs those people are on. I don't want to be anywhere near that kind of drug, 'cause that's not practical.
Tomasz
00:34:44.179
we won't, get there. The supply chains are simply not ready. The electrical grids are simply not ready. politically it's, becoming more challenging
Viktor
00:34:51.456
But on the other hand it feels, and please correct me, that actually many of those challenges China already solved in a way, right? They don't have such issues with energy, with the rate of production, and so on and so forth, right? Does that mean that actually this might-- is this-- This almost feels like the first time that actually China is almost dominating it in a way. Not necessarily in all the aspects, but in many of the aspects that or components, let's say, that are needed for this to work
Tomasz
00:35:29.207
energy is the most important difference. If we look at US energy production over the last 30 years, it's effectively flat. Population has increased. We use electric cars. We have more data centers. That has increased consumption. Our refrigerators, washers and dryers, dishwashers have become significantly more efficient with electricity. Those effects have offset each other, which means very few new power plants, not a lot of investment in the electrical grid. Everything stays the same. China, on the other hand, has been producing huge numbers of new nuclear reactors, coal power plants, and other kinds of energy producer-- clean technology. And so the country and the populace is accustomed to increasing electricity production. That hasn't happened in the US basically since the early 1970s. And so we have to change, I think, we have to change in the United States the way that we view electricity production, and that's because just the way that when we had an electric motor, we converted electricity into work now we have a computer motor that converts electrons into a different kind of work. And if we wanna become a lot more productive we need a lot more electricity to do it
Darin
00:36:42.387
Let me get political for just a second, 'cause we never get political here, 'cause we don't. It's political, not political. Is solar or wind even an option for power gen in the US? Like a legitimate move the needle solution
Tomasz
00:36:56.737
So energy is not really my domain, but um, solar is really important. it is incredibly important. It… Solar plus battery
Tomasz
00:37:05.621
important, and the major reason is the US electrical grid is quite fragile. So in the US, we use alternating current, which means the current changes polarity several times a, a second, and it means that everything in the US electrical grid has to be synchronized. I remember I was an intern in a building in New York City in 2001, and there was a -- or in 2002, and there was a massive power outage that started in Ohio and cascaded all the way through the eastern seaboard. And the origin of that was machines along the electrical grid were slightly out of phase. They were slightly s- one was slightly slower, and it creates a cascading failure. And so the electrical grid is brittle in the US, aside from Texas, but everything else is interconnected. solar provides a lot of anti-brittleness or antifragility because I can stand up solar in my apartment or my house with a battery and be self-sustaining. And so I think that's really important, that's where a lot of th-the renewables will push
Tomasz
00:38:12.677
you would need enormous batteries. But the great part about data centers is the electricity c- demands are extremely predictable. They're like Uber. You have a surge at 9:00 in the morning, you have a surge at 5:00 in the afternoon
Darin
00:38:25.159
true, but I could argue that unless you're already-- ' Cause as, as the planning engineer, I know I'm gonna have 60,000 H100s in there, so it's easy to do the math. But what happens when the next chip comes out and now we're starting to go through and now my capacity is too low? I mean, Hopefully at that point it's, it's like that leveling would hopefully come in back into play again.
Viktor
00:38:52.046
But that als- you're almost giving argument to solar and others rather than-- Because kind of if my demands increase, I cannot increase as the owner of the factor. No, I cannot increase the grid of the country, but I can cr- put more pon- pan- solar panels. I can put more mini nuclear reactors or whatever, right? I can do stuff on my own
Tomasz
00:39:16.986
This is the very interesting thing about data centers. It's one of the largest consumers of electricity where the electricity be- can be consumed in place. Most of our power plants ship electricity. They distribute it elsewhere. But a data center consumes the electricity where it's produced, and that's a big difference. It doesn't seem like it's a big difference, but it's an enormous difference, and that's why you can stand up a solar farm someplace or a collection of different reciprocating motors or a collection of jet turbines in the middle of nowhere, and it works just provided that it's connected to the internet
Darin
00:39:55.452
To the internet? No, not to the internet. We can't put the power on the internet. That's where China comes back into play at least from a US perspective or-or the other players. for somebody that didn't know energy you did a very good job there towards the end of like oh okay bring bringing out all the jet engines and everything. If somebody starting a startup Like they got laid off They decided that they got a good payout And it's "Okay I am gonna go do this." So imma go do X"and they are wanting to get funds reality pill Do they need to take in this market
Tomasz
00:40:25.559
you're likely starting a company at the application layer, so helping somebody solve a workflow, I think the most important thing there is ability to sell quickly, as we talked about. Like in that first quarter, it's really important to be able to grow and establish yourself in that space, and finding authenticity in how to sell and speak to that buyer effectively is essential
Darin
00:40:48.759
Okay. Same story as it ever was. People just ignored it though, right? That's the problem?
Tomasz
00:40:54.341
Yeah. Well, so, you know, I had this Belgian professor of marketing in grad school. He had this equation that I love, which is innovation equals invention plus distribution. So in order to change the world, you have to make something awesome and then convince people to buy it.
Darin
00:41:08.081
I actually think that formula should be instead of plus, it should be times distribution.
Darin
00:41:20.031
Whereas I built something plus zero is like I still have something, but it's not going anywhere. It's like, okay, yet another home lab project
Tomasz
00:41:28.116
that is a great-- evolution of that. And then the question is, what are the scalars associated with it? How do you weight each component? And initially, when I first started as an investor, I thought most of it was the invention part. But as the cost of inventing has decreased, the weight actually shifts pretty meaningfully to the
Darin
00:41:47.297
I could argue the distribution was always 80 and the product was 20. Because if you couldn't sell it… you, you get a bunch of nerds around writing code all day in their mom's basement, we can do that all day long. But we can't talk to any- we can't even talk to the kiosk at McDonald's to get our order in Sales, to me has always been the big deal which is why I've never been a salesperson. Always let somebody else do that. think distribution has always been a big deal and I think my opinion not being a VC it's look if you-- If you've got something that you want to sell start selling it before you create the stupid product please because then you're not wasting any tokens or minimal tokens That's the good thing now as we could waste minimal tokens and have something to really show versus showing slides. I think that's the true jump now from the twenty eleven, twenty twelve when we were doing slideware or maybe some clickware And that was it Now we can actually see something really working
Darin
00:42:47.631
Yes. Well, Okay, that was the very first thing is if you buy it, we'll make it. Hmm, Okay. Oh, hang on a second. That won't work." But I think now we can actually show them something legitimate. Viktor, you talked about this too. It's i- it may not have all the bells and whistles, it doesn't have all the error checking blah, blah, blah, but it works right? Does this cal-- solve your problem?" We'd like this one other thing." "Okay go eat lunch come back in an hour," And now you've got it for him in the next hour We couldn't have done that two years ago
Darin
00:43:21.051
They just don't want to. Some of them don't want to. the enterprise don't want to. if people are… Okay, they believe that. They've got the sales down. What do they really have to do? What's the reality? It's like, okay, now I'm… We talked a little bit about it, but okay, I wanna get some funding. I don't need the 50 million, 100. I, I d- I want the classical seed round. Can we still get that today?
Tomasz
00:43:49.650
They're looking for businesses that grow much faster than in the last generation. That's definitely
Darin
00:43:54.852
that's the hard part then. So it's harder to get old school money, I guess, is how I'm classifying it
Tomasz
00:44:00.198
That's the trade. The trade is Boy, 15 years ago a company that went one to three million to seven, eight million was probably like top quartile in terms of growth. Five years ago it was one five fifteen and then today, top 1% of companies are growing zero to 100 in a year.
Tomasz
00:44:22.962
And so okay that-that's the competitive set So the competitive set has changed It's not every company Many companies are financed with growth trajectories that are still incredibly attractive Like one to seven One to ten Then 10 to twenty-five To thirty-five That's also eminently financeable In this market But It's very hard to raise money if your plan is one to three Or something like that
Viktor
00:44:47.790
assuming that the amount of money is more or less constant, those rates cannot continue in a way, right? Because the… How many companies can actually-- how many sales can you make of 100 million in parallel? I mean, What I'm trying to say is that there is-- then that means that there is a limited number of those startups that will ever reach those numbers to begin with, right?
Tomasz
00:45:11.606
Absolutely, right? And so how do you back into it where you look at overall US GDP growth? That's the numbers we're talking about. And so the idea is that let's see if AI can actually-- AI data center build-outs are now responsible for two thirds of U.S. GDP growth. So that's all net new money, let's say, that's being created that can be sold. And if we start to grow with 30 trillion USD GP growing 30%, You know, you do the math. It's $900 billion of new revenue each year. That's a lot
Darin
00:45:43.416
It's not bad except we're at 40, 40 trillion in debt, so the math keeps going the other
Darin
00:45:53.684
It's… If I could only do math or accounting math the same way as the US government, my life would be so much easier. Have money, great, I'll just print it.
Darin
00:46:03.768
go tell somebody, "You need to give me your money," 'cause that'll make it better.
Viktor
00:46:16.589
it's almost, it's very similar to US government, right? " I need more money. Let's be printed." Printing in this case means kind of VC are
Darin
00:46:24.656
back in the VC. in your, either your portfolio companies or ones you're looking at, what are the things that you're seeing that you're having to tell those teams, " Look, you need to rethink what you're doing"? Are they doing l- is there a common boneheaded thing that they're doing?
Tomasz
00:46:45.676
I think the most important thing is, to grow this quickly you need to be-- It's called the token path. You need to be either in the business of reselling inference in some form, packaging inference and AI in your product or you need to be benefiting from customers… Or you need to be part of the tools that allow inference to be sold. And an example there is like Twilio. So if I wanna have a voice agent and I need to talk to the customer through phone, Twilio doesn't sell voice inference but they provide the infrastructure to connect an agent to a person over the phone. And so to grow at these new astronomical rates one of those things needs to be true and that's different than it was in the past and that's why the per seat model of buying software is dying because you can't grow at these rates without being involved in the token path in some form
Viktor
00:47:41.054
I guess when, when you invest s- sometimes the investment goes into somebody will buy this company. But I g-- I'm get… This is not my area so correct me at any moment, right? Much of investment goes to, "Hey yeah, we want to go IPO," right we wanna become public that will give us infinitely more money and whatso not. But if OpenAI and Anthropic are going to go public for probably biggest bigger numbers than anybody before, doesn't that mean that basically the, the pool of money available is going to dry up for anybody else
Tomasz
00:48:17.614
Good question. Okay, so a couple of points. The first is about 80% to 85% of venture exits when we sell our s- our shares are in M&A
Tomasz
00:48:28.482
selling to other investor. Only about 15%, by count not by dollars. By dollars it's different because the IPOs are bigger. The other thing I'd say is w-when Anthropic and OpenAI and now that SpaceX has gone public, there is much more money coming into startups. Because let's say I have… Let's say I'm a big endowment and I have $100 million to put into Venture, and as a result of any one of these or all of these IPOs, I now five X the dollars that I have. For sure I'm going to reinvest my hundred and then I'll probably invest maybe 150 or 250 million back into the ecosystem because I generate a great return and I might take the other half and give it away or pay for somebody's education. And so now all of a sudden for that investor in venture capital, the amount of money in the ecosystem has now gone up by two and a half-- one and a half to two and a half times. So when we have these big waves of liquidity it's called, when people can sell their shares, the ecosystem surges and that's what we're seeing. You can see it in the prices of real estate in San Francisco. So that's, that's what we'll see
Viktor
00:49:37.555
I, I was referring more to the uh, available capital, public capital for those IPOs,
Tomasz
00:49:44.383
Oh, good question. Okay, so yeah. Okay, so let's put it in perspective. Largest IPO before SpaceX in the US was m- Facebook at the time, now Meta, roughly 15 billion. The SpaceX IPO was much, much larger, I think 75. And the SpaceX IPO, in terms of dollars raised, was larger than the sum of all dollars raised in the previous decade across all IPOs So massive, and there were questions during the SpaceX IPO of could they raise that amount of money? They were many times oversubscribed. And so I think the answer is it seems like the capital markets are much bigger than we thought, and the rumors are OpenAI will raise more than SpaceX, let's call it 100 billion, and that money will come from big institutions, banks, endowments, foundations. But a lot of it will come from retail, you and I, who either are trying to fight inflation in the US or we really wanna generate a great retirement, will buy some of those shares. And the contribution of both of them is much larger than I think most people expected
Viktor
00:50:50.840
Okay, so if I understood what you're saying, basically the cup is much bigger than what was thought
Tomasz
00:50:58.720
Yeah. A-another way of thinking about this is let's look at the S&P 500. There are 500 stocks in that index, and when you buy a share of the S&P, you buy a piece of each. There are 10 companies that are growing really fast and whose stock price is going up, and then there are 490 companies that are b-basically flat. And so what's happening is people are moving the money that goes to those 490 and pushing it to the 10 or the 15
Darin
00:51:25.663
They're not being lazy investors like I am. Being retired, it's like index funds are great. I mean, A- again, the S&P has done well over long term. Like step back, look at it. It's… March was horrible, but you can't look at March of this year.
Tomasz
00:51:39.983
No, but, and if you own the index well, you own big, I think it's, I calculated the top 10 companies represent 40% of the market cap. half of your S&P share is going to effectively AI in some form.
Darin
00:51:52.924
Wh- what would you tell somebody that wants to be a VC? let's say they're getting an exit from fill in the blank, any of those top 10 companies from the S&P 500, and now they're just flush with cash and they wanna do some VC. What do you tell them? What can you tell them? Let's put it that way, 'cause you're under probably regulations here
Tomasz
00:52:08.903
an esteemed venture capitalist gave me this advice which was "Venture capital is a sales job." Imagine you're selling a multimillion-dollar contract. That's VC except instead of doing it twice… Let's say once, you do it twice. I do it first when I raise money to invest from people and that's a big enterprise contract that take 18 to 24 months to develop trust maybe longer. And then when I meet an exciting company we decide to work together, that's again a $15 to a $40 million enterprise contract that will last 10 years. And so both sides of the business are sales. I learned to sell well
Darin
00:52:50.071
sorry, it's we're back to distribution. All All roads lead to sales whether we like it or not. I have one other question here because I've been on the backside of, working with VCs on that side. What's it like to actually create a fund and ra- Is that just… You said 18 to 24 months. Again, that sounds like true enterprise sale cycles. Is it more painful today than it was? I don't-- W- when did Theory start?
Darin
00:53:17.741
Three years old. But you've probably talked to other people that have had funds 10, 15 years back. Has raising money to build the fund become a lot more painful, or has it become reasonable?
Tomasz
00:53:30.870
It has become more standardized. So 20 years ago, venture capital was very small and there were maybe 10 firms, and so it was an exotic product. Very few people understood what venture capital was and understood how to research the investment and how they would generate returns. And when I started in 2008, I think one of the VCs I looked up to still to this day, he had the first billion dollar exit in software ever. And so now, you know, the US venture capital ecosystem has grown from about eight billion in 2008 to about 300 billion in the early 2020s. So we've seen absolutely tremendous growth. And so with that growth comes standardization. So that's a very big change. It's much more well understood by many more investors. And because of how exciting technology has become, a lot more people want to invest. So in that sense, it's easier. On the other hand, it's also more competitive because there are so many firms and so many great firms, and so in that sense it's harder
Tomasz
00:54:40.740
I really like that multiplication. That, that is a brilliant insight. I'm gonna go back. I wrote a blog post. I think I, I am going to update it because you're right, it's multiplicative. It is not additive
Darin
00:54:56.932
Okay. I'm glad I could help today. that's your next blog post, and hopefully AI will not say, "No, that's a stupid thing. You should leave it at Plus."
Darin
00:55:05.542
So all of Tamas's information is gonna be down in the episode description. Tamas, thanks for hanging out with us today.
Tomasz
00:55:12.106
Pleasure was mine. Thank you both We hope this episode was helpful to you. If you want to discuss it or ask a question, please reach out to us. Our contact information and a link to the Slack workspace are at devopsparadox.com/contact. If you subscribe through Apple Podcast, be sure to leave us a review there. That helps other people discover this podcast. Go sign up right now at devopsparadox.com to receive an email whenever we drop the latest episode. Thank you for listening to DevOps Paradox